commit 18d72bf

sewn  ·  2026-05-07 20:01:39 +0000 UTC
parent 96aabc6
import stb_image_resize2
2 files changed,  +10730, -0
+10679, -0
    1@@ -0,0 +1,10679 @@
    2+/* stb_image_resize2 - v2.18 - public domain image resizing
    3+
    4+   by Jeff Roberts (v2) and Jorge L Rodriguez
    5+   http://github.com/nothings/stb
    6+
    7+   Can be threaded with the extended API. SSE2, AVX, Neon and WASM SIMD support. Only
    8+   scaling and translation is supported, no rotations or shears.
    9+
   10+   COMPILING & LINKING
   11+      In one C/C++ file that #includes this file, do this:
   12+         #define STB_IMAGE_RESIZE_IMPLEMENTATION
   13+      before the #include. That will create the implementation in that file.
   14+
   15+   EASY API CALLS:
   16+     Easy API downsamples w/Mitchell filter, upsamples w/cubic interpolation, clamps to edge.
   17+
   18+     stbir_resize_uint8_srgb( input_pixels,  input_w,  input_h,  input_stride_in_bytes,
   19+                              output_pixels, output_w, output_h, output_stride_in_bytes,
   20+                              pixel_layout_enum )
   21+
   22+     stbir_resize_uint8_linear( input_pixels,  input_w,  input_h,  input_stride_in_bytes,
   23+                                output_pixels, output_w, output_h, output_stride_in_bytes,
   24+                                pixel_layout_enum )
   25+
   26+     stbir_resize_float_linear( input_pixels,  input_w,  input_h,  input_stride_in_bytes,
   27+                                output_pixels, output_w, output_h, output_stride_in_bytes,
   28+                                pixel_layout_enum )
   29+
   30+     If you pass NULL or zero for the output_pixels, we will allocate the output buffer
   31+     for you and return it from the function (free with free() or STBIR_FREE).
   32+     As a special case, XX_stride_in_bytes of 0 means packed continuously in memory.
   33+
   34+   API LEVELS
   35+      There are three levels of API - easy-to-use, medium-complexity and extended-complexity.
   36+
   37+      See the "header file" section of the source for API documentation.
   38+
   39+   ADDITIONAL DOCUMENTATION
   40+
   41+      MEMORY ALLOCATION
   42+         By default, we use malloc and free for memory allocation.  To override the
   43+         memory allocation, before the implementation #include, add a:
   44+
   45+            #define STBIR_MALLOC(size,user_data) ...
   46+            #define STBIR_FREE(ptr,user_data)   ...
   47+
   48+         Each resize makes exactly one call to malloc/free (unless you use the
   49+         extended API where you can do one allocation for many resizes). Under
   50+         address sanitizer, we do separate allocations to find overread/writes.
   51+
   52+      PERFORMANCE
   53+         This library was written with an emphasis on performance. When testing
   54+         stb_image_resize with RGBA, the fastest mode is STBIR_4CHANNEL with
   55+         STBIR_TYPE_UINT8 pixels and CLAMPed edges (which is what many other resize
   56+         libs do by default). Also, make sure SIMD is turned on of course (default
   57+         for 64-bit targets). Avoid WRAP edge mode if you want the fastest speed.
   58+
   59+         This library also comes with profiling built-in. If you define STBIR_PROFILE,
   60+         you can use the advanced API and get low-level profiling information by
   61+         calling stbir_resize_extended_profile_info() or stbir_resize_split_profile_info()
   62+         after a resize.
   63+
   64+      SIMD
   65+         Most of the routines have optimized SSE2, AVX, NEON and WASM versions.
   66+
   67+         On Microsoft compilers, we automatically turn on SIMD for 64-bit x64 and
   68+         ARM; for 32-bit x86 and ARM, you select SIMD mode by defining STBIR_SSE2 or
   69+         STBIR_NEON. For AVX and AVX2, we auto-select it by detecting the /arch:AVX
   70+         or /arch:AVX2 switches. You can also always manually turn SSE2, AVX or AVX2
   71+         support on by defining STBIR_SSE2, STBIR_AVX or STBIR_AVX2.
   72+
   73+         On Linux, SSE2 and Neon is on by default for 64-bit x64 or ARM64. For 32-bit,
   74+         we select x86 SIMD mode by whether you have -msse2, -mavx or -mavx2 enabled
   75+         on the command line. For 32-bit ARM, you must pass -mfpu=neon-vfpv4 for both
   76+         clang and GCC, but GCC also requires an additional -mfp16-format=ieee to
   77+         automatically enable NEON.
   78+
   79+         On x86 platforms, you can also define STBIR_FP16C to turn on FP16C instructions
   80+         for converting back and forth to half-floats. This is autoselected when we
   81+         are using AVX2. Clang and GCC also require the -mf16c switch. ARM always uses
   82+         the built-in half float hardware NEON instructions.
   83+
   84+         You can also tell us to use multiply-add instructions with STBIR_USE_FMA.
   85+         Because x86 doesn't always have fma, we turn it off by default to maintain
   86+         determinism across all platforms. If you don't care about non-FMA determinism
   87+         and are willing to restrict yourself to more recent x86 CPUs (around the AVX
   88+         timeframe), then fma will give you around a 15% speedup.
   89+
   90+         You can force off SIMD in all cases by defining STBIR_NO_SIMD. You can turn
   91+         off AVX or AVX2 specifically with STBIR_NO_AVX or STBIR_NO_AVX2. AVX is 10%
   92+         to 40% faster, and AVX2 is generally another 12%.
   93+
   94+      ALPHA CHANNEL
   95+         Most of the resizing functions provide the ability to control how the alpha
   96+         channel of an image is processed.
   97+
   98+         When alpha represents transparency, it is important that when combining
   99+         colors with filtering, the pixels should not be treated equally; they
  100+         should use a weighted average based on their alpha values. For example,
  101+         if a pixel is 1% opaque bright green and another pixel is 99% opaque
  102+         black and you average them, the average will be 50% opaque, but the
  103+         unweighted average and will be a middling green color, while the weighted
  104+         average will be nearly black. This means the unweighted version introduced
  105+         green energy that didn't exist in the source image.
  106+
  107+         (If you want to know why this makes sense, you can work out the math for
  108+         the following: consider what happens if you alpha composite a source image
  109+         over a fixed color and then average the output, vs. if you average the
  110+         source image pixels and then composite that over the same fixed color.
  111+         Only the weighted average produces the same result as the ground truth
  112+         composite-then-average result.)
  113+
  114+         Therefore, it is in general best to "alpha weight" the pixels when applying
  115+         filters to them. This essentially means multiplying the colors by the alpha
  116+         values before combining them, and then dividing by the alpha value at the
  117+         end.
  118+
  119+         The computer graphics industry introduced a technique called "premultiplied
  120+         alpha" or "associated alpha" in which image colors are stored in image files
  121+         already multiplied by their alpha. This saves some math when compositing,
  122+         and also avoids the need to divide by the alpha at the end (which is quite
  123+         inefficient). However, while premultiplied alpha is common in the movie CGI
  124+         industry, it is not commonplace in other industries like videogames, and most
  125+         consumer file formats are generally expected to contain not-premultiplied
  126+         colors. For example, Photoshop saves PNG files "unpremultiplied", and web
  127+         browsers like Chrome and Firefox expect PNG images to be unpremultiplied.
  128+
  129+         Note that there are three possibilities that might describe your image
  130+         and resize expectation:
  131+
  132+             1. images are not premultiplied, alpha weighting is desired
  133+             2. images are not premultiplied, alpha weighting is not desired
  134+             3. images are premultiplied
  135+
  136+         Both case #2 and case #3 require the exact same math: no alpha weighting
  137+         should be applied or removed. Only case 1 requires extra math operations;
  138+         the other two cases can be handled identically.
  139+
  140+         stb_image_resize expects case #1 by default, applying alpha weighting to
  141+         images, expecting the input images to be unpremultiplied. This is what the
  142+         COLOR+ALPHA buffer types tell the resizer to do.
  143+
  144+         When you use the pixel layouts STBIR_RGBA, STBIR_BGRA, STBIR_ARGB,
  145+         STBIR_ABGR, STBIR_RA, or STBIR_AR you are telling us that the pixels are
  146+         non-premultiplied. In these cases, the resizer will alpha weight the colors
  147+         (effectively creating the premultiplied image), do the filtering, and then
  148+         convert back to non-premult on exit.
  149+
  150+         When you use the pixel layouts STBIR_RGBA_PM, STBIR_BGRA_PM, STBIR_ARGB_PM,
  151+         STBIR_ABGR_PM, STBIR_RA_PM or STBIR_AR_PM, you are telling that the pixels
  152+         ARE premultiplied. In this case, the resizer doesn't have to do the
  153+         premultipling - it can filter directly on the input. This about twice as
  154+         fast as the non-premultiplied case, so it's the right option if your data is
  155+         already setup correctly.
  156+
  157+         When you use the pixel layout STBIR_4CHANNEL or STBIR_2CHANNEL, you are
  158+         telling us that there is no channel that represents transparency; it may be
  159+         RGB and some unrelated fourth channel that has been stored in the alpha
  160+         channel, but it is actually not alpha. No special processing will be
  161+         performed.
  162+
  163+         The difference between the generic 4 or 2 channel layouts, and the
  164+         specialized _PM versions is with the _PM versions you are telling us that
  165+         the data *is* alpha, just don't premultiply it. That's important when
  166+         using SRGB pixel formats, we need to know where the alpha is, because
  167+         it is converted linearly (rather than with the SRGB converters).
  168+
  169+         Because alpha weighting produces the same effect as premultiplying, you
  170+         even have the option with non-premultiplied inputs to let the resizer
  171+         produce a premultiplied output. Because the intially computed alpha-weighted
  172+         output image is effectively premultiplied, this is actually more performant
  173+         than the normal path which un-premultiplies the output image as a final step.
  174+
  175+         Finally, when converting both in and out of non-premulitplied space (for
  176+         example, when using STBIR_RGBA), we go to somewhat heroic measures to
  177+         ensure that areas with zero alpha value pixels get something reasonable
  178+         in the RGB values. If you don't care about the RGB values of zero alpha
  179+         pixels, you can call the stbir_set_non_pm_alpha_speed_over_quality()
  180+         function - this runs a premultiplied resize about 25% faster. That said,
  181+         when you really care about speed, using premultiplied pixels for both in
  182+         and out (STBIR_RGBA_PM, etc) much faster than both of these premultiplied
  183+         options.
  184+
  185+      PIXEL LAYOUT CONVERSION
  186+         The resizer can convert from some pixel layouts to others. When using the
  187+         stbir_set_pixel_layouts(), you can, for example, specify STBIR_RGBA
  188+         on input, and STBIR_ARGB on output, and it will re-organize the channels
  189+         during the resize. Currently, you can only convert between two pixel
  190+         layouts with the same number of channels.
  191+
  192+      DETERMINISM
  193+         We commit to being deterministic (from x64 to ARM to scalar to SIMD, etc).
  194+         This requires compiling with fast-math off (using at least /fp:precise).
  195+         Also, you must turn off fp-contracting (which turns mult+adds into fmas)!
  196+         We attempt to do this with pragmas, but with Clang, you usually want to add
  197+         -ffp-contract=off to the command line as well.
  198+
  199+         For 32-bit x86, you must use SSE and SSE2 codegen for determinism. That is,
  200+         if the scalar x87 unit gets used at all, we immediately lose determinism.
  201+         On Microsoft Visual Studio 2008 and earlier, from what we can tell there is
  202+         no way to be deterministic in 32-bit x86 (some x87 always leaks in, even
  203+         with fp:strict). On 32-bit x86 GCC, determinism requires both -msse2 and
  204+         -fpmath=sse.
  205+
  206+         Note that we will not be deterministic with float data containing NaNs -
  207+         the NaNs will propagate differently on different SIMD and platforms.
  208+
  209+         If you turn on STBIR_USE_FMA, then we will be deterministic with other
  210+         fma targets, but we will differ from non-fma targets (this is unavoidable,
  211+         because a fma isn't simply an add with a mult - it also introduces a
  212+         rounding difference compared to non-fma instruction sequences.
  213+
  214+      FLOAT PIXEL FORMAT RANGE
  215+         Any range of values can be used for the non-alpha float data that you pass
  216+         in (0 to 1, -1 to 1, whatever). However, if you are inputting float values
  217+         but *outputting* bytes or shorts, you must use a range of 0 to 1 so that we
  218+         scale back properly. The alpha channel must also be 0 to 1 for any format
  219+         that does premultiplication prior to resizing.
  220+
  221+         Note also that with float output, using filters with negative lobes, the
  222+         output filtered values might go slightly out of range. You can define
  223+         STBIR_FLOAT_LOW_CLAMP and/or STBIR_FLOAT_HIGH_CLAMP to specify the range
  224+         to clamp to on output, if that's important.
  225+
  226+      MAX/MIN SCALE FACTORS
  227+         The input pixel resolutions are in integers, and we do the internal pointer
  228+         resolution in size_t sized integers. However, the scale ratio from input
  229+         resolution to output resolution is calculated in float form. This means
  230+         the effective possible scale ratio is limited to 24 bits (or 16 million
  231+         to 1). As you get close to the size of the float resolution (again, 16
  232+         million pixels wide or high), you might start seeing float inaccuracy
  233+         issues in general in the pipeline. If you have to do extreme resizes,
  234+         you can usually do this is multiple stages (using float intermediate
  235+         buffers).
  236+
  237+      FLIPPED IMAGES
  238+         Stride is just the delta from one scanline to the next. This means you can
  239+         use a negative stride to handle inverted images (point to the final
  240+         scanline and use a negative stride). You can invert the input or output,
  241+         using negative strides.
  242+
  243+      DEFAULT FILTERS
  244+         For functions which don't provide explicit control over what filters to
  245+         use, you can change the compile-time defaults with:
  246+
  247+            #define STBIR_DEFAULT_FILTER_UPSAMPLE     STBIR_FILTER_something
  248+            #define STBIR_DEFAULT_FILTER_DOWNSAMPLE   STBIR_FILTER_something
  249+
  250+         See stbir_filter in the header-file section for the list of filters.
  251+
  252+      NEW FILTERS
  253+         A number of 1D filter kernels are supplied. For a list of supported
  254+         filters, see the stbir_filter enum. You can install your own filters by
  255+         using the stbir_set_filter_callbacks function.
  256+
  257+      PROGRESS
  258+         For interactive use with slow resize operations, you can use the 
  259+         scanline callbacks in the extended API. It would have to be a *very* large
  260+         image resample to need progress though - we're very fast.
  261+
  262+      CEIL and FLOOR
  263+         In scalar mode, the only functions we use from math.h are ceilf and floorf,
  264+         but if you have your own versions, you can define the STBIR_CEILF(v) and
  265+         STBIR_FLOORF(v) macros and we'll use them instead. In SIMD, we just use
  266+         our own versions.
  267+
  268+      ASSERT
  269+         Define STBIR_ASSERT(boolval) to override assert() and not use assert.h
  270+
  271+     PORTING FROM VERSION 1
  272+        The API has changed. You can continue to use the old version of stb_image_resize.h,
  273+        which is available in the "deprecated/" directory.
  274+
  275+        If you're using the old simple-to-use API, porting is straightforward.
  276+        (For more advanced APIs, read the documentation.)
  277+
  278+          stbir_resize_uint8():
  279+            - call `stbir_resize_uint8_linear`, cast channel count to `stbir_pixel_layout`
  280+
  281+          stbir_resize_float():
  282+            - call `stbir_resize_float_linear`, cast channel count to `stbir_pixel_layout`
  283+
  284+          stbir_resize_uint8_srgb():
  285+            - function name is unchanged
  286+            - cast channel count to `stbir_pixel_layout`
  287+            - above is sufficient unless your image has alpha and it's not RGBA/BGRA
  288+              - in that case, follow the below instructions for stbir_resize_uint8_srgb_edgemode
  289+
  290+          stbir_resize_uint8_srgb_edgemode()
  291+            - switch to the "medium complexity" API
  292+            - stbir_resize(), very similar API but a few more parameters:
  293+              - pixel_layout: cast channel count to `stbir_pixel_layout`
  294+              - data_type:    STBIR_TYPE_UINT8_SRGB
  295+              - edge:         unchanged (STBIR_EDGE_WRAP, etc.)
  296+              - filter:       STBIR_FILTER_DEFAULT
  297+            - which channel is alpha is specified in stbir_pixel_layout, see enum for details
  298+
  299+      FUTURE TODOS
  300+        *  For polyphase integral filters, we just memcpy the coeffs to dupe
  301+           them, but we should indirect and use the same coeff memory.
  302+        *  Add pixel layout conversions for sensible different channel counts
  303+           (maybe, 1->3/4, 3->4, 4->1, 3->1).
  304+         * For SIMD encode and decode scanline routines, do any pre-aligning
  305+           for bad input/output buffer alignments and pitch?
  306+         * For very wide scanlines, we should we do vertical strips to stay within
  307+           L2 cache. Maybe do chunks of 1K pixels at a time. There would be
  308+           some pixel reconversion, but probably dwarfed by things falling out
  309+           of cache. Probably also something possible with alternating between
  310+           scattering and gathering at high resize scales?
  311+         * Should we have a multiple MIPs at the same time function (could keep
  312+           more memory in cache during multiple resizes)?
  313+         * Rewrite the coefficient generator to do many at once.
  314+         * AVX-512 vertical kernels - worried about downclocking here.
  315+         * Convert the reincludes to macros when we know they aren't changing.
  316+         * Experiment with pivoting the horizontal and always using the
  317+           vertical filters (which are faster, but perhaps not enough to overcome
  318+           the pivot cost and the extra memory touches). Need to buffer the whole
  319+           image so have to balance memory use.
  320+         * Most of our code is internally function pointers, should we compile
  321+           all the SIMD stuff always and dynamically dispatch?
  322+
  323+   CONTRIBUTORS
  324+      Jeff Roberts: 2.0 implementation, optimizations, SIMD
  325+      Martins Mozeiko: NEON simd, WASM simd, clang and GCC whisperer
  326+      Fabian Giesen: half float and srgb converters
  327+      Sean Barrett: API design, optimizations
  328+      Jorge L Rodriguez: Original 1.0 implementation
  329+      Aras Pranckevicius: bugfixes
  330+      Nathan Reed: warning fixes for 1.0
  331+
  332+   REVISIONS
  333+      2.18 (2026-03-25) fixed coefficient calculation when skipping a coefficient off 
  334+                          the left side of the window, added non-aligned access safe 
  335+                          memcpy mode for scalar path, fixed various typos, and fixed 
  336+                          define error in the float clamp output mode. 
  337+      2.17 (2025-10-25) silly format bug in easy-to-use APIs.
  338+      2.16 (2025-10-21) fixed the easy-to-use APIs to allow inverted bitmaps (negative
  339+                          strides), fix vertical filter kernel callback, fix threaded
  340+                          gather buffer priming (and assert).
  341+                          (thanks adipose, TainZerL, and Harrison Green)
  342+      2.15 (2025-07-17) fixed an assert in debug mode when using floats with input
  343+                          callbacks, work around GCC warning when adding to null ptr
  344+                          (thanks Johannes Spohr and Pyry Kovanen).
  345+      2.14 (2025-05-09) fixed a bug using downsampling gather horizontal first, and 
  346+                          scatter with vertical first.
  347+      2.13 (2025-02-27) fixed a bug when using input callbacks, turned off simd for 
  348+                          tiny-c, fixed some variables that should have been static,
  349+                          fixes a bug when calculating temp memory with resizes that
  350+                          exceed 2GB of temp memory (very large resizes).
  351+      2.12 (2024-10-18) fix incorrect use of user_data with STBIR_FREE
  352+      2.11 (2024-09-08) fix harmless asan warnings in 2-channel and 3-channel mode
  353+                          with AVX-2, fix some weird scaling edge conditions with
  354+                          point sample mode.
  355+      2.10 (2024-07-27) fix the defines GCC and mingw for loop unroll control,
  356+                          fix MSVC 32-bit arm half float routines.
  357+      2.09 (2024-06-19) fix the defines for 32-bit ARM GCC builds (was selecting
  358+                          hardware half floats).
  359+      2.08 (2024-06-10) fix for RGB->BGR three channel flips and add SIMD (thanks
  360+                          to Ryan Salsbury), fix for sub-rect resizes, use the
  361+                          pragmas to control unrolling when they are available.
  362+      2.07 (2024-05-24) fix for slow final split during threaded conversions of very 
  363+                          wide scanlines when downsampling (caused by extra input 
  364+                          converting), fix for wide scanline resamples with many 
  365+                          splits (int overflow), fix GCC warning.
  366+      2.06 (2024-02-10) fix for identical width/height 3x or more down-scaling 
  367+                          undersampling a single row on rare resize ratios (about 1%).
  368+      2.05 (2024-02-07) fix for 2 pixel to 1 pixel resizes with wrap (thanks Aras),
  369+                        fix for output callback (thanks Julien Koenen).
  370+      2.04 (2023-11-17) fix for rare AVX bug, shadowed symbol (thanks Nikola Smiljanic).
  371+      2.03 (2023-11-01) ASAN and TSAN warnings fixed, minor tweaks.
  372+      2.00 (2023-10-10) mostly new source: new api, optimizations, simd, vertical-first, etc
  373+                          2x-5x faster without simd, 4x-12x faster with simd,
  374+                          in some cases, 20x to 40x faster esp resizing large to very small.
  375+      0.96 (2019-03-04) fixed warnings
  376+      0.95 (2017-07-23) fixed warnings
  377+      0.94 (2017-03-18) fixed warnings
  378+      0.93 (2017-03-03) fixed bug with certain combinations of heights
  379+      0.92 (2017-01-02) fix integer overflow on large (>2GB) images
  380+      0.91 (2016-04-02) fix warnings; fix handling of subpixel regions
  381+      0.90 (2014-09-17) first released version
  382+
  383+   LICENSE
  384+     See end of file for license information.
  385+*/
  386+
  387+#if !defined(STB_IMAGE_RESIZE_DO_HORIZONTALS) && !defined(STB_IMAGE_RESIZE_DO_VERTICALS) && !defined(STB_IMAGE_RESIZE_DO_CODERS)   // for internal re-includes
  388+
  389+#ifndef STBIR_INCLUDE_STB_IMAGE_RESIZE2_H
  390+#define STBIR_INCLUDE_STB_IMAGE_RESIZE2_H
  391+
  392+#include <stddef.h>
  393+#ifdef _MSC_VER
  394+typedef unsigned char    stbir_uint8;
  395+typedef unsigned short   stbir_uint16;
  396+typedef unsigned int     stbir_uint32;
  397+typedef unsigned __int64 stbir_uint64;
  398+#else
  399+#include <stdint.h>
  400+typedef uint8_t  stbir_uint8;
  401+typedef uint16_t stbir_uint16;
  402+typedef uint32_t stbir_uint32;
  403+typedef uint64_t stbir_uint64;
  404+#endif
  405+
  406+#ifndef STBIRDEF
  407+#ifdef STB_IMAGE_RESIZE_STATIC
  408+#define STBIRDEF static
  409+#else
  410+#ifdef __cplusplus
  411+#define STBIRDEF extern "C"
  412+#else
  413+#define STBIRDEF extern
  414+#endif
  415+#endif
  416+#endif
  417+
  418+//////////////////////////////////////////////////////////////////////////////
  419+////   start "header file" ///////////////////////////////////////////////////
  420+//
  421+// Easy-to-use API:
  422+//
  423+//     * stride is the offset between successive rows of image data
  424+//        in memory, in bytes. specify 0 for packed continuously in memory
  425+//     * colorspace is linear or sRGB as specified by function name
  426+//     * Uses the default filters
  427+//     * Uses edge mode clamped
  428+//     * returned result is 1 for success or 0 in case of an error.
  429+
  430+
  431+// stbir_pixel_layout specifies:
  432+//   number of channels
  433+//   order of channels
  434+//   whether color is premultiplied by alpha
  435+// for back compatibility, you can cast the old channel count to an stbir_pixel_layout
  436+typedef enum
  437+{
  438+  STBIR_1CHANNEL = 1,
  439+  STBIR_2CHANNEL = 2,
  440+  STBIR_RGB      = 3,               // 3-chan, with order specified (for channel flipping)
  441+  STBIR_BGR      = 0,               // 3-chan, with order specified (for channel flipping)
  442+  STBIR_4CHANNEL = 5,
  443+
  444+  STBIR_RGBA = 4,                   // alpha formats, where alpha is NOT premultiplied into color channels
  445+  STBIR_BGRA = 6,
  446+  STBIR_ARGB = 7,
  447+  STBIR_ABGR = 8,
  448+  STBIR_RA   = 9,
  449+  STBIR_AR   = 10,
  450+
  451+  STBIR_RGBA_PM = 11,               // alpha formats, where alpha is premultiplied into color channels
  452+  STBIR_BGRA_PM = 12,
  453+  STBIR_ARGB_PM = 13,
  454+  STBIR_ABGR_PM = 14,
  455+  STBIR_RA_PM   = 15,
  456+  STBIR_AR_PM   = 16,
  457+
  458+  STBIR_RGBA_NO_AW = 11,            // alpha formats, where NO alpha weighting is applied at all!
  459+  STBIR_BGRA_NO_AW = 12,            //   these are just synonyms for the _PM flags (which also do
  460+  STBIR_ARGB_NO_AW = 13,            //   no alpha weighting). These names just make it more clear
  461+  STBIR_ABGR_NO_AW = 14,            //   for some folks).
  462+  STBIR_RA_NO_AW   = 15,
  463+  STBIR_AR_NO_AW   = 16,
  464+
  465+} stbir_pixel_layout;
  466+
  467+//===============================================================
  468+//  Simple-complexity API
  469+//
  470+//    If output_pixels is NULL (0), then we will allocate the buffer and return it to you.
  471+//--------------------------------
  472+
  473+STBIRDEF unsigned char * stbir_resize_uint8_srgb( const unsigned char *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
  474+                                                        unsigned char *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
  475+                                                        stbir_pixel_layout pixel_type );
  476+
  477+STBIRDEF unsigned char * stbir_resize_uint8_linear( const unsigned char *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
  478+                                                          unsigned char *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
  479+                                                          stbir_pixel_layout pixel_type );
  480+
  481+STBIRDEF float * stbir_resize_float_linear( const float *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
  482+                                                  float *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
  483+                                                  stbir_pixel_layout pixel_type );
  484+//===============================================================
  485+
  486+//===============================================================
  487+// Medium-complexity API
  488+//
  489+// This extends the easy-to-use API as follows:
  490+//
  491+//     * Can specify the datatype - U8, U8_SRGB, U16, FLOAT, HALF_FLOAT
  492+//     * Edge wrap can selected explicitly
  493+//     * Filter can be selected explicitly
  494+//--------------------------------
  495+
  496+typedef enum
  497+{
  498+  STBIR_EDGE_CLAMP   = 0,
  499+  STBIR_EDGE_REFLECT = 1,
  500+  STBIR_EDGE_WRAP    = 2,  // this edge mode is slower and uses more memory
  501+  STBIR_EDGE_ZERO    = 3,
  502+} stbir_edge;
  503+
  504+typedef enum
  505+{
  506+  STBIR_FILTER_DEFAULT      = 0,  // use same filter type that easy-to-use API chooses
  507+  STBIR_FILTER_BOX          = 1,  // A trapezoid w/1-pixel wide ramps, same result as box for integer scale ratios
  508+  STBIR_FILTER_TRIANGLE     = 2,  // On upsampling, produces same results as bilinear texture filtering
  509+  STBIR_FILTER_CUBICBSPLINE = 3,  // The cubic b-spline (aka Mitchell-Netrevalli with B=1,C=0), gaussian-esque
  510+  STBIR_FILTER_CATMULLROM   = 4,  // An interpolating cubic spline
  511+  STBIR_FILTER_MITCHELL     = 5,  // Mitchell-Netrevalli filter with B=1/3, C=1/3
  512+  STBIR_FILTER_POINT_SAMPLE = 6,  // Simple point sampling
  513+  STBIR_FILTER_OTHER        = 7,  // User callback specified
  514+} stbir_filter;
  515+
  516+typedef enum
  517+{
  518+  STBIR_TYPE_UINT8            = 0,
  519+  STBIR_TYPE_UINT8_SRGB       = 1,
  520+  STBIR_TYPE_UINT8_SRGB_ALPHA = 2,  // alpha channel, when present, should also be SRGB (this is very unusual)
  521+  STBIR_TYPE_UINT16           = 3,
  522+  STBIR_TYPE_FLOAT            = 4,
  523+  STBIR_TYPE_HALF_FLOAT       = 5
  524+} stbir_datatype;
  525+
  526+// medium api
  527+STBIRDEF void *  stbir_resize( const void *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
  528+                                     void *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
  529+                               stbir_pixel_layout pixel_layout, stbir_datatype data_type,
  530+                               stbir_edge edge, stbir_filter filter );
  531+//===============================================================
  532+
  533+
  534+
  535+//===============================================================
  536+// Extended-complexity API
  537+//
  538+// This API exposes all resize functionality.
  539+//
  540+//     * Separate filter types for each axis
  541+//     * Separate edge modes for each axis
  542+//     * Separate input and output data types
  543+//     * Can specify regions with subpixel correctness
  544+//     * Can specify alpha flags
  545+//     * Can specify a memory callback
  546+//     * Can specify a callback data type for pixel input and output
  547+//     * Can be threaded for a single resize
  548+//     * Can be used to resize many frames without recalculating the sampler info
  549+//
  550+//  Use this API as follows:
  551+//     1) Call the stbir_resize_init function on a local STBIR_RESIZE structure
  552+//     2) Call any of the stbir_set functions
  553+//     3) Optionally call stbir_build_samplers() if you are going to resample multiple times
  554+//        with the same input and output dimensions (like resizing video frames)
  555+//     4) Resample by calling stbir_resize_extended().
  556+//     5) Call stbir_free_samplers() if you called stbir_build_samplers()
  557+//--------------------------------
  558+
  559+
  560+// Types:
  561+
  562+// INPUT CALLBACK: this callback is used for input scanlines
  563+typedef void const * stbir_input_callback( void * optional_output, void const * input_ptr, int num_pixels, int x, int y, void * context );
  564+
  565+// OUTPUT CALLBACK: this callback is used for output scanlines
  566+typedef void stbir_output_callback( void const * output_ptr, int num_pixels, int y, void * context );
  567+
  568+// callbacks for user installed filters
  569+typedef float stbir__kernel_callback( float x, float scale, void * user_data ); // centered at zero
  570+typedef float stbir__support_callback( float scale, void * user_data );
  571+
  572+// internal structure with precomputed scaling
  573+typedef struct stbir__info stbir__info;
  574+
  575+typedef struct STBIR_RESIZE  // use the stbir_resize_init and stbir_override functions to set these values for future compatibility
  576+{
  577+  void * user_data;
  578+  void const * input_pixels;
  579+  int input_w, input_h;
  580+  double input_s0, input_t0, input_s1, input_t1;
  581+  stbir_input_callback * input_cb;
  582+  void * output_pixels;
  583+  int output_w, output_h;
  584+  int output_subx, output_suby, output_subw, output_subh;
  585+  stbir_output_callback * output_cb;
  586+  int input_stride_in_bytes;
  587+  int output_stride_in_bytes;
  588+  int splits;
  589+  int fast_alpha;
  590+  int needs_rebuild;
  591+  int called_alloc;
  592+  stbir_pixel_layout input_pixel_layout_public;
  593+  stbir_pixel_layout output_pixel_layout_public;
  594+  stbir_datatype input_data_type;
  595+  stbir_datatype output_data_type;
  596+  stbir_filter horizontal_filter, vertical_filter;
  597+  stbir_edge horizontal_edge, vertical_edge;
  598+  stbir__kernel_callback * horizontal_filter_kernel; stbir__support_callback * horizontal_filter_support;
  599+  stbir__kernel_callback * vertical_filter_kernel; stbir__support_callback * vertical_filter_support;
  600+  stbir__info * samplers;
  601+} STBIR_RESIZE;
  602+
  603+// extended complexity api
  604+
  605+
  606+// First off, you must ALWAYS call stbir_resize_init on your resize structure before any of the other calls!
  607+STBIRDEF void stbir_resize_init( STBIR_RESIZE * resize,
  608+                                 const void *input_pixels,  int input_w,  int input_h, int input_stride_in_bytes, // stride can be zero
  609+                                       void *output_pixels, int output_w, int output_h, int output_stride_in_bytes, // stride can be zero
  610+                                 stbir_pixel_layout pixel_layout, stbir_datatype data_type );
  611+
  612+//===============================================================
  613+// You can update these parameters any time after resize_init and there is no cost
  614+//--------------------------------
  615+
  616+STBIRDEF void stbir_set_datatypes( STBIR_RESIZE * resize, stbir_datatype input_type, stbir_datatype output_type );
  617+STBIRDEF void stbir_set_pixel_callbacks( STBIR_RESIZE * resize, stbir_input_callback * input_cb, stbir_output_callback * output_cb );   // no callbacks by default
  618+STBIRDEF void stbir_set_user_data( STBIR_RESIZE * resize, void * user_data );                                               // pass back STBIR_RESIZE* by default
  619+STBIRDEF void stbir_set_buffer_ptrs( STBIR_RESIZE * resize, const void * input_pixels, int input_stride_in_bytes, void * output_pixels, int output_stride_in_bytes );
  620+
  621+//===============================================================
  622+
  623+
  624+//===============================================================
  625+// If you call any of these functions, you will trigger a sampler rebuild!
  626+//--------------------------------
  627+
  628+STBIRDEF int stbir_set_pixel_layouts( STBIR_RESIZE * resize, stbir_pixel_layout input_pixel_layout, stbir_pixel_layout output_pixel_layout );  // sets new buffer layouts
  629+STBIRDEF int stbir_set_edgemodes( STBIR_RESIZE * resize, stbir_edge horizontal_edge, stbir_edge vertical_edge );       // CLAMP by default
  630+
  631+STBIRDEF int stbir_set_filters( STBIR_RESIZE * resize, stbir_filter horizontal_filter, stbir_filter vertical_filter ); // STBIR_DEFAULT_FILTER_UPSAMPLE/DOWNSAMPLE by default
  632+STBIRDEF int stbir_set_filter_callbacks( STBIR_RESIZE * resize, stbir__kernel_callback * horizontal_filter, stbir__support_callback * horizontal_support, stbir__kernel_callback * vertical_filter, stbir__support_callback * vertical_support );
  633+
  634+STBIRDEF int stbir_set_pixel_subrect( STBIR_RESIZE * resize, int subx, int suby, int subw, int subh );        // sets both sub-regions (full regions by default)
  635+STBIRDEF int stbir_set_input_subrect( STBIR_RESIZE * resize, double s0, double t0, double s1, double t1 );    // sets input sub-region (full region by default)
  636+STBIRDEF int stbir_set_output_pixel_subrect( STBIR_RESIZE * resize, int subx, int suby, int subw, int subh ); // sets output sub-region (full region by default)
  637+
  638+// when inputting AND outputting non-premultiplied alpha pixels, we use a slower but higher quality technique
  639+//   that fills the zero alpha pixel's RGB values with something plausible.  If you don't care about areas of
  640+//   zero alpha, you can call this function to get about a 25% speed improvement for STBIR_RGBA to STBIR_RGBA
  641+//   types of resizes.
  642+STBIRDEF int stbir_set_non_pm_alpha_speed_over_quality( STBIR_RESIZE * resize, int non_pma_alpha_speed_over_quality );
  643+//===============================================================
  644+
  645+
  646+//===============================================================
  647+// You can call build_samplers to prebuild all the internal data we need to resample.
  648+//   Then, if you call resize_extended many times with the same resize, you only pay the
  649+//   cost once.
  650+// If you do call build_samplers, you MUST call free_samplers eventually.
  651+//--------------------------------
  652+
  653+// This builds the samplers and does one allocation
  654+STBIRDEF int stbir_build_samplers( STBIR_RESIZE * resize );
  655+
  656+// You MUST call this, if you call stbir_build_samplers or stbir_build_samplers_with_splits
  657+STBIRDEF void stbir_free_samplers( STBIR_RESIZE * resize );
  658+//===============================================================
  659+
  660+
  661+// And this is the main function to perform the resize synchronously on one thread.
  662+STBIRDEF int stbir_resize_extended( STBIR_RESIZE * resize );
  663+
  664+
  665+//===============================================================
  666+// Use these functions for multithreading.
  667+//   1) You call stbir_build_samplers_with_splits first on the main thread
  668+//   2) Then stbir_resize_with_split on each thread
  669+//   3) stbir_free_samplers when done on the main thread
  670+//--------------------------------
  671+
  672+// This will build samplers for threading.
  673+//   You can pass in the number of threads you'd like to use (try_splits).
  674+//   It returns the number of splits (threads) that you can call it with.
  675+///  It might be less if the image resize can't be split up that many ways.
  676+
  677+STBIRDEF int stbir_build_samplers_with_splits( STBIR_RESIZE * resize, int try_splits );
  678+
  679+// This function does a split of the resizing (you call this fuction for each
  680+// split, on multiple threads). A split is a piece of the output resize pixel space.
  681+
  682+// Note that you MUST call stbir_build_samplers_with_splits before stbir_resize_extended_split!
  683+
  684+// Usually, you will always call stbir_resize_split with split_start as the thread_index
  685+//   and "1" for the split_count.
  686+// But, if you have a weird situation where you MIGHT want 8 threads, but sometimes
  687+//   only 4 threads, you can use 0,2,4,6 for the split_start's and use "2" for the
  688+//   split_count each time to turn in into a 4 thread resize. (This is unusual).
  689+
  690+STBIRDEF int stbir_resize_extended_split( STBIR_RESIZE * resize, int split_start, int split_count );
  691+//===============================================================
  692+
  693+
  694+//===============================================================
  695+// Pixel Callbacks info:
  696+//--------------------------------
  697+
  698+//   The input callback is super flexible - it calls you with the input address
  699+//   (based on the stride and base pointer), it gives you an optional_output
  700+//   pointer that you can fill, or you can just return your own pointer into
  701+//   your own data.
  702+//
  703+//   You can also do conversion from non-supported data types if necessary - in
  704+//   this case, you ignore the input_ptr and just use the x and y parameters to
  705+//   calculate your own input_ptr based on the size of each non-supported pixel.
  706+//   (Something like the third example below.)
  707+//
  708+//   You can also install just an input or just an output callback by setting the
  709+//   callback that you don't want to zero.
  710+//
  711+//     First example, progress: (getting a callback that you can monitor the progress):
  712+//        void const * my_callback( void * optional_output, void const * input_ptr, int num_pixels, int x, int y, void * context )
  713+//        {
  714+//           percentage_done = y / input_height;
  715+//           return input_ptr;  // use buffer from call
  716+//        }
  717+//
  718+//     Next example, copying: (copy from some other buffer or stream):
  719+//        void const * my_callback( void * optional_output, void const * input_ptr, int num_pixels, int x, int y, void * context )
  720+//        {
  721+//           CopyOrStreamData( optional_output, other_data_src, num_pixels * pixel_width_in_bytes );
  722+//           return optional_output;  // return the optional buffer that we filled
  723+//        }
  724+//
  725+//     Third example, input another buffer without copying: (zero-copy from other buffer):
  726+//        void const * my_callback( void * optional_output, void const * input_ptr, int num_pixels, int x, int y, void * context )
  727+//        {
  728+//           void * pixels = ( (char*) other_image_base ) + ( y * other_image_stride ) + ( x * other_pixel_width_in_bytes );
  729+//           return pixels;       // return pointer to your data without copying
  730+//        }
  731+//
  732+//
  733+//   The output callback is considerably simpler - it just calls you so that you can dump
  734+//   out each scanline. You could even directly copy out to disk if you have a simple format
  735+//   like TGA or BMP. You can also convert to other output types here if you want.
  736+//
  737+//   Simple example:
  738+//        void const * my_output( void * output_ptr, int num_pixels, int y, void * context )
  739+//        {
  740+//           percentage_done = y / output_height;
  741+//           fwrite( output_ptr, pixel_width_in_bytes, num_pixels, output_file );
  742+//        }
  743+//===============================================================
  744+
  745+
  746+
  747+
  748+//===============================================================
  749+// optional built-in profiling API
  750+//--------------------------------
  751+
  752+#ifdef STBIR_PROFILE
  753+
  754+typedef struct STBIR_PROFILE_INFO
  755+{
  756+  stbir_uint64 total_clocks;
  757+
  758+  // how many clocks spent (of total_clocks) in the various resize routines, along with a string description
  759+  //    there are "resize_count" number of zones
  760+  stbir_uint64 clocks[ 8 ];
  761+  char const ** descriptions;
  762+
  763+  // count of clocks and descriptions
  764+  stbir_uint32 count;
  765+} STBIR_PROFILE_INFO;
  766+
  767+// use after calling stbir_resize_extended (or stbir_build_samplers or stbir_build_samplers_with_splits)
  768+STBIRDEF void stbir_resize_build_profile_info( STBIR_PROFILE_INFO * out_info, STBIR_RESIZE const * resize );
  769+
  770+// use after calling stbir_resize_extended
  771+STBIRDEF void stbir_resize_extended_profile_info( STBIR_PROFILE_INFO * out_info, STBIR_RESIZE const * resize );
  772+
  773+// use after calling stbir_resize_extended_split
  774+STBIRDEF void stbir_resize_split_profile_info( STBIR_PROFILE_INFO * out_info, STBIR_RESIZE const * resize, int split_start, int split_num );
  775+
  776+//===============================================================
  777+
  778+#endif
  779+
  780+
  781+////   end header file   /////////////////////////////////////////////////////
  782+#endif // STBIR_INCLUDE_STB_IMAGE_RESIZE2_H
  783+
  784+#if defined(STB_IMAGE_RESIZE_IMPLEMENTATION) || defined(STB_IMAGE_RESIZE2_IMPLEMENTATION)
  785+
  786+#ifndef STBIR_ASSERT
  787+#include <assert.h>
  788+#define STBIR_ASSERT(x) assert(x)
  789+#endif
  790+
  791+#ifndef STBIR_MALLOC
  792+#include <stdlib.h>
  793+#define STBIR_MALLOC(size,user_data) ((void)(user_data), malloc(size))
  794+#define STBIR_FREE(ptr,user_data)    ((void)(user_data), free(ptr))
  795+// (we used the comma operator to evaluate user_data, to avoid "unused parameter" warnings)
  796+#endif
  797+
  798+#ifdef _MSC_VER
  799+
  800+#define stbir__inline __forceinline
  801+
  802+#else
  803+
  804+#define stbir__inline __inline__
  805+
  806+// Clang address sanitizer
  807+#if defined(__has_feature)
  808+  #if __has_feature(address_sanitizer) || __has_feature(memory_sanitizer)
  809+    #ifndef STBIR__SEPARATE_ALLOCATIONS
  810+      #define STBIR__SEPARATE_ALLOCATIONS
  811+    #endif
  812+  #endif
  813+#endif
  814+
  815+#endif
  816+
  817+// GCC and MSVC
  818+#if defined(__SANITIZE_ADDRESS__)
  819+  #ifndef STBIR__SEPARATE_ALLOCATIONS
  820+    #define STBIR__SEPARATE_ALLOCATIONS
  821+  #endif
  822+#endif
  823+
  824+// Always turn off automatic FMA use - use STBIR_USE_FMA if you want.
  825+// Otherwise, this is a determinism disaster.
  826+#ifndef STBIR_DONT_CHANGE_FP_CONTRACT  // override in case you don't want this behavior
  827+#if defined(_MSC_VER) && !defined(__clang__)
  828+#if _MSC_VER > 1200
  829+#pragma fp_contract(off)
  830+#endif
  831+#elif defined(__GNUC__) &&  !defined(__clang__)
  832+#pragma GCC optimize("fp-contract=off")
  833+#else
  834+#pragma STDC FP_CONTRACT OFF
  835+#endif
  836+#endif
  837+
  838+#ifdef _MSC_VER
  839+#define STBIR__UNUSED(v)  (void)(v)
  840+#else
  841+#define STBIR__UNUSED(v)  (void)sizeof(v)
  842+#endif
  843+
  844+#define STBIR__ARRAY_SIZE(a) (sizeof((a))/sizeof((a)[0]))
  845+
  846+
  847+#ifndef STBIR_DEFAULT_FILTER_UPSAMPLE
  848+#define STBIR_DEFAULT_FILTER_UPSAMPLE    STBIR_FILTER_CATMULLROM
  849+#endif
  850+
  851+#ifndef STBIR_DEFAULT_FILTER_DOWNSAMPLE
  852+#define STBIR_DEFAULT_FILTER_DOWNSAMPLE  STBIR_FILTER_MITCHELL
  853+#endif
  854+
  855+
  856+#ifndef STBIR__HEADER_FILENAME
  857+#define STBIR__HEADER_FILENAME "stb_image_resize2.h"
  858+#endif
  859+
  860+// the internal pixel layout enums are in a different order, so we can easily do range comparisons of types
  861+//   the public pixel layout is ordered in a way that if you cast num_channels (1-4) to the enum, you get something sensible
  862+typedef enum
  863+{
  864+  STBIRI_1CHANNEL = 0,
  865+  STBIRI_2CHANNEL = 1,
  866+  STBIRI_RGB      = 2,
  867+  STBIRI_BGR      = 3,
  868+  STBIRI_4CHANNEL = 4,
  869+
  870+  STBIRI_RGBA = 5,
  871+  STBIRI_BGRA = 6,
  872+  STBIRI_ARGB = 7,
  873+  STBIRI_ABGR = 8,
  874+  STBIRI_RA   = 9,
  875+  STBIRI_AR   = 10,
  876+
  877+  STBIRI_RGBA_PM = 11,
  878+  STBIRI_BGRA_PM = 12,
  879+  STBIRI_ARGB_PM = 13,
  880+  STBIRI_ABGR_PM = 14,
  881+  STBIRI_RA_PM   = 15,
  882+  STBIRI_AR_PM   = 16,
  883+} stbir_internal_pixel_layout;
  884+
  885+// define the public pixel layouts to not compile inside the implementation (to avoid accidental use)
  886+#define STBIR_BGR bad_dont_use_in_implementation
  887+#define STBIR_1CHANNEL STBIR_BGR
  888+#define STBIR_2CHANNEL STBIR_BGR
  889+#define STBIR_RGB STBIR_BGR
  890+#define STBIR_RGBA STBIR_BGR
  891+#define STBIR_4CHANNEL STBIR_BGR
  892+#define STBIR_BGRA STBIR_BGR
  893+#define STBIR_ARGB STBIR_BGR
  894+#define STBIR_ABGR STBIR_BGR
  895+#define STBIR_RA STBIR_BGR
  896+#define STBIR_AR STBIR_BGR
  897+#define STBIR_RGBA_PM STBIR_BGR
  898+#define STBIR_BGRA_PM STBIR_BGR
  899+#define STBIR_ARGB_PM STBIR_BGR
  900+#define STBIR_ABGR_PM STBIR_BGR
  901+#define STBIR_RA_PM STBIR_BGR
  902+#define STBIR_AR_PM STBIR_BGR
  903+
  904+// must match stbir_datatype
  905+static unsigned char stbir__type_size[] = {
  906+  1,1,1,2,4,2 // STBIR_TYPE_UINT8,STBIR_TYPE_UINT8_SRGB,STBIR_TYPE_UINT8_SRGB_ALPHA,STBIR_TYPE_UINT16,STBIR_TYPE_FLOAT,STBIR_TYPE_HALF_FLOAT
  907+};
  908+
  909+// When gathering, the contributors are which source pixels contribute.
  910+// When scattering, the contributors are which destination pixels are contributed to.
  911+typedef struct
  912+{
  913+  int n0; // First contributing pixel
  914+  int n1; // Last contributing pixel
  915+} stbir__contributors;
  916+
  917+typedef struct
  918+{
  919+  int lowest;    // First sample index for whole filter
  920+  int highest;   // Last sample index for whole filter
  921+  int widest;    // widest single set of samples for an output
  922+} stbir__filter_extent_info;
  923+
  924+typedef struct
  925+{
  926+  int n0; // First pixel of decode buffer to write to
  927+  int n1; // Last pixel of decode that will be written to
  928+  int pixel_offset_for_input;  // Pixel offset into input_scanline
  929+} stbir__span;
  930+
  931+typedef struct stbir__scale_info
  932+{
  933+  int input_full_size;
  934+  int output_sub_size;
  935+  float scale;
  936+  float inv_scale;
  937+  float pixel_shift; // starting shift in output pixel space (in pixels)
  938+  int scale_is_rational;
  939+  stbir_uint32 scale_numerator, scale_denominator;
  940+} stbir__scale_info;
  941+
  942+typedef struct
  943+{
  944+  stbir__contributors * contributors;
  945+  float* coefficients;
  946+  stbir__contributors * gather_prescatter_contributors;
  947+  float * gather_prescatter_coefficients;
  948+  stbir__scale_info scale_info;
  949+  float support;
  950+  stbir_filter filter_enum;
  951+  stbir__kernel_callback * filter_kernel;
  952+  stbir__support_callback * filter_support;
  953+  stbir_edge edge;
  954+  int coefficient_width;
  955+  int filter_pixel_width;
  956+  int filter_pixel_margin;
  957+  int num_contributors;
  958+  int contributors_size;
  959+  int coefficients_size;
  960+  stbir__filter_extent_info extent_info;
  961+  int is_gather;  // 0 = scatter, 1 = gather with scale >= 1, 2 = gather with scale < 1
  962+  int gather_prescatter_num_contributors;
  963+  int gather_prescatter_coefficient_width;
  964+  int gather_prescatter_contributors_size;
  965+  int gather_prescatter_coefficients_size;
  966+} stbir__sampler;
  967+
  968+typedef struct
  969+{
  970+  stbir__contributors conservative;
  971+  int edge_sizes[2];    // this can be less than filter_pixel_margin, if the filter and scaling falls off
  972+  stbir__span spans[2]; // can be two spans, if doing input subrect with clamp mode WRAP
  973+} stbir__extents;
  974+
  975+typedef struct
  976+{
  977+#ifdef STBIR_PROFILE
  978+  union
  979+  {
  980+    struct { stbir_uint64 total, looping, vertical, horizontal, decode, encode, alpha, unalpha; } named;
  981+    stbir_uint64 array[8];
  982+  } profile;
  983+  stbir_uint64 * current_zone_excluded_ptr;
  984+#endif
  985+  float* decode_buffer;
  986+
  987+  int ring_buffer_first_scanline;
  988+  int ring_buffer_last_scanline;
  989+  int ring_buffer_begin_index;    // first_scanline is at this index in the ring buffer
  990+  int start_output_y, end_output_y;
  991+  int start_input_y, end_input_y;  // used in scatter only
  992+
  993+  #ifdef STBIR__SEPARATE_ALLOCATIONS
  994+    float** ring_buffers; // one pointer for each ring buffer
  995+  #else
  996+    float* ring_buffer;  // one big buffer that we index into
  997+  #endif
  998+
  999+  float* vertical_buffer;
 1000+
 1001+  char no_cache_straddle[64];
 1002+} stbir__per_split_info;
 1003+
 1004+typedef float * stbir__decode_pixels_func( float * decode, int width_times_channels, void const * input );
 1005+typedef void stbir__alpha_weight_func( float * decode_buffer, int width_times_channels );
 1006+typedef void stbir__horizontal_gather_channels_func( float * output_buffer, unsigned int output_sub_size, float const * decode_buffer,
 1007+  stbir__contributors const * horizontal_contributors, float const * horizontal_coefficients, int coefficient_width );
 1008+typedef void stbir__alpha_unweight_func(float * encode_buffer, int width_times_channels );
 1009+typedef void stbir__encode_pixels_func( void * output, int width_times_channels, float const * encode );
 1010+
 1011+struct stbir__info
 1012+{
 1013+#ifdef STBIR_PROFILE
 1014+  union
 1015+  {
 1016+    struct { stbir_uint64 total, build, alloc, horizontal, vertical, cleanup, pivot; } named;
 1017+    stbir_uint64 array[7];
 1018+  } profile;
 1019+  stbir_uint64 * current_zone_excluded_ptr;
 1020+#endif
 1021+  stbir__sampler horizontal;
 1022+  stbir__sampler vertical;
 1023+
 1024+  void const * input_data;
 1025+  void * output_data;
 1026+
 1027+  int input_stride_bytes;
 1028+  int output_stride_bytes;
 1029+  int ring_buffer_length_bytes;   // The length of an individual entry in the ring buffer. The total number of ring buffers is stbir__get_filter_pixel_width(filter)
 1030+  int ring_buffer_num_entries;    // Total number of entries in the ring buffer.
 1031+
 1032+  stbir_datatype input_type;
 1033+  stbir_datatype output_type;
 1034+
 1035+  stbir_input_callback * in_pixels_cb;
 1036+  void * user_data;
 1037+  stbir_output_callback * out_pixels_cb;
 1038+
 1039+  stbir__extents scanline_extents;
 1040+
 1041+  void * alloced_mem;
 1042+  stbir__per_split_info * split_info;  // by default 1, but there will be N of these allocated based on the thread init you did
 1043+
 1044+  stbir__decode_pixels_func * decode_pixels;
 1045+  stbir__alpha_weight_func * alpha_weight;
 1046+  stbir__horizontal_gather_channels_func * horizontal_gather_channels;
 1047+  stbir__alpha_unweight_func * alpha_unweight;
 1048+  stbir__encode_pixels_func * encode_pixels;
 1049+
 1050+  int alloc_ring_buffer_num_entries;    // Number of entries in the ring buffer that will be allocated
 1051+  int splits; // count of splits
 1052+
 1053+  stbir_internal_pixel_layout input_pixel_layout_internal;
 1054+  stbir_internal_pixel_layout output_pixel_layout_internal;
 1055+
 1056+  int input_color_and_type;
 1057+  int offset_x, offset_y; // offset within output_data
 1058+  int vertical_first;
 1059+  int channels;
 1060+  int effective_channels; // same as channels, except on RGBA/ARGB (7), or XA/AX (3)
 1061+  size_t alloced_total;
 1062+};
 1063+
 1064+
 1065+#define stbir__max_uint8_as_float             255.0f
 1066+#define stbir__max_uint16_as_float            65535.0f
 1067+#define stbir__max_uint8_as_float_inverted    3.9215689e-03f     // (1.0f/255.0f)
 1068+#define stbir__max_uint16_as_float_inverted   1.5259022e-05f     // (1.0f/65535.0f)
 1069+#define stbir__small_float ((float)1 / (1 << 20) / (1 << 20) / (1 << 20) / (1 << 20) / (1 << 20) / (1 << 20))
 1070+
 1071+// min/max friendly
 1072+#define STBIR_CLAMP(x, xmin, xmax) for(;;) { \
 1073+  if ( (x) < (xmin) ) (x) = (xmin);     \
 1074+  if ( (x) > (xmax) ) (x) = (xmax);     \
 1075+  break;                                \
 1076+}
 1077+
 1078+static stbir__inline int stbir__min(int a, int b)
 1079+{
 1080+  return a < b ? a : b;
 1081+}
 1082+
 1083+static stbir__inline int stbir__max(int a, int b)
 1084+{
 1085+  return a > b ? a : b;
 1086+}
 1087+
 1088+static float stbir__srgb_uchar_to_linear_float[256] = {
 1089+  0.000000f, 0.000304f, 0.000607f, 0.000911f, 0.001214f, 0.001518f, 0.001821f, 0.002125f, 0.002428f, 0.002732f, 0.003035f,
 1090+  0.003347f, 0.003677f, 0.004025f, 0.004391f, 0.004777f, 0.005182f, 0.005605f, 0.006049f, 0.006512f, 0.006995f, 0.007499f,
 1091+  0.008023f, 0.008568f, 0.009134f, 0.009721f, 0.010330f, 0.010960f, 0.011612f, 0.012286f, 0.012983f, 0.013702f, 0.014444f,
 1092+  0.015209f, 0.015996f, 0.016807f, 0.017642f, 0.018500f, 0.019382f, 0.020289f, 0.021219f, 0.022174f, 0.023153f, 0.024158f,
 1093+  0.025187f, 0.026241f, 0.027321f, 0.028426f, 0.029557f, 0.030713f, 0.031896f, 0.033105f, 0.034340f, 0.035601f, 0.036889f,
 1094+  0.038204f, 0.039546f, 0.040915f, 0.042311f, 0.043735f, 0.045186f, 0.046665f, 0.048172f, 0.049707f, 0.051269f, 0.052861f,
 1095+  0.054480f, 0.056128f, 0.057805f, 0.059511f, 0.061246f, 0.063010f, 0.064803f, 0.066626f, 0.068478f, 0.070360f, 0.072272f,
 1096+  0.074214f, 0.076185f, 0.078187f, 0.080220f, 0.082283f, 0.084376f, 0.086500f, 0.088656f, 0.090842f, 0.093059f, 0.095307f,
 1097+  0.097587f, 0.099899f, 0.102242f, 0.104616f, 0.107023f, 0.109462f, 0.111932f, 0.114435f, 0.116971f, 0.119538f, 0.122139f,
 1098+  0.124772f, 0.127438f, 0.130136f, 0.132868f, 0.135633f, 0.138432f, 0.141263f, 0.144128f, 0.147027f, 0.149960f, 0.152926f,
 1099+  0.155926f, 0.158961f, 0.162029f, 0.165132f, 0.168269f, 0.171441f, 0.174647f, 0.177888f, 0.181164f, 0.184475f, 0.187821f,
 1100+  0.191202f, 0.194618f, 0.198069f, 0.201556f, 0.205079f, 0.208637f, 0.212231f, 0.215861f, 0.219526f, 0.223228f, 0.226966f,
 1101+  0.230740f, 0.234551f, 0.238398f, 0.242281f, 0.246201f, 0.250158f, 0.254152f, 0.258183f, 0.262251f, 0.266356f, 0.270498f,
 1102+  0.274677f, 0.278894f, 0.283149f, 0.287441f, 0.291771f, 0.296138f, 0.300544f, 0.304987f, 0.309469f, 0.313989f, 0.318547f,
 1103+  0.323143f, 0.327778f, 0.332452f, 0.337164f, 0.341914f, 0.346704f, 0.351533f, 0.356400f, 0.361307f, 0.366253f, 0.371238f,
 1104+  0.376262f, 0.381326f, 0.386430f, 0.391573f, 0.396755f, 0.401978f, 0.407240f, 0.412543f, 0.417885f, 0.423268f, 0.428691f,
 1105+  0.434154f, 0.439657f, 0.445201f, 0.450786f, 0.456411f, 0.462077f, 0.467784f, 0.473532f, 0.479320f, 0.485150f, 0.491021f,
 1106+  0.496933f, 0.502887f, 0.508881f, 0.514918f, 0.520996f, 0.527115f, 0.533276f, 0.539480f, 0.545725f, 0.552011f, 0.558340f,
 1107+  0.564712f, 0.571125f, 0.577581f, 0.584078f, 0.590619f, 0.597202f, 0.603827f, 0.610496f, 0.617207f, 0.623960f, 0.630757f,
 1108+  0.637597f, 0.644480f, 0.651406f, 0.658375f, 0.665387f, 0.672443f, 0.679543f, 0.686685f, 0.693872f, 0.701102f, 0.708376f,
 1109+  0.715694f, 0.723055f, 0.730461f, 0.737911f, 0.745404f, 0.752942f, 0.760525f, 0.768151f, 0.775822f, 0.783538f, 0.791298f,
 1110+  0.799103f, 0.806952f, 0.814847f, 0.822786f, 0.830770f, 0.838799f, 0.846873f, 0.854993f, 0.863157f, 0.871367f, 0.879622f,
 1111+  0.887923f, 0.896269f, 0.904661f, 0.913099f, 0.921582f, 0.930111f, 0.938686f, 0.947307f, 0.955974f, 0.964686f, 0.973445f,
 1112+  0.982251f, 0.991102f, 1.0f
 1113+};
 1114+
 1115+typedef union
 1116+{
 1117+  unsigned int u;
 1118+  float f;
 1119+} stbir__FP32;
 1120+
 1121+// From https://gist.github.com/rygorous/2203834
 1122+
 1123+static const stbir_uint32 fp32_to_srgb8_tab4[104] = {
 1124+  0x0073000d, 0x007a000d, 0x0080000d, 0x0087000d, 0x008d000d, 0x0094000d, 0x009a000d, 0x00a1000d,
 1125+  0x00a7001a, 0x00b4001a, 0x00c1001a, 0x00ce001a, 0x00da001a, 0x00e7001a, 0x00f4001a, 0x0101001a,
 1126+  0x010e0033, 0x01280033, 0x01410033, 0x015b0033, 0x01750033, 0x018f0033, 0x01a80033, 0x01c20033,
 1127+  0x01dc0067, 0x020f0067, 0x02430067, 0x02760067, 0x02aa0067, 0x02dd0067, 0x03110067, 0x03440067,
 1128+  0x037800ce, 0x03df00ce, 0x044600ce, 0x04ad00ce, 0x051400ce, 0x057b00c5, 0x05dd00bc, 0x063b00b5,
 1129+  0x06970158, 0x07420142, 0x07e30130, 0x087b0120, 0x090b0112, 0x09940106, 0x0a1700fc, 0x0a9500f2,
 1130+  0x0b0f01cb, 0x0bf401ae, 0x0ccb0195, 0x0d950180, 0x0e56016e, 0x0f0d015e, 0x0fbc0150, 0x10630143,
 1131+  0x11070264, 0x1238023e, 0x1357021d, 0x14660201, 0x156601e9, 0x165a01d3, 0x174401c0, 0x182401af,
 1132+  0x18fe0331, 0x1a9602fe, 0x1c1502d2, 0x1d7e02ad, 0x1ed4028d, 0x201a0270, 0x21520256, 0x227d0240,
 1133+  0x239f0443, 0x25c003fe, 0x27bf03c4, 0x29a10392, 0x2b6a0367, 0x2d1d0341, 0x2ebe031f, 0x304d0300,
 1134+  0x31d105b0, 0x34a80555, 0x37520507, 0x39d504c5, 0x3c37048b, 0x3e7c0458, 0x40a8042a, 0x42bd0401,
 1135+  0x44c20798, 0x488e071e, 0x4c1c06b6, 0x4f76065d, 0x52a50610, 0x55ac05cc, 0x5892058f, 0x5b590559,
 1136+  0x5e0c0a23, 0x631c0980, 0x67db08f6, 0x6c55087f, 0x70940818, 0x74a007bd, 0x787d076c, 0x7c330723,
 1137+};
 1138+
 1139+static stbir__inline stbir_uint8 stbir__linear_to_srgb_uchar(float in)
 1140+{
 1141+  static const stbir__FP32 almostone = { 0x3f7fffff }; // 1-eps
 1142+  static const stbir__FP32 minval = { (127-13) << 23 };
 1143+  stbir_uint32 tab,bias,scale,t;
 1144+  stbir__FP32 f;
 1145+
 1146+  // Clamp to [2^(-13), 1-eps]; these two values map to 0 and 1, respectively.
 1147+  // The tests are carefully written so that NaNs map to 0, same as in the reference
 1148+  // implementation.
 1149+  if (!(in > minval.f)) // written this way to catch NaNs
 1150+      return 0;
 1151+  if (in > almostone.f)
 1152+      return 255;
 1153+
 1154+  // Do the table lookup and unpack bias, scale
 1155+  f.f = in;
 1156+  tab = fp32_to_srgb8_tab4[(f.u - minval.u) >> 20];
 1157+  bias = (tab >> 16) << 9;
 1158+  scale = tab & 0xffff;
 1159+
 1160+  // Grab next-highest mantissa bits and perform linear interpolation
 1161+  t = (f.u >> 12) & 0xff;
 1162+  return (unsigned char) ((bias + scale*t) >> 16);
 1163+}
 1164+
 1165+#ifndef STBIR_FORCE_GATHER_FILTER_SCANLINES_AMOUNT
 1166+#define STBIR_FORCE_GATHER_FILTER_SCANLINES_AMOUNT 32 // when downsampling and <= 32 scanlines of buffering, use gather. gather used down to 1/8th scaling for 25% win.
 1167+#endif
 1168+
 1169+#ifndef STBIR_FORCE_MINIMUM_SCANLINES_FOR_SPLITS
 1170+#define STBIR_FORCE_MINIMUM_SCANLINES_FOR_SPLITS 4 // when threading, what is the minimum number of scanlines for a split?
 1171+#endif
 1172+
 1173+#define STBIR_INPUT_CALLBACK_PADDING 3
 1174+
 1175+#ifdef _M_IX86_FP
 1176+#if ( _M_IX86_FP >= 1 )
 1177+#ifndef STBIR_SSE
 1178+#define STBIR_SSE
 1179+#endif
 1180+#endif
 1181+#endif
 1182+
 1183+#ifdef __TINYC__
 1184+  // tiny c has no intrinsics yet - this can become a version check if they add them
 1185+  #define STBIR_NO_SIMD
 1186+#endif
 1187+
 1188+#if defined(_x86_64) || defined( __x86_64__ ) || defined( _M_X64 ) || defined(__x86_64) || defined(_M_AMD64) || defined(__SSE2__) || defined(STBIR_SSE) || defined(STBIR_SSE2)
 1189+  #ifndef STBIR_SSE2
 1190+    #define STBIR_SSE2
 1191+  #endif
 1192+  #if defined(__AVX__) || defined(STBIR_AVX2)
 1193+    #ifndef STBIR_AVX
 1194+      #ifndef STBIR_NO_AVX
 1195+        #define STBIR_AVX
 1196+      #endif
 1197+    #endif
 1198+  #endif
 1199+  #if defined(__AVX2__) || defined(STBIR_AVX2)
 1200+    #ifndef STBIR_NO_AVX2
 1201+      #ifndef STBIR_AVX2
 1202+        #define STBIR_AVX2
 1203+      #endif
 1204+      #if defined( _MSC_VER ) && !defined(__clang__)
 1205+        #ifndef STBIR_FP16C  // FP16C instructions are on all AVX2 cpus, so we can autoselect it here on microsoft - clang needs -mf16c
 1206+          #define STBIR_FP16C
 1207+        #endif
 1208+      #endif
 1209+    #endif
 1210+  #endif
 1211+  #ifdef __F16C__
 1212+    #ifndef STBIR_FP16C  // turn on FP16C instructions if the define is set (for clang and gcc)
 1213+      #define STBIR_FP16C
 1214+    #endif
 1215+  #endif
 1216+#endif
 1217+
 1218+#if defined( _M_ARM64 ) || defined( __aarch64__ ) || defined( __arm64__ ) || ((__ARM_NEON_FP & 4) != 0) || defined(__ARM_NEON__)
 1219+#ifndef STBIR_NEON
 1220+#define STBIR_NEON
 1221+#endif
 1222+#endif
 1223+
 1224+#if defined(_M_ARM) || defined(__arm__)
 1225+#ifdef STBIR_USE_FMA
 1226+#undef STBIR_USE_FMA // no FMA for 32-bit arm on MSVC
 1227+#endif
 1228+#endif
 1229+
 1230+#if defined(__wasm__) && defined(__wasm_simd128__)
 1231+#ifndef STBIR_WASM
 1232+#define STBIR_WASM
 1233+#endif
 1234+#endif
 1235+
 1236+// restrict pointers for the output pointers, other loop and unroll control
 1237+#if defined( _MSC_VER ) && !defined(__clang__)
 1238+  #define STBIR_STREAMOUT_PTR( star ) star __restrict
 1239+  #define STBIR_NO_UNROLL( ptr ) __assume(ptr) // this oddly keeps msvc from unrolling a loop
 1240+  #if _MSC_VER >= 1900
 1241+    #define STBIR_NO_UNROLL_LOOP_START __pragma(loop( no_vector )) 
 1242+  #else
 1243+    #define STBIR_NO_UNROLL_LOOP_START 
 1244+  #endif
 1245+#elif defined( __clang__ )
 1246+  #define STBIR_STREAMOUT_PTR( star ) star __restrict__
 1247+  #define STBIR_NO_UNROLL( ptr ) __asm__ (""::"r"(ptr)) 
 1248+  #if ( __clang_major__ >= 4 ) || ( ( __clang_major__ >= 3 ) && ( __clang_minor__ >= 5 ) )
 1249+    #define STBIR_NO_UNROLL_LOOP_START _Pragma("clang loop unroll(disable)") _Pragma("clang loop vectorize(disable)")
 1250+  #else
 1251+    #define STBIR_NO_UNROLL_LOOP_START
 1252+  #endif 
 1253+#elif defined( __GNUC__ )
 1254+  #define STBIR_STREAMOUT_PTR( star ) star __restrict__
 1255+  #define STBIR_NO_UNROLL( ptr ) __asm__ (""::"r"(ptr))
 1256+  #if __GNUC__ >= 14
 1257+    #define STBIR_NO_UNROLL_LOOP_START _Pragma("GCC unroll 0") _Pragma("GCC novector")
 1258+  #else
 1259+    #define STBIR_NO_UNROLL_LOOP_START
 1260+  #endif
 1261+  #define STBIR_NO_UNROLL_LOOP_START_INF_FOR
 1262+#else
 1263+  #define STBIR_STREAMOUT_PTR( star ) star
 1264+  #define STBIR_NO_UNROLL( ptr )
 1265+  #define STBIR_NO_UNROLL_LOOP_START
 1266+#endif
 1267+
 1268+#ifndef STBIR_NO_UNROLL_LOOP_START_INF_FOR
 1269+#define STBIR_NO_UNROLL_LOOP_START_INF_FOR STBIR_NO_UNROLL_LOOP_START
 1270+#endif
 1271+
 1272+#ifdef STBIR_NO_SIMD // force simd off for whatever reason
 1273+
 1274+// force simd off overrides everything else, so clear it all
 1275+
 1276+#ifdef STBIR_SSE2
 1277+#undef STBIR_SSE2
 1278+#endif
 1279+
 1280+#ifdef STBIR_AVX
 1281+#undef STBIR_AVX
 1282+#endif
 1283+
 1284+#ifdef STBIR_NEON
 1285+#undef STBIR_NEON
 1286+#endif
 1287+
 1288+#ifdef STBIR_AVX2
 1289+#undef STBIR_AVX2
 1290+#endif
 1291+
 1292+#ifdef STBIR_FP16C
 1293+#undef STBIR_FP16C
 1294+#endif
 1295+
 1296+#ifdef STBIR_WASM
 1297+#undef STBIR_WASM
 1298+#endif
 1299+
 1300+#ifdef STBIR_SIMD
 1301+#undef STBIR_SIMD
 1302+#endif
 1303+
 1304+#else // STBIR_SIMD
 1305+
 1306+#ifdef STBIR_SSE2
 1307+  #include <emmintrin.h>
 1308+
 1309+  #define stbir__simdf __m128
 1310+  #define stbir__simdi __m128i
 1311+
 1312+  #define stbir_simdi_castf( reg ) _mm_castps_si128(reg)
 1313+  #define stbir_simdf_casti( reg ) _mm_castsi128_ps(reg)
 1314+
 1315+  #define stbir__simdf_load( reg, ptr ) (reg) = _mm_loadu_ps( (float const*)(ptr) )
 1316+  #define stbir__simdi_load( reg, ptr ) (reg) = _mm_loadu_si128 ( (stbir__simdi const*)(ptr) )
 1317+  #define stbir__simdf_load1( out, ptr ) (out) = _mm_load_ss( (float const*)(ptr) )  // top values can be random (not denormal or nan for perf)
 1318+  #define stbir__simdi_load1( out, ptr ) (out) = _mm_castps_si128( _mm_load_ss( (float const*)(ptr) ))
 1319+  #define stbir__simdf_load1z( out, ptr ) (out) = _mm_load_ss( (float const*)(ptr) )  // top values must be zero
 1320+  #define stbir__simdf_frep4( fvar ) _mm_set_ps1( fvar )
 1321+  #define stbir__simdf_load1frep4( out, fvar ) (out) = _mm_set_ps1( fvar )
 1322+  #define stbir__simdf_load2( out, ptr ) (out) = _mm_castsi128_ps( _mm_loadl_epi64( (__m128i*)(ptr)) ) // top values can be random (not denormal or nan for perf)
 1323+  #define stbir__simdf_load2z( out, ptr ) (out) = _mm_castsi128_ps( _mm_loadl_epi64( (__m128i*)(ptr)) ) // top values must be zero
 1324+  #define stbir__simdf_load2hmerge( out, reg, ptr ) (out) = _mm_castpd_ps(_mm_loadh_pd( _mm_castps_pd(reg), (double*)(ptr) ))
 1325+
 1326+  #define stbir__simdf_zeroP() _mm_setzero_ps()
 1327+  #define stbir__simdf_zero( reg ) (reg) = _mm_setzero_ps()
 1328+
 1329+  #define stbir__simdf_store( ptr, reg )  _mm_storeu_ps( (float*)(ptr), reg )
 1330+  #define stbir__simdf_store1( ptr, reg ) _mm_store_ss( (float*)(ptr), reg )
 1331+  #define stbir__simdf_store2( ptr, reg ) _mm_storel_epi64( (__m128i*)(ptr), _mm_castps_si128(reg) )
 1332+  #define stbir__simdf_store2h( ptr, reg ) _mm_storeh_pd( (double*)(ptr), _mm_castps_pd(reg) )
 1333+
 1334+  #define stbir__simdi_store( ptr, reg )  _mm_storeu_si128( (__m128i*)(ptr), reg )
 1335+  #define stbir__simdi_store1( ptr, reg ) _mm_store_ss( (float*)(ptr), _mm_castsi128_ps(reg) )
 1336+  #define stbir__simdi_store2( ptr, reg ) _mm_storel_epi64( (__m128i*)(ptr), (reg) )
 1337+
 1338+  #define stbir__prefetch( ptr ) _mm_prefetch((char*)(ptr), _MM_HINT_T0 )
 1339+
 1340+  #define stbir__simdi_expand_u8_to_u32(out0,out1,out2,out3,ireg) \
 1341+  { \
 1342+    stbir__simdi zero = _mm_setzero_si128(); \
 1343+    out2 = _mm_unpacklo_epi8( ireg, zero ); \
 1344+    out3 = _mm_unpackhi_epi8( ireg, zero ); \
 1345+    out0 = _mm_unpacklo_epi16( out2, zero ); \
 1346+    out1 = _mm_unpackhi_epi16( out2, zero ); \
 1347+    out2 = _mm_unpacklo_epi16( out3, zero ); \
 1348+    out3 = _mm_unpackhi_epi16( out3, zero ); \
 1349+  }
 1350+
 1351+#define stbir__simdi_expand_u8_to_1u32(out,ireg) \
 1352+  { \
 1353+    stbir__simdi zero = _mm_setzero_si128(); \
 1354+    out = _mm_unpacklo_epi8( ireg, zero ); \
 1355+    out = _mm_unpacklo_epi16( out, zero ); \
 1356+  }
 1357+
 1358+  #define stbir__simdi_expand_u16_to_u32(out0,out1,ireg) \
 1359+  { \
 1360+    stbir__simdi zero = _mm_setzero_si128(); \
 1361+    out0 = _mm_unpacklo_epi16( ireg, zero ); \
 1362+    out1 = _mm_unpackhi_epi16( ireg, zero ); \
 1363+  }
 1364+
 1365+  #define stbir__simdf_convert_float_to_i32( i, f ) (i) = _mm_cvttps_epi32(f)
 1366+  #define stbir__simdf_convert_float_to_int( f ) _mm_cvtt_ss2si(f)
 1367+  #define stbir__simdf_convert_float_to_uint8( f ) ((unsigned char)_mm_cvtsi128_si32(_mm_cvttps_epi32(_mm_max_ps(_mm_min_ps(f,STBIR__CONSTF(STBIR_max_uint8_as_float)),_mm_setzero_ps()))))
 1368+  #define stbir__simdf_convert_float_to_short( f ) ((unsigned short)_mm_cvtsi128_si32(_mm_cvttps_epi32(_mm_max_ps(_mm_min_ps(f,STBIR__CONSTF(STBIR_max_uint16_as_float)),_mm_setzero_ps()))))
 1369+
 1370+  #define stbir__simdi_to_int( i ) _mm_cvtsi128_si32(i)
 1371+  #define stbir__simdi_convert_i32_to_float(out, ireg) (out) = _mm_cvtepi32_ps( ireg )
 1372+  #define stbir__simdf_add( out, reg0, reg1 ) (out) = _mm_add_ps( reg0, reg1 )
 1373+  #define stbir__simdf_mult( out, reg0, reg1 ) (out) = _mm_mul_ps( reg0, reg1 )
 1374+  #define stbir__simdf_mult_mem( out, reg, ptr ) (out) = _mm_mul_ps( reg, _mm_loadu_ps( (float const*)(ptr) ) )
 1375+  #define stbir__simdf_mult1_mem( out, reg, ptr ) (out) = _mm_mul_ss( reg, _mm_load_ss( (float const*)(ptr) ) )
 1376+  #define stbir__simdf_add_mem( out, reg, ptr ) (out) = _mm_add_ps( reg, _mm_loadu_ps( (float const*)(ptr) ) )
 1377+  #define stbir__simdf_add1_mem( out, reg, ptr ) (out) = _mm_add_ss( reg, _mm_load_ss( (float const*)(ptr) ) )
 1378+
 1379+  #ifdef STBIR_USE_FMA           // not on by default to maintain bit identical simd to non-simd
 1380+  #include <immintrin.h>
 1381+  #define stbir__simdf_madd( out, add, mul1, mul2 ) (out) = _mm_fmadd_ps( mul1, mul2, add )
 1382+  #define stbir__simdf_madd1( out, add, mul1, mul2 ) (out) = _mm_fmadd_ss( mul1, mul2, add )
 1383+  #define stbir__simdf_madd_mem( out, add, mul, ptr ) (out) = _mm_fmadd_ps( mul, _mm_loadu_ps( (float const*)(ptr) ), add )
 1384+  #define stbir__simdf_madd1_mem( out, add, mul, ptr ) (out) = _mm_fmadd_ss( mul, _mm_load_ss( (float const*)(ptr) ), add )
 1385+  #else
 1386+  #define stbir__simdf_madd( out, add, mul1, mul2 ) (out) = _mm_add_ps( add, _mm_mul_ps( mul1, mul2 ) )
 1387+  #define stbir__simdf_madd1( out, add, mul1, mul2 ) (out) = _mm_add_ss( add, _mm_mul_ss( mul1, mul2 ) )
 1388+  #define stbir__simdf_madd_mem( out, add, mul, ptr ) (out) = _mm_add_ps( add, _mm_mul_ps( mul, _mm_loadu_ps( (float const*)(ptr) ) ) )
 1389+  #define stbir__simdf_madd1_mem( out, add, mul, ptr ) (out) = _mm_add_ss( add, _mm_mul_ss( mul, _mm_load_ss( (float const*)(ptr) ) ) )
 1390+  #endif
 1391+
 1392+  #define stbir__simdf_add1( out, reg0, reg1 ) (out) = _mm_add_ss( reg0, reg1 )
 1393+  #define stbir__simdf_mult1( out, reg0, reg1 ) (out) = _mm_mul_ss( reg0, reg1 )
 1394+
 1395+  #define stbir__simdf_and( out, reg0, reg1 ) (out) = _mm_and_ps( reg0, reg1 )
 1396+  #define stbir__simdf_or( out, reg0, reg1 ) (out) = _mm_or_ps( reg0, reg1 )
 1397+
 1398+  #define stbir__simdf_min( out, reg0, reg1 ) (out) = _mm_min_ps( reg0, reg1 )
 1399+  #define stbir__simdf_max( out, reg0, reg1 ) (out) = _mm_max_ps( reg0, reg1 )
 1400+  #define stbir__simdf_min1( out, reg0, reg1 ) (out) = _mm_min_ss( reg0, reg1 )
 1401+  #define stbir__simdf_max1( out, reg0, reg1 ) (out) = _mm_max_ss( reg0, reg1 )
 1402+
 1403+  #define stbir__simdf_0123ABCDto3ABx( out, reg0, reg1 ) (out)=_mm_castsi128_ps( _mm_shuffle_epi32( _mm_castps_si128( _mm_shuffle_ps( reg1,reg0, (0<<0) + (1<<2) + (2<<4) + (3<<6) )), (3<<0) + (0<<2) + (1<<4) + (2<<6) ) )
 1404+  #define stbir__simdf_0123ABCDto23Ax( out, reg0, reg1 ) (out)=_mm_castsi128_ps( _mm_shuffle_epi32( _mm_castps_si128( _mm_shuffle_ps( reg1,reg0, (0<<0) + (1<<2) + (2<<4) + (3<<6) )), (2<<0) + (3<<2) + (0<<4) + (1<<6) ) )
 1405+
 1406+  static const stbir__simdf STBIR_zeroones = { 0.0f,1.0f,0.0f,1.0f };
 1407+  static const stbir__simdf STBIR_onezeros = { 1.0f,0.0f,1.0f,0.0f };
 1408+  #define stbir__simdf_aaa1( out, alp, ones ) (out)=_mm_castsi128_ps( _mm_shuffle_epi32( _mm_castps_si128( _mm_movehl_ps( ones, alp ) ), (1<<0) + (1<<2) + (1<<4) + (2<<6) ) )
 1409+  #define stbir__simdf_1aaa( out, alp, ones ) (out)=_mm_castsi128_ps( _mm_shuffle_epi32( _mm_castps_si128( _mm_movelh_ps( ones, alp ) ), (0<<0) + (2<<2) + (2<<4) + (2<<6) ) )
 1410+  #define stbir__simdf_a1a1( out, alp, ones) (out) = _mm_or_ps( _mm_castsi128_ps( _mm_srli_epi64( _mm_castps_si128(alp), 32 ) ), STBIR_zeroones )
 1411+  #define stbir__simdf_1a1a( out, alp, ones) (out) = _mm_or_ps( _mm_castsi128_ps( _mm_slli_epi64( _mm_castps_si128(alp), 32 ) ), STBIR_onezeros )
 1412+
 1413+  #define stbir__simdf_swiz( reg, one, two, three, four ) _mm_castsi128_ps( _mm_shuffle_epi32( _mm_castps_si128( reg ), (one<<0) + (two<<2) + (three<<4) + (four<<6) ) )
 1414+
 1415+  #define stbir__simdi_and( out, reg0, reg1 ) (out) = _mm_and_si128( reg0, reg1 )
 1416+  #define stbir__simdi_or( out, reg0, reg1 ) (out) = _mm_or_si128( reg0, reg1 )
 1417+  #define stbir__simdi_16madd( out, reg0, reg1 ) (out) = _mm_madd_epi16( reg0, reg1 )
 1418+
 1419+  #define stbir__simdf_pack_to_8bytes(out,aa,bb) \
 1420+  { \
 1421+    stbir__simdf af,bf; \
 1422+    stbir__simdi a,b; \
 1423+    af = _mm_min_ps( aa, STBIR_max_uint8_as_float ); \
 1424+    bf = _mm_min_ps( bb, STBIR_max_uint8_as_float ); \
 1425+    af = _mm_max_ps( af, _mm_setzero_ps() ); \
 1426+    bf = _mm_max_ps( bf, _mm_setzero_ps() ); \
 1427+    a = _mm_cvttps_epi32( af ); \
 1428+    b = _mm_cvttps_epi32( bf ); \
 1429+    a = _mm_packs_epi32( a, b ); \
 1430+    out = _mm_packus_epi16( a, a ); \
 1431+  }
 1432+
 1433+  #define stbir__simdf_load4_transposed( o0, o1, o2, o3, ptr ) \
 1434+      stbir__simdf_load( o0, (ptr) );    \
 1435+      stbir__simdf_load( o1, (ptr)+4 );  \
 1436+      stbir__simdf_load( o2, (ptr)+8 );  \
 1437+      stbir__simdf_load( o3, (ptr)+12 ); \
 1438+      {                                  \
 1439+        __m128 tmp0, tmp1, tmp2, tmp3;   \
 1440+        tmp0 = _mm_unpacklo_ps(o0, o1);  \
 1441+        tmp2 = _mm_unpacklo_ps(o2, o3);  \
 1442+        tmp1 = _mm_unpackhi_ps(o0, o1);  \
 1443+        tmp3 = _mm_unpackhi_ps(o2, o3);  \
 1444+        o0 = _mm_movelh_ps(tmp0, tmp2);  \
 1445+        o1 = _mm_movehl_ps(tmp2, tmp0);  \
 1446+        o2 = _mm_movelh_ps(tmp1, tmp3);  \
 1447+        o3 = _mm_movehl_ps(tmp3, tmp1);  \
 1448+      }
 1449+
 1450+  #define stbir__interleave_pack_and_store_16_u8( ptr, r0, r1, r2, r3 ) \
 1451+      r0 = _mm_packs_epi32( r0, r1 ); \
 1452+      r2 = _mm_packs_epi32( r2, r3 ); \
 1453+      r1 = _mm_unpacklo_epi16( r0, r2 ); \
 1454+      r3 = _mm_unpackhi_epi16( r0, r2 ); \
 1455+      r0 = _mm_unpacklo_epi16( r1, r3 ); \
 1456+      r2 = _mm_unpackhi_epi16( r1, r3 ); \
 1457+      r0 = _mm_packus_epi16( r0, r2 ); \
 1458+      stbir__simdi_store( ptr, r0 ); \
 1459+
 1460+  #define stbir__simdi_32shr( out, reg, imm ) out = _mm_srli_epi32( reg, imm )
 1461+
 1462+  #if defined(_MSC_VER) && !defined(__clang__)
 1463+    // msvc inits with 8 bytes
 1464+    #define STBIR__CONST_32_TO_8( v ) (char)(unsigned char)((v)&255),(char)(unsigned char)(((v)>>8)&255),(char)(unsigned char)(((v)>>16)&255),(char)(unsigned char)(((v)>>24)&255)
 1465+    #define STBIR__CONST_4_32i( v ) STBIR__CONST_32_TO_8( v ), STBIR__CONST_32_TO_8( v ), STBIR__CONST_32_TO_8( v ), STBIR__CONST_32_TO_8( v )
 1466+    #define STBIR__CONST_4d_32i( v0, v1, v2, v3 ) STBIR__CONST_32_TO_8( v0 ), STBIR__CONST_32_TO_8( v1 ), STBIR__CONST_32_TO_8( v2 ), STBIR__CONST_32_TO_8( v3 )
 1467+  #else
 1468+    // everything else inits with long long's
 1469+    #define STBIR__CONST_4_32i( v ) (long long)((((stbir_uint64)(stbir_uint32)(v))<<32)|((stbir_uint64)(stbir_uint32)(v))),(long long)((((stbir_uint64)(stbir_uint32)(v))<<32)|((stbir_uint64)(stbir_uint32)(v)))
 1470+    #define STBIR__CONST_4d_32i( v0, v1, v2, v3 ) (long long)((((stbir_uint64)(stbir_uint32)(v1))<<32)|((stbir_uint64)(stbir_uint32)(v0))),(long long)((((stbir_uint64)(stbir_uint32)(v3))<<32)|((stbir_uint64)(stbir_uint32)(v2)))
 1471+  #endif
 1472+
 1473+  #define STBIR__SIMDF_CONST(var, x) stbir__simdf var = { x, x, x, x }
 1474+  #define STBIR__SIMDI_CONST(var, x) stbir__simdi var = { STBIR__CONST_4_32i(x) }
 1475+  #define STBIR__CONSTF(var) (var)
 1476+  #define STBIR__CONSTI(var) (var)
 1477+
 1478+  #if defined(STBIR_AVX) || defined(__SSE4_1__)
 1479+    #include <smmintrin.h>
 1480+    #define stbir__simdf_pack_to_8words(out,reg0,reg1) out = _mm_packus_epi32(_mm_cvttps_epi32(_mm_max_ps(_mm_min_ps(reg0,STBIR__CONSTF(STBIR_max_uint16_as_float)),_mm_setzero_ps())), _mm_cvttps_epi32(_mm_max_ps(_mm_min_ps(reg1,STBIR__CONSTF(STBIR_max_uint16_as_float)),_mm_setzero_ps())))
 1481+  #else
 1482+    static STBIR__SIMDI_CONST(stbir__s32_32768, 32768);
 1483+    static STBIR__SIMDI_CONST(stbir__s16_32768, ((32768<<16)|32768));
 1484+
 1485+    #define stbir__simdf_pack_to_8words(out,reg0,reg1) \
 1486+      { \
 1487+        stbir__simdi tmp0,tmp1; \
 1488+        tmp0 = _mm_cvttps_epi32(_mm_max_ps(_mm_min_ps(reg0,STBIR__CONSTF(STBIR_max_uint16_as_float)),_mm_setzero_ps())); \
 1489+        tmp1 = _mm_cvttps_epi32(_mm_max_ps(_mm_min_ps(reg1,STBIR__CONSTF(STBIR_max_uint16_as_float)),_mm_setzero_ps())); \
 1490+        tmp0 = _mm_sub_epi32( tmp0, stbir__s32_32768 ); \
 1491+        tmp1 = _mm_sub_epi32( tmp1, stbir__s32_32768 ); \
 1492+        out = _mm_packs_epi32( tmp0, tmp1 ); \
 1493+        out = _mm_sub_epi16( out, stbir__s16_32768 ); \
 1494+      }
 1495+
 1496+  #endif
 1497+
 1498+  #define STBIR_SIMD
 1499+
 1500+  // if we detect AVX, set the simd8 defines
 1501+  #ifdef STBIR_AVX
 1502+    #include <immintrin.h>
 1503+    #define STBIR_SIMD8
 1504+    #define stbir__simdf8 __m256
 1505+    #define stbir__simdi8 __m256i
 1506+    #define stbir__simdf8_load( out, ptr ) (out) = _mm256_loadu_ps( (float const *)(ptr) )
 1507+    #define stbir__simdi8_load( out, ptr ) (out) = _mm256_loadu_si256( (__m256i const *)(ptr) )
 1508+    #define stbir__simdf8_mult( out, a, b ) (out) = _mm256_mul_ps( (a), (b) )
 1509+    #define stbir__simdf8_store( ptr, out ) _mm256_storeu_ps( (float*)(ptr), out )
 1510+    #define stbir__simdi8_store( ptr, reg )  _mm256_storeu_si256( (__m256i*)(ptr), reg )
 1511+    #define stbir__simdf8_frep8( fval ) _mm256_set1_ps( fval )
 1512+
 1513+    #define stbir__simdf8_min( out, reg0, reg1 ) (out) = _mm256_min_ps( reg0, reg1 )
 1514+    #define stbir__simdf8_max( out, reg0, reg1 ) (out) = _mm256_max_ps( reg0, reg1 )
 1515+
 1516+    #define stbir__simdf8_add4halves( out, bot4, top8 ) (out) = _mm_add_ps( bot4, _mm256_extractf128_ps( top8, 1 ) )
 1517+    #define stbir__simdf8_mult_mem( out, reg, ptr ) (out) = _mm256_mul_ps( reg, _mm256_loadu_ps( (float const*)(ptr) ) )
 1518+    #define stbir__simdf8_add_mem( out, reg, ptr ) (out) = _mm256_add_ps( reg, _mm256_loadu_ps( (float const*)(ptr) ) )
 1519+    #define stbir__simdf8_add( out, a, b ) (out) = _mm256_add_ps( a, b )
 1520+    #define stbir__simdf8_load1b( out, ptr ) (out) = _mm256_broadcast_ss( ptr )
 1521+    #define stbir__simdf_load1rep4( out, ptr ) (out) = _mm_broadcast_ss( ptr )  // avx load instruction
 1522+
 1523+    #define stbir__simdi8_convert_i32_to_float(out, ireg) (out) = _mm256_cvtepi32_ps( ireg )
 1524+    #define stbir__simdf8_convert_float_to_i32( i, f ) (i) = _mm256_cvttps_epi32(f)
 1525+
 1526+    #define stbir__simdf8_bot4s( out, a, b ) (out) = _mm256_permute2f128_ps(a,b, (0<<0)+(2<<4) )
 1527+    #define stbir__simdf8_top4s( out, a, b ) (out) = _mm256_permute2f128_ps(a,b, (1<<0)+(3<<4) )
 1528+
 1529+    #define stbir__simdf8_gettop4( reg ) _mm256_extractf128_ps(reg,1)
 1530+
 1531+    #ifdef STBIR_AVX2
 1532+
 1533+    #define stbir__simdi8_expand_u8_to_u32(out0,out1,ireg) \
 1534+    { \
 1535+      stbir__simdi8 a, zero  =_mm256_setzero_si256();\
 1536+      a = _mm256_permute4x64_epi64( _mm256_unpacklo_epi8( _mm256_permute4x64_epi64(_mm256_castsi128_si256(ireg),(0<<0)+(2<<2)+(1<<4)+(3<<6)), zero ),(0<<0)+(2<<2)+(1<<4)+(3<<6)); \
 1537+      out0 = _mm256_unpacklo_epi16( a, zero ); \
 1538+      out1 = _mm256_unpackhi_epi16( a, zero ); \
 1539+    }
 1540+
 1541+    #define stbir__simdf8_pack_to_16bytes(out,aa,bb) \
 1542+    { \
 1543+      stbir__simdi8 t; \
 1544+      stbir__simdf8 af,bf; \
 1545+      stbir__simdi8 a,b; \
 1546+      af = _mm256_min_ps( aa, STBIR_max_uint8_as_floatX ); \
 1547+      bf = _mm256_min_ps( bb, STBIR_max_uint8_as_floatX ); \
 1548+      af = _mm256_max_ps( af, _mm256_setzero_ps() ); \
 1549+      bf = _mm256_max_ps( bf, _mm256_setzero_ps() ); \
 1550+      a = _mm256_cvttps_epi32( af ); \
 1551+      b = _mm256_cvttps_epi32( bf ); \
 1552+      t = _mm256_permute4x64_epi64( _mm256_packs_epi32( a, b ), (0<<0)+(2<<2)+(1<<4)+(3<<6) ); \
 1553+      out = _mm256_castsi256_si128( _mm256_permute4x64_epi64( _mm256_packus_epi16( t, t ), (0<<0)+(2<<2)+(1<<4)+(3<<6) ) ); \
 1554+    }
 1555+
 1556+    #define stbir__simdi8_expand_u16_to_u32(out,ireg) out = _mm256_unpacklo_epi16( _mm256_permute4x64_epi64(_mm256_castsi128_si256(ireg),(0<<0)+(2<<2)+(1<<4)+(3<<6)), _mm256_setzero_si256() );
 1557+
 1558+    #define stbir__simdf8_pack_to_16words(out,aa,bb) \
 1559+      { \
 1560+        stbir__simdf8 af,bf; \
 1561+        stbir__simdi8 a,b; \
 1562+        af = _mm256_min_ps( aa, STBIR_max_uint16_as_floatX ); \
 1563+        bf = _mm256_min_ps( bb, STBIR_max_uint16_as_floatX ); \
 1564+        af = _mm256_max_ps( af, _mm256_setzero_ps() ); \
 1565+        bf = _mm256_max_ps( bf, _mm256_setzero_ps() ); \
 1566+        a = _mm256_cvttps_epi32( af ); \
 1567+        b = _mm256_cvttps_epi32( bf ); \
 1568+        (out) = _mm256_permute4x64_epi64( _mm256_packus_epi32(a, b), (0<<0)+(2<<2)+(1<<4)+(3<<6) ); \
 1569+      }
 1570+
 1571+    #else
 1572+
 1573+    #define stbir__simdi8_expand_u8_to_u32(out0,out1,ireg) \
 1574+    { \
 1575+      stbir__simdi a,zero = _mm_setzero_si128(); \
 1576+      a = _mm_unpacklo_epi8( ireg, zero ); \
 1577+      out0 = _mm256_setr_m128i( _mm_unpacklo_epi16( a, zero ), _mm_unpackhi_epi16( a, zero ) ); \
 1578+      a = _mm_unpackhi_epi8( ireg, zero ); \
 1579+      out1 = _mm256_setr_m128i( _mm_unpacklo_epi16( a, zero ), _mm_unpackhi_epi16( a, zero ) ); \
 1580+    }
 1581+
 1582+    #define stbir__simdf8_pack_to_16bytes(out,aa,bb) \
 1583+    { \
 1584+      stbir__simdi t; \
 1585+      stbir__simdf8 af,bf; \
 1586+      stbir__simdi8 a,b; \
 1587+      af = _mm256_min_ps( aa, STBIR_max_uint8_as_floatX ); \
 1588+      bf = _mm256_min_ps( bb, STBIR_max_uint8_as_floatX ); \
 1589+      af = _mm256_max_ps( af, _mm256_setzero_ps() ); \
 1590+      bf = _mm256_max_ps( bf, _mm256_setzero_ps() ); \
 1591+      a = _mm256_cvttps_epi32( af ); \
 1592+      b = _mm256_cvttps_epi32( bf ); \
 1593+      out = _mm_packs_epi32( _mm256_castsi256_si128(a), _mm256_extractf128_si256( a, 1 ) ); \
 1594+      out = _mm_packus_epi16( out, out ); \
 1595+      t = _mm_packs_epi32( _mm256_castsi256_si128(b), _mm256_extractf128_si256( b, 1 ) ); \
 1596+      t = _mm_packus_epi16( t, t ); \
 1597+      out = _mm_castps_si128( _mm_shuffle_ps( _mm_castsi128_ps(out), _mm_castsi128_ps(t), (0<<0)+(1<<2)+(0<<4)+(1<<6) ) ); \
 1598+    }
 1599+
 1600+    #define stbir__simdi8_expand_u16_to_u32(out,ireg) \
 1601+    { \
 1602+      stbir__simdi a,b,zero = _mm_setzero_si128(); \
 1603+      a = _mm_unpacklo_epi16( ireg, zero ); \
 1604+      b = _mm_unpackhi_epi16( ireg, zero ); \
 1605+      out = _mm256_insertf128_si256( _mm256_castsi128_si256( a ), b, 1 ); \
 1606+    }
 1607+
 1608+    #define stbir__simdf8_pack_to_16words(out,aa,bb) \
 1609+      { \
 1610+        stbir__simdi t0,t1; \
 1611+        stbir__simdf8 af,bf; \
 1612+        stbir__simdi8 a,b; \
 1613+        af = _mm256_min_ps( aa, STBIR_max_uint16_as_floatX ); \
 1614+        bf = _mm256_min_ps( bb, STBIR_max_uint16_as_floatX ); \
 1615+        af = _mm256_max_ps( af, _mm256_setzero_ps() ); \
 1616+        bf = _mm256_max_ps( bf, _mm256_setzero_ps() ); \
 1617+        a = _mm256_cvttps_epi32( af ); \
 1618+        b = _mm256_cvttps_epi32( bf ); \
 1619+        t0 = _mm_packus_epi32( _mm256_castsi256_si128(a), _mm256_extractf128_si256( a, 1 ) ); \
 1620+        t1 = _mm_packus_epi32( _mm256_castsi256_si128(b), _mm256_extractf128_si256( b, 1 ) ); \
 1621+        out = _mm256_setr_m128i( t0, t1 ); \
 1622+      }
 1623+
 1624+    #endif
 1625+
 1626+    static __m256i stbir_00001111 = { STBIR__CONST_4d_32i( 0, 0, 0, 0 ), STBIR__CONST_4d_32i( 1, 1, 1, 1 ) };
 1627+    #define stbir__simdf8_0123to00001111( out, in ) (out) = _mm256_permutevar_ps ( in, stbir_00001111 )
 1628+
 1629+    static __m256i stbir_22223333 = { STBIR__CONST_4d_32i( 2, 2, 2, 2 ), STBIR__CONST_4d_32i( 3, 3, 3, 3 ) };
 1630+    #define stbir__simdf8_0123to22223333( out, in ) (out) = _mm256_permutevar_ps ( in, stbir_22223333 )
 1631+
 1632+    #define stbir__simdf8_0123to2222( out, in ) (out) = stbir__simdf_swiz(_mm256_castps256_ps128(in), 2,2,2,2 )
 1633+
 1634+    #define stbir__simdf8_load4b( out, ptr ) (out) = _mm256_broadcast_ps( (__m128 const *)(ptr) )
 1635+
 1636+    static __m256i stbir_00112233 = { STBIR__CONST_4d_32i( 0, 0, 1, 1 ), STBIR__CONST_4d_32i( 2, 2, 3, 3 ) };
 1637+    #define stbir__simdf8_0123to00112233( out, in ) (out) = _mm256_permutevar_ps ( in, stbir_00112233 )
 1638+    #define stbir__simdf8_add4( out, a8, b ) (out) = _mm256_add_ps( a8,  _mm256_castps128_ps256( b ) )
 1639+
 1640+    static __m256i stbir_load6 = { STBIR__CONST_4_32i( 0x80000000 ), STBIR__CONST_4d_32i(  0x80000000,  0x80000000, 0, 0 ) };
 1641+    #define stbir__simdf8_load6z( out, ptr ) (out) = _mm256_maskload_ps( ptr, stbir_load6 )
 1642+
 1643+    #define stbir__simdf8_0123to00000000( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (0<<0)+(0<<2)+(0<<4)+(0<<6) )
 1644+    #define stbir__simdf8_0123to11111111( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (1<<0)+(1<<2)+(1<<4)+(1<<6) )
 1645+    #define stbir__simdf8_0123to22222222( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (2<<0)+(2<<2)+(2<<4)+(2<<6) )
 1646+    #define stbir__simdf8_0123to33333333( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (3<<0)+(3<<2)+(3<<4)+(3<<6) )
 1647+    #define stbir__simdf8_0123to21032103( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (2<<0)+(1<<2)+(0<<4)+(3<<6) )
 1648+    #define stbir__simdf8_0123to32103210( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (3<<0)+(2<<2)+(1<<4)+(0<<6) )
 1649+    #define stbir__simdf8_0123to12301230( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (1<<0)+(2<<2)+(3<<4)+(0<<6) )
 1650+    #define stbir__simdf8_0123to10321032( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (1<<0)+(0<<2)+(3<<4)+(2<<6) )
 1651+    #define stbir__simdf8_0123to30123012( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (3<<0)+(0<<2)+(1<<4)+(2<<6) )
 1652+
 1653+    #define stbir__simdf8_0123to11331133( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (1<<0)+(1<<2)+(3<<4)+(3<<6) )
 1654+    #define stbir__simdf8_0123to00220022( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (0<<0)+(0<<2)+(2<<4)+(2<<6) )
 1655+
 1656+    #define stbir__simdf8_aaa1( out, alp, ones ) (out) = _mm256_blend_ps( alp, ones, (1<<0)+(1<<1)+(1<<2)+(0<<3)+(1<<4)+(1<<5)+(1<<6)+(0<<7)); (out)=_mm256_shuffle_ps( out,out, (3<<0) + (3<<2) + (3<<4) + (0<<6) )
 1657+    #define stbir__simdf8_1aaa( out, alp, ones ) (out) = _mm256_blend_ps( alp, ones, (0<<0)+(1<<1)+(1<<2)+(1<<3)+(0<<4)+(1<<5)+(1<<6)+(1<<7)); (out)=_mm256_shuffle_ps( out,out, (1<<0) + (0<<2) + (0<<4) + (0<<6) )
 1658+    #define stbir__simdf8_a1a1( out, alp, ones) (out) = _mm256_blend_ps( alp, ones, (1<<0)+(0<<1)+(1<<2)+(0<<3)+(1<<4)+(0<<5)+(1<<6)+(0<<7)); (out)=_mm256_shuffle_ps( out,out, (1<<0) + (0<<2) + (3<<4) + (2<<6) )
 1659+    #define stbir__simdf8_1a1a( out, alp, ones) (out) = _mm256_blend_ps( alp, ones, (0<<0)+(1<<1)+(0<<2)+(1<<3)+(0<<4)+(1<<5)+(0<<6)+(1<<7)); (out)=_mm256_shuffle_ps( out,out, (1<<0) + (0<<2) + (3<<4) + (2<<6) )
 1660+
 1661+    #define stbir__simdf8_zero( reg ) (reg) = _mm256_setzero_ps()
 1662+
 1663+    #ifdef STBIR_USE_FMA           // not on by default to maintain bit identical simd to non-simd
 1664+    #define stbir__simdf8_madd( out, add, mul1, mul2 ) (out) = _mm256_fmadd_ps( mul1, mul2, add )
 1665+    #define stbir__simdf8_madd_mem( out, add, mul, ptr ) (out) = _mm256_fmadd_ps( mul, _mm256_loadu_ps( (float const*)(ptr) ), add )
 1666+    #define stbir__simdf8_madd_mem4( out, add, mul, ptr )(out) = _mm256_fmadd_ps( _mm256_setr_m128( mul, _mm_setzero_ps() ), _mm256_setr_m128( _mm_loadu_ps( (float const*)(ptr) ), _mm_setzero_ps() ), add )
 1667+    #else
 1668+    #define stbir__simdf8_madd( out, add, mul1, mul2 ) (out) = _mm256_add_ps( add, _mm256_mul_ps( mul1, mul2 ) )
 1669+    #define stbir__simdf8_madd_mem( out, add, mul, ptr ) (out) = _mm256_add_ps( add, _mm256_mul_ps( mul, _mm256_loadu_ps( (float const*)(ptr) ) ) )
 1670+    #define stbir__simdf8_madd_mem4( out, add, mul, ptr )  (out) = _mm256_add_ps( add, _mm256_setr_m128( _mm_mul_ps( mul, _mm_loadu_ps( (float const*)(ptr) ) ), _mm_setzero_ps() ) )
 1671+    #endif
 1672+    #define stbir__if_simdf8_cast_to_simdf4( val ) _mm256_castps256_ps128( val )
 1673+
 1674+  #endif
 1675+
 1676+  #ifdef STBIR_FLOORF
 1677+  #undef STBIR_FLOORF
 1678+  #endif
 1679+  #define STBIR_FLOORF stbir_simd_floorf
 1680+  static stbir__inline float stbir_simd_floorf(float x)  // martins floorf
 1681+  {
 1682+    #if defined(STBIR_AVX) || defined(__SSE4_1__) || defined(STBIR_SSE41)
 1683+    __m128 t = _mm_set_ss(x);
 1684+    return _mm_cvtss_f32( _mm_floor_ss(t, t) );
 1685+    #else
 1686+    __m128 f = _mm_set_ss(x);
 1687+    __m128 t = _mm_cvtepi32_ps(_mm_cvttps_epi32(f));
 1688+    __m128 r = _mm_add_ss(t, _mm_and_ps(_mm_cmplt_ss(f, t), _mm_set_ss(-1.0f)));
 1689+    return _mm_cvtss_f32(r);
 1690+    #endif
 1691+  }
 1692+
 1693+  #ifdef STBIR_CEILF
 1694+  #undef STBIR_CEILF
 1695+  #endif
 1696+  #define STBIR_CEILF stbir_simd_ceilf
 1697+  static stbir__inline float stbir_simd_ceilf(float x)  // martins ceilf
 1698+  {
 1699+    #if defined(STBIR_AVX) || defined(__SSE4_1__) || defined(STBIR_SSE41)
 1700+    __m128 t = _mm_set_ss(x);
 1701+    return _mm_cvtss_f32( _mm_ceil_ss(t, t) );
 1702+    #else
 1703+    __m128 f = _mm_set_ss(x);
 1704+    __m128 t = _mm_cvtepi32_ps(_mm_cvttps_epi32(f));
 1705+    __m128 r = _mm_add_ss(t, _mm_and_ps(_mm_cmplt_ss(t, f), _mm_set_ss(1.0f)));
 1706+    return _mm_cvtss_f32(r);
 1707+    #endif
 1708+  }
 1709+
 1710+#elif defined(STBIR_NEON)
 1711+
 1712+  #include <arm_neon.h>
 1713+
 1714+  #define stbir__simdf float32x4_t
 1715+  #define stbir__simdi uint32x4_t
 1716+
 1717+  #define stbir_simdi_castf( reg ) vreinterpretq_u32_f32(reg)
 1718+  #define stbir_simdf_casti( reg ) vreinterpretq_f32_u32(reg)
 1719+
 1720+  #define stbir__simdf_load( reg, ptr ) (reg) = vld1q_f32( (float const*)(ptr) )
 1721+  #define stbir__simdi_load( reg, ptr ) (reg) = vld1q_u32( (uint32_t const*)(ptr) )
 1722+  #define stbir__simdf_load1( out, ptr ) (out) = vld1q_dup_f32( (float const*)(ptr) ) // top values can be random (not denormal or nan for perf)
 1723+  #define stbir__simdi_load1( out, ptr ) (out) = vld1q_dup_u32( (uint32_t const*)(ptr) )
 1724+  #define stbir__simdf_load1z( out, ptr ) (out) = vld1q_lane_f32( (float const*)(ptr), vdupq_n_f32(0), 0 ) // top values must be zero
 1725+  #define stbir__simdf_frep4( fvar ) vdupq_n_f32( fvar )
 1726+  #define stbir__simdf_load1frep4( out, fvar ) (out) = vdupq_n_f32( fvar )
 1727+  #define stbir__simdf_load2( out, ptr ) (out) = vcombine_f32( vld1_f32( (float const*)(ptr) ), vcreate_f32(0) ) // top values can be random (not denormal or nan for perf)
 1728+  #define stbir__simdf_load2z( out, ptr ) (out) = vcombine_f32( vld1_f32( (float const*)(ptr) ), vcreate_f32(0) )  // top values must be zero
 1729+  #define stbir__simdf_load2hmerge( out, reg, ptr ) (out) = vcombine_f32( vget_low_f32(reg), vld1_f32( (float const*)(ptr) ) )
 1730+
 1731+  #define stbir__simdf_zeroP() vdupq_n_f32(0)
 1732+  #define stbir__simdf_zero( reg ) (reg) = vdupq_n_f32(0)
 1733+
 1734+  #define stbir__simdf_store( ptr, reg )  vst1q_f32( (float*)(ptr), reg )
 1735+  #define stbir__simdf_store1( ptr, reg ) vst1q_lane_f32( (float*)(ptr), reg, 0)
 1736+  #define stbir__simdf_store2( ptr, reg ) vst1_f32( (float*)(ptr), vget_low_f32(reg) )
 1737+  #define stbir__simdf_store2h( ptr, reg ) vst1_f32( (float*)(ptr), vget_high_f32(reg) )
 1738+
 1739+  #define stbir__simdi_store( ptr, reg )  vst1q_u32( (uint32_t*)(ptr), reg )
 1740+  #define stbir__simdi_store1( ptr, reg ) vst1q_lane_u32( (uint32_t*)(ptr), reg, 0 )
 1741+  #define stbir__simdi_store2( ptr, reg ) vst1_u32( (uint32_t*)(ptr), vget_low_u32(reg) )
 1742+
 1743+  #define stbir__prefetch( ptr )
 1744+
 1745+  #define stbir__simdi_expand_u8_to_u32(out0,out1,out2,out3,ireg) \
 1746+  { \
 1747+    uint16x8_t l = vmovl_u8( vget_low_u8 ( vreinterpretq_u8_u32(ireg) ) ); \
 1748+    uint16x8_t h = vmovl_u8( vget_high_u8( vreinterpretq_u8_u32(ireg) ) ); \
 1749+    out0 = vmovl_u16( vget_low_u16 ( l ) ); \
 1750+    out1 = vmovl_u16( vget_high_u16( l ) ); \
 1751+    out2 = vmovl_u16( vget_low_u16 ( h ) ); \
 1752+    out3 = vmovl_u16( vget_high_u16( h ) ); \
 1753+  }
 1754+
 1755+  #define stbir__simdi_expand_u8_to_1u32(out,ireg) \
 1756+  { \
 1757+    uint16x8_t tmp = vmovl_u8( vget_low_u8( vreinterpretq_u8_u32(ireg) ) ); \
 1758+    out = vmovl_u16( vget_low_u16( tmp ) ); \
 1759+  }
 1760+
 1761+  #define stbir__simdi_expand_u16_to_u32(out0,out1,ireg) \
 1762+  { \
 1763+    uint16x8_t tmp = vreinterpretq_u16_u32(ireg); \
 1764+    out0 = vmovl_u16( vget_low_u16 ( tmp ) ); \
 1765+    out1 = vmovl_u16( vget_high_u16( tmp ) ); \
 1766+  }
 1767+
 1768+  #define stbir__simdf_convert_float_to_i32( i, f ) (i) = vreinterpretq_u32_s32( vcvtq_s32_f32(f) )
 1769+  #define stbir__simdf_convert_float_to_int( f ) vgetq_lane_s32(vcvtq_s32_f32(f), 0)
 1770+  #define stbir__simdi_to_int( i ) (int)vgetq_lane_u32(i, 0)
 1771+  #define stbir__simdf_convert_float_to_uint8( f ) ((unsigned char)vgetq_lane_s32(vcvtq_s32_f32(vmaxq_f32(vminq_f32(f,STBIR__CONSTF(STBIR_max_uint8_as_float)),vdupq_n_f32(0))), 0))
 1772+  #define stbir__simdf_convert_float_to_short( f ) ((unsigned short)vgetq_lane_s32(vcvtq_s32_f32(vmaxq_f32(vminq_f32(f,STBIR__CONSTF(STBIR_max_uint16_as_float)),vdupq_n_f32(0))), 0))
 1773+  #define stbir__simdi_convert_i32_to_float(out, ireg) (out) = vcvtq_f32_s32( vreinterpretq_s32_u32(ireg) )
 1774+  #define stbir__simdf_add( out, reg0, reg1 ) (out) = vaddq_f32( reg0, reg1 )
 1775+  #define stbir__simdf_mult( out, reg0, reg1 ) (out) = vmulq_f32( reg0, reg1 )
 1776+  #define stbir__simdf_mult_mem( out, reg, ptr ) (out) = vmulq_f32( reg, vld1q_f32( (float const*)(ptr) ) )
 1777+  #define stbir__simdf_mult1_mem( out, reg, ptr ) (out) = vmulq_f32( reg, vld1q_dup_f32( (float const*)(ptr) ) )
 1778+  #define stbir__simdf_add_mem( out, reg, ptr ) (out) = vaddq_f32( reg, vld1q_f32( (float const*)(ptr) ) )
 1779+  #define stbir__simdf_add1_mem( out, reg, ptr ) (out) = vaddq_f32( reg, vld1q_dup_f32( (float const*)(ptr) ) )
 1780+
 1781+  #ifdef STBIR_USE_FMA           // not on by default to maintain bit identical simd to non-simd (and also x64 no madd to arm madd)
 1782+  #define stbir__simdf_madd( out, add, mul1, mul2 ) (out) = vfmaq_f32( add, mul1, mul2 )
 1783+  #define stbir__simdf_madd1( out, add, mul1, mul2 ) (out) = vfmaq_f32( add, mul1, mul2 )
 1784+  #define stbir__simdf_madd_mem( out, add, mul, ptr ) (out) = vfmaq_f32( add, mul, vld1q_f32( (float const*)(ptr) ) )
 1785+  #define stbir__simdf_madd1_mem( out, add, mul, ptr ) (out) = vfmaq_f32( add, mul, vld1q_dup_f32( (float const*)(ptr) ) )
 1786+  #else
 1787+  #define stbir__simdf_madd( out, add, mul1, mul2 ) (out) = vaddq_f32( add, vmulq_f32( mul1, mul2 ) )
 1788+  #define stbir__simdf_madd1( out, add, mul1, mul2 ) (out) = vaddq_f32( add, vmulq_f32( mul1, mul2 ) )
 1789+  #define stbir__simdf_madd_mem( out, add, mul, ptr ) (out) = vaddq_f32( add, vmulq_f32( mul, vld1q_f32( (float const*)(ptr) ) ) )
 1790+  #define stbir__simdf_madd1_mem( out, add, mul, ptr ) (out) = vaddq_f32( add, vmulq_f32( mul, vld1q_dup_f32( (float const*)(ptr) ) ) )
 1791+  #endif
 1792+
 1793+  #define stbir__simdf_add1( out, reg0, reg1 ) (out) = vaddq_f32( reg0, reg1 )
 1794+  #define stbir__simdf_mult1( out, reg0, reg1 ) (out) = vmulq_f32( reg0, reg1 )
 1795+
 1796+  #define stbir__simdf_and( out, reg0, reg1 ) (out) = vreinterpretq_f32_u32( vandq_u32( vreinterpretq_u32_f32(reg0), vreinterpretq_u32_f32(reg1) ) )
 1797+  #define stbir__simdf_or( out, reg0, reg1 ) (out) = vreinterpretq_f32_u32( vorrq_u32( vreinterpretq_u32_f32(reg0), vreinterpretq_u32_f32(reg1) ) )
 1798+
 1799+  #define stbir__simdf_min( out, reg0, reg1 ) (out) = vminq_f32( reg0, reg1 )
 1800+  #define stbir__simdf_max( out, reg0, reg1 ) (out) = vmaxq_f32( reg0, reg1 )
 1801+  #define stbir__simdf_min1( out, reg0, reg1 ) (out) = vminq_f32( reg0, reg1 )
 1802+  #define stbir__simdf_max1( out, reg0, reg1 ) (out) = vmaxq_f32( reg0, reg1 )
 1803+
 1804+  #define stbir__simdf_0123ABCDto3ABx( out, reg0, reg1 ) (out) = vextq_f32( reg0, reg1, 3 )
 1805+  #define stbir__simdf_0123ABCDto23Ax( out, reg0, reg1 ) (out) = vextq_f32( reg0, reg1, 2 )
 1806+
 1807+  #define stbir__simdf_a1a1( out, alp, ones ) (out) = vzipq_f32(vuzpq_f32(alp, alp).val[1], ones).val[0]
 1808+  #define stbir__simdf_1a1a( out, alp, ones ) (out) = vzipq_f32(ones, vuzpq_f32(alp, alp).val[0]).val[0]
 1809+
 1810+  #if defined( _M_ARM64 ) || defined( __aarch64__ ) || defined( __arm64__ )
 1811+
 1812+    #define stbir__simdf_aaa1( out, alp, ones ) (out) = vcopyq_laneq_f32(vdupq_n_f32(vgetq_lane_f32(alp, 3)), 3, ones, 3)
 1813+    #define stbir__simdf_1aaa( out, alp, ones ) (out) = vcopyq_laneq_f32(vdupq_n_f32(vgetq_lane_f32(alp, 0)), 0, ones, 0)
 1814+
 1815+    #if defined( _MSC_VER ) && !defined(__clang__)
 1816+      #define stbir_make16(a,b,c,d) vcombine_u8( \
 1817+        vcreate_u8( (4*a+0) | ((4*a+1)<<8) | ((4*a+2)<<16) | ((4*a+3)<<24) | \
 1818+          ((stbir_uint64)(4*b+0)<<32) | ((stbir_uint64)(4*b+1)<<40) | ((stbir_uint64)(4*b+2)<<48) | ((stbir_uint64)(4*b+3)<<56)), \
 1819+        vcreate_u8( (4*c+0) | ((4*c+1)<<8) | ((4*c+2)<<16) | ((4*c+3)<<24) | \
 1820+          ((stbir_uint64)(4*d+0)<<32) | ((stbir_uint64)(4*d+1)<<40) | ((stbir_uint64)(4*d+2)<<48) | ((stbir_uint64)(4*d+3)<<56) ) )
 1821+
 1822+      static stbir__inline uint8x16x2_t stbir_make16x2(float32x4_t rega,float32x4_t regb)
 1823+      {
 1824+        uint8x16x2_t r = { vreinterpretq_u8_f32(rega), vreinterpretq_u8_f32(regb) };
 1825+        return r;
 1826+      }
 1827+    #else
 1828+      #define stbir_make16(a,b,c,d) (uint8x16_t){4*a+0,4*a+1,4*a+2,4*a+3,4*b+0,4*b+1,4*b+2,4*b+3,4*c+0,4*c+1,4*c+2,4*c+3,4*d+0,4*d+1,4*d+2,4*d+3}
 1829+      #define stbir_make16x2(a,b) (uint8x16x2_t){{vreinterpretq_u8_f32(a),vreinterpretq_u8_f32(b)}}
 1830+    #endif
 1831+
 1832+    #define stbir__simdf_swiz( reg, one, two, three, four ) vreinterpretq_f32_u8( vqtbl1q_u8( vreinterpretq_u8_f32(reg), stbir_make16(one, two, three, four) ) )
 1833+    #define stbir__simdf_swiz2( rega, regb, one, two, three, four ) vreinterpretq_f32_u8( vqtbl2q_u8( stbir_make16x2(rega,regb), stbir_make16(one, two, three, four) ) )
 1834+
 1835+    #define stbir__simdi_16madd( out, reg0, reg1 ) \
 1836+    { \
 1837+      int16x8_t r0 = vreinterpretq_s16_u32(reg0); \
 1838+      int16x8_t r1 = vreinterpretq_s16_u32(reg1); \
 1839+      int32x4_t tmp0 = vmull_s16( vget_low_s16(r0), vget_low_s16(r1) ); \
 1840+      int32x4_t tmp1 = vmull_s16( vget_high_s16(r0), vget_high_s16(r1) ); \
 1841+      (out) = vreinterpretq_u32_s32( vpaddq_s32(tmp0, tmp1) ); \
 1842+    }
 1843+
 1844+  #else
 1845+
 1846+    #define stbir__simdf_aaa1( out, alp, ones ) (out) = vsetq_lane_f32(1.0f, vdupq_n_f32(vgetq_lane_f32(alp, 3)), 3)
 1847+    #define stbir__simdf_1aaa( out, alp, ones ) (out) = vsetq_lane_f32(1.0f, vdupq_n_f32(vgetq_lane_f32(alp, 0)), 0)
 1848+
 1849+    #if defined( _MSC_VER ) && !defined(__clang__)
 1850+      static stbir__inline uint8x8x2_t stbir_make8x2(float32x4_t reg)
 1851+      {
 1852+        uint8x8x2_t r = { { vget_low_u8(vreinterpretq_u8_f32(reg)), vget_high_u8(vreinterpretq_u8_f32(reg)) } };
 1853+        return r;
 1854+      }
 1855+      #define stbir_make8(a,b) vcreate_u8( \
 1856+        (4*a+0) | ((4*a+1)<<8) | ((4*a+2)<<16) | ((4*a+3)<<24) | \
 1857+        ((stbir_uint64)(4*b+0)<<32) | ((stbir_uint64)(4*b+1)<<40) | ((stbir_uint64)(4*b+2)<<48) | ((stbir_uint64)(4*b+3)<<56) )
 1858+    #else
 1859+      #define stbir_make8x2(reg) (uint8x8x2_t){ { vget_low_u8(vreinterpretq_u8_f32(reg)), vget_high_u8(vreinterpretq_u8_f32(reg)) } }
 1860+      #define stbir_make8(a,b) (uint8x8_t){4*a+0,4*a+1,4*a+2,4*a+3,4*b+0,4*b+1,4*b+2,4*b+3}
 1861+    #endif
 1862+
 1863+    #define stbir__simdf_swiz( reg, one, two, three, four ) vreinterpretq_f32_u8( vcombine_u8( \
 1864+        vtbl2_u8( stbir_make8x2( reg ), stbir_make8( one, two ) ), \
 1865+        vtbl2_u8( stbir_make8x2( reg ), stbir_make8( three, four ) ) ) )
 1866+
 1867+    #define stbir__simdi_16madd( out, reg0, reg1 ) \
 1868+    { \
 1869+      int16x8_t r0 = vreinterpretq_s16_u32(reg0); \
 1870+      int16x8_t r1 = vreinterpretq_s16_u32(reg1); \
 1871+      int32x4_t tmp0 = vmull_s16( vget_low_s16(r0), vget_low_s16(r1) ); \
 1872+      int32x4_t tmp1 = vmull_s16( vget_high_s16(r0), vget_high_s16(r1) ); \
 1873+      int32x2_t out0 = vpadd_s32( vget_low_s32(tmp0), vget_high_s32(tmp0) ); \
 1874+      int32x2_t out1 = vpadd_s32( vget_low_s32(tmp1), vget_high_s32(tmp1) ); \
 1875+      (out) = vreinterpretq_u32_s32( vcombine_s32(out0, out1) ); \
 1876+    }
 1877+
 1878+  #endif
 1879+
 1880+  #define stbir__simdi_and( out, reg0, reg1 ) (out) = vandq_u32( reg0, reg1 )
 1881+  #define stbir__simdi_or( out, reg0, reg1 ) (out) = vorrq_u32( reg0, reg1 )
 1882+
 1883+  #define stbir__simdf_pack_to_8bytes(out,aa,bb) \
 1884+  { \
 1885+    float32x4_t af = vmaxq_f32( vminq_f32(aa,STBIR__CONSTF(STBIR_max_uint8_as_float) ), vdupq_n_f32(0) ); \
 1886+    float32x4_t bf = vmaxq_f32( vminq_f32(bb,STBIR__CONSTF(STBIR_max_uint8_as_float) ), vdupq_n_f32(0) ); \
 1887+    int16x4_t ai = vqmovn_s32( vcvtq_s32_f32( af ) ); \
 1888+    int16x4_t bi = vqmovn_s32( vcvtq_s32_f32( bf ) ); \
 1889+    uint8x8_t out8 = vqmovun_s16( vcombine_s16(ai, bi) ); \
 1890+    out = vreinterpretq_u32_u8( vcombine_u8(out8, out8) ); \
 1891+  }
 1892+
 1893+  #define stbir__simdf_pack_to_8words(out,aa,bb) \
 1894+  { \
 1895+    float32x4_t af = vmaxq_f32( vminq_f32(aa,STBIR__CONSTF(STBIR_max_uint16_as_float) ), vdupq_n_f32(0) ); \
 1896+    float32x4_t bf = vmaxq_f32( vminq_f32(bb,STBIR__CONSTF(STBIR_max_uint16_as_float) ), vdupq_n_f32(0) ); \
 1897+    int32x4_t ai = vcvtq_s32_f32( af ); \
 1898+    int32x4_t bi = vcvtq_s32_f32( bf ); \
 1899+    out = vreinterpretq_u32_u16( vcombine_u16(vqmovun_s32(ai), vqmovun_s32(bi)) ); \
 1900+  }
 1901+
 1902+  #define stbir__interleave_pack_and_store_16_u8( ptr, r0, r1, r2, r3 ) \
 1903+  { \
 1904+    int16x4x2_t tmp0 = vzip_s16( vqmovn_s32(vreinterpretq_s32_u32(r0)), vqmovn_s32(vreinterpretq_s32_u32(r2)) ); \
 1905+    int16x4x2_t tmp1 = vzip_s16( vqmovn_s32(vreinterpretq_s32_u32(r1)), vqmovn_s32(vreinterpretq_s32_u32(r3)) ); \
 1906+    uint8x8x2_t out = \
 1907+    { { \
 1908+      vqmovun_s16( vcombine_s16(tmp0.val[0], tmp0.val[1]) ), \
 1909+      vqmovun_s16( vcombine_s16(tmp1.val[0], tmp1.val[1]) ), \
 1910+    } }; \
 1911+    vst2_u8(ptr, out); \
 1912+  }
 1913+
 1914+  #define stbir__simdf_load4_transposed( o0, o1, o2, o3, ptr ) \
 1915+  { \
 1916+    float32x4x4_t tmp = vld4q_f32(ptr); \
 1917+    o0 = tmp.val[0]; \
 1918+    o1 = tmp.val[1]; \
 1919+    o2 = tmp.val[2]; \
 1920+    o3 = tmp.val[3]; \
 1921+  }
 1922+
 1923+  #define stbir__simdi_32shr( out, reg, imm ) out = vshrq_n_u32( reg, imm )
 1924+
 1925+  #if defined( _MSC_VER ) && !defined(__clang__)
 1926+    #define STBIR__SIMDF_CONST(var, x) __declspec(align(8)) float var[] = { x, x, x, x }
 1927+    #define STBIR__SIMDI_CONST(var, x) __declspec(align(8)) uint32_t var[] = { x, x, x, x }
 1928+    #define STBIR__CONSTF(var) (*(const float32x4_t*)var)
 1929+    #define STBIR__CONSTI(var) (*(const uint32x4_t*)var)
 1930+  #else
 1931+    #define STBIR__SIMDF_CONST(var, x) stbir__simdf var = { x, x, x, x }
 1932+    #define STBIR__SIMDI_CONST(var, x) stbir__simdi var = { x, x, x, x }
 1933+    #define STBIR__CONSTF(var) (var)
 1934+    #define STBIR__CONSTI(var) (var)
 1935+  #endif
 1936+
 1937+  #ifdef STBIR_FLOORF
 1938+  #undef STBIR_FLOORF
 1939+  #endif
 1940+  #define STBIR_FLOORF stbir_simd_floorf
 1941+  static stbir__inline float stbir_simd_floorf(float x)
 1942+  {
 1943+    #if defined( _M_ARM64 ) || defined( __aarch64__ ) || defined( __arm64__ )
 1944+    return vget_lane_f32( vrndm_f32( vdup_n_f32(x) ), 0);
 1945+    #else
 1946+    float32x2_t f = vdup_n_f32(x);
 1947+    float32x2_t t = vcvt_f32_s32(vcvt_s32_f32(f));
 1948+    uint32x2_t a = vclt_f32(f, t);
 1949+    uint32x2_t b = vreinterpret_u32_f32(vdup_n_f32(-1.0f));
 1950+    float32x2_t r = vadd_f32(t, vreinterpret_f32_u32(vand_u32(a, b)));
 1951+    return vget_lane_f32(r, 0);
 1952+    #endif
 1953+  }
 1954+
 1955+  #ifdef STBIR_CEILF
 1956+  #undef STBIR_CEILF
 1957+  #endif
 1958+  #define STBIR_CEILF stbir_simd_ceilf
 1959+  static stbir__inline float stbir_simd_ceilf(float x)
 1960+  {
 1961+    #if defined( _M_ARM64 ) || defined( __aarch64__ ) || defined( __arm64__ )
 1962+    return vget_lane_f32( vrndp_f32( vdup_n_f32(x) ), 0);
 1963+    #else
 1964+    float32x2_t f = vdup_n_f32(x);
 1965+    float32x2_t t = vcvt_f32_s32(vcvt_s32_f32(f));
 1966+    uint32x2_t a = vclt_f32(t, f);
 1967+    uint32x2_t b = vreinterpret_u32_f32(vdup_n_f32(1.0f));
 1968+    float32x2_t r = vadd_f32(t, vreinterpret_f32_u32(vand_u32(a, b)));
 1969+    return vget_lane_f32(r, 0);
 1970+    #endif
 1971+  }
 1972+
 1973+  #define STBIR_SIMD
 1974+
 1975+#elif defined(STBIR_WASM)
 1976+
 1977+  #include <wasm_simd128.h>
 1978+
 1979+  #define stbir__simdf v128_t
 1980+  #define stbir__simdi v128_t
 1981+
 1982+  #define stbir_simdi_castf( reg ) (reg)
 1983+  #define stbir_simdf_casti( reg ) (reg)
 1984+
 1985+  #define stbir__simdf_load( reg, ptr )             (reg) = wasm_v128_load( (void const*)(ptr) )
 1986+  #define stbir__simdi_load( reg, ptr )             (reg) = wasm_v128_load( (void const*)(ptr) )
 1987+  #define stbir__simdf_load1( out, ptr )            (out) = wasm_v128_load32_splat( (void const*)(ptr) ) // top values can be random (not denormal or nan for perf)
 1988+  #define stbir__simdi_load1( out, ptr )            (out) = wasm_v128_load32_splat( (void const*)(ptr) )
 1989+  #define stbir__simdf_load1z( out, ptr )           (out) = wasm_v128_load32_zero( (void const*)(ptr) ) // top values must be zero
 1990+  #define stbir__simdf_frep4( fvar )                wasm_f32x4_splat( fvar )
 1991+  #define stbir__simdf_load1frep4( out, fvar )      (out) = wasm_f32x4_splat( fvar )
 1992+  #define stbir__simdf_load2( out, ptr )            (out) = wasm_v128_load64_splat( (void const*)(ptr) ) // top values can be random (not denormal or nan for perf)
 1993+  #define stbir__simdf_load2z( out, ptr )           (out) = wasm_v128_load64_zero( (void const*)(ptr) ) // top values must be zero
 1994+  #define stbir__simdf_load2hmerge( out, reg, ptr ) (out) = wasm_v128_load64_lane( (void const*)(ptr), reg, 1 )
 1995+
 1996+  #define stbir__simdf_zeroP() wasm_f32x4_const_splat(0)
 1997+  #define stbir__simdf_zero( reg ) (reg) = wasm_f32x4_const_splat(0)
 1998+
 1999+  #define stbir__simdf_store( ptr, reg )   wasm_v128_store( (void*)(ptr), reg )
 2000+  #define stbir__simdf_store1( ptr, reg )  wasm_v128_store32_lane( (void*)(ptr), reg, 0 )
 2001+  #define stbir__simdf_store2( ptr, reg )  wasm_v128_store64_lane( (void*)(ptr), reg, 0 )
 2002+  #define stbir__simdf_store2h( ptr, reg ) wasm_v128_store64_lane( (void*)(ptr), reg, 1 )
 2003+
 2004+  #define stbir__simdi_store( ptr, reg )  wasm_v128_store( (void*)(ptr), reg )
 2005+  #define stbir__simdi_store1( ptr, reg ) wasm_v128_store32_lane( (void*)(ptr), reg, 0 )
 2006+  #define stbir__simdi_store2( ptr, reg ) wasm_v128_store64_lane( (void*)(ptr), reg, 0 )
 2007+
 2008+  #define stbir__prefetch( ptr )
 2009+
 2010+  #define stbir__simdi_expand_u8_to_u32(out0,out1,out2,out3,ireg) \
 2011+  { \
 2012+    v128_t l = wasm_u16x8_extend_low_u8x16 ( ireg ); \
 2013+    v128_t h = wasm_u16x8_extend_high_u8x16( ireg ); \
 2014+    out0 = wasm_u32x4_extend_low_u16x8 ( l ); \
 2015+    out1 = wasm_u32x4_extend_high_u16x8( l ); \
 2016+    out2 = wasm_u32x4_extend_low_u16x8 ( h ); \
 2017+    out3 = wasm_u32x4_extend_high_u16x8( h ); \
 2018+  }
 2019+
 2020+  #define stbir__simdi_expand_u8_to_1u32(out,ireg) \
 2021+  { \
 2022+    v128_t tmp = wasm_u16x8_extend_low_u8x16(ireg); \
 2023+    out = wasm_u32x4_extend_low_u16x8(tmp); \
 2024+  }
 2025+
 2026+  #define stbir__simdi_expand_u16_to_u32(out0,out1,ireg) \
 2027+  { \
 2028+    out0 = wasm_u32x4_extend_low_u16x8 ( ireg ); \
 2029+    out1 = wasm_u32x4_extend_high_u16x8( ireg ); \
 2030+  }
 2031+
 2032+  #define stbir__simdf_convert_float_to_i32( i, f )    (i) = wasm_i32x4_trunc_sat_f32x4(f)
 2033+  #define stbir__simdf_convert_float_to_int( f )       wasm_i32x4_extract_lane(wasm_i32x4_trunc_sat_f32x4(f), 0)
 2034+  #define stbir__simdi_to_int( i )                     wasm_i32x4_extract_lane(i, 0)
 2035+  #define stbir__simdf_convert_float_to_uint8( f )     ((unsigned char)wasm_i32x4_extract_lane(wasm_i32x4_trunc_sat_f32x4(wasm_f32x4_max(wasm_f32x4_min(f,STBIR_max_uint8_as_float),wasm_f32x4_const_splat(0))), 0))
 2036+  #define stbir__simdf_convert_float_to_short( f )     ((unsigned short)wasm_i32x4_extract_lane(wasm_i32x4_trunc_sat_f32x4(wasm_f32x4_max(wasm_f32x4_min(f,STBIR_max_uint16_as_float),wasm_f32x4_const_splat(0))), 0))
 2037+  #define stbir__simdi_convert_i32_to_float(out, ireg) (out) = wasm_f32x4_convert_i32x4(ireg)
 2038+  #define stbir__simdf_add( out, reg0, reg1 )          (out) = wasm_f32x4_add( reg0, reg1 )
 2039+  #define stbir__simdf_mult( out, reg0, reg1 )         (out) = wasm_f32x4_mul( reg0, reg1 )
 2040+  #define stbir__simdf_mult_mem( out, reg, ptr )       (out) = wasm_f32x4_mul( reg, wasm_v128_load( (void const*)(ptr) ) )
 2041+  #define stbir__simdf_mult1_mem( out, reg, ptr )      (out) = wasm_f32x4_mul( reg, wasm_v128_load32_splat( (void const*)(ptr) ) )
 2042+  #define stbir__simdf_add_mem( out, reg, ptr )        (out) = wasm_f32x4_add( reg, wasm_v128_load( (void const*)(ptr) ) )
 2043+  #define stbir__simdf_add1_mem( out, reg, ptr )       (out) = wasm_f32x4_add( reg, wasm_v128_load32_splat( (void const*)(ptr) ) )
 2044+
 2045+  #define stbir__simdf_madd( out, add, mul1, mul2 )    (out) = wasm_f32x4_add( add, wasm_f32x4_mul( mul1, mul2 ) )
 2046+  #define stbir__simdf_madd1( out, add, mul1, mul2 )   (out) = wasm_f32x4_add( add, wasm_f32x4_mul( mul1, mul2 ) )
 2047+  #define stbir__simdf_madd_mem( out, add, mul, ptr )  (out) = wasm_f32x4_add( add, wasm_f32x4_mul( mul, wasm_v128_load( (void const*)(ptr) ) ) )
 2048+  #define stbir__simdf_madd1_mem( out, add, mul, ptr ) (out) = wasm_f32x4_add( add, wasm_f32x4_mul( mul, wasm_v128_load32_splat( (void const*)(ptr) ) ) )
 2049+
 2050+  #define stbir__simdf_add1( out, reg0, reg1 )  (out) = wasm_f32x4_add( reg0, reg1 )
 2051+  #define stbir__simdf_mult1( out, reg0, reg1 ) (out) = wasm_f32x4_mul( reg0, reg1 )
 2052+
 2053+  #define stbir__simdf_and( out, reg0, reg1 ) (out) = wasm_v128_and( reg0, reg1 )
 2054+  #define stbir__simdf_or( out, reg0, reg1 )  (out) = wasm_v128_or( reg0, reg1 )
 2055+
 2056+  #define stbir__simdf_min( out, reg0, reg1 ) (out) = wasm_f32x4_min( reg0, reg1 )
 2057+  #define stbir__simdf_max( out, reg0, reg1 ) (out) = wasm_f32x4_max( reg0, reg1 )
 2058+  #define stbir__simdf_min1( out, reg0, reg1 ) (out) = wasm_f32x4_min( reg0, reg1 )
 2059+  #define stbir__simdf_max1( out, reg0, reg1 ) (out) = wasm_f32x4_max( reg0, reg1 )
 2060+
 2061+  #define stbir__simdf_0123ABCDto3ABx( out, reg0, reg1 ) (out) = wasm_i32x4_shuffle( reg0, reg1, 3, 4, 5, -1 )
 2062+  #define stbir__simdf_0123ABCDto23Ax( out, reg0, reg1 ) (out) = wasm_i32x4_shuffle( reg0, reg1, 2, 3, 4, -1 )
 2063+
 2064+  #define stbir__simdf_aaa1(out,alp,ones) (out) = wasm_i32x4_shuffle(alp, ones, 3, 3, 3, 4)
 2065+  #define stbir__simdf_1aaa(out,alp,ones) (out) = wasm_i32x4_shuffle(alp, ones, 4, 0, 0, 0)
 2066+  #define stbir__simdf_a1a1(out,alp,ones) (out) = wasm_i32x4_shuffle(alp, ones, 1, 4, 3, 4)
 2067+  #define stbir__simdf_1a1a(out,alp,ones) (out) = wasm_i32x4_shuffle(alp, ones, 4, 0, 4, 2)
 2068+
 2069+  #define stbir__simdf_swiz( reg, one, two, three, four ) wasm_i32x4_shuffle(reg, reg, one, two, three, four)
 2070+
 2071+  #define stbir__simdi_and( out, reg0, reg1 )    (out) = wasm_v128_and( reg0, reg1 )
 2072+  #define stbir__simdi_or( out, reg0, reg1 )     (out) = wasm_v128_or( reg0, reg1 )
 2073+  #define stbir__simdi_16madd( out, reg0, reg1 ) (out) = wasm_i32x4_dot_i16x8( reg0, reg1 )
 2074+
 2075+  #define stbir__simdf_pack_to_8bytes(out,aa,bb) \
 2076+  { \
 2077+    v128_t af = wasm_f32x4_max( wasm_f32x4_min(aa, STBIR_max_uint8_as_float), wasm_f32x4_const_splat(0) ); \
 2078+    v128_t bf = wasm_f32x4_max( wasm_f32x4_min(bb, STBIR_max_uint8_as_float), wasm_f32x4_const_splat(0) ); \
 2079+    v128_t ai = wasm_i32x4_trunc_sat_f32x4( af ); \
 2080+    v128_t bi = wasm_i32x4_trunc_sat_f32x4( bf ); \
 2081+    v128_t out16 = wasm_i16x8_narrow_i32x4( ai, bi ); \
 2082+    out = wasm_u8x16_narrow_i16x8( out16, out16 ); \
 2083+  }
 2084+
 2085+  #define stbir__simdf_pack_to_8words(out,aa,bb) \
 2086+  { \
 2087+    v128_t af = wasm_f32x4_max( wasm_f32x4_min(aa, STBIR_max_uint16_as_float), wasm_f32x4_const_splat(0)); \
 2088+    v128_t bf = wasm_f32x4_max( wasm_f32x4_min(bb, STBIR_max_uint16_as_float), wasm_f32x4_const_splat(0)); \
 2089+    v128_t ai = wasm_i32x4_trunc_sat_f32x4( af ); \
 2090+    v128_t bi = wasm_i32x4_trunc_sat_f32x4( bf ); \
 2091+    out = wasm_u16x8_narrow_i32x4( ai, bi ); \
 2092+  }
 2093+
 2094+  #define stbir__interleave_pack_and_store_16_u8( ptr, r0, r1, r2, r3 ) \
 2095+  { \
 2096+    v128_t tmp0 = wasm_i16x8_narrow_i32x4(r0, r1); \
 2097+    v128_t tmp1 = wasm_i16x8_narrow_i32x4(r2, r3); \
 2098+    v128_t tmp = wasm_u8x16_narrow_i16x8(tmp0, tmp1); \
 2099+    tmp = wasm_i8x16_shuffle(tmp, tmp, 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15); \
 2100+    wasm_v128_store( (void*)(ptr), tmp); \
 2101+  }
 2102+
 2103+  #define stbir__simdf_load4_transposed( o0, o1, o2, o3, ptr ) \
 2104+  { \
 2105+    v128_t t0 = wasm_v128_load( ptr    ); \
 2106+    v128_t t1 = wasm_v128_load( ptr+4  ); \
 2107+    v128_t t2 = wasm_v128_load( ptr+8  ); \
 2108+    v128_t t3 = wasm_v128_load( ptr+12 ); \
 2109+    v128_t s0 = wasm_i32x4_shuffle(t0, t1, 0, 4, 2, 6); \
 2110+    v128_t s1 = wasm_i32x4_shuffle(t0, t1, 1, 5, 3, 7); \
 2111+    v128_t s2 = wasm_i32x4_shuffle(t2, t3, 0, 4, 2, 6); \
 2112+    v128_t s3 = wasm_i32x4_shuffle(t2, t3, 1, 5, 3, 7); \
 2113+    o0 = wasm_i32x4_shuffle(s0, s2, 0, 1, 4, 5); \
 2114+    o1 = wasm_i32x4_shuffle(s1, s3, 0, 1, 4, 5); \
 2115+    o2 = wasm_i32x4_shuffle(s0, s2, 2, 3, 6, 7); \
 2116+    o3 = wasm_i32x4_shuffle(s1, s3, 2, 3, 6, 7); \
 2117+  }
 2118+
 2119+  #define stbir__simdi_32shr( out, reg, imm ) out = wasm_u32x4_shr( reg, imm )
 2120+
 2121+  typedef float stbir__f32x4 __attribute__((__vector_size__(16), __aligned__(16)));
 2122+  #define STBIR__SIMDF_CONST(var, x) stbir__simdf var = (v128_t)(stbir__f32x4){ x, x, x, x }
 2123+  #define STBIR__SIMDI_CONST(var, x) stbir__simdi var = { x, x, x, x }
 2124+  #define STBIR__CONSTF(var) (var)
 2125+  #define STBIR__CONSTI(var) (var)
 2126+
 2127+  #ifdef STBIR_FLOORF
 2128+  #undef STBIR_FLOORF
 2129+  #endif
 2130+  #define STBIR_FLOORF stbir_simd_floorf
 2131+  static stbir__inline float stbir_simd_floorf(float x)
 2132+  {
 2133+    return wasm_f32x4_extract_lane( wasm_f32x4_floor( wasm_f32x4_splat(x) ), 0);
 2134+  }
 2135+
 2136+  #ifdef STBIR_CEILF
 2137+  #undef STBIR_CEILF
 2138+  #endif
 2139+  #define STBIR_CEILF stbir_simd_ceilf
 2140+  static stbir__inline float stbir_simd_ceilf(float x)
 2141+  {
 2142+    return wasm_f32x4_extract_lane( wasm_f32x4_ceil( wasm_f32x4_splat(x) ), 0);
 2143+  }
 2144+
 2145+  #define STBIR_SIMD
 2146+
 2147+#endif  // SSE2/NEON/WASM
 2148+
 2149+#endif // NO SIMD
 2150+
 2151+#ifdef STBIR_SIMD8
 2152+  #define stbir__simdfX stbir__simdf8
 2153+  #define stbir__simdiX stbir__simdi8
 2154+  #define stbir__simdfX_load stbir__simdf8_load
 2155+  #define stbir__simdiX_load stbir__simdi8_load
 2156+  #define stbir__simdfX_mult stbir__simdf8_mult
 2157+  #define stbir__simdfX_add_mem stbir__simdf8_add_mem
 2158+  #define stbir__simdfX_madd_mem stbir__simdf8_madd_mem
 2159+  #define stbir__simdfX_store stbir__simdf8_store
 2160+  #define stbir__simdiX_store stbir__simdi8_store
 2161+  #define stbir__simdf_frepX  stbir__simdf8_frep8
 2162+  #define stbir__simdfX_madd stbir__simdf8_madd
 2163+  #define stbir__simdfX_min stbir__simdf8_min
 2164+  #define stbir__simdfX_max stbir__simdf8_max
 2165+  #define stbir__simdfX_aaa1 stbir__simdf8_aaa1
 2166+  #define stbir__simdfX_1aaa stbir__simdf8_1aaa
 2167+  #define stbir__simdfX_a1a1 stbir__simdf8_a1a1
 2168+  #define stbir__simdfX_1a1a stbir__simdf8_1a1a
 2169+  #define stbir__simdfX_convert_float_to_i32 stbir__simdf8_convert_float_to_i32
 2170+  #define stbir__simdfX_pack_to_words stbir__simdf8_pack_to_16words
 2171+  #define stbir__simdfX_zero stbir__simdf8_zero
 2172+  #define STBIR_onesX STBIR_ones8
 2173+  #define STBIR_max_uint8_as_floatX STBIR_max_uint8_as_float8
 2174+  #define STBIR_max_uint16_as_floatX STBIR_max_uint16_as_float8
 2175+  #define STBIR_simd_point5X STBIR_simd_point58
 2176+  #define stbir__simdfX_float_count 8
 2177+  #define stbir__simdfX_0123to1230 stbir__simdf8_0123to12301230
 2178+  #define stbir__simdfX_0123to2103 stbir__simdf8_0123to21032103
 2179+  static const stbir__simdf8 STBIR_max_uint16_as_float_inverted8 = { stbir__max_uint16_as_float_inverted,stbir__max_uint16_as_float_inverted,stbir__max_uint16_as_float_inverted,stbir__max_uint16_as_float_inverted,stbir__max_uint16_as_float_inverted,stbir__max_uint16_as_float_inverted,stbir__max_uint16_as_float_inverted,stbir__max_uint16_as_float_inverted };
 2180+  static const stbir__simdf8 STBIR_max_uint8_as_float_inverted8 = { stbir__max_uint8_as_float_inverted,stbir__max_uint8_as_float_inverted,stbir__max_uint8_as_float_inverted,stbir__max_uint8_as_float_inverted,stbir__max_uint8_as_float_inverted,stbir__max_uint8_as_float_inverted,stbir__max_uint8_as_float_inverted,stbir__max_uint8_as_float_inverted };
 2181+  static const stbir__simdf8 STBIR_ones8 = { 1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0 };
 2182+  static const stbir__simdf8 STBIR_simd_point58 = { 0.5,0.5,0.5,0.5,0.5,0.5,0.5,0.5 };
 2183+  static const stbir__simdf8 STBIR_max_uint8_as_float8 = { stbir__max_uint8_as_float,stbir__max_uint8_as_float,stbir__max_uint8_as_float,stbir__max_uint8_as_float, stbir__max_uint8_as_float,stbir__max_uint8_as_float,stbir__max_uint8_as_float,stbir__max_uint8_as_float };
 2184+  static const stbir__simdf8 STBIR_max_uint16_as_float8 = { stbir__max_uint16_as_float,stbir__max_uint16_as_float,stbir__max_uint16_as_float,stbir__max_uint16_as_float, stbir__max_uint16_as_float,stbir__max_uint16_as_float,stbir__max_uint16_as_float,stbir__max_uint16_as_float };
 2185+#else
 2186+  #define stbir__simdfX stbir__simdf
 2187+  #define stbir__simdiX stbir__simdi
 2188+  #define stbir__simdfX_load stbir__simdf_load
 2189+  #define stbir__simdiX_load stbir__simdi_load
 2190+  #define stbir__simdfX_mult stbir__simdf_mult
 2191+  #define stbir__simdfX_add_mem stbir__simdf_add_mem
 2192+  #define stbir__simdfX_madd_mem stbir__simdf_madd_mem
 2193+  #define stbir__simdfX_store stbir__simdf_store
 2194+  #define stbir__simdiX_store stbir__simdi_store
 2195+  #define stbir__simdf_frepX  stbir__simdf_frep4
 2196+  #define stbir__simdfX_madd stbir__simdf_madd
 2197+  #define stbir__simdfX_min stbir__simdf_min
 2198+  #define stbir__simdfX_max stbir__simdf_max
 2199+  #define stbir__simdfX_aaa1 stbir__simdf_aaa1
 2200+  #define stbir__simdfX_1aaa stbir__simdf_1aaa
 2201+  #define stbir__simdfX_a1a1 stbir__simdf_a1a1
 2202+  #define stbir__simdfX_1a1a stbir__simdf_1a1a
 2203+  #define stbir__simdfX_convert_float_to_i32 stbir__simdf_convert_float_to_i32
 2204+  #define stbir__simdfX_pack_to_words stbir__simdf_pack_to_8words
 2205+  #define stbir__simdfX_zero stbir__simdf_zero
 2206+  #define STBIR_onesX STBIR__CONSTF(STBIR_ones)
 2207+  #define STBIR_simd_point5X STBIR__CONSTF(STBIR_simd_point5)
 2208+  #define STBIR_max_uint8_as_floatX STBIR__CONSTF(STBIR_max_uint8_as_float)
 2209+  #define STBIR_max_uint16_as_floatX STBIR__CONSTF(STBIR_max_uint16_as_float)
 2210+  #define stbir__simdfX_float_count 4
 2211+  #define stbir__if_simdf8_cast_to_simdf4( val ) ( val )
 2212+  #define stbir__simdfX_0123to1230 stbir__simdf_0123to1230
 2213+  #define stbir__simdfX_0123to2103 stbir__simdf_0123to2103
 2214+#endif
 2215+
 2216+
 2217+#if defined(STBIR_NEON) && !defined(_M_ARM) && !defined(__arm__)
 2218+
 2219+  #if defined( _MSC_VER ) && !defined(__clang__)
 2220+  typedef __int16 stbir__FP16;
 2221+  #else
 2222+  typedef float16_t stbir__FP16;
 2223+  #endif
 2224+
 2225+#else // no NEON, or 32-bit ARM for MSVC
 2226+
 2227+  typedef union stbir__FP16
 2228+  {
 2229+    unsigned short u;
 2230+  } stbir__FP16;
 2231+
 2232+#endif
 2233+
 2234+#if (!defined(STBIR_NEON) && !defined(STBIR_FP16C)) || (defined(STBIR_NEON) && defined(_M_ARM)) || (defined(STBIR_NEON) && defined(__arm__))
 2235+
 2236+  // Fabian's half float routines, see: https://gist.github.com/rygorous/2156668
 2237+
 2238+  static stbir__inline float stbir__half_to_float( stbir__FP16 h )
 2239+  {
 2240+    static const stbir__FP32 magic = { (254 - 15) << 23 };
 2241+    static const stbir__FP32 was_infnan = { (127 + 16) << 23 };
 2242+    stbir__FP32 o;
 2243+
 2244+    o.u = (h.u & 0x7fff) << 13;     // exponent/mantissa bits
 2245+    o.f *= magic.f;                 // exponent adjust
 2246+    if (o.f >= was_infnan.f)        // make sure Inf/NaN survive
 2247+      o.u |= 255 << 23;
 2248+    o.u |= (h.u & 0x8000) << 16;    // sign bit
 2249+    return o.f;
 2250+  }
 2251+
 2252+  static stbir__inline stbir__FP16 stbir__float_to_half(float val)
 2253+  {
 2254+    stbir__FP32 f32infty = { 255 << 23 };
 2255+    stbir__FP32 f16max   = { (127 + 16) << 23 };
 2256+    stbir__FP32 denorm_magic = { ((127 - 15) + (23 - 10) + 1) << 23 };
 2257+    unsigned int sign_mask = 0x80000000u;
 2258+    stbir__FP16 o = { 0 };
 2259+    stbir__FP32 f;
 2260+    unsigned int sign;
 2261+
 2262+    f.f = val;
 2263+    sign = f.u & sign_mask;
 2264+    f.u ^= sign;
 2265+
 2266+    if (f.u >= f16max.u) // result is Inf or NaN (all exponent bits set)
 2267+      o.u = (f.u > f32infty.u) ? 0x7e00 : 0x7c00; // NaN->qNaN and Inf->Inf
 2268+    else // (De)normalized number or zero
 2269+    {
 2270+      if (f.u < (113 << 23)) // resulting FP16 is subnormal or zero
 2271+      {
 2272+        // use a magic value to align our 10 mantissa bits at the bottom of
 2273+        // the float. as long as FP addition is round-to-nearest-even this
 2274+        // just works.
 2275+        f.f += denorm_magic.f;
 2276+        // and one integer subtract of the bias later, we have our final float!
 2277+        o.u = (unsigned short) ( f.u - denorm_magic.u );
 2278+      }
 2279+      else
 2280+      {
 2281+        unsigned int mant_odd = (f.u >> 13) & 1; // resulting mantissa is odd
 2282+        // update exponent, rounding bias part 1
 2283+        f.u = f.u + ((15u - 127) << 23) + 0xfff;
 2284+        // rounding bias part 2
 2285+        f.u += mant_odd;
 2286+        // take the bits!
 2287+        o.u = (unsigned short) ( f.u >> 13 );
 2288+      }
 2289+    }
 2290+
 2291+    o.u |= sign >> 16;
 2292+    return o;
 2293+  }
 2294+
 2295+#endif
 2296+
 2297+
 2298+#if defined(STBIR_FP16C)
 2299+
 2300+  #include <immintrin.h>
 2301+
 2302+  static stbir__inline void stbir__half_to_float_SIMD(float * output, stbir__FP16 const * input)
 2303+  {
 2304+    _mm256_storeu_ps( (float*)output, _mm256_cvtph_ps( _mm_loadu_si128( (__m128i const* )input ) ) );
 2305+  }
 2306+
 2307+  static stbir__inline void stbir__float_to_half_SIMD(stbir__FP16 * output, float const * input)
 2308+  {
 2309+    _mm_storeu_si128( (__m128i*)output, _mm256_cvtps_ph( _mm256_loadu_ps( input ), 0 ) );
 2310+  }
 2311+
 2312+  static stbir__inline float stbir__half_to_float( stbir__FP16 h )
 2313+  {
 2314+    return _mm_cvtss_f32( _mm_cvtph_ps( _mm_cvtsi32_si128( (int)h.u ) ) );
 2315+  }
 2316+
 2317+  static stbir__inline stbir__FP16 stbir__float_to_half( float f )
 2318+  {
 2319+    stbir__FP16 h;
 2320+    h.u = (unsigned short) _mm_cvtsi128_si32( _mm_cvtps_ph( _mm_set_ss( f ), 0 ) );
 2321+    return h;
 2322+  }
 2323+
 2324+#elif defined(STBIR_SSE2)
 2325+
 2326+  // Fabian's half float routines, see: https://gist.github.com/rygorous/2156668
 2327+  stbir__inline static void stbir__half_to_float_SIMD(float * output, void const * input)
 2328+  {
 2329+    static const STBIR__SIMDI_CONST(mask_nosign,      0x7fff);
 2330+    static const STBIR__SIMDI_CONST(smallest_normal,  0x0400);
 2331+    static const STBIR__SIMDI_CONST(infinity,         0x7c00);
 2332+    static const STBIR__SIMDI_CONST(expadjust_normal, (127 - 15) << 23);
 2333+    static const STBIR__SIMDI_CONST(magic_denorm,     113 << 23);
 2334+
 2335+    __m128i i = _mm_loadu_si128 ( (__m128i const*)(input) );
 2336+    __m128i h = _mm_unpacklo_epi16 ( i, _mm_setzero_si128() );
 2337+    __m128i mnosign     = STBIR__CONSTI(mask_nosign);
 2338+    __m128i eadjust     = STBIR__CONSTI(expadjust_normal);
 2339+    __m128i smallest    = STBIR__CONSTI(smallest_normal);
 2340+    __m128i infty       = STBIR__CONSTI(infinity);
 2341+    __m128i expmant     = _mm_and_si128(mnosign, h);
 2342+    __m128i justsign    = _mm_xor_si128(h, expmant);
 2343+    __m128i b_notinfnan = _mm_cmpgt_epi32(infty, expmant);
 2344+    __m128i b_isdenorm  = _mm_cmpgt_epi32(smallest, expmant);
 2345+    __m128i shifted     = _mm_slli_epi32(expmant, 13);
 2346+    __m128i adj_infnan  = _mm_andnot_si128(b_notinfnan, eadjust);
 2347+    __m128i adjusted    = _mm_add_epi32(eadjust, shifted);
 2348+    __m128i den1        = _mm_add_epi32(shifted, STBIR__CONSTI(magic_denorm));
 2349+    __m128i adjusted2   = _mm_add_epi32(adjusted, adj_infnan);
 2350+    __m128  den2        = _mm_sub_ps(_mm_castsi128_ps(den1), *(const __m128 *)&magic_denorm);
 2351+    __m128  adjusted3   = _mm_and_ps(den2, _mm_castsi128_ps(b_isdenorm));
 2352+    __m128  adjusted4   = _mm_andnot_ps(_mm_castsi128_ps(b_isdenorm), _mm_castsi128_ps(adjusted2));
 2353+    __m128  adjusted5   = _mm_or_ps(adjusted3, adjusted4);
 2354+    __m128i sign        = _mm_slli_epi32(justsign, 16);
 2355+    __m128  final       = _mm_or_ps(adjusted5, _mm_castsi128_ps(sign));
 2356+    stbir__simdf_store( output + 0,  final );
 2357+
 2358+    h = _mm_unpackhi_epi16 ( i, _mm_setzero_si128() );
 2359+    expmant     = _mm_and_si128(mnosign, h);
 2360+    justsign    = _mm_xor_si128(h, expmant);
 2361+    b_notinfnan = _mm_cmpgt_epi32(infty, expmant);
 2362+    b_isdenorm  = _mm_cmpgt_epi32(smallest, expmant);
 2363+    shifted     = _mm_slli_epi32(expmant, 13);
 2364+    adj_infnan  = _mm_andnot_si128(b_notinfnan, eadjust);
 2365+    adjusted    = _mm_add_epi32(eadjust, shifted);
 2366+    den1        = _mm_add_epi32(shifted, STBIR__CONSTI(magic_denorm));
 2367+    adjusted2   = _mm_add_epi32(adjusted, adj_infnan);
 2368+    den2        = _mm_sub_ps(_mm_castsi128_ps(den1), *(const __m128 *)&magic_denorm);
 2369+    adjusted3   = _mm_and_ps(den2, _mm_castsi128_ps(b_isdenorm));
 2370+    adjusted4   = _mm_andnot_ps(_mm_castsi128_ps(b_isdenorm), _mm_castsi128_ps(adjusted2));
 2371+    adjusted5   = _mm_or_ps(adjusted3, adjusted4);
 2372+    sign        = _mm_slli_epi32(justsign, 16);
 2373+    final       = _mm_or_ps(adjusted5, _mm_castsi128_ps(sign));
 2374+    stbir__simdf_store( output + 4,  final );
 2375+
 2376+    // ~38 SSE2 ops for 8 values
 2377+  }
 2378+
 2379+  // Fabian's round-to-nearest-even float to half
 2380+  // ~48 SSE2 ops for 8 output
 2381+  stbir__inline static void stbir__float_to_half_SIMD(void * output, float const * input)
 2382+  {
 2383+    static const STBIR__SIMDI_CONST(mask_sign,      0x80000000u);
 2384+    static const STBIR__SIMDI_CONST(c_f16max,       (127 + 16) << 23); // all FP32 values >=this round to +inf
 2385+    static const STBIR__SIMDI_CONST(c_nanbit,        0x200);
 2386+    static const STBIR__SIMDI_CONST(c_infty_as_fp16, 0x7c00);
 2387+    static const STBIR__SIMDI_CONST(c_min_normal,    (127 - 14) << 23); // smallest FP32 that yields a normalized FP16
 2388+    static const STBIR__SIMDI_CONST(c_subnorm_magic, ((127 - 15) + (23 - 10) + 1) << 23);
 2389+    static const STBIR__SIMDI_CONST(c_normal_bias,    0xfff - ((127 - 15) << 23)); // adjust exponent and add mantissa rounding
 2390+
 2391+    __m128  f           =  _mm_loadu_ps(input);
 2392+    __m128  msign       = _mm_castsi128_ps(STBIR__CONSTI(mask_sign));
 2393+    __m128  justsign    = _mm_and_ps(msign, f);
 2394+    __m128  absf        = _mm_xor_ps(f, justsign);
 2395+    __m128i absf_int    = _mm_castps_si128(absf); // the cast is "free" (extra bypass latency, but no thruput hit)
 2396+    __m128i f16max      = STBIR__CONSTI(c_f16max);
 2397+    __m128  b_isnan     = _mm_cmpunord_ps(absf, absf); // is this a NaN?
 2398+    __m128i b_isregular = _mm_cmpgt_epi32(f16max, absf_int); // (sub)normalized or special?
 2399+    __m128i nanbit      = _mm_and_si128(_mm_castps_si128(b_isnan), STBIR__CONSTI(c_nanbit));
 2400+    __m128i inf_or_nan  = _mm_or_si128(nanbit, STBIR__CONSTI(c_infty_as_fp16)); // output for specials
 2401+
 2402+    __m128i min_normal  = STBIR__CONSTI(c_min_normal);
 2403+    __m128i b_issub     = _mm_cmpgt_epi32(min_normal, absf_int);
 2404+
 2405+    // "result is subnormal" path
 2406+    __m128  subnorm1    = _mm_add_ps(absf, _mm_castsi128_ps(STBIR__CONSTI(c_subnorm_magic))); // magic value to round output mantissa
 2407+    __m128i subnorm2    = _mm_sub_epi32(_mm_castps_si128(subnorm1), STBIR__CONSTI(c_subnorm_magic)); // subtract out bias
 2408+
 2409+    // "result is normal" path
 2410+    __m128i mantoddbit  = _mm_slli_epi32(absf_int, 31 - 13); // shift bit 13 (mantissa LSB) to sign
 2411+    __m128i mantodd     = _mm_srai_epi32(mantoddbit, 31); // -1 if FP16 mantissa odd, else 0
 2412+
 2413+    __m128i round1      = _mm_add_epi32(absf_int, STBIR__CONSTI(c_normal_bias));
 2414+    __m128i round2      = _mm_sub_epi32(round1, mantodd); // if mantissa LSB odd, bias towards rounding up (RTNE)
 2415+    __m128i normal      = _mm_srli_epi32(round2, 13); // rounded result
 2416+
 2417+    // combine the two non-specials
 2418+    __m128i nonspecial  = _mm_or_si128(_mm_and_si128(subnorm2, b_issub), _mm_andnot_si128(b_issub, normal));
 2419+
 2420+    // merge in specials as well
 2421+    __m128i joined      = _mm_or_si128(_mm_and_si128(nonspecial, b_isregular), _mm_andnot_si128(b_isregular, inf_or_nan));
 2422+
 2423+    __m128i sign_shift  = _mm_srai_epi32(_mm_castps_si128(justsign), 16);
 2424+    __m128i final2, final= _mm_or_si128(joined, sign_shift);
 2425+
 2426+    f           =  _mm_loadu_ps(input+4);
 2427+    justsign    = _mm_and_ps(msign, f);
 2428+    absf        = _mm_xor_ps(f, justsign);
 2429+    absf_int    = _mm_castps_si128(absf); // the cast is "free" (extra bypass latency, but no thruput hit)
 2430+    b_isnan     = _mm_cmpunord_ps(absf, absf); // is this a NaN?
 2431+    b_isregular = _mm_cmpgt_epi32(f16max, absf_int); // (sub)normalized or special?
 2432+    nanbit      = _mm_and_si128(_mm_castps_si128(b_isnan), c_nanbit);
 2433+    inf_or_nan  = _mm_or_si128(nanbit, STBIR__CONSTI(c_infty_as_fp16)); // output for specials
 2434+
 2435+    b_issub     = _mm_cmpgt_epi32(min_normal, absf_int);
 2436+
 2437+    // "result is subnormal" path
 2438+    subnorm1    = _mm_add_ps(absf, _mm_castsi128_ps(STBIR__CONSTI(c_subnorm_magic))); // magic value to round output mantissa
 2439+    subnorm2    = _mm_sub_epi32(_mm_castps_si128(subnorm1), STBIR__CONSTI(c_subnorm_magic)); // subtract out bias
 2440+
 2441+    // "result is normal" path
 2442+    mantoddbit  = _mm_slli_epi32(absf_int, 31 - 13); // shift bit 13 (mantissa LSB) to sign
 2443+    mantodd     = _mm_srai_epi32(mantoddbit, 31); // -1 if FP16 mantissa odd, else 0
 2444+
 2445+    round1      = _mm_add_epi32(absf_int, STBIR__CONSTI(c_normal_bias));
 2446+    round2      = _mm_sub_epi32(round1, mantodd); // if mantissa LSB odd, bias towards rounding up (RTNE)
 2447+    normal      = _mm_srli_epi32(round2, 13); // rounded result
 2448+
 2449+    // combine the two non-specials
 2450+    nonspecial  = _mm_or_si128(_mm_and_si128(subnorm2, b_issub), _mm_andnot_si128(b_issub, normal));
 2451+
 2452+    // merge in specials as well
 2453+    joined      = _mm_or_si128(_mm_and_si128(nonspecial, b_isregular), _mm_andnot_si128(b_isregular, inf_or_nan));
 2454+
 2455+    sign_shift  = _mm_srai_epi32(_mm_castps_si128(justsign), 16);
 2456+    final2      = _mm_or_si128(joined, sign_shift);
 2457+    final       = _mm_packs_epi32(final, final2);
 2458+    stbir__simdi_store( output,final );
 2459+  }
 2460+
 2461+#elif defined(STBIR_NEON) && defined(_MSC_VER) && defined(_M_ARM64) && !defined(__clang__) // 64-bit ARM on MSVC (not clang)
 2462+
 2463+  static stbir__inline void stbir__half_to_float_SIMD(float * output, stbir__FP16 const * input)
 2464+  {
 2465+    float16x4_t in0 = vld1_f16(input + 0);
 2466+    float16x4_t in1 = vld1_f16(input + 4);
 2467+    vst1q_f32(output + 0, vcvt_f32_f16(in0));
 2468+    vst1q_f32(output + 4, vcvt_f32_f16(in1));
 2469+  }
 2470+
 2471+  static stbir__inline void stbir__float_to_half_SIMD(stbir__FP16 * output, float const * input)
 2472+  {
 2473+    float16x4_t out0 = vcvt_f16_f32(vld1q_f32(input + 0));
 2474+    float16x4_t out1 = vcvt_f16_f32(vld1q_f32(input + 4));
 2475+    vst1_f16(output+0, out0);
 2476+    vst1_f16(output+4, out1);
 2477+  }
 2478+
 2479+  static stbir__inline float stbir__half_to_float( stbir__FP16 h )
 2480+  {
 2481+    return vgetq_lane_f32(vcvt_f32_f16(vld1_dup_f16(&h)), 0);
 2482+  }
 2483+
 2484+  static stbir__inline stbir__FP16 stbir__float_to_half( float f )
 2485+  {
 2486+    return vget_lane_f16(vcvt_f16_f32(vdupq_n_f32(f)), 0).n16_u16[0];
 2487+  }
 2488+
 2489+#elif defined(STBIR_NEON) && ( defined( _M_ARM64 ) || defined( __aarch64__ ) || defined( __arm64__ ) ) // 64-bit ARM
 2490+
 2491+  static stbir__inline void stbir__half_to_float_SIMD(float * output, stbir__FP16 const * input)
 2492+  {
 2493+    float16x8_t in = vld1q_f16(input);
 2494+    vst1q_f32(output + 0, vcvt_f32_f16(vget_low_f16(in)));
 2495+    vst1q_f32(output + 4, vcvt_f32_f16(vget_high_f16(in)));
 2496+  }
 2497+
 2498+  static stbir__inline void stbir__float_to_half_SIMD(stbir__FP16 * output, float const * input)
 2499+  {
 2500+    float16x4_t out0 = vcvt_f16_f32(vld1q_f32(input + 0));
 2501+    float16x4_t out1 = vcvt_f16_f32(vld1q_f32(input + 4));
 2502+    vst1q_f16(output, vcombine_f16(out0, out1));
 2503+  }
 2504+
 2505+  static stbir__inline float stbir__half_to_float( stbir__FP16 h )
 2506+  {
 2507+    return vgetq_lane_f32(vcvt_f32_f16(vdup_n_f16(h)), 0);
 2508+  }
 2509+
 2510+  static stbir__inline stbir__FP16 stbir__float_to_half( float f )
 2511+  {
 2512+    return vget_lane_f16(vcvt_f16_f32(vdupq_n_f32(f)), 0);
 2513+  }
 2514+
 2515+#elif defined(STBIR_WASM) || (defined(STBIR_NEON) && (defined(_MSC_VER) || defined(_M_ARM) || defined(__arm__))) // WASM or 32-bit ARM on MSVC/clang
 2516+
 2517+  static stbir__inline void stbir__half_to_float_SIMD(float * output, stbir__FP16 const * input)
 2518+  {
 2519+    for (int i=0; i<8; i++)
 2520+    {
 2521+      output[i] = stbir__half_to_float(input[i]);
 2522+    }
 2523+  }
 2524+  static stbir__inline void stbir__float_to_half_SIMD(stbir__FP16 * output, float const * input)
 2525+  {
 2526+    for (int i=0; i<8; i++)
 2527+    {
 2528+      output[i] = stbir__float_to_half(input[i]);
 2529+    }
 2530+  }
 2531+
 2532+#endif
 2533+
 2534+
 2535+#ifdef STBIR_SIMD
 2536+
 2537+#define stbir__simdf_0123to3333( out, reg ) (out) = stbir__simdf_swiz( reg, 3,3,3,3 )
 2538+#define stbir__simdf_0123to2222( out, reg ) (out) = stbir__simdf_swiz( reg, 2,2,2,2 )
 2539+#define stbir__simdf_0123to1111( out, reg ) (out) = stbir__simdf_swiz( reg, 1,1,1,1 )
 2540+#define stbir__simdf_0123to0000( out, reg ) (out) = stbir__simdf_swiz( reg, 0,0,0,0 )
 2541+#define stbir__simdf_0123to0003( out, reg ) (out) = stbir__simdf_swiz( reg, 0,0,0,3 )
 2542+#define stbir__simdf_0123to0001( out, reg ) (out) = stbir__simdf_swiz( reg, 0,0,0,1 )
 2543+#define stbir__simdf_0123to1122( out, reg ) (out) = stbir__simdf_swiz( reg, 1,1,2,2 )
 2544+#define stbir__simdf_0123to2333( out, reg ) (out) = stbir__simdf_swiz( reg, 2,3,3,3 )
 2545+#define stbir__simdf_0123to0023( out, reg ) (out) = stbir__simdf_swiz( reg, 0,0,2,3 )
 2546+#define stbir__simdf_0123to1230( out, reg ) (out) = stbir__simdf_swiz( reg, 1,2,3,0 )
 2547+#define stbir__simdf_0123to2103( out, reg ) (out) = stbir__simdf_swiz( reg, 2,1,0,3 )
 2548+#define stbir__simdf_0123to3210( out, reg ) (out) = stbir__simdf_swiz( reg, 3,2,1,0 )
 2549+#define stbir__simdf_0123to2301( out, reg ) (out) = stbir__simdf_swiz( reg, 2,3,0,1 )
 2550+#define stbir__simdf_0123to3012( out, reg ) (out) = stbir__simdf_swiz( reg, 3,0,1,2 )
 2551+#define stbir__simdf_0123to0011( out, reg ) (out) = stbir__simdf_swiz( reg, 0,0,1,1 )
 2552+#define stbir__simdf_0123to1100( out, reg ) (out) = stbir__simdf_swiz( reg, 1,1,0,0 )
 2553+#define stbir__simdf_0123to2233( out, reg ) (out) = stbir__simdf_swiz( reg, 2,2,3,3 )
 2554+#define stbir__simdf_0123to1133( out, reg ) (out) = stbir__simdf_swiz( reg, 1,1,3,3 )
 2555+#define stbir__simdf_0123to0022( out, reg ) (out) = stbir__simdf_swiz( reg, 0,0,2,2 )
 2556+#define stbir__simdf_0123to1032( out, reg ) (out) = stbir__simdf_swiz( reg, 1,0,3,2 )
 2557+
 2558+typedef union stbir__simdi_u32
 2559+{
 2560+  stbir_uint32 m128i_u32[4];
 2561+  int m128i_i32[4];
 2562+  stbir__simdi m128i_i128;
 2563+} stbir__simdi_u32;
 2564+
 2565+static const int STBIR_mask[9] = { 0,0,0,-1,-1,-1,0,0,0 };
 2566+
 2567+static const STBIR__SIMDF_CONST(STBIR_max_uint8_as_float,           stbir__max_uint8_as_float);
 2568+static const STBIR__SIMDF_CONST(STBIR_max_uint16_as_float,          stbir__max_uint16_as_float);
 2569+static const STBIR__SIMDF_CONST(STBIR_max_uint8_as_float_inverted,  stbir__max_uint8_as_float_inverted);
 2570+static const STBIR__SIMDF_CONST(STBIR_max_uint16_as_float_inverted, stbir__max_uint16_as_float_inverted);
 2571+
 2572+static const STBIR__SIMDF_CONST(STBIR_simd_point5,   0.5f);
 2573+static const STBIR__SIMDF_CONST(STBIR_ones,          1.0f);
 2574+static const STBIR__SIMDI_CONST(STBIR_almost_zero,   (127 - 13) << 23);
 2575+static const STBIR__SIMDI_CONST(STBIR_almost_one,    0x3f7fffff);
 2576+static const STBIR__SIMDI_CONST(STBIR_mantissa_mask, 0xff);
 2577+static const STBIR__SIMDI_CONST(STBIR_topscale,      0x02000000);
 2578+
 2579+//   Basically, in simd mode, we unroll the proper amount, and we don't want
 2580+//   the non-simd remnant loops to be unroll because they only run a few times
 2581+//   Adding this switch saves about 5K on clang which is Captain Unroll the 3rd.
 2582+#define STBIR_SIMD_STREAMOUT_PTR( star )  STBIR_STREAMOUT_PTR( star )
 2583+#define STBIR_SIMD_NO_UNROLL(ptr) STBIR_NO_UNROLL(ptr)
 2584+#define STBIR_SIMD_NO_UNROLL_LOOP_START STBIR_NO_UNROLL_LOOP_START
 2585+#define STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR STBIR_NO_UNROLL_LOOP_START_INF_FOR
 2586+
 2587+#ifdef STBIR_MEMCPY
 2588+#undef STBIR_MEMCPY
 2589+#endif
 2590+#define STBIR_MEMCPY stbir_simd_memcpy
 2591+
 2592+// override normal use of memcpy with much simpler copy (faster and smaller with our sized copies)
 2593+static void stbir_simd_memcpy( void * dest, void const * src, size_t bytes )
 2594+{
 2595+  char STBIR_SIMD_STREAMOUT_PTR (*) d = (char*) dest;
 2596+  char STBIR_SIMD_STREAMOUT_PTR( * ) d_end = ((char*) dest) + bytes;
 2597+  ptrdiff_t ofs_to_src = (char*)src - (char*)dest;
 2598+
 2599+  // check overlaps
 2600+  STBIR_ASSERT( ( ( d >= ( (char*)src) + bytes ) ) || ( ( d + bytes ) <= (char*)src ) );
 2601+
 2602+  if ( bytes < (16*stbir__simdfX_float_count) )
 2603+  {
 2604+    if ( bytes < 16 )
 2605+    {
 2606+      if ( bytes )
 2607+      {
 2608+        STBIR_SIMD_NO_UNROLL_LOOP_START
 2609+        do
 2610+        {
 2611+          STBIR_SIMD_NO_UNROLL(d);
 2612+          d[ 0 ] = d[ ofs_to_src ];
 2613+          ++d;
 2614+        } while ( d < d_end );
 2615+      }
 2616+    }
 2617+    else
 2618+    {
 2619+      stbir__simdf x;
 2620+      // do one unaligned to get us aligned for the stream out below
 2621+      stbir__simdf_load( x, ( d + ofs_to_src ) );
 2622+      stbir__simdf_store( d, x );
 2623+      d = (char*)( ( ( (size_t)d ) + 16 ) & ~15 );
 2624+
 2625+      STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 2626+      for(;;)
 2627+      {
 2628+        STBIR_SIMD_NO_UNROLL(d);
 2629+
 2630+        if ( d > ( d_end - 16 ) )
 2631+        {
 2632+          if ( d == d_end )
 2633+            return;
 2634+          d = d_end - 16;
 2635+        }
 2636+
 2637+        stbir__simdf_load( x, ( d + ofs_to_src ) );
 2638+        stbir__simdf_store( d, x );
 2639+        d += 16;
 2640+      }
 2641+    }
 2642+  }
 2643+  else
 2644+  {
 2645+    stbir__simdfX x0,x1,x2,x3;
 2646+
 2647+    // do one unaligned to get us aligned for the stream out below
 2648+    stbir__simdfX_load( x0, ( d + ofs_to_src ) +  0*stbir__simdfX_float_count );
 2649+    stbir__simdfX_load( x1, ( d + ofs_to_src ) +  4*stbir__simdfX_float_count );
 2650+    stbir__simdfX_load( x2, ( d + ofs_to_src ) +  8*stbir__simdfX_float_count );
 2651+    stbir__simdfX_load( x3, ( d + ofs_to_src ) + 12*stbir__simdfX_float_count );
 2652+    stbir__simdfX_store( d +  0*stbir__simdfX_float_count, x0 );
 2653+    stbir__simdfX_store( d +  4*stbir__simdfX_float_count, x1 );
 2654+    stbir__simdfX_store( d +  8*stbir__simdfX_float_count, x2 );
 2655+    stbir__simdfX_store( d + 12*stbir__simdfX_float_count, x3 );
 2656+    d = (char*)( ( ( (size_t)d ) + (16*stbir__simdfX_float_count) ) & ~((16*stbir__simdfX_float_count)-1) );
 2657+
 2658+    STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 2659+    for(;;)
 2660+    {
 2661+      STBIR_SIMD_NO_UNROLL(d);
 2662+
 2663+      if ( d > ( d_end - (16*stbir__simdfX_float_count) ) )
 2664+      {
 2665+        if ( d == d_end )
 2666+          return;
 2667+        d = d_end - (16*stbir__simdfX_float_count);
 2668+      }
 2669+
 2670+      stbir__simdfX_load( x0, ( d + ofs_to_src ) +  0*stbir__simdfX_float_count );
 2671+      stbir__simdfX_load( x1, ( d + ofs_to_src ) +  4*stbir__simdfX_float_count );
 2672+      stbir__simdfX_load( x2, ( d + ofs_to_src ) +  8*stbir__simdfX_float_count );
 2673+      stbir__simdfX_load( x3, ( d + ofs_to_src ) + 12*stbir__simdfX_float_count );
 2674+      stbir__simdfX_store( d +  0*stbir__simdfX_float_count, x0 );
 2675+      stbir__simdfX_store( d +  4*stbir__simdfX_float_count, x1 );
 2676+      stbir__simdfX_store( d +  8*stbir__simdfX_float_count, x2 );
 2677+      stbir__simdfX_store( d + 12*stbir__simdfX_float_count, x3 );
 2678+      d += (16*stbir__simdfX_float_count);
 2679+    }
 2680+  }
 2681+}
 2682+
 2683+// memcpy that is specically intentionally overlapping (src is smaller then dest, so can be
 2684+//   a normal forward copy, bytes is divisible by 4 and bytes is greater than or equal to
 2685+//   the diff between dest and src)
 2686+static void stbir_overlapping_memcpy( void * dest, void const * src, size_t bytes )
 2687+{
 2688+  char STBIR_SIMD_STREAMOUT_PTR (*) sd = (char*) src;
 2689+  char STBIR_SIMD_STREAMOUT_PTR( * ) s_end = ((char*) src) + bytes;
 2690+  ptrdiff_t ofs_to_dest = (char*)dest - (char*)src;
 2691+
 2692+  if ( ofs_to_dest >= 16 ) // is the overlap more than 16 away?
 2693+  {
 2694+    char STBIR_SIMD_STREAMOUT_PTR( * ) s_end16 = ((char*) src) + (bytes&~15);
 2695+    STBIR_SIMD_NO_UNROLL_LOOP_START
 2696+    do
 2697+    {
 2698+      stbir__simdf x;
 2699+      STBIR_SIMD_NO_UNROLL(sd);
 2700+      stbir__simdf_load( x, sd );
 2701+      stbir__simdf_store(  ( sd + ofs_to_dest ), x );
 2702+      sd += 16;
 2703+    } while ( sd < s_end16 );
 2704+
 2705+    if ( sd == s_end )
 2706+      return;
 2707+  }
 2708+
 2709+  do
 2710+  {
 2711+    STBIR_SIMD_NO_UNROLL(sd);
 2712+    *(int*)( sd + ofs_to_dest ) = *(int*) sd;
 2713+    sd += 4;
 2714+  } while ( sd < s_end );
 2715+}
 2716+
 2717+#else // no SSE2
 2718+
 2719+// when in scalar mode, we let unrolling happen, so this macro just does the __restrict
 2720+#define STBIR_SIMD_STREAMOUT_PTR( star ) STBIR_STREAMOUT_PTR( star )
 2721+#define STBIR_SIMD_NO_UNROLL(ptr)
 2722+#define STBIR_SIMD_NO_UNROLL_LOOP_START
 2723+#define STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 2724+
 2725+#endif // SSE2
 2726+
 2727+
 2728+#ifdef STBIR_PROFILE
 2729+
 2730+#ifndef STBIR_PROFILE_FUNC
 2731+
 2732+#if defined(_x86_64) || defined( __x86_64__ ) || defined( _M_X64 ) || defined(__x86_64) || defined(__SSE2__) || defined(STBIR_SSE) || defined( _M_IX86_FP ) || defined(__i386) || defined( __i386__ ) || defined( _M_IX86 ) || defined( _X86_ )
 2733+
 2734+#ifdef _MSC_VER
 2735+
 2736+  STBIRDEF stbir_uint64 __rdtsc();
 2737+  #define STBIR_PROFILE_FUNC() __rdtsc()
 2738+
 2739+#else // non msvc
 2740+
 2741+  static stbir__inline stbir_uint64 STBIR_PROFILE_FUNC()
 2742+  {
 2743+    stbir_uint32 lo, hi;
 2744+    asm volatile ("rdtsc" : "=a" (lo), "=d" (hi) );
 2745+    return ( ( (stbir_uint64) hi ) << 32 ) | ( (stbir_uint64) lo );
 2746+  }
 2747+
 2748+#endif  // msvc
 2749+
 2750+#elif defined( _M_ARM64 ) || defined( __aarch64__ ) || defined( __arm64__ ) || defined(__ARM_NEON__)
 2751+
 2752+#if defined( _MSC_VER ) && !defined(__clang__)
 2753+
 2754+  #define STBIR_PROFILE_FUNC() _ReadStatusReg(ARM64_CNTVCT)
 2755+
 2756+#else
 2757+
 2758+  static stbir__inline stbir_uint64 STBIR_PROFILE_FUNC()
 2759+  {
 2760+    stbir_uint64 tsc;
 2761+    asm volatile("mrs %0, cntvct_el0" : "=r" (tsc));
 2762+    return tsc;
 2763+  }
 2764+
 2765+#endif
 2766+
 2767+#else // x64, arm
 2768+
 2769+#error Unknown platform for profiling.
 2770+
 2771+#endif  // x64, arm
 2772+
 2773+#endif // STBIR_PROFILE_FUNC
 2774+
 2775+#define STBIR_ONLY_PROFILE_GET_SPLIT_INFO ,stbir__per_split_info * split_info
 2776+#define STBIR_ONLY_PROFILE_SET_SPLIT_INFO ,split_info
 2777+
 2778+#define STBIR_ONLY_PROFILE_BUILD_GET_INFO ,stbir__info * profile_info
 2779+#define STBIR_ONLY_PROFILE_BUILD_SET_INFO ,profile_info
 2780+
 2781+// super light-weight micro profiler
 2782+#define STBIR_PROFILE_START_ll( info, wh ) { stbir_uint64 wh##thiszonetime = STBIR_PROFILE_FUNC(); stbir_uint64 * wh##save_parent_excluded_ptr = info->current_zone_excluded_ptr; stbir_uint64 wh##current_zone_excluded = 0; info->current_zone_excluded_ptr = &wh##current_zone_excluded;
 2783+#define STBIR_PROFILE_END_ll( info, wh ) wh##thiszonetime = STBIR_PROFILE_FUNC() - wh##thiszonetime; info->profile.named.wh += wh##thiszonetime - wh##current_zone_excluded; *wh##save_parent_excluded_ptr += wh##thiszonetime; info->current_zone_excluded_ptr = wh##save_parent_excluded_ptr; }
 2784+#define STBIR_PROFILE_FIRST_START_ll( info, wh ) { int i; info->current_zone_excluded_ptr = &info->profile.named.total; for(i=0;i<STBIR__ARRAY_SIZE(info->profile.array);i++) info->profile.array[i]=0; } STBIR_PROFILE_START_ll( info, wh );
 2785+#define STBIR_PROFILE_CLEAR_EXTRAS_ll( info, num ) { int extra; for(extra=1;extra<(num);extra++) { int i; for(i=0;i<STBIR__ARRAY_SIZE((info)->profile.array);i++) (info)[extra].profile.array[i]=0; } }
 2786+
 2787+// for thread data
 2788+#define STBIR_PROFILE_START( wh ) STBIR_PROFILE_START_ll( split_info, wh )
 2789+#define STBIR_PROFILE_END( wh ) STBIR_PROFILE_END_ll( split_info, wh )
 2790+#define STBIR_PROFILE_FIRST_START( wh ) STBIR_PROFILE_FIRST_START_ll( split_info, wh )
 2791+#define STBIR_PROFILE_CLEAR_EXTRAS() STBIR_PROFILE_CLEAR_EXTRAS_ll( split_info, split_count )
 2792+
 2793+// for build data
 2794+#define STBIR_PROFILE_BUILD_START( wh ) STBIR_PROFILE_START_ll( profile_info, wh )
 2795+#define STBIR_PROFILE_BUILD_END( wh ) STBIR_PROFILE_END_ll( profile_info, wh )
 2796+#define STBIR_PROFILE_BUILD_FIRST_START( wh ) STBIR_PROFILE_FIRST_START_ll( profile_info, wh )
 2797+#define STBIR_PROFILE_BUILD_CLEAR( info ) { int i; for(i=0;i<STBIR__ARRAY_SIZE(info->profile.array);i++) info->profile.array[i]=0; }
 2798+
 2799+#else  // no profile
 2800+
 2801+#define STBIR_ONLY_PROFILE_GET_SPLIT_INFO
 2802+#define STBIR_ONLY_PROFILE_SET_SPLIT_INFO
 2803+
 2804+#define STBIR_ONLY_PROFILE_BUILD_GET_INFO
 2805+#define STBIR_ONLY_PROFILE_BUILD_SET_INFO
 2806+
 2807+#define STBIR_PROFILE_START( wh )
 2808+#define STBIR_PROFILE_END( wh )
 2809+#define STBIR_PROFILE_FIRST_START( wh )
 2810+#define STBIR_PROFILE_CLEAR_EXTRAS( )
 2811+
 2812+#define STBIR_PROFILE_BUILD_START( wh )
 2813+#define STBIR_PROFILE_BUILD_END( wh )
 2814+#define STBIR_PROFILE_BUILD_FIRST_START( wh )
 2815+#define STBIR_PROFILE_BUILD_CLEAR( info )
 2816+
 2817+#endif  // stbir_profile
 2818+
 2819+#ifndef STBIR_CEILF
 2820+#include <math.h>
 2821+#if _MSC_VER <= 1200 // support VC6 for Sean
 2822+#define STBIR_CEILF(x) ((float)ceil((float)(x)))
 2823+#define STBIR_FLOORF(x) ((float)floor((float)(x)))
 2824+#else
 2825+#define STBIR_CEILF(x) ceilf(x)
 2826+#define STBIR_FLOORF(x) floorf(x)
 2827+#endif
 2828+#endif
 2829+
 2830+#ifndef STBIR_MEMCPY
 2831+// For memcpy
 2832+#include <string.h>
 2833+#define STBIR_MEMCPY( dest, src, len ) memcpy( dest, src, len )
 2834+#endif
 2835+
 2836+#ifndef STBIR_SIMD
 2837+
 2838+// memcpy that is specifically intentionally overlapping (src is smaller then dest, so can be
 2839+//   a normal forward copy, bytes is divisible by 4 and bytes is greater than or equal to
 2840+//   the diff between dest and src)
 2841+static void stbir_overlapping_memcpy( void * dest, void const * src, size_t bytes )
 2842+{
 2843+  char STBIR_SIMD_STREAMOUT_PTR (*) sd = (char*) src;
 2844+  char STBIR_SIMD_STREAMOUT_PTR( * ) s_end = ((char*) src) + bytes;
 2845+  ptrdiff_t ofs_to_dest = (char*)dest - (char*)src;
 2846+
 2847+  if ( ofs_to_dest >= 8 ) // is the overlap more than 8 away
 2848+  {
 2849+    char STBIR_SIMD_STREAMOUT_PTR( * ) s_end8 = ((char*) src) + (bytes&~7);
 2850+
 2851+    if ( ( ( ((ptrdiff_t)dest)|((ptrdiff_t)src) ) & 7 ) == 0 ) // is it 8byte aligned?
 2852+    {
 2853+      STBIR_NO_UNROLL_LOOP_START
 2854+      do
 2855+      {
 2856+        STBIR_NO_UNROLL(sd);
 2857+        *(stbir_uint64*)( sd + ofs_to_dest ) = *(stbir_uint64*) sd;
 2858+        sd += 8;
 2859+      } while ( sd < s_end8 );
 2860+    }
 2861+    else
 2862+    {
 2863+      STBIR_NO_UNROLL_LOOP_START
 2864+      do
 2865+      {
 2866+        int a,b;
 2867+        STBIR_NO_UNROLL(sd);
 2868+        a = ((int*)sd)[0];
 2869+        b = ((int*)sd)[1];
 2870+        ((int*)( sd + ofs_to_dest ))[0] = a;
 2871+        ((int*)( sd + ofs_to_dest ))[1] = b;
 2872+        sd += 8;
 2873+      } while ( sd < s_end8 );
 2874+    }
 2875+
 2876+    if ( sd == s_end )
 2877+      return;
 2878+  }
 2879+
 2880+  STBIR_NO_UNROLL_LOOP_START
 2881+  do
 2882+  {
 2883+    STBIR_NO_UNROLL(sd);
 2884+    *(int*)( sd + ofs_to_dest ) = *(int*) sd;
 2885+    sd += 4;
 2886+  } while ( sd < s_end );
 2887+}
 2888+
 2889+#endif
 2890+
 2891+static float stbir__filter_trapezoid(float x, float scale, void * user_data)
 2892+{
 2893+  float halfscale = scale / 2;
 2894+  float t = 0.5f + halfscale;
 2895+  STBIR_ASSERT(scale <= 1);
 2896+  STBIR__UNUSED(user_data);
 2897+
 2898+  if ( x < 0.0f ) x = -x;
 2899+
 2900+  if (x >= t)
 2901+    return 0.0f;
 2902+  else
 2903+  {
 2904+    float r = 0.5f - halfscale;
 2905+    if (x <= r)
 2906+      return 1.0f;
 2907+    else
 2908+      return (t - x) / scale;
 2909+  }
 2910+}
 2911+
 2912+static float stbir__support_trapezoid(float scale, void * user_data)
 2913+{
 2914+  STBIR__UNUSED(user_data);
 2915+  return 0.5f + scale / 2.0f;
 2916+}
 2917+
 2918+static float stbir__filter_triangle(float x, float s, void * user_data)
 2919+{
 2920+  STBIR__UNUSED(s);
 2921+  STBIR__UNUSED(user_data);
 2922+
 2923+  if ( x < 0.0f ) x = -x;
 2924+
 2925+  if (x <= 1.0f)
 2926+    return 1.0f - x;
 2927+  else
 2928+    return 0.0f;
 2929+}
 2930+
 2931+static float stbir__filter_point(float x, float s, void * user_data)
 2932+{
 2933+  STBIR__UNUSED(x);
 2934+  STBIR__UNUSED(s);
 2935+  STBIR__UNUSED(user_data);
 2936+
 2937+  return 1.0f;
 2938+}
 2939+
 2940+static float stbir__filter_cubic(float x, float s, void * user_data)
 2941+{
 2942+  STBIR__UNUSED(s);
 2943+  STBIR__UNUSED(user_data);
 2944+
 2945+  if ( x < 0.0f ) x = -x;
 2946+
 2947+  if (x < 1.0f)
 2948+    return (4.0f + x*x*(3.0f*x - 6.0f))/6.0f;
 2949+  else if (x < 2.0f)
 2950+    return (8.0f + x*(-12.0f + x*(6.0f - x)))/6.0f;
 2951+
 2952+  return (0.0f);
 2953+}
 2954+
 2955+static float stbir__filter_catmullrom(float x, float s, void * user_data)
 2956+{
 2957+  STBIR__UNUSED(s);
 2958+  STBIR__UNUSED(user_data);
 2959+
 2960+  if ( x < 0.0f ) x = -x;
 2961+
 2962+  if (x < 1.0f)
 2963+    return 1.0f - x*x*(2.5f - 1.5f*x);
 2964+  else if (x < 2.0f)
 2965+    return 2.0f - x*(4.0f + x*(0.5f*x - 2.5f));
 2966+
 2967+  return (0.0f);
 2968+}
 2969+
 2970+static float stbir__filter_mitchell(float x, float s, void * user_data)
 2971+{
 2972+  STBIR__UNUSED(s);
 2973+  STBIR__UNUSED(user_data);
 2974+
 2975+  if ( x < 0.0f ) x = -x;
 2976+
 2977+  if (x < 1.0f)
 2978+    return (16.0f + x*x*(21.0f * x - 36.0f))/18.0f;
 2979+  else if (x < 2.0f)
 2980+    return (32.0f + x*(-60.0f + x*(36.0f - 7.0f*x)))/18.0f;
 2981+
 2982+  return (0.0f);
 2983+}
 2984+
 2985+static float stbir__support_zeropoint5(float s, void * user_data)
 2986+{
 2987+  STBIR__UNUSED(s);
 2988+  STBIR__UNUSED(user_data);
 2989+  return 0.5f;
 2990+}
 2991+
 2992+static float stbir__support_one(float s, void * user_data)
 2993+{
 2994+  STBIR__UNUSED(s);
 2995+  STBIR__UNUSED(user_data);
 2996+  return 1;
 2997+}
 2998+
 2999+static float stbir__support_two(float s, void * user_data)
 3000+{
 3001+  STBIR__UNUSED(s);
 3002+  STBIR__UNUSED(user_data);
 3003+  return 2;
 3004+}
 3005+
 3006+// This is the maximum number of input samples that can affect an output sample
 3007+// with the given filter from the output pixel's perspective
 3008+static int stbir__get_filter_pixel_width(stbir__support_callback * support, float scale, void * user_data)
 3009+{
 3010+  STBIR_ASSERT(support != 0);
 3011+
 3012+  if ( scale >= ( 1.0f-stbir__small_float ) ) // upscale
 3013+    return (int)STBIR_CEILF(support(1.0f/scale,user_data) * 2.0f);
 3014+  else
 3015+    return (int)STBIR_CEILF(support(scale,user_data) * 2.0f / scale);
 3016+}
 3017+
 3018+// this is how many coefficents per run of the filter (which is different
 3019+//   from the filter_pixel_width depending on if we are scattering or gathering)
 3020+static int stbir__get_coefficient_width(stbir__sampler * samp, int is_gather, void * user_data)
 3021+{
 3022+  float scale = samp->scale_info.scale;
 3023+  stbir__support_callback * support = samp->filter_support;
 3024+
 3025+  switch( is_gather )
 3026+  {
 3027+    case 1:
 3028+      return (int)STBIR_CEILF(support(1.0f / scale, user_data) * 2.0f);
 3029+    case 2:
 3030+      return (int)STBIR_CEILF(support(scale, user_data) * 2.0f / scale);
 3031+    case 0:
 3032+      return (int)STBIR_CEILF(support(scale, user_data) * 2.0f);
 3033+    default:
 3034+      STBIR_ASSERT( (is_gather >= 0 ) && (is_gather <= 2 ) );
 3035+      return 0;
 3036+  }
 3037+}
 3038+
 3039+static int stbir__get_contributors(stbir__sampler * samp, int is_gather)
 3040+{
 3041+  if (is_gather)
 3042+      return samp->scale_info.output_sub_size;
 3043+  else
 3044+      return (samp->scale_info.input_full_size + samp->filter_pixel_margin * 2);
 3045+}
 3046+
 3047+static int stbir__edge_zero_full( int n, int max )
 3048+{
 3049+  STBIR__UNUSED(n);
 3050+  STBIR__UNUSED(max);
 3051+  return 0; // NOTREACHED
 3052+}
 3053+
 3054+static int stbir__edge_clamp_full( int n, int max )
 3055+{
 3056+  if (n < 0)
 3057+    return 0;
 3058+
 3059+  if (n >= max)
 3060+    return max - 1;
 3061+
 3062+  return n; // NOTREACHED
 3063+}
 3064+
 3065+static int stbir__edge_reflect_full( int n, int max )
 3066+{
 3067+  if (n < 0)
 3068+  {
 3069+    if (n > -max)
 3070+      return -n;
 3071+    else
 3072+      return max - 1;
 3073+  }
 3074+
 3075+  if (n >= max)
 3076+  {
 3077+    int max2 = max * 2;
 3078+    if (n >= max2)
 3079+      return 0;
 3080+    else
 3081+      return max2 - n - 1;
 3082+  }
 3083+
 3084+  return n; // NOTREACHED
 3085+}
 3086+
 3087+static int stbir__edge_wrap_full( int n, int max )
 3088+{
 3089+  if (n >= 0)
 3090+    return (n % max);
 3091+  else
 3092+  {
 3093+    int m = (-n) % max;
 3094+
 3095+    if (m != 0)
 3096+      m = max - m;
 3097+
 3098+    return (m);
 3099+  }
 3100+}
 3101+
 3102+typedef int stbir__edge_wrap_func( int n, int max );
 3103+static stbir__edge_wrap_func * stbir__edge_wrap_slow[] =
 3104+{
 3105+  stbir__edge_clamp_full,    // STBIR_EDGE_CLAMP
 3106+  stbir__edge_reflect_full,  // STBIR_EDGE_REFLECT
 3107+  stbir__edge_wrap_full,     // STBIR_EDGE_WRAP
 3108+  stbir__edge_zero_full,     // STBIR_EDGE_ZERO
 3109+};
 3110+
 3111+stbir__inline static int stbir__edge_wrap(stbir_edge edge, int n, int max)
 3112+{
 3113+  // avoid per-pixel switch
 3114+  if (n >= 0 && n < max)
 3115+      return n;
 3116+  return stbir__edge_wrap_slow[edge]( n, max );
 3117+}
 3118+
 3119+#define STBIR__MERGE_RUNS_PIXEL_THRESHOLD 16
 3120+
 3121+// get information on the extents of a sampler
 3122+static void stbir__get_extents( stbir__sampler * samp, stbir__extents * scanline_extents )
 3123+{
 3124+  int j, stop;
 3125+  int left_margin, right_margin;
 3126+  int min_n = 0x7fffffff, max_n = -0x7fffffff;
 3127+  int min_left = 0x7fffffff, max_left = -0x7fffffff;
 3128+  int min_right = 0x7fffffff, max_right = -0x7fffffff;
 3129+  stbir_edge edge = samp->edge;
 3130+  stbir__contributors* contributors = samp->contributors;
 3131+  int output_sub_size = samp->scale_info.output_sub_size;
 3132+  int input_full_size = samp->scale_info.input_full_size;
 3133+  int filter_pixel_margin = samp->filter_pixel_margin;
 3134+
 3135+  STBIR_ASSERT( samp->is_gather );
 3136+
 3137+  stop = output_sub_size;
 3138+  for (j = 0; j < stop; j++ )
 3139+  {
 3140+    STBIR_ASSERT( contributors[j].n1 >= contributors[j].n0 );
 3141+    if ( contributors[j].n0 < min_n )
 3142+    {
 3143+      min_n = contributors[j].n0;
 3144+      stop = j + filter_pixel_margin;  // if we find a new min, only scan another filter width
 3145+      if ( stop > output_sub_size ) stop = output_sub_size;
 3146+    }
 3147+  }
 3148+
 3149+  stop = 0;
 3150+  for (j = output_sub_size - 1; j >= stop; j-- )
 3151+  {
 3152+    STBIR_ASSERT( contributors[j].n1 >= contributors[j].n0 );
 3153+    if ( contributors[j].n1 > max_n )
 3154+    {
 3155+      max_n = contributors[j].n1;
 3156+      stop = j - filter_pixel_margin;  // if we find a new max, only scan another filter width
 3157+      if (stop<0) stop = 0;
 3158+    }
 3159+  }
 3160+
 3161+  STBIR_ASSERT( scanline_extents->conservative.n0 <= min_n );
 3162+  STBIR_ASSERT( scanline_extents->conservative.n1 >= max_n );
 3163+
 3164+  // now calculate how much into the margins we really read
 3165+  left_margin = 0;
 3166+  if ( min_n < 0 )
 3167+  {
 3168+    left_margin = -min_n;
 3169+    min_n = 0;
 3170+  }
 3171+
 3172+  right_margin = 0;
 3173+  if ( max_n >= input_full_size )
 3174+  {
 3175+    right_margin = max_n - input_full_size + 1;
 3176+    max_n = input_full_size - 1;
 3177+  }
 3178+
 3179+  // index 1 is margin pixel extents (how many pixels we hang over the edge)
 3180+  scanline_extents->edge_sizes[0] = left_margin;
 3181+  scanline_extents->edge_sizes[1] = right_margin;
 3182+
 3183+  // index 2 is pixels read from the input
 3184+  scanline_extents->spans[0].n0 = min_n;
 3185+  scanline_extents->spans[0].n1 = max_n;
 3186+  scanline_extents->spans[0].pixel_offset_for_input = min_n;
 3187+
 3188+  // default to no other input range
 3189+  scanline_extents->spans[1].n0 = 0;
 3190+  scanline_extents->spans[1].n1 = -1;
 3191+  scanline_extents->spans[1].pixel_offset_for_input = 0;
 3192+
 3193+  // don't have to do edge calc for zero clamp
 3194+  if ( edge == STBIR_EDGE_ZERO )
 3195+    return;
 3196+
 3197+  // convert margin pixels to the pixels within the input (min and max)
 3198+  for( j = -left_margin ; j < 0 ; j++ )
 3199+  {
 3200+      int p = stbir__edge_wrap( edge, j, input_full_size );
 3201+      if ( p < min_left )
 3202+        min_left = p;
 3203+      if ( p > max_left )
 3204+        max_left = p;
 3205+  }
 3206+
 3207+  for( j = input_full_size ; j < (input_full_size + right_margin) ; j++ )
 3208+  {
 3209+      int p = stbir__edge_wrap( edge, j, input_full_size );
 3210+      if ( p < min_right )
 3211+        min_right = p;
 3212+      if ( p > max_right )
 3213+        max_right = p;
 3214+  }
 3215+
 3216+  // merge the left margin pixel region if it connects within 4 pixels of main pixel region
 3217+  if ( min_left != 0x7fffffff )
 3218+  {
 3219+    if ( ( ( min_left <= min_n ) && ( ( max_left  + STBIR__MERGE_RUNS_PIXEL_THRESHOLD ) >= min_n ) ) ||
 3220+         ( ( min_n <= min_left ) && ( ( max_n  + STBIR__MERGE_RUNS_PIXEL_THRESHOLD ) >= max_left ) ) )
 3221+    {
 3222+      scanline_extents->spans[0].n0 = min_n = stbir__min( min_n, min_left );
 3223+      scanline_extents->spans[0].n1 = max_n = stbir__max( max_n, max_left );
 3224+      scanline_extents->spans[0].pixel_offset_for_input = min_n;
 3225+      left_margin = 0;
 3226+    }
 3227+  }
 3228+
 3229+  // merge the right margin pixel region if it connects within 4 pixels of main pixel region
 3230+  if ( min_right != 0x7fffffff )
 3231+  {
 3232+    if ( ( ( min_right <= min_n ) && ( ( max_right  + STBIR__MERGE_RUNS_PIXEL_THRESHOLD ) >= min_n ) ) ||
 3233+         ( ( min_n <= min_right ) && ( ( max_n  + STBIR__MERGE_RUNS_PIXEL_THRESHOLD ) >= max_right ) ) )
 3234+    {
 3235+      scanline_extents->spans[0].n0 = min_n = stbir__min( min_n, min_right );
 3236+      scanline_extents->spans[0].n1 = max_n = stbir__max( max_n, max_right );
 3237+      scanline_extents->spans[0].pixel_offset_for_input = min_n;
 3238+      right_margin = 0;
 3239+    }
 3240+  }
 3241+
 3242+  STBIR_ASSERT( scanline_extents->conservative.n0 <= min_n );
 3243+  STBIR_ASSERT( scanline_extents->conservative.n1 >= max_n );
 3244+
 3245+  // you get two ranges when you have the WRAP edge mode and you are doing just the a piece of the resize
 3246+  //   so you need to get a second run of pixels from the opposite side of the scanline (which you
 3247+  //   wouldn't need except for WRAP)
 3248+
 3249+
 3250+  // if we can't merge the min_left range, add it as a second range
 3251+  if ( ( left_margin ) && ( min_left != 0x7fffffff ) )
 3252+  {
 3253+    stbir__span * newspan = scanline_extents->spans + 1;
 3254+    STBIR_ASSERT( right_margin == 0 );
 3255+    if ( min_left < scanline_extents->spans[0].n0 )
 3256+    {
 3257+      scanline_extents->spans[1].pixel_offset_for_input = scanline_extents->spans[0].n0;
 3258+      scanline_extents->spans[1].n0 = scanline_extents->spans[0].n0;
 3259+      scanline_extents->spans[1].n1 = scanline_extents->spans[0].n1;
 3260+      --newspan;
 3261+    }
 3262+    newspan->pixel_offset_for_input = min_left;
 3263+    newspan->n0 = -left_margin;
 3264+    newspan->n1 = ( max_left - min_left ) - left_margin;
 3265+    scanline_extents->edge_sizes[0] = 0;  // don't need to copy the left margin, since we are directly decoding into the margin
 3266+  }
 3267+  // if we can't merge the min_right range, add it as a second range
 3268+  else  
 3269+  if ( ( right_margin ) && ( min_right != 0x7fffffff ) )
 3270+  {
 3271+    stbir__span * newspan = scanline_extents->spans + 1;
 3272+    if ( min_right < scanline_extents->spans[0].n0 )
 3273+    {
 3274+      scanline_extents->spans[1].pixel_offset_for_input = scanline_extents->spans[0].n0;
 3275+      scanline_extents->spans[1].n0 = scanline_extents->spans[0].n0;
 3276+      scanline_extents->spans[1].n1 = scanline_extents->spans[0].n1;
 3277+      --newspan;
 3278+    }
 3279+    newspan->pixel_offset_for_input = min_right;
 3280+    newspan->n0 = scanline_extents->spans[1].n1 + 1;
 3281+    newspan->n1 = scanline_extents->spans[1].n1 + 1 + ( max_right - min_right );
 3282+    scanline_extents->edge_sizes[1] = 0;  // don't need to copy the right margin, since we are directly decoding into the margin
 3283+  }
 3284+
 3285+  // sort the spans into write output order
 3286+  if ( ( scanline_extents->spans[1].n1 > scanline_extents->spans[1].n0 ) && ( scanline_extents->spans[0].n0 > scanline_extents->spans[1].n0 ) )
 3287+  {
 3288+    stbir__span tspan = scanline_extents->spans[0];
 3289+    scanline_extents->spans[0] = scanline_extents->spans[1];
 3290+    scanline_extents->spans[1] = tspan;
 3291+  }
 3292+}
 3293+
 3294+static void stbir__calculate_in_pixel_range( int * first_pixel, int * last_pixel, float out_pixel_center, float out_filter_radius, float inv_scale, float out_shift, int input_size, stbir_edge edge )
 3295+{
 3296+  int first, last;
 3297+  float out_pixel_influence_lowerbound = out_pixel_center - out_filter_radius;
 3298+  float out_pixel_influence_upperbound = out_pixel_center + out_filter_radius;
 3299+
 3300+  float in_pixel_influence_lowerbound = (out_pixel_influence_lowerbound + out_shift) * inv_scale;
 3301+  float in_pixel_influence_upperbound = (out_pixel_influence_upperbound + out_shift) * inv_scale;
 3302+
 3303+  first = (int)(STBIR_FLOORF(in_pixel_influence_lowerbound + 0.5f));
 3304+  last = (int)(STBIR_FLOORF(in_pixel_influence_upperbound - 0.5f));
 3305+  if ( last < first ) last = first; // point sample mode can span a value *right* at 0.5, and cause these to cross
 3306+
 3307+  if ( edge == STBIR_EDGE_WRAP )
 3308+  {
 3309+    if ( first < -input_size )
 3310+      first = -input_size;
 3311+    if ( last >= (input_size*2))
 3312+      last = (input_size*2) - 1;
 3313+  }
 3314+
 3315+  *first_pixel = first;
 3316+  *last_pixel = last;
 3317+}
 3318+
 3319+static void stbir__calculate_coefficients_for_gather_upsample( float out_filter_radius, stbir__kernel_callback * kernel, stbir__scale_info * scale_info, int num_contributors, stbir__contributors* contributors, float* coefficient_group, int coefficient_width, stbir_edge edge, void * user_data )
 3320+{
 3321+  int n, end;
 3322+  float inv_scale = scale_info->inv_scale;
 3323+  float out_shift = scale_info->pixel_shift;
 3324+  int input_size  = scale_info->input_full_size;
 3325+  int numerator = scale_info->scale_numerator;
 3326+  int polyphase = ( ( scale_info->scale_is_rational ) && ( numerator < num_contributors ) );
 3327+
 3328+  // Looping through out pixels
 3329+  end = num_contributors; if ( polyphase ) end = numerator;
 3330+  for (n = 0; n < end; n++)
 3331+  {
 3332+    int i;
 3333+    int last_non_zero;
 3334+    float out_pixel_center = (float)n + 0.5f;
 3335+    float in_center_of_out = (out_pixel_center + out_shift) * inv_scale;
 3336+
 3337+    int in_first_pixel, in_last_pixel;
 3338+
 3339+    stbir__calculate_in_pixel_range( &in_first_pixel, &in_last_pixel, out_pixel_center, out_filter_radius, inv_scale, out_shift, input_size, edge );
 3340+
 3341+    // make sure we never generate a range larger than our precalculated coeff width
 3342+    //   this only happens in point sample mode, but it's a good safe thing to do anyway
 3343+    if ( ( in_last_pixel - in_first_pixel + 1 ) > coefficient_width )
 3344+      in_last_pixel = in_first_pixel + coefficient_width - 1;
 3345+
 3346+    last_non_zero = -1;
 3347+    for (i = 0; i <= in_last_pixel - in_first_pixel; i++)
 3348+    {
 3349+      float in_pixel_center = (float)(i + in_first_pixel) + 0.5f;
 3350+      float coeff = kernel(in_center_of_out - in_pixel_center, inv_scale, user_data);
 3351+
 3352+      // kill denormals
 3353+      if ( ( ( coeff < stbir__small_float ) && ( coeff > -stbir__small_float ) ) )
 3354+      {
 3355+        if ( i == 0 )  // if we're at the front, just eat zero contributors
 3356+        {
 3357+          STBIR_ASSERT ( ( in_last_pixel - in_first_pixel ) != 0 ); // there should be at least one contrib
 3358+          ++in_first_pixel;
 3359+          i--;
 3360+          continue;
 3361+        }
 3362+        coeff = 0;  // make sure is fully zero (should keep denormals away)
 3363+      }
 3364+      else
 3365+        last_non_zero = i;
 3366+
 3367+      coefficient_group[i] = coeff;
 3368+    }
 3369+
 3370+    in_last_pixel = last_non_zero+in_first_pixel; // kills trailing zeros
 3371+    contributors->n0 = in_first_pixel;
 3372+    contributors->n1 = in_last_pixel;
 3373+
 3374+    STBIR_ASSERT(contributors->n1 >= contributors->n0);
 3375+
 3376+    ++contributors;
 3377+    coefficient_group += coefficient_width;
 3378+  }
 3379+}
 3380+
 3381+static void stbir__insert_coeff( stbir__contributors * contribs, float * coeffs, int new_pixel, float new_coeff, int max_width )
 3382+{
 3383+  if ( contribs->n1 < contribs->n0 ) // this first clause should never happen, but handle in case
 3384+  {
 3385+    contribs->n0 = contribs->n1 = new_pixel;
 3386+    coeffs[0] = new_coeff;
 3387+  }
 3388+  else if ( new_pixel <= contribs->n1 )  // before the end
 3389+  {
 3390+    if ( new_pixel < contribs->n0 ) // before the front?
 3391+    {
 3392+      if ( ( contribs->n1 - new_pixel + 1 ) <= max_width )
 3393+      { 
 3394+        int j, o = contribs->n0 - new_pixel;
 3395+        for ( j = contribs->n1 - contribs->n0 ; j >= 0 ; j-- )
 3396+          coeffs[ j + o ] = coeffs[ j ];
 3397+        for ( j = 1 ; j < o ; j++ )
 3398+          coeffs[ j ] = 0;
 3399+        coeffs[ 0 ] = new_coeff;
 3400+        contribs->n0 = new_pixel;
 3401+      }
 3402+    }
 3403+    else
 3404+    {
 3405+      // add new weight to existing coeff if already there
 3406+      coeffs[ new_pixel - contribs->n0 ] += new_coeff;
 3407+    }
 3408+  }
 3409+  else
 3410+  {
 3411+    if ( ( new_pixel - contribs->n0 + 1 ) <= max_width )
 3412+    {
 3413+      int j, e = new_pixel - contribs->n0;
 3414+      for( j = ( contribs->n1 - contribs->n0 ) + 1 ; j < e ; j++ ) // clear in-betweens coeffs if there are any
 3415+        coeffs[j] = 0;
 3416+
 3417+      coeffs[ e ] = new_coeff;
 3418+      contribs->n1 = new_pixel;
 3419+    }
 3420+  }
 3421+}
 3422+
 3423+static void stbir__calculate_out_pixel_range( int * first_pixel, int * last_pixel, float in_pixel_center, float in_pixels_radius, float scale, float out_shift, int out_size )
 3424+{
 3425+  float in_pixel_influence_lowerbound = in_pixel_center - in_pixels_radius;
 3426+  float in_pixel_influence_upperbound = in_pixel_center + in_pixels_radius;
 3427+  float out_pixel_influence_lowerbound = in_pixel_influence_lowerbound * scale - out_shift;
 3428+  float out_pixel_influence_upperbound = in_pixel_influence_upperbound * scale - out_shift;
 3429+  int out_first_pixel = (int)(STBIR_FLOORF(out_pixel_influence_lowerbound + 0.5f));
 3430+  int out_last_pixel = (int)(STBIR_FLOORF(out_pixel_influence_upperbound - 0.5f));
 3431+
 3432+  if ( out_first_pixel < 0 )
 3433+    out_first_pixel = 0;
 3434+  if ( out_last_pixel >= out_size )
 3435+    out_last_pixel = out_size - 1;
 3436+  *first_pixel = out_first_pixel;
 3437+  *last_pixel = out_last_pixel;
 3438+}
 3439+
 3440+static void stbir__calculate_coefficients_for_gather_downsample( int start, int end, float in_pixels_radius, stbir__kernel_callback * kernel, stbir__scale_info * scale_info, int coefficient_width, int num_contributors, stbir__contributors * contributors, float * coefficient_group, void * user_data )
 3441+{
 3442+  int in_pixel;
 3443+  int i;
 3444+  int first_out_inited = -1;
 3445+  float scale = scale_info->scale;
 3446+  float out_shift = scale_info->pixel_shift;
 3447+  int out_size = scale_info->output_sub_size;
 3448+  int numerator = scale_info->scale_numerator;
 3449+  int polyphase = ( ( scale_info->scale_is_rational ) && ( numerator < out_size ) );
 3450+
 3451+  STBIR__UNUSED(num_contributors);
 3452+
 3453+  // Loop through the input pixels
 3454+  for (in_pixel = start; in_pixel < end; in_pixel++)
 3455+  {
 3456+    float in_pixel_center = (float)in_pixel + 0.5f;
 3457+    float out_center_of_in = in_pixel_center * scale - out_shift;
 3458+    int out_first_pixel, out_last_pixel;
 3459+
 3460+    stbir__calculate_out_pixel_range( &out_first_pixel, &out_last_pixel, in_pixel_center, in_pixels_radius, scale, out_shift, out_size );
 3461+
 3462+    if ( out_first_pixel > out_last_pixel )
 3463+      continue;
 3464+
 3465+    // clamp or exit if we are using polyphase filtering, and the limit is up
 3466+    if ( polyphase )
 3467+    {
 3468+      // when polyphase, you only have to do coeffs up to the numerator count
 3469+      if ( out_first_pixel == numerator )
 3470+        break;
 3471+
 3472+      // don't do any extra work, clamp last pixel at numerator too
 3473+      if ( out_last_pixel >= numerator )
 3474+        out_last_pixel = numerator - 1;
 3475+    }
 3476+
 3477+    for (i = 0; i <= out_last_pixel - out_first_pixel; i++)
 3478+    {
 3479+      float out_pixel_center = (float)(i + out_first_pixel) + 0.5f;
 3480+      float x = out_pixel_center - out_center_of_in;
 3481+      float coeff = kernel(x, scale, user_data) * scale;
 3482+
 3483+      // kill the coeff if it's too small (avoid denormals)
 3484+      if ( ( ( coeff < stbir__small_float ) && ( coeff > -stbir__small_float ) ) )
 3485+        coeff = 0.0f;
 3486+
 3487+      {
 3488+        int out = i + out_first_pixel;
 3489+        float * coeffs = coefficient_group + out * coefficient_width;
 3490+        stbir__contributors * contribs = contributors + out;
 3491+
 3492+        // is this the first time this output pixel has been seen?  Init it.
 3493+        if ( out > first_out_inited )
 3494+        {
 3495+          STBIR_ASSERT( out == ( first_out_inited + 1 ) ); // ensure we have only advanced one at time
 3496+          first_out_inited = out;
 3497+          contribs->n0 = in_pixel;
 3498+          contribs->n1 = in_pixel;
 3499+          coeffs[0]  = coeff;
 3500+        }
 3501+        else
 3502+        {
 3503+          // insert on end (always in order)
 3504+          if ( coeffs[0] == 0.0f )  // if the first coefficent is zero, then zap it for this coeffs
 3505+          {
 3506+            STBIR_ASSERT( ( in_pixel - contribs->n0 ) == 1 ); // ensure that when we zap, we're at the 2nd pos
 3507+            contribs->n0 = in_pixel;
 3508+          }
 3509+          contribs->n1 = in_pixel;
 3510+          STBIR_ASSERT( ( in_pixel - contribs->n0 ) < coefficient_width );
 3511+          coeffs[in_pixel - contribs->n0]  = coeff;
 3512+        }
 3513+      }
 3514+    }
 3515+  }
 3516+}
 3517+
 3518+#ifdef STBIR_RENORMALIZE_IN_FLOAT
 3519+#define STBIR_RENORM_TYPE float
 3520+#else
 3521+#define STBIR_RENORM_TYPE double
 3522+#endif
 3523+
 3524+static void stbir__cleanup_gathered_coefficients( stbir_edge edge, stbir__filter_extent_info* filter_info, stbir__scale_info * scale_info, int num_contributors, stbir__contributors* contributors, float * coefficient_group, int coefficient_width )
 3525+{
 3526+  int input_size = scale_info->input_full_size;
 3527+  int input_last_n1 = input_size - 1;
 3528+  int n, end;
 3529+  int lowest = 0x7fffffff;
 3530+  int highest = -0x7fffffff;
 3531+  int widest = -1;
 3532+  int numerator = scale_info->scale_numerator;
 3533+  int denominator = scale_info->scale_denominator;
 3534+  int polyphase = ( ( scale_info->scale_is_rational ) && ( numerator < num_contributors ) );
 3535+  float * coeffs;
 3536+  stbir__contributors * contribs;
 3537+
 3538+  // weight all the coeffs for each sample
 3539+  coeffs = coefficient_group;
 3540+  contribs = contributors;
 3541+  end = num_contributors; if ( polyphase ) end = numerator;
 3542+  for (n = 0; n < end; n++)
 3543+  {
 3544+    int i;
 3545+    STBIR_RENORM_TYPE filter_scale, total_filter = 0;
 3546+    int e;
 3547+
 3548+    // add all contribs
 3549+    e = contribs->n1 - contribs->n0;
 3550+    for( i = 0 ; i <= e ; i++ )
 3551+    {
 3552+      total_filter += (STBIR_RENORM_TYPE) coeffs[i];
 3553+      STBIR_ASSERT( ( coeffs[i] >= -2.0f ) && ( coeffs[i] <= 2.0f )  ); // check for wonky weights
 3554+    }
 3555+
 3556+    // rescale
 3557+    if ( ( total_filter < stbir__small_float ) && ( total_filter > -stbir__small_float ) )
 3558+    {
 3559+      // all coeffs are extremely small, just zero it
 3560+      contribs->n1 = contribs->n0;
 3561+      coeffs[0] = 0.0f;
 3562+    }
 3563+    else
 3564+    {
 3565+      // if the total isn't 1.0, rescale everything
 3566+      if ( ( total_filter < (1.0f-stbir__small_float) ) || ( total_filter > (1.0f+stbir__small_float) ) )
 3567+      {
 3568+        filter_scale = ((STBIR_RENORM_TYPE)1.0) / total_filter;
 3569+
 3570+        // scale them all
 3571+        for (i = 0; i <= e; i++)
 3572+          coeffs[i] = (float) ( coeffs[i] * filter_scale );
 3573+      }
 3574+    }
 3575+    ++contribs;
 3576+    coeffs += coefficient_width;
 3577+  }
 3578+
 3579+  // if we have a rational for the scale, we can exploit the polyphaseness to not calculate
 3580+  //   most of the coefficients, so we copy them here
 3581+  if ( polyphase )
 3582+  {
 3583+    stbir__contributors * prev_contribs = contributors;
 3584+    stbir__contributors * cur_contribs = contributors + numerator;
 3585+
 3586+    for( n = numerator ; n < num_contributors ; n++ )
 3587+    {
 3588+      cur_contribs->n0 = prev_contribs->n0 + denominator;
 3589+      cur_contribs->n1 = prev_contribs->n1 + denominator;
 3590+      ++cur_contribs;
 3591+      ++prev_contribs;
 3592+    }
 3593+    stbir_overlapping_memcpy( coefficient_group + numerator * coefficient_width, coefficient_group, ( num_contributors - numerator ) * coefficient_width * sizeof( coeffs[ 0 ] ) );
 3594+  }
 3595+
 3596+  coeffs = coefficient_group;
 3597+  contribs = contributors;
 3598+
 3599+  for (n = 0; n < num_contributors; n++)
 3600+  {
 3601+    int i;
 3602+
 3603+    // in zero edge mode, just remove out of bounds contribs completely (since their weights are accounted for now)
 3604+    if ( edge == STBIR_EDGE_ZERO )
 3605+    {
 3606+      // shrink the right side if necessary
 3607+      if ( contribs->n1 > input_last_n1 )
 3608+        contribs->n1 = input_last_n1;
 3609+
 3610+      // shrink the left side
 3611+      if ( contribs->n0 < 0 )
 3612+      {
 3613+        int j, left, skips = 0;
 3614+
 3615+        skips = -contribs->n0;
 3616+        contribs->n0 = 0;
 3617+
 3618+        // now move down the weights
 3619+        left = contribs->n1 - contribs->n0 + 1;
 3620+        if ( left > 0 )
 3621+        {
 3622+          for( j = 0 ; j < left ; j++ )
 3623+            coeffs[ j ] = coeffs[ j + skips ];
 3624+        }
 3625+      }
 3626+    }
 3627+    else if ( ( edge == STBIR_EDGE_CLAMP ) || ( edge == STBIR_EDGE_REFLECT ) )
 3628+    {
 3629+      // for clamp and reflect, calculate the true inbounds position (based on edge type) and just add that to the existing weight
 3630+
 3631+      // right hand side first
 3632+      if ( contribs->n1 > input_last_n1 )
 3633+      {
 3634+        int start = contribs->n0;
 3635+        int endi = contribs->n1;
 3636+        contribs->n1 = input_last_n1;
 3637+        for( i = input_size; i <= endi; i++ )
 3638+          stbir__insert_coeff( contribs, coeffs, stbir__edge_wrap_slow[edge]( i, input_size ), coeffs[i-start], coefficient_width );
 3639+      }
 3640+
 3641+      // now check left hand edge
 3642+      if ( contribs->n0 < 0 )
 3643+      {
 3644+        int save_n0;
 3645+        float save_n0_coeff;
 3646+        float * c = coeffs - ( contribs->n0 + 1 );
 3647+
 3648+        // reinsert the coeffs with it reflected or clamped (insert accumulates, if the coeffs exist)
 3649+        for( i = -1 ; i > contribs->n0 ; i-- )
 3650+          stbir__insert_coeff( contribs, coeffs, stbir__edge_wrap_slow[edge]( i, input_size ), *c--, coefficient_width );
 3651+        save_n0 = contribs->n0;
 3652+        save_n0_coeff = c[0]; // save it, since we didn't do the final one (i==n0), because there might be too many coeffs to hold (before we resize)!
 3653+
 3654+        // now slide all the coeffs down (since we have accumulated them in the positive contribs) and reset the first contrib
 3655+        contribs->n0 = 0;
 3656+        for(i = 0 ; i <= contribs->n1 ; i++ )
 3657+          coeffs[i] = coeffs[i-save_n0];
 3658+
 3659+        // now that we have shrunk down the contribs, we insert the first one safely
 3660+        stbir__insert_coeff( contribs, coeffs, stbir__edge_wrap_slow[edge]( save_n0, input_size ), save_n0_coeff, coefficient_width );
 3661+      }
 3662+    }
 3663+
 3664+    if ( contribs->n0 <= contribs->n1 )
 3665+    {
 3666+      int diff = contribs->n1 - contribs->n0 + 1;
 3667+      while ( diff && ( coeffs[ diff-1 ] == 0.0f ) )
 3668+        --diff;
 3669+
 3670+      contribs->n1 = contribs->n0 + diff - 1;
 3671+
 3672+      if ( contribs->n0 <= contribs->n1 )
 3673+      {
 3674+        if ( contribs->n0 < lowest )
 3675+          lowest = contribs->n0;
 3676+        if ( contribs->n1 > highest )
 3677+          highest = contribs->n1;
 3678+        if ( diff > widest )
 3679+          widest = diff;
 3680+      }
 3681+
 3682+      // re-zero out unused coefficients (if any)
 3683+      for( i = diff ; i < coefficient_width ; i++ )
 3684+        coeffs[i] = 0.0f;
 3685+    }
 3686+
 3687+    ++contribs;
 3688+    coeffs += coefficient_width;
 3689+  }
 3690+  filter_info->lowest = lowest;
 3691+  filter_info->highest = highest;
 3692+  filter_info->widest = widest;
 3693+}
 3694+
 3695+#undef STBIR_RENORM_TYPE 
 3696+
 3697+static int stbir__pack_coefficients( int num_contributors, stbir__contributors* contributors, float * coefficents, int coefficient_width, int widest, int row0, int row1 ) 
 3698+{
 3699+  #define STBIR_MOVE_1( dest, src ) { STBIR_NO_UNROLL(dest); ((stbir_uint32*)(dest))[0] = ((stbir_uint32*)(src))[0]; }
 3700+  #define STBIR_MOVE_2( dest, src ) { STBIR_NO_UNROLL(dest); ((stbir_uint64*)(dest))[0] = ((stbir_uint64*)(src))[0]; }
 3701+  #ifdef STBIR_SIMD
 3702+  #define STBIR_MOVE_4( dest, src ) { stbir__simdf t; STBIR_NO_UNROLL(dest); stbir__simdf_load( t, src ); stbir__simdf_store( dest, t ); }
 3703+  #else
 3704+  #define STBIR_MOVE_4( dest, src ) { STBIR_NO_UNROLL(dest); ((stbir_uint64*)(dest))[0] = ((stbir_uint64*)(src))[0]; ((stbir_uint64*)(dest))[1] = ((stbir_uint64*)(src))[1]; }
 3705+  #endif
 3706+
 3707+  int row_end = row1 + 1;
 3708+  STBIR__UNUSED( row0 ); // only used in an assert
 3709+
 3710+  if ( coefficient_width != widest )
 3711+  {
 3712+    float * pc = coefficents;
 3713+    float * coeffs = coefficents;
 3714+    float * pc_end = coefficents + num_contributors * widest;
 3715+    switch( widest )
 3716+    {
 3717+      case 1:
 3718+        STBIR_NO_UNROLL_LOOP_START
 3719+        do {
 3720+          STBIR_MOVE_1( pc, coeffs );
 3721+          ++pc;
 3722+          coeffs += coefficient_width;
 3723+        } while ( pc < pc_end );
 3724+        break;
 3725+      case 2:
 3726+        STBIR_NO_UNROLL_LOOP_START
 3727+        do {
 3728+          STBIR_MOVE_2( pc, coeffs );
 3729+          pc += 2;
 3730+          coeffs += coefficient_width;
 3731+        } while ( pc < pc_end );
 3732+        break;
 3733+      case 3:
 3734+        STBIR_NO_UNROLL_LOOP_START
 3735+        do {
 3736+          STBIR_MOVE_2( pc, coeffs );
 3737+          STBIR_MOVE_1( pc+2, coeffs+2 );
 3738+          pc += 3;
 3739+          coeffs += coefficient_width;
 3740+        } while ( pc < pc_end );
 3741+        break;
 3742+      case 4:
 3743+        STBIR_NO_UNROLL_LOOP_START
 3744+        do {
 3745+          STBIR_MOVE_4( pc, coeffs );
 3746+          pc += 4;
 3747+          coeffs += coefficient_width;
 3748+        } while ( pc < pc_end );
 3749+        break;
 3750+      case 5:
 3751+        STBIR_NO_UNROLL_LOOP_START
 3752+        do {
 3753+          STBIR_MOVE_4( pc, coeffs );
 3754+          STBIR_MOVE_1( pc+4, coeffs+4 );
 3755+          pc += 5;
 3756+          coeffs += coefficient_width;
 3757+        } while ( pc < pc_end );
 3758+        break;
 3759+      case 6:
 3760+        STBIR_NO_UNROLL_LOOP_START
 3761+        do {
 3762+          STBIR_MOVE_4( pc, coeffs );
 3763+          STBIR_MOVE_2( pc+4, coeffs+4 );
 3764+          pc += 6;
 3765+          coeffs += coefficient_width;
 3766+        } while ( pc < pc_end );
 3767+        break;
 3768+      case 7:
 3769+        STBIR_NO_UNROLL_LOOP_START
 3770+        do {
 3771+          STBIR_MOVE_4( pc, coeffs );
 3772+          STBIR_MOVE_2( pc+4, coeffs+4 );
 3773+          STBIR_MOVE_1( pc+6, coeffs+6 );
 3774+          pc += 7;
 3775+          coeffs += coefficient_width;
 3776+        } while ( pc < pc_end );
 3777+        break;
 3778+      case 8:
 3779+        STBIR_NO_UNROLL_LOOP_START
 3780+        do {
 3781+          STBIR_MOVE_4( pc, coeffs );
 3782+          STBIR_MOVE_4( pc+4, coeffs+4 );
 3783+          pc += 8;
 3784+          coeffs += coefficient_width;
 3785+        } while ( pc < pc_end );
 3786+        break;
 3787+      case 9:
 3788+        STBIR_NO_UNROLL_LOOP_START
 3789+        do {
 3790+          STBIR_MOVE_4( pc, coeffs );
 3791+          STBIR_MOVE_4( pc+4, coeffs+4 );
 3792+          STBIR_MOVE_1( pc+8, coeffs+8 );
 3793+          pc += 9;
 3794+          coeffs += coefficient_width;
 3795+        } while ( pc < pc_end );
 3796+        break;
 3797+      case 10:
 3798+        STBIR_NO_UNROLL_LOOP_START
 3799+        do {
 3800+          STBIR_MOVE_4( pc, coeffs );
 3801+          STBIR_MOVE_4( pc+4, coeffs+4 );
 3802+          STBIR_MOVE_2( pc+8, coeffs+8 );
 3803+          pc += 10;
 3804+          coeffs += coefficient_width;
 3805+        } while ( pc < pc_end );
 3806+        break;
 3807+      case 11:
 3808+        STBIR_NO_UNROLL_LOOP_START
 3809+        do {
 3810+          STBIR_MOVE_4( pc, coeffs );
 3811+          STBIR_MOVE_4( pc+4, coeffs+4 );
 3812+          STBIR_MOVE_2( pc+8, coeffs+8 );
 3813+          STBIR_MOVE_1( pc+10, coeffs+10 );
 3814+          pc += 11;
 3815+          coeffs += coefficient_width;
 3816+        } while ( pc < pc_end );
 3817+        break;
 3818+      case 12:
 3819+        STBIR_NO_UNROLL_LOOP_START
 3820+        do {
 3821+          STBIR_MOVE_4( pc, coeffs );
 3822+          STBIR_MOVE_4( pc+4, coeffs+4 );
 3823+          STBIR_MOVE_4( pc+8, coeffs+8 );
 3824+          pc += 12;
 3825+          coeffs += coefficient_width;
 3826+        } while ( pc < pc_end );
 3827+        break;
 3828+      default:
 3829+        STBIR_NO_UNROLL_LOOP_START
 3830+        do {
 3831+          float * copy_end = pc + widest - 4;
 3832+          float * c = coeffs;
 3833+          do {
 3834+            STBIR_NO_UNROLL( pc );
 3835+            STBIR_MOVE_4( pc, c );
 3836+            pc += 4;
 3837+            c += 4;
 3838+          } while ( pc <= copy_end );
 3839+          copy_end += 4;
 3840+          STBIR_NO_UNROLL_LOOP_START
 3841+          while ( pc < copy_end )
 3842+          {
 3843+            STBIR_MOVE_1( pc, c );
 3844+            ++pc; ++c;
 3845+          }
 3846+          coeffs += coefficient_width;
 3847+        } while ( pc < pc_end );
 3848+        break;
 3849+    }
 3850+  }
 3851+
 3852+  // some horizontal routines read one float off the end (which is then masked off), so put in a sentinel so we don't read an snan or denormal
 3853+  coefficents[ widest * num_contributors ] = 8888.0f;
 3854+
 3855+  // the minimum we might read for unrolled filters widths is 12. So, we need to
 3856+  //   make sure we never read outside the decode buffer, by possibly moving
 3857+  //   the sample area back into the scanline, and putting zeros weights first.
 3858+  // we start on the right edge and check until we're well past the possible
 3859+  //   clip area (2*widest).
 3860+  {
 3861+    stbir__contributors * contribs = contributors + num_contributors - 1;
 3862+    float * coeffs = coefficents + widest * ( num_contributors - 1 );
 3863+
 3864+    // go until no chance of clipping (this is usually less than 8 lops)
 3865+    while ( ( contribs >= contributors ) && ( ( contribs->n0 + widest*2 ) >= row_end ) )
 3866+    {
 3867+      // might we clip??
 3868+      if ( ( contribs->n0 + widest ) > row_end )
 3869+      {
 3870+        int stop_range = widest;
 3871+
 3872+        // if range is larger than 12, it will be handled by generic loops that can terminate on the exact length
 3873+        //   of this contrib n1, instead of a fixed widest amount - so calculate this
 3874+        if ( widest > 12 )
 3875+        {
 3876+          int mod;
 3877+
 3878+          // how far will be read in the n_coeff loop (which depends on the widest count mod4);
 3879+          mod = widest & 3;
 3880+          stop_range = ( ( ( contribs->n1 - contribs->n0 + 1 ) - mod + 3 ) & ~3 ) + mod;
 3881+
 3882+          // the n_coeff loops do a minimum amount of coeffs, so factor that in!
 3883+          if ( stop_range < ( 8 + mod ) ) stop_range = 8 + mod;
 3884+        }
 3885+
 3886+        // now see if we still clip with the refined range
 3887+        if ( ( contribs->n0 + stop_range ) > row_end )
 3888+        {
 3889+          int new_n0 = row_end - stop_range;
 3890+          int num = contribs->n1 - contribs->n0 + 1;
 3891+          int backup = contribs->n0 - new_n0;
 3892+          float * from_co = coeffs + num - 1;
 3893+          float * to_co = from_co + backup;
 3894+
 3895+          STBIR_ASSERT( ( new_n0 >= row0 ) && ( new_n0 < contribs->n0 ) );
 3896+
 3897+          // move the coeffs over
 3898+          while( num )
 3899+          {
 3900+            *to_co-- = *from_co--;
 3901+            --num;
 3902+          }
 3903+          // zero new positions
 3904+          while ( to_co >= coeffs )
 3905+            *to_co-- = 0;
 3906+          // set new start point
 3907+          contribs->n0 = new_n0;
 3908+          if ( widest > 12 )
 3909+          {
 3910+            int mod;
 3911+
 3912+            // how far will be read in the n_coeff loop (which depends on the widest count mod4);
 3913+            mod = widest & 3;
 3914+            stop_range = ( ( ( contribs->n1 - contribs->n0 + 1 ) - mod + 3 ) & ~3 ) + mod;
 3915+
 3916+            // the n_coeff loops do a minimum amount of coeffs, so factor that in!
 3917+            if ( stop_range < ( 8 + mod ) ) stop_range = 8 + mod;
 3918+          }
 3919+        }
 3920+      }
 3921+      --contribs;
 3922+      coeffs -= widest;
 3923+    }
 3924+  }
 3925+
 3926+  return widest;
 3927+  #undef STBIR_MOVE_1
 3928+  #undef STBIR_MOVE_2
 3929+  #undef STBIR_MOVE_4
 3930+}
 3931+
 3932+static void stbir__calculate_filters( stbir__sampler * samp, stbir__sampler * other_axis_for_pivot, void * user_data STBIR_ONLY_PROFILE_BUILD_GET_INFO )
 3933+{
 3934+  int n;
 3935+  float scale = samp->scale_info.scale;
 3936+  stbir__kernel_callback * kernel = samp->filter_kernel;
 3937+  stbir__support_callback * support = samp->filter_support;
 3938+  float inv_scale = samp->scale_info.inv_scale;
 3939+  int input_full_size = samp->scale_info.input_full_size;
 3940+  int gather_num_contributors = samp->num_contributors;
 3941+  stbir__contributors* gather_contributors = samp->contributors;
 3942+  float * gather_coeffs = samp->coefficients;
 3943+  int gather_coefficient_width = samp->coefficient_width;
 3944+
 3945+  switch ( samp->is_gather )
 3946+  {
 3947+    case 1: // gather upsample
 3948+    {
 3949+      float out_pixels_radius = support(inv_scale,user_data) * scale;
 3950+
 3951+      stbir__calculate_coefficients_for_gather_upsample( out_pixels_radius, kernel, &samp->scale_info, gather_num_contributors, gather_contributors, gather_coeffs, gather_coefficient_width, samp->edge, user_data );
 3952+
 3953+      STBIR_PROFILE_BUILD_START( cleanup );
 3954+      stbir__cleanup_gathered_coefficients( samp->edge, &samp->extent_info, &samp->scale_info, gather_num_contributors, gather_contributors, gather_coeffs, gather_coefficient_width );
 3955+      STBIR_PROFILE_BUILD_END( cleanup );
 3956+    }
 3957+    break;
 3958+
 3959+    case 0: // scatter downsample (only on vertical)
 3960+    case 2: // gather downsample
 3961+    {
 3962+      float in_pixels_radius = support(scale,user_data) * inv_scale;
 3963+      int filter_pixel_margin = samp->filter_pixel_margin;
 3964+      int input_end = input_full_size + filter_pixel_margin;
 3965+
 3966+      // if this is a scatter, we do a downsample gather to get the coeffs, and then pivot after
 3967+      if ( !samp->is_gather )
 3968+      {
 3969+        // check if we are using the same gather downsample on the horizontal as this vertical,
 3970+        //   if so, then we don't have to generate them, we can just pivot from the horizontal.
 3971+        if ( other_axis_for_pivot )
 3972+        {
 3973+          gather_contributors = other_axis_for_pivot->contributors;
 3974+          gather_coeffs = other_axis_for_pivot->coefficients;
 3975+          gather_coefficient_width = other_axis_for_pivot->coefficient_width;
 3976+          gather_num_contributors = other_axis_for_pivot->num_contributors;
 3977+          samp->extent_info.lowest = other_axis_for_pivot->extent_info.lowest;
 3978+          samp->extent_info.highest = other_axis_for_pivot->extent_info.highest;
 3979+          samp->extent_info.widest = other_axis_for_pivot->extent_info.widest;
 3980+          goto jump_right_to_pivot;
 3981+        }
 3982+
 3983+        gather_contributors = samp->gather_prescatter_contributors;
 3984+        gather_coeffs = samp->gather_prescatter_coefficients;
 3985+        gather_coefficient_width = samp->gather_prescatter_coefficient_width;
 3986+        gather_num_contributors = samp->gather_prescatter_num_contributors;
 3987+      }
 3988+
 3989+      stbir__calculate_coefficients_for_gather_downsample( -filter_pixel_margin, input_end, in_pixels_radius, kernel, &samp->scale_info, gather_coefficient_width, gather_num_contributors, gather_contributors, gather_coeffs, user_data );
 3990+
 3991+      STBIR_PROFILE_BUILD_START( cleanup );
 3992+      stbir__cleanup_gathered_coefficients( samp->edge, &samp->extent_info, &samp->scale_info, gather_num_contributors, gather_contributors, gather_coeffs, gather_coefficient_width );
 3993+      STBIR_PROFILE_BUILD_END( cleanup );
 3994+
 3995+      if ( !samp->is_gather )
 3996+      {
 3997+        // if this is a scatter (vertical only), then we need to pivot the coeffs
 3998+        stbir__contributors * scatter_contributors;
 3999+        int highest_set;
 4000+
 4001+        jump_right_to_pivot:
 4002+
 4003+        STBIR_PROFILE_BUILD_START( pivot );
 4004+
 4005+        highest_set = (-filter_pixel_margin) - 1;
 4006+        for (n = 0; n < gather_num_contributors; n++)
 4007+        {
 4008+          int k;
 4009+          int gn0 = gather_contributors->n0, gn1 = gather_contributors->n1;
 4010+          int scatter_coefficient_width = samp->coefficient_width;
 4011+          float * scatter_coeffs = samp->coefficients + ( gn0 + filter_pixel_margin ) * scatter_coefficient_width;
 4012+          float * g_coeffs = gather_coeffs;
 4013+          scatter_contributors = samp->contributors + ( gn0 + filter_pixel_margin );
 4014+
 4015+          for (k = gn0 ; k <= gn1 ; k++ )
 4016+          {
 4017+            float gc = *g_coeffs++;
 4018+            
 4019+            // skip zero and denormals - must skip zeros to avoid adding coeffs beyond scatter_coefficient_width
 4020+            //   (which happens when pivoting from horizontal, which might have dummy zeros)
 4021+            if ( ( ( gc >= stbir__small_float ) || ( gc <= -stbir__small_float ) ) )
 4022+            {
 4023+              if ( ( k > highest_set ) || ( scatter_contributors->n0 > scatter_contributors->n1 ) )
 4024+              {
 4025+                {
 4026+                  // if we are skipping over several contributors, we need to clear the skipped ones
 4027+                  stbir__contributors * clear_contributors = samp->contributors + ( highest_set + filter_pixel_margin + 1);
 4028+                  while ( clear_contributors < scatter_contributors )
 4029+                  {
 4030+                    clear_contributors->n0 = 0;
 4031+                    clear_contributors->n1 = -1;
 4032+                    ++clear_contributors;
 4033+                  }
 4034+                }
 4035+                scatter_contributors->n0 = n;
 4036+                scatter_contributors->n1 = n;
 4037+                scatter_coeffs[0]  = gc;
 4038+                highest_set = k;
 4039+              }
 4040+              else
 4041+              {
 4042+                stbir__insert_coeff( scatter_contributors, scatter_coeffs, n, gc, scatter_coefficient_width );
 4043+              }
 4044+              STBIR_ASSERT( ( scatter_contributors->n1 - scatter_contributors->n0 + 1 ) <= scatter_coefficient_width );
 4045+            }
 4046+            ++scatter_contributors;
 4047+            scatter_coeffs += scatter_coefficient_width;
 4048+          }
 4049+
 4050+          ++gather_contributors;
 4051+          gather_coeffs += gather_coefficient_width;
 4052+        }
 4053+
 4054+        // now clear any unset contribs
 4055+        {
 4056+          stbir__contributors * clear_contributors = samp->contributors + ( highest_set + filter_pixel_margin + 1);
 4057+          stbir__contributors * end_contributors = samp->contributors + samp->num_contributors;
 4058+          while ( clear_contributors < end_contributors )
 4059+          {
 4060+            clear_contributors->n0 = 0;
 4061+            clear_contributors->n1 = -1;
 4062+            ++clear_contributors;
 4063+          }
 4064+        }
 4065+
 4066+        STBIR_PROFILE_BUILD_END( pivot );
 4067+      }
 4068+    }
 4069+    break;
 4070+  }
 4071+}
 4072+
 4073+
 4074+//========================================================================================================
 4075+// scanline decoders and encoders
 4076+
 4077+#define stbir__coder_min_num 1
 4078+#define STB_IMAGE_RESIZE_DO_CODERS
 4079+#include STBIR__HEADER_FILENAME
 4080+
 4081+#define stbir__decode_suffix BGRA
 4082+#define stbir__decode_swizzle
 4083+#define stbir__decode_order0  2
 4084+#define stbir__decode_order1  1
 4085+#define stbir__decode_order2  0
 4086+#define stbir__decode_order3  3
 4087+#define stbir__encode_order0  2
 4088+#define stbir__encode_order1  1
 4089+#define stbir__encode_order2  0
 4090+#define stbir__encode_order3  3
 4091+#define stbir__coder_min_num 4
 4092+#define STB_IMAGE_RESIZE_DO_CODERS
 4093+#include STBIR__HEADER_FILENAME
 4094+
 4095+#define stbir__decode_suffix ARGB
 4096+#define stbir__decode_swizzle
 4097+#define stbir__decode_order0  1
 4098+#define stbir__decode_order1  2
 4099+#define stbir__decode_order2  3
 4100+#define stbir__decode_order3  0
 4101+#define stbir__encode_order0  3
 4102+#define stbir__encode_order1  0
 4103+#define stbir__encode_order2  1
 4104+#define stbir__encode_order3  2
 4105+#define stbir__coder_min_num 4
 4106+#define STB_IMAGE_RESIZE_DO_CODERS
 4107+#include STBIR__HEADER_FILENAME
 4108+
 4109+#define stbir__decode_suffix ABGR
 4110+#define stbir__decode_swizzle
 4111+#define stbir__decode_order0  3
 4112+#define stbir__decode_order1  2
 4113+#define stbir__decode_order2  1
 4114+#define stbir__decode_order3  0
 4115+#define stbir__encode_order0  3
 4116+#define stbir__encode_order1  2
 4117+#define stbir__encode_order2  1
 4118+#define stbir__encode_order3  0
 4119+#define stbir__coder_min_num 4
 4120+#define STB_IMAGE_RESIZE_DO_CODERS
 4121+#include STBIR__HEADER_FILENAME
 4122+
 4123+#define stbir__decode_suffix AR
 4124+#define stbir__decode_swizzle
 4125+#define stbir__decode_order0  1
 4126+#define stbir__decode_order1  0
 4127+#define stbir__decode_order2  3
 4128+#define stbir__decode_order3  2
 4129+#define stbir__encode_order0  1
 4130+#define stbir__encode_order1  0
 4131+#define stbir__encode_order2  3
 4132+#define stbir__encode_order3  2
 4133+#define stbir__coder_min_num 2
 4134+#define STB_IMAGE_RESIZE_DO_CODERS
 4135+#include STBIR__HEADER_FILENAME
 4136+
 4137+
 4138+// fancy alpha means we expand to keep both premultipied and non-premultiplied color channels
 4139+static void stbir__fancy_alpha_weight_4ch( float * out_buffer, int width_times_channels )
 4140+{
 4141+  float STBIR_STREAMOUT_PTR(*) out = out_buffer;
 4142+  float const * end_decode = out_buffer + ( width_times_channels / 4 ) * 7;  // decode buffer aligned to end of out_buffer
 4143+  float STBIR_STREAMOUT_PTR(*) decode = (float*)end_decode - width_times_channels;
 4144+
 4145+  // fancy alpha is stored internally as R G B A Rpm Gpm Bpm
 4146+
 4147+  #ifdef STBIR_SIMD
 4148+
 4149+  #ifdef STBIR_SIMD8
 4150+  decode += 16;
 4151+  STBIR_NO_UNROLL_LOOP_START
 4152+  while ( decode <= end_decode )
 4153+  {
 4154+    stbir__simdf8 d0,d1,a0,a1,p0,p1;
 4155+    STBIR_NO_UNROLL(decode);
 4156+    stbir__simdf8_load( d0, decode-16 );
 4157+    stbir__simdf8_load( d1, decode-16+8 );
 4158+    stbir__simdf8_0123to33333333( a0, d0 );
 4159+    stbir__simdf8_0123to33333333( a1, d1 );
 4160+    stbir__simdf8_mult( p0, a0, d0 );
 4161+    stbir__simdf8_mult( p1, a1, d1 );
 4162+    stbir__simdf8_bot4s( a0, d0, p0 );
 4163+    stbir__simdf8_bot4s( a1, d1, p1 );
 4164+    stbir__simdf8_top4s( d0, d0, p0 );
 4165+    stbir__simdf8_top4s( d1, d1, p1 );
 4166+    stbir__simdf8_store ( out, a0 );
 4167+    stbir__simdf8_store ( out+7, d0 );
 4168+    stbir__simdf8_store ( out+14, a1 );
 4169+    stbir__simdf8_store ( out+21, d1 );
 4170+    decode += 16;
 4171+    out += 28;
 4172+  }
 4173+  decode -= 16;
 4174+  #else
 4175+  decode += 8;
 4176+  STBIR_NO_UNROLL_LOOP_START
 4177+  while ( decode <= end_decode )
 4178+  {
 4179+    stbir__simdf d0,a0,d1,a1,p0,p1;
 4180+    STBIR_NO_UNROLL(decode);
 4181+    stbir__simdf_load( d0, decode-8 );
 4182+    stbir__simdf_load( d1, decode-8+4 );
 4183+    stbir__simdf_0123to3333( a0, d0 );
 4184+    stbir__simdf_0123to3333( a1, d1 );
 4185+    stbir__simdf_mult( p0, a0, d0 );
 4186+    stbir__simdf_mult( p1, a1, d1 );
 4187+    stbir__simdf_store ( out, d0 );
 4188+    stbir__simdf_store ( out+4, p0 );
 4189+    stbir__simdf_store ( out+7, d1 );
 4190+    stbir__simdf_store ( out+7+4, p1 );
 4191+    decode += 8;
 4192+    out += 14;
 4193+  }
 4194+  decode -= 8;
 4195+  #endif
 4196+
 4197+  // might be one last odd pixel
 4198+  #ifdef STBIR_SIMD8
 4199+  STBIR_NO_UNROLL_LOOP_START
 4200+  while ( decode < end_decode )
 4201+  #else
 4202+  if ( decode < end_decode )
 4203+  #endif
 4204+  {
 4205+    stbir__simdf d,a,p;
 4206+    STBIR_NO_UNROLL(decode);
 4207+    stbir__simdf_load( d, decode );
 4208+    stbir__simdf_0123to3333( a, d );
 4209+    stbir__simdf_mult( p, a, d );
 4210+    stbir__simdf_store ( out, d );
 4211+    stbir__simdf_store ( out+4, p );
 4212+    decode += 4;
 4213+    out += 7;
 4214+  }
 4215+
 4216+  #else
 4217+
 4218+  while( decode < end_decode )
 4219+  {
 4220+    float r = decode[0], g = decode[1], b = decode[2], alpha = decode[3];
 4221+    out[0] = r;
 4222+    out[1] = g;
 4223+    out[2] = b;
 4224+    out[3] = alpha;
 4225+    out[4] = r * alpha;
 4226+    out[5] = g * alpha;
 4227+    out[6] = b * alpha;
 4228+    out += 7;
 4229+    decode += 4;
 4230+  }
 4231+
 4232+  #endif
 4233+}
 4234+
 4235+static void stbir__fancy_alpha_weight_2ch( float * out_buffer, int width_times_channels )
 4236+{
 4237+  float STBIR_STREAMOUT_PTR(*) out = out_buffer;
 4238+  float const * end_decode = out_buffer + ( width_times_channels / 2 ) * 3;
 4239+  float STBIR_STREAMOUT_PTR(*) decode = (float*)end_decode - width_times_channels;
 4240+
 4241+  //  for fancy alpha, turns into: [X A Xpm][X A Xpm],etc
 4242+
 4243+  #ifdef STBIR_SIMD
 4244+
 4245+  decode += 8;
 4246+  if ( decode <= end_decode )
 4247+  {
 4248+    STBIR_NO_UNROLL_LOOP_START
 4249+    do {
 4250+      #ifdef STBIR_SIMD8
 4251+      stbir__simdf8 d0,a0,p0;
 4252+      STBIR_NO_UNROLL(decode);
 4253+      stbir__simdf8_load( d0, decode-8 );
 4254+      stbir__simdf8_0123to11331133( p0, d0 );
 4255+      stbir__simdf8_0123to00220022( a0, d0 );
 4256+      stbir__simdf8_mult( p0, p0, a0 );
 4257+
 4258+      stbir__simdf_store2( out, stbir__if_simdf8_cast_to_simdf4( d0 ) );
 4259+      stbir__simdf_store( out+2, stbir__if_simdf8_cast_to_simdf4( p0 ) );
 4260+      stbir__simdf_store2h( out+3, stbir__if_simdf8_cast_to_simdf4( d0 ) );
 4261+
 4262+      stbir__simdf_store2( out+6, stbir__simdf8_gettop4( d0 ) );
 4263+      stbir__simdf_store( out+8, stbir__simdf8_gettop4( p0 ) );
 4264+      stbir__simdf_store2h( out+9, stbir__simdf8_gettop4( d0 ) );
 4265+      #else
 4266+      stbir__simdf d0,a0,d1,a1,p0,p1;
 4267+      STBIR_NO_UNROLL(decode);
 4268+      stbir__simdf_load( d0, decode-8 );
 4269+      stbir__simdf_load( d1, decode-8+4 );
 4270+      stbir__simdf_0123to1133( p0, d0 );
 4271+      stbir__simdf_0123to1133( p1, d1 );
 4272+      stbir__simdf_0123to0022( a0, d0 );
 4273+      stbir__simdf_0123to0022( a1, d1 );
 4274+      stbir__simdf_mult( p0, p0, a0 );
 4275+      stbir__simdf_mult( p1, p1, a1 );
 4276+
 4277+      stbir__simdf_store2( out, d0 );
 4278+      stbir__simdf_store( out+2, p0 );
 4279+      stbir__simdf_store2h( out+3, d0 );
 4280+
 4281+      stbir__simdf_store2( out+6, d1 );
 4282+      stbir__simdf_store( out+8, p1 );
 4283+      stbir__simdf_store2h( out+9, d1 );
 4284+      #endif
 4285+      decode += 8;
 4286+      out += 12;
 4287+    } while ( decode <= end_decode );
 4288+  }
 4289+  decode -= 8;
 4290+  #endif
 4291+
 4292+  STBIR_SIMD_NO_UNROLL_LOOP_START
 4293+  while( decode < end_decode )
 4294+  {
 4295+    float x = decode[0], y = decode[1];
 4296+    STBIR_SIMD_NO_UNROLL(decode);
 4297+    out[0] = x;
 4298+    out[1] = y;
 4299+    out[2] = x * y;
 4300+    out += 3;
 4301+    decode += 2;
 4302+  }
 4303+}
 4304+
 4305+static void stbir__fancy_alpha_unweight_4ch( float * encode_buffer, int width_times_channels )
 4306+{
 4307+  float STBIR_SIMD_STREAMOUT_PTR(*) encode = encode_buffer;
 4308+  float STBIR_SIMD_STREAMOUT_PTR(*) input = encode_buffer;
 4309+  float const * end_output = encode_buffer + width_times_channels;
 4310+
 4311+  // fancy RGBA is stored internally as R G B A Rpm Gpm Bpm
 4312+
 4313+  STBIR_SIMD_NO_UNROLL_LOOP_START
 4314+  do {
 4315+    float alpha = input[3];
 4316+#ifdef STBIR_SIMD
 4317+    stbir__simdf i,ia;
 4318+    STBIR_SIMD_NO_UNROLL(encode);
 4319+    if ( alpha < stbir__small_float )
 4320+    {
 4321+      stbir__simdf_load( i, input );
 4322+      stbir__simdf_store( encode, i );
 4323+    }
 4324+    else
 4325+    {
 4326+      stbir__simdf_load1frep4( ia, 1.0f / alpha );
 4327+      stbir__simdf_load( i, input+4 );
 4328+      stbir__simdf_mult( i, i, ia );
 4329+      stbir__simdf_store( encode, i );
 4330+      encode[3] = alpha;
 4331+    }
 4332+#else
 4333+    if ( alpha < stbir__small_float )
 4334+    {
 4335+      encode[0] = input[0];
 4336+      encode[1] = input[1];
 4337+      encode[2] = input[2];
 4338+    }
 4339+    else
 4340+    {
 4341+      float ialpha = 1.0f / alpha;
 4342+      encode[0] = input[4] * ialpha;
 4343+      encode[1] = input[5] * ialpha;
 4344+      encode[2] = input[6] * ialpha;
 4345+    }
 4346+    encode[3] = alpha;
 4347+#endif
 4348+
 4349+    input += 7;
 4350+    encode += 4;
 4351+  } while ( encode < end_output );
 4352+}
 4353+
 4354+//  format: [X A Xpm][X A Xpm] etc
 4355+static void stbir__fancy_alpha_unweight_2ch( float * encode_buffer, int width_times_channels )
 4356+{
 4357+  float STBIR_SIMD_STREAMOUT_PTR(*) encode = encode_buffer;
 4358+  float STBIR_SIMD_STREAMOUT_PTR(*) input = encode_buffer;
 4359+  float const * end_output = encode_buffer + width_times_channels;
 4360+
 4361+  do {
 4362+    float alpha = input[1];
 4363+    encode[0] = input[0];
 4364+    if ( alpha >= stbir__small_float )
 4365+      encode[0] = input[2] / alpha;
 4366+    encode[1] = alpha;
 4367+
 4368+    input += 3;
 4369+    encode += 2;
 4370+  } while ( encode < end_output );
 4371+}
 4372+
 4373+static void stbir__simple_alpha_weight_4ch( float * decode_buffer, int width_times_channels )
 4374+{
 4375+  float STBIR_STREAMOUT_PTR(*) decode = decode_buffer;
 4376+  float const * end_decode = decode_buffer + width_times_channels;
 4377+
 4378+  #ifdef STBIR_SIMD
 4379+  {
 4380+    decode += 2 * stbir__simdfX_float_count;
 4381+    STBIR_NO_UNROLL_LOOP_START
 4382+    while ( decode <= end_decode )
 4383+    {
 4384+      stbir__simdfX d0,a0,d1,a1;
 4385+      STBIR_NO_UNROLL(decode);
 4386+      stbir__simdfX_load( d0, decode-2*stbir__simdfX_float_count );
 4387+      stbir__simdfX_load( d1, decode-2*stbir__simdfX_float_count+stbir__simdfX_float_count );
 4388+      stbir__simdfX_aaa1( a0, d0, STBIR_onesX );
 4389+      stbir__simdfX_aaa1( a1, d1, STBIR_onesX );
 4390+      stbir__simdfX_mult( d0, d0, a0 );
 4391+      stbir__simdfX_mult( d1, d1, a1 );
 4392+      stbir__simdfX_store ( decode-2*stbir__simdfX_float_count, d0 );
 4393+      stbir__simdfX_store ( decode-2*stbir__simdfX_float_count+stbir__simdfX_float_count, d1 );
 4394+      decode += 2 * stbir__simdfX_float_count;
 4395+    }
 4396+    decode -= 2 * stbir__simdfX_float_count;
 4397+
 4398+    // few last pixels remnants
 4399+    #ifdef STBIR_SIMD8
 4400+    STBIR_NO_UNROLL_LOOP_START
 4401+    while ( decode < end_decode )
 4402+    #else
 4403+    if ( decode < end_decode )
 4404+    #endif
 4405+    {
 4406+      stbir__simdf d,a;
 4407+      stbir__simdf_load( d, decode );
 4408+      stbir__simdf_aaa1( a, d, STBIR__CONSTF(STBIR_ones) );
 4409+      stbir__simdf_mult( d, d, a );
 4410+      stbir__simdf_store ( decode, d );
 4411+      decode += 4;
 4412+    }
 4413+  }
 4414+
 4415+  #else
 4416+
 4417+  while( decode < end_decode )
 4418+  {
 4419+    float alpha = decode[3];
 4420+    decode[0] *= alpha;
 4421+    decode[1] *= alpha;
 4422+    decode[2] *= alpha;
 4423+    decode += 4;
 4424+  }
 4425+
 4426+  #endif
 4427+}
 4428+
 4429+static void stbir__simple_alpha_weight_2ch( float * decode_buffer, int width_times_channels )
 4430+{
 4431+  float STBIR_STREAMOUT_PTR(*) decode = decode_buffer;
 4432+  float const * end_decode = decode_buffer + width_times_channels;
 4433+
 4434+  #ifdef STBIR_SIMD
 4435+  decode += 2 * stbir__simdfX_float_count;
 4436+  STBIR_NO_UNROLL_LOOP_START
 4437+  while ( decode <= end_decode )
 4438+  {
 4439+    stbir__simdfX d0,a0,d1,a1;
 4440+    STBIR_NO_UNROLL(decode);
 4441+    stbir__simdfX_load( d0, decode-2*stbir__simdfX_float_count );
 4442+    stbir__simdfX_load( d1, decode-2*stbir__simdfX_float_count+stbir__simdfX_float_count );
 4443+    stbir__simdfX_a1a1( a0, d0, STBIR_onesX );
 4444+    stbir__simdfX_a1a1( a1, d1, STBIR_onesX );
 4445+    stbir__simdfX_mult( d0, d0, a0 );
 4446+    stbir__simdfX_mult( d1, d1, a1 );
 4447+    stbir__simdfX_store ( decode-2*stbir__simdfX_float_count, d0 );
 4448+    stbir__simdfX_store ( decode-2*stbir__simdfX_float_count+stbir__simdfX_float_count, d1 );
 4449+    decode += 2 * stbir__simdfX_float_count;
 4450+  }
 4451+  decode -= 2 * stbir__simdfX_float_count;
 4452+  #endif
 4453+
 4454+  STBIR_SIMD_NO_UNROLL_LOOP_START
 4455+  while( decode < end_decode )
 4456+  {
 4457+    float alpha = decode[1];
 4458+    STBIR_SIMD_NO_UNROLL(decode);
 4459+    decode[0] *= alpha;
 4460+    decode += 2;
 4461+  }
 4462+}
 4463+
 4464+static void stbir__simple_alpha_unweight_4ch( float * encode_buffer, int width_times_channels )
 4465+{
 4466+  float STBIR_SIMD_STREAMOUT_PTR(*) encode = encode_buffer;
 4467+  float const * end_output = encode_buffer + width_times_channels;
 4468+
 4469+  STBIR_SIMD_NO_UNROLL_LOOP_START
 4470+  do {
 4471+    float alpha = encode[3];
 4472+
 4473+#ifdef STBIR_SIMD
 4474+    stbir__simdf i,ia;
 4475+    STBIR_SIMD_NO_UNROLL(encode);
 4476+    if ( alpha >= stbir__small_float )
 4477+    {
 4478+      stbir__simdf_load1frep4( ia, 1.0f / alpha );
 4479+      stbir__simdf_load( i, encode );
 4480+      stbir__simdf_mult( i, i, ia );
 4481+      stbir__simdf_store( encode, i );
 4482+      encode[3] = alpha;
 4483+    }
 4484+#else
 4485+    if ( alpha >= stbir__small_float )
 4486+    {
 4487+      float ialpha = 1.0f / alpha;
 4488+      encode[0] *= ialpha;
 4489+      encode[1] *= ialpha;
 4490+      encode[2] *= ialpha;
 4491+    }
 4492+#endif
 4493+    encode += 4;
 4494+  } while ( encode < end_output );
 4495+}
 4496+
 4497+static void stbir__simple_alpha_unweight_2ch( float * encode_buffer, int width_times_channels )
 4498+{
 4499+  float STBIR_SIMD_STREAMOUT_PTR(*) encode = encode_buffer;
 4500+  float const * end_output = encode_buffer + width_times_channels;
 4501+
 4502+  do {
 4503+    float alpha = encode[1];
 4504+    if ( alpha >= stbir__small_float )
 4505+      encode[0] /= alpha;
 4506+    encode += 2;
 4507+  } while ( encode < end_output );
 4508+}
 4509+
 4510+
 4511+// only used in RGB->BGR or BGR->RGB
 4512+static void stbir__simple_flip_3ch( float * decode_buffer, int width_times_channels )
 4513+{
 4514+  float STBIR_STREAMOUT_PTR(*) decode = decode_buffer;
 4515+  float const * end_decode = decode_buffer + width_times_channels;
 4516+
 4517+#ifdef STBIR_SIMD
 4518+    #ifdef stbir__simdf_swiz2 // do we have two argument swizzles?
 4519+      end_decode -= 12; 
 4520+      STBIR_NO_UNROLL_LOOP_START
 4521+      while( decode <= end_decode )
 4522+      {
 4523+        // on arm64 8 instructions, no overlapping stores
 4524+        stbir__simdf a,b,c,na,nb;
 4525+        STBIR_SIMD_NO_UNROLL(decode);
 4526+        stbir__simdf_load( a, decode );
 4527+        stbir__simdf_load( b, decode+4 );
 4528+        stbir__simdf_load( c, decode+8 );
 4529+
 4530+        na = stbir__simdf_swiz2( a, b, 2, 1, 0, 5 );   
 4531+        b  = stbir__simdf_swiz2( a, b, 4, 3, 6, 7 );   
 4532+        nb = stbir__simdf_swiz2( b, c, 0, 1, 4, 3 );   
 4533+        c  = stbir__simdf_swiz2( b, c, 2, 7, 6, 5 );   
 4534+
 4535+        stbir__simdf_store( decode, na );
 4536+        stbir__simdf_store( decode+4, nb ); 
 4537+        stbir__simdf_store( decode+8, c );
 4538+        decode += 12;
 4539+      }
 4540+      end_decode += 12;
 4541+    #else
 4542+      end_decode -= 24;
 4543+      STBIR_NO_UNROLL_LOOP_START
 4544+      while( decode <= end_decode )
 4545+      {
 4546+        // 26 instructions on x64
 4547+        stbir__simdf a,b,c,d,e,f,g;
 4548+        float i21, i23;
 4549+        STBIR_SIMD_NO_UNROLL(decode);
 4550+        stbir__simdf_load( a, decode );
 4551+        stbir__simdf_load( b, decode+3 );
 4552+        stbir__simdf_load( c, decode+6 );
 4553+        stbir__simdf_load( d, decode+9 );
 4554+        stbir__simdf_load( e, decode+12 );
 4555+        stbir__simdf_load( f, decode+15 );
 4556+        stbir__simdf_load( g, decode+18 );
 4557+
 4558+        a = stbir__simdf_swiz( a, 2, 1, 0, 3 );   
 4559+        b = stbir__simdf_swiz( b, 2, 1, 0, 3 );   
 4560+        c = stbir__simdf_swiz( c, 2, 1, 0, 3 );   
 4561+        d = stbir__simdf_swiz( d, 2, 1, 0, 3 );   
 4562+        e = stbir__simdf_swiz( e, 2, 1, 0, 3 );   
 4563+        f = stbir__simdf_swiz( f, 2, 1, 0, 3 );   
 4564+        g = stbir__simdf_swiz( g, 2, 1, 0, 3 );   
 4565+
 4566+        // stores overlap, need to be in order, 
 4567+        stbir__simdf_store( decode,    a );
 4568+        i21 = decode[21];
 4569+        stbir__simdf_store( decode+3,  b ); 
 4570+        i23 = decode[23];
 4571+        stbir__simdf_store( decode+6,  c );
 4572+        stbir__simdf_store( decode+9,  d );
 4573+        stbir__simdf_store( decode+12, e );
 4574+        stbir__simdf_store( decode+15, f );
 4575+        stbir__simdf_store( decode+18, g );
 4576+        decode[21] = i23;
 4577+        decode[23] = i21;
 4578+        decode += 24;
 4579+      }
 4580+      end_decode += 24;
 4581+    #endif
 4582+#else
 4583+  end_decode -= 12;
 4584+  STBIR_NO_UNROLL_LOOP_START
 4585+  while( decode <= end_decode )
 4586+  {
 4587+    // 16 instructions
 4588+    float t0,t1,t2,t3;
 4589+    STBIR_NO_UNROLL(decode);
 4590+    t0 = decode[0]; t1 = decode[3]; t2 = decode[6]; t3 = decode[9];
 4591+    decode[0] = decode[2]; decode[3] = decode[5]; decode[6] = decode[8]; decode[9] = decode[11];
 4592+    decode[2] = t0; decode[5] = t1; decode[8] = t2; decode[11] = t3;
 4593+    decode += 12;
 4594+  }
 4595+  end_decode += 12;
 4596+#endif
 4597+
 4598+  STBIR_NO_UNROLL_LOOP_START
 4599+  while( decode < end_decode )
 4600+  {
 4601+    float t = decode[0];
 4602+    STBIR_NO_UNROLL(decode);
 4603+    decode[0] = decode[2];
 4604+    decode[2] = t;
 4605+    decode += 3;
 4606+  }
 4607+}
 4608+
 4609+
 4610+
 4611+static void stbir__decode_scanline(stbir__info const * stbir_info, int n, float * output_buffer STBIR_ONLY_PROFILE_GET_SPLIT_INFO )
 4612+{
 4613+  int channels = stbir_info->channels;
 4614+  int effective_channels = stbir_info->effective_channels;
 4615+  int input_sample_in_bytes = stbir__type_size[stbir_info->input_type] * channels;
 4616+  stbir_edge edge_horizontal = stbir_info->horizontal.edge;
 4617+  stbir_edge edge_vertical = stbir_info->vertical.edge;
 4618+  int row = stbir__edge_wrap(edge_vertical, n, stbir_info->vertical.scale_info.input_full_size);
 4619+  const void* input_plane_data = ( (char *) stbir_info->input_data ) + (size_t)row * (size_t) stbir_info->input_stride_bytes;
 4620+  stbir__span const * spans = stbir_info->scanline_extents.spans;
 4621+  float * full_decode_buffer = output_buffer - stbir_info->scanline_extents.conservative.n0 * effective_channels;
 4622+  float * last_decoded = 0;
 4623+
 4624+  // if we are on edge_zero, and we get in here with an out of bounds n, then the calculate filters has failed
 4625+  STBIR_ASSERT( !(edge_vertical == STBIR_EDGE_ZERO && (n < 0 || n >= stbir_info->vertical.scale_info.input_full_size)) );
 4626+
 4627+  do
 4628+  {
 4629+    float * decode_buffer;
 4630+    void const * input_data;
 4631+    float * end_decode;
 4632+    int width_times_channels;
 4633+    int width;
 4634+
 4635+    if ( spans->n1 < spans->n0 )
 4636+      break;
 4637+
 4638+    width = spans->n1 + 1 - spans->n0;
 4639+    decode_buffer = full_decode_buffer + spans->n0 * effective_channels;
 4640+    end_decode = full_decode_buffer + ( spans->n1 + 1 ) * effective_channels;
 4641+    width_times_channels = width * channels;
 4642+
 4643+    // read directly out of input plane by default
 4644+    input_data = ( (char*)input_plane_data ) + spans->pixel_offset_for_input * input_sample_in_bytes;
 4645+
 4646+    // if we have an input callback, call it to get the input data
 4647+    if ( stbir_info->in_pixels_cb )
 4648+    {
 4649+      // call the callback with a temp buffer (that they can choose to use or not).  the temp is just right aligned memory in the decode_buffer itself
 4650+      input_data = stbir_info->in_pixels_cb( ( (char*) end_decode ) - ( width * input_sample_in_bytes ) + ( ( stbir_info->input_type != STBIR_TYPE_FLOAT ) ? ( sizeof(float)*STBIR_INPUT_CALLBACK_PADDING ) : 0 ), input_plane_data, width, spans->pixel_offset_for_input, row, stbir_info->user_data );
 4651+    }
 4652+
 4653+    STBIR_PROFILE_START( decode );
 4654+    // convert the pixels info the float decode_buffer, (we index from end_decode, so that when channels<effective_channels, we are right justified in the buffer)
 4655+    last_decoded = stbir_info->decode_pixels( (float*)end_decode - width_times_channels, width_times_channels, input_data );
 4656+    STBIR_PROFILE_END( decode );
 4657+
 4658+    if (stbir_info->alpha_weight)
 4659+    {
 4660+      STBIR_PROFILE_START( alpha );
 4661+      stbir_info->alpha_weight( decode_buffer, width_times_channels );
 4662+      STBIR_PROFILE_END( alpha );
 4663+    }
 4664+
 4665+    ++spans;
 4666+  } while ( spans <= ( &stbir_info->scanline_extents.spans[1] ) );
 4667+
 4668+  // handle the edge_wrap filter (all other types are handled back out at the calculate_filter stage)
 4669+  // basically the idea here is that if we have the whole scanline in memory, we don't redecode the
 4670+  //   wrapped edge pixels, and instead just memcpy them from the scanline into the edge positions
 4671+  if ( ( edge_horizontal == STBIR_EDGE_WRAP ) && ( stbir_info->scanline_extents.edge_sizes[0] | stbir_info->scanline_extents.edge_sizes[1] ) )
 4672+  {
 4673+    // this code only runs if we're in edge_wrap, and we're doing the entire scanline
 4674+    int e, start_x[2];
 4675+    int input_full_size = stbir_info->horizontal.scale_info.input_full_size;
 4676+
 4677+    start_x[0] = -stbir_info->scanline_extents.edge_sizes[0];  // left edge start x
 4678+    start_x[1] =  input_full_size;                             // right edge
 4679+
 4680+    for( e = 0; e < 2 ; e++ )
 4681+    {
 4682+      // do each margin
 4683+      int margin = stbir_info->scanline_extents.edge_sizes[e];
 4684+      if ( margin )
 4685+      {
 4686+        int x = start_x[e];
 4687+        float * marg = full_decode_buffer + x * effective_channels;
 4688+        float const * src = full_decode_buffer + stbir__edge_wrap(edge_horizontal, x, input_full_size) * effective_channels;
 4689+        STBIR_MEMCPY( marg, src, margin * effective_channels * sizeof(float) );
 4690+        if ( e == 1 ) last_decoded = marg + margin * effective_channels;
 4691+      }
 4692+    }
 4693+  }
 4694+  
 4695+  // some of the horizontal gathers read one float off the edge (which is masked out), but we force a zero here to make sure no NaNs leak in
 4696+  //   (we can't pre-zero it, because the input callback can use that area as padding)
 4697+  last_decoded[0] = 0.0f; 
 4698+
 4699+  // we clear this extra float, because the final output pixel filter kernel might have used one less coeff than the max filter width
 4700+  //   when this happens, we do read that pixel from the input, so it too could be Nan, so just zero an extra one.
 4701+  //   this fits because each scanline is padded by three floats (STBIR_INPUT_CALLBACK_PADDING)
 4702+  last_decoded[1] = 0.0f;
 4703+}
 4704+
 4705+
 4706+//=================
 4707+// Do 1 channel horizontal routines
 4708+
 4709+#ifdef STBIR_SIMD
 4710+
 4711+#define stbir__1_coeff_only()          \
 4712+    stbir__simdf tot,c;                \
 4713+    STBIR_SIMD_NO_UNROLL(decode);      \
 4714+    stbir__simdf_load1( c, hc );       \
 4715+    stbir__simdf_mult1_mem( tot, c, decode );
 4716+
 4717+#define stbir__2_coeff_only()          \
 4718+    stbir__simdf tot,c,d;              \
 4719+    STBIR_SIMD_NO_UNROLL(decode);      \
 4720+    stbir__simdf_load2z( c, hc );      \
 4721+    stbir__simdf_load2( d, decode );   \
 4722+    stbir__simdf_mult( tot, c, d );    \
 4723+    stbir__simdf_0123to1230( c, tot ); \
 4724+    stbir__simdf_add1( tot, tot, c );
 4725+
 4726+#define stbir__3_coeff_only()                  \
 4727+    stbir__simdf tot,c,t;                      \
 4728+    STBIR_SIMD_NO_UNROLL(decode);              \
 4729+    stbir__simdf_load( c, hc );                \
 4730+    stbir__simdf_mult_mem( tot, c, decode );   \
 4731+    stbir__simdf_0123to1230( c, tot );         \
 4732+    stbir__simdf_0123to2301( t, tot );         \
 4733+    stbir__simdf_add1( tot, tot, c );          \
 4734+    stbir__simdf_add1( tot, tot, t );
 4735+
 4736+#define stbir__store_output_tiny()                \
 4737+    stbir__simdf_store1( output, tot );           \
 4738+    horizontal_coefficients += coefficient_width; \
 4739+    ++horizontal_contributors;                    \
 4740+    output += 1;
 4741+
 4742+#define stbir__4_coeff_start()                 \
 4743+    stbir__simdf tot,c;                        \
 4744+    STBIR_SIMD_NO_UNROLL(decode);              \
 4745+    stbir__simdf_load( c, hc );                \
 4746+    stbir__simdf_mult_mem( tot, c, decode );   \
 4747+
 4748+#define stbir__4_coeff_continue_from_4( ofs )  \
 4749+    STBIR_SIMD_NO_UNROLL(decode);              \
 4750+    stbir__simdf_load( c, hc + (ofs) );        \
 4751+    stbir__simdf_madd_mem( tot, tot, c, decode+(ofs) );
 4752+
 4753+#define stbir__1_coeff_remnant( ofs )          \
 4754+    { stbir__simdf d;                          \
 4755+    stbir__simdf_load1z( c, hc + (ofs) );      \
 4756+    stbir__simdf_load1( d, decode + (ofs) );   \
 4757+    stbir__simdf_madd( tot, tot, d, c ); }
 4758+
 4759+#define stbir__2_coeff_remnant( ofs )          \
 4760+    { stbir__simdf d;                          \
 4761+    stbir__simdf_load2z( c, hc+(ofs) );        \
 4762+    stbir__simdf_load2( d, decode+(ofs) );     \
 4763+    stbir__simdf_madd( tot, tot, d, c ); }
 4764+
 4765+#define stbir__3_coeff_setup()                 \
 4766+    stbir__simdf mask;                         \
 4767+    stbir__simdf_load( mask, STBIR_mask + 3 );
 4768+
 4769+#define stbir__3_coeff_remnant( ofs )                  \
 4770+    stbir__simdf_load( c, hc+(ofs) );                  \
 4771+    stbir__simdf_and( c, c, mask );                    \
 4772+    stbir__simdf_madd_mem( tot, tot, c, decode+(ofs) );
 4773+
 4774+#define stbir__store_output()                     \
 4775+    stbir__simdf_0123to2301( c, tot );            \
 4776+    stbir__simdf_add( tot, tot, c );              \
 4777+    stbir__simdf_0123to1230( c, tot );            \
 4778+    stbir__simdf_add1( tot, tot, c );             \
 4779+    stbir__simdf_store1( output, tot );           \
 4780+    horizontal_coefficients += coefficient_width; \
 4781+    ++horizontal_contributors;                    \
 4782+    output += 1;
 4783+
 4784+#else
 4785+
 4786+#define stbir__1_coeff_only()  \
 4787+    float tot;                 \
 4788+    tot = decode[0]*hc[0];
 4789+
 4790+#define stbir__2_coeff_only()  \
 4791+    float tot;                 \
 4792+    tot = decode[0] * hc[0];   \
 4793+    tot += decode[1] * hc[1];
 4794+
 4795+#define stbir__3_coeff_only()  \
 4796+    float tot;                 \
 4797+    tot = decode[0] * hc[0];   \
 4798+    tot += decode[1] * hc[1];  \
 4799+    tot += decode[2] * hc[2];
 4800+
 4801+#define stbir__store_output_tiny()                \
 4802+    output[0] = tot;                              \
 4803+    horizontal_coefficients += coefficient_width; \
 4804+    ++horizontal_contributors;                    \
 4805+    output += 1;
 4806+
 4807+#define stbir__4_coeff_start()  \
 4808+    float tot0,tot1,tot2,tot3;  \
 4809+    tot0 = decode[0] * hc[0];   \
 4810+    tot1 = decode[1] * hc[1];   \
 4811+    tot2 = decode[2] * hc[2];   \
 4812+    tot3 = decode[3] * hc[3];
 4813+
 4814+#define stbir__4_coeff_continue_from_4( ofs )  \
 4815+    tot0 += decode[0+(ofs)] * hc[0+(ofs)];     \
 4816+    tot1 += decode[1+(ofs)] * hc[1+(ofs)];     \
 4817+    tot2 += decode[2+(ofs)] * hc[2+(ofs)];     \
 4818+    tot3 += decode[3+(ofs)] * hc[3+(ofs)];
 4819+
 4820+#define stbir__1_coeff_remnant( ofs )        \
 4821+    tot0 += decode[0+(ofs)] * hc[0+(ofs)];
 4822+
 4823+#define stbir__2_coeff_remnant( ofs )        \
 4824+    tot0 += decode[0+(ofs)] * hc[0+(ofs)];   \
 4825+    tot1 += decode[1+(ofs)] * hc[1+(ofs)];   \
 4826+
 4827+#define stbir__3_coeff_remnant( ofs )        \
 4828+    tot0 += decode[0+(ofs)] * hc[0+(ofs)];   \
 4829+    tot1 += decode[1+(ofs)] * hc[1+(ofs)];   \
 4830+    tot2 += decode[2+(ofs)] * hc[2+(ofs)];
 4831+
 4832+#define stbir__store_output()                     \
 4833+    output[0] = (tot0+tot2)+(tot1+tot3);          \
 4834+    horizontal_coefficients += coefficient_width; \
 4835+    ++horizontal_contributors;                    \
 4836+    output += 1;
 4837+
 4838+#endif
 4839+
 4840+#define STBIR__horizontal_channels 1
 4841+#define STB_IMAGE_RESIZE_DO_HORIZONTALS
 4842+#include STBIR__HEADER_FILENAME
 4843+
 4844+
 4845+//=================
 4846+// Do 2 channel horizontal routines
 4847+
 4848+#ifdef STBIR_SIMD
 4849+
 4850+#define stbir__1_coeff_only()         \
 4851+    stbir__simdf tot,c,d;             \
 4852+    STBIR_SIMD_NO_UNROLL(decode);     \
 4853+    stbir__simdf_load1z( c, hc );     \
 4854+    stbir__simdf_0123to0011( c, c );  \
 4855+    stbir__simdf_load2( d, decode );  \
 4856+    stbir__simdf_mult( tot, d, c );
 4857+
 4858+#define stbir__2_coeff_only()         \
 4859+    stbir__simdf tot,c;               \
 4860+    STBIR_SIMD_NO_UNROLL(decode);     \
 4861+    stbir__simdf_load2( c, hc );      \
 4862+    stbir__simdf_0123to0011( c, c );  \
 4863+    stbir__simdf_mult_mem( tot, c, decode );
 4864+
 4865+#define stbir__3_coeff_only()                \
 4866+    stbir__simdf tot,c,cs,d;                 \
 4867+    STBIR_SIMD_NO_UNROLL(decode);            \
 4868+    stbir__simdf_load( cs, hc );             \
 4869+    stbir__simdf_0123to0011( c, cs );        \
 4870+    stbir__simdf_mult_mem( tot, c, decode ); \
 4871+    stbir__simdf_0123to2222( c, cs );        \
 4872+    stbir__simdf_load2z( d, decode+4 );      \
 4873+    stbir__simdf_madd( tot, tot, d, c );
 4874+
 4875+#define stbir__store_output_tiny()                \
 4876+    stbir__simdf_0123to2301( c, tot );            \
 4877+    stbir__simdf_add( tot, tot, c );              \
 4878+    stbir__simdf_store2( output, tot );           \
 4879+    horizontal_coefficients += coefficient_width; \
 4880+    ++horizontal_contributors;                    \
 4881+    output += 2;
 4882+
 4883+#ifdef STBIR_SIMD8
 4884+
 4885+#define stbir__4_coeff_start()                    \
 4886+    stbir__simdf8 tot0,c,cs;                      \
 4887+    STBIR_SIMD_NO_UNROLL(decode);                 \
 4888+    stbir__simdf8_load4b( cs, hc );               \
 4889+    stbir__simdf8_0123to00112233( c, cs );        \
 4890+    stbir__simdf8_mult_mem( tot0, c, decode );
 4891+
 4892+#define stbir__4_coeff_continue_from_4( ofs )        \
 4893+    STBIR_SIMD_NO_UNROLL(decode);                    \
 4894+    stbir__simdf8_load4b( cs, hc + (ofs) );          \
 4895+    stbir__simdf8_0123to00112233( c, cs );           \
 4896+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*2 );
 4897+
 4898+#define stbir__1_coeff_remnant( ofs )                \
 4899+    { stbir__simdf t,d;                              \
 4900+    stbir__simdf_load1z( t, hc + (ofs) );            \
 4901+    stbir__simdf_load2( d, decode + (ofs) * 2 );     \
 4902+    stbir__simdf_0123to0011( t, t );                 \
 4903+    stbir__simdf_mult( t, t, d );                    \
 4904+    stbir__simdf8_add4( tot0, tot0, t ); }
 4905+ 
 4906+#define stbir__2_coeff_remnant( ofs )                \
 4907+    { stbir__simdf t;                                \
 4908+    stbir__simdf_load2( t, hc + (ofs) );             \
 4909+    stbir__simdf_0123to0011( t, t );                 \
 4910+    stbir__simdf_mult_mem( t, t, decode+(ofs)*2 );   \
 4911+    stbir__simdf8_add4( tot0, tot0, t ); }
 4912+
 4913+#define stbir__3_coeff_remnant( ofs )                \
 4914+    { stbir__simdf8 d;                               \
 4915+    stbir__simdf8_load4b( cs, hc + (ofs) );          \
 4916+    stbir__simdf8_0123to00112233( c, cs );           \
 4917+    stbir__simdf8_load6z( d, decode+(ofs)*2 );       \
 4918+    stbir__simdf8_madd( tot0, tot0, c, d ); }
 4919+
 4920+#define stbir__store_output()                     \
 4921+    { stbir__simdf t,d;                           \
 4922+    stbir__simdf8_add4halves( t, stbir__if_simdf8_cast_to_simdf4(tot0), tot0 );    \
 4923+    stbir__simdf_0123to2301( d, t );              \
 4924+    stbir__simdf_add( t, t, d );                  \
 4925+    stbir__simdf_store2( output, t );             \
 4926+    horizontal_coefficients += coefficient_width; \
 4927+    ++horizontal_contributors;                    \
 4928+    output += 2; }
 4929+
 4930+#else
 4931+
 4932+#define stbir__4_coeff_start()                   \
 4933+    stbir__simdf tot0,tot1,c,cs;                 \
 4934+    STBIR_SIMD_NO_UNROLL(decode);                \
 4935+    stbir__simdf_load( cs, hc );                 \
 4936+    stbir__simdf_0123to0011( c, cs );            \
 4937+    stbir__simdf_mult_mem( tot0, c, decode );    \
 4938+    stbir__simdf_0123to2233( c, cs );            \
 4939+    stbir__simdf_mult_mem( tot1, c, decode+4 );
 4940+
 4941+#define stbir__4_coeff_continue_from_4( ofs )                \
 4942+    STBIR_SIMD_NO_UNROLL(decode);                            \
 4943+    stbir__simdf_load( cs, hc + (ofs) );                     \
 4944+    stbir__simdf_0123to0011( c, cs );                        \
 4945+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*2 );  \
 4946+    stbir__simdf_0123to2233( c, cs );                        \
 4947+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*2+4 );
 4948+
 4949+#define stbir__1_coeff_remnant( ofs )            \
 4950+    { stbir__simdf d;                            \
 4951+    stbir__simdf_load1z( cs, hc + (ofs) );       \
 4952+    stbir__simdf_0123to0011( c, cs );            \
 4953+    stbir__simdf_load2( d, decode + (ofs) * 2 ); \
 4954+    stbir__simdf_madd( tot0, tot0, d, c ); }
 4955+
 4956+#define stbir__2_coeff_remnant( ofs )                      \
 4957+    stbir__simdf_load2( cs, hc + (ofs) );                  \
 4958+    stbir__simdf_0123to0011( c, cs );                      \
 4959+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*2 );
 4960+
 4961+#define stbir__3_coeff_remnant( ofs )                       \
 4962+    { stbir__simdf d;                                       \
 4963+    stbir__simdf_load( cs, hc + (ofs) );                    \
 4964+    stbir__simdf_0123to0011( c, cs );                       \
 4965+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*2 ); \
 4966+    stbir__simdf_0123to2222( c, cs );                       \
 4967+    stbir__simdf_load2z( d, decode + (ofs) * 2 + 4 );       \
 4968+    stbir__simdf_madd( tot1, tot1, d, c ); }
 4969+
 4970+#define stbir__store_output()                     \
 4971+    stbir__simdf_add( tot0, tot0, tot1 );         \
 4972+    stbir__simdf_0123to2301( c, tot0 );           \
 4973+    stbir__simdf_add( tot0, tot0, c );            \
 4974+    stbir__simdf_store2( output, tot0 );          \
 4975+    horizontal_coefficients += coefficient_width; \
 4976+    ++horizontal_contributors;                    \
 4977+    output += 2;
 4978+
 4979+#endif
 4980+
 4981+#else
 4982+
 4983+#define stbir__1_coeff_only()  \
 4984+    float tota,totb,c;         \
 4985+    c = hc[0];                 \
 4986+    tota = decode[0]*c;        \
 4987+    totb = decode[1]*c;
 4988+
 4989+#define stbir__2_coeff_only()  \
 4990+    float tota,totb,c;         \
 4991+    c = hc[0];                 \
 4992+    tota = decode[0]*c;        \
 4993+    totb = decode[1]*c;        \
 4994+    c = hc[1];                 \
 4995+    tota += decode[2]*c;       \
 4996+    totb += decode[3]*c;
 4997+
 4998+// this weird order of add matches the simd
 4999+#define stbir__3_coeff_only()  \
 5000+    float tota,totb,c;         \
 5001+    c = hc[0];                 \
 5002+    tota = decode[0]*c;        \
 5003+    totb = decode[1]*c;        \
 5004+    c = hc[2];                 \
 5005+    tota += decode[4]*c;       \
 5006+    totb += decode[5]*c;       \
 5007+    c = hc[1];                 \
 5008+    tota += decode[2]*c;       \
 5009+    totb += decode[3]*c;
 5010+
 5011+#define stbir__store_output_tiny()                \
 5012+    output[0] = tota;                             \
 5013+    output[1] = totb;                             \
 5014+    horizontal_coefficients += coefficient_width; \
 5015+    ++horizontal_contributors;                    \
 5016+    output += 2;
 5017+
 5018+#define stbir__4_coeff_start()      \
 5019+    float tota0,tota1,tota2,tota3,totb0,totb1,totb2,totb3,c;  \
 5020+    c = hc[0];                      \
 5021+    tota0 = decode[0]*c;            \
 5022+    totb0 = decode[1]*c;            \
 5023+    c = hc[1];                      \
 5024+    tota1 = decode[2]*c;            \
 5025+    totb1 = decode[3]*c;            \
 5026+    c = hc[2];                      \
 5027+    tota2 = decode[4]*c;            \
 5028+    totb2 = decode[5]*c;            \
 5029+    c = hc[3];                      \
 5030+    tota3 = decode[6]*c;            \
 5031+    totb3 = decode[7]*c;
 5032+
 5033+#define stbir__4_coeff_continue_from_4( ofs )  \
 5034+    c = hc[0+(ofs)];                           \
 5035+    tota0 += decode[0+(ofs)*2]*c;              \
 5036+    totb0 += decode[1+(ofs)*2]*c;              \
 5037+    c = hc[1+(ofs)];                           \
 5038+    tota1 += decode[2+(ofs)*2]*c;              \
 5039+    totb1 += decode[3+(ofs)*2]*c;              \
 5040+    c = hc[2+(ofs)];                           \
 5041+    tota2 += decode[4+(ofs)*2]*c;              \
 5042+    totb2 += decode[5+(ofs)*2]*c;              \
 5043+    c = hc[3+(ofs)];                           \
 5044+    tota3 += decode[6+(ofs)*2]*c;              \
 5045+    totb3 += decode[7+(ofs)*2]*c;
 5046+
 5047+#define stbir__1_coeff_remnant( ofs )  \
 5048+    c = hc[0+(ofs)];                   \
 5049+    tota0 += decode[0+(ofs)*2] * c;    \
 5050+    totb0 += decode[1+(ofs)*2] * c;
 5051+
 5052+#define stbir__2_coeff_remnant( ofs )  \
 5053+    c = hc[0+(ofs)];                   \
 5054+    tota0 += decode[0+(ofs)*2] * c;    \
 5055+    totb0 += decode[1+(ofs)*2] * c;    \
 5056+    c = hc[1+(ofs)];                   \
 5057+    tota1 += decode[2+(ofs)*2] * c;    \
 5058+    totb1 += decode[3+(ofs)*2] * c;
 5059+
 5060+#define stbir__3_coeff_remnant( ofs )  \
 5061+    c = hc[0+(ofs)];                   \
 5062+    tota0 += decode[0+(ofs)*2] * c;    \
 5063+    totb0 += decode[1+(ofs)*2] * c;    \
 5064+    c = hc[1+(ofs)];                   \
 5065+    tota1 += decode[2+(ofs)*2] * c;    \
 5066+    totb1 += decode[3+(ofs)*2] * c;    \
 5067+    c = hc[2+(ofs)];                   \
 5068+    tota2 += decode[4+(ofs)*2] * c;    \
 5069+    totb2 += decode[5+(ofs)*2] * c;
 5070+
 5071+#define stbir__store_output()                     \
 5072+    output[0] = (tota0+tota2)+(tota1+tota3);      \
 5073+    output[1] = (totb0+totb2)+(totb1+totb3);      \
 5074+    horizontal_coefficients += coefficient_width; \
 5075+    ++horizontal_contributors;                    \
 5076+    output += 2;
 5077+
 5078+#endif
 5079+
 5080+#define STBIR__horizontal_channels 2
 5081+#define STB_IMAGE_RESIZE_DO_HORIZONTALS
 5082+#include STBIR__HEADER_FILENAME
 5083+
 5084+
 5085+//=================
 5086+// Do 3 channel horizontal routines
 5087+
 5088+#ifdef STBIR_SIMD
 5089+
 5090+#define stbir__1_coeff_only()         \
 5091+    stbir__simdf tot,c,d;             \
 5092+    STBIR_SIMD_NO_UNROLL(decode);     \
 5093+    stbir__simdf_load1z( c, hc );     \
 5094+    stbir__simdf_0123to0001( c, c );  \
 5095+    stbir__simdf_load( d, decode );   \
 5096+    stbir__simdf_mult( tot, d, c );
 5097+
 5098+#define stbir__2_coeff_only()         \
 5099+    stbir__simdf tot,c,cs,d;          \
 5100+    STBIR_SIMD_NO_UNROLL(decode);     \
 5101+    stbir__simdf_load2( cs, hc );     \
 5102+    stbir__simdf_0123to0000( c, cs ); \
 5103+    stbir__simdf_load( d, decode );   \
 5104+    stbir__simdf_mult( tot, d, c );   \
 5105+    stbir__simdf_0123to1111( c, cs ); \
 5106+    stbir__simdf_load( d, decode+3 ); \
 5107+    stbir__simdf_madd( tot, tot, d, c );
 5108+
 5109+#define stbir__3_coeff_only()            \
 5110+    stbir__simdf tot,c,d,cs;             \
 5111+    STBIR_SIMD_NO_UNROLL(decode);        \
 5112+    stbir__simdf_load( cs, hc );         \
 5113+    stbir__simdf_0123to0000( c, cs );    \
 5114+    stbir__simdf_load( d, decode );      \
 5115+    stbir__simdf_mult( tot, d, c );      \
 5116+    stbir__simdf_0123to1111( c, cs );    \
 5117+    stbir__simdf_load( d, decode+3 );    \
 5118+    stbir__simdf_madd( tot, tot, d, c ); \
 5119+    stbir__simdf_0123to2222( c, cs );    \
 5120+    stbir__simdf_load( d, decode+6 );    \
 5121+    stbir__simdf_madd( tot, tot, d, c );
 5122+
 5123+#define stbir__store_output_tiny()                \
 5124+    stbir__simdf_store2( output, tot );           \
 5125+    stbir__simdf_0123to2301( tot, tot );          \
 5126+    stbir__simdf_store1( output+2, tot );         \
 5127+    horizontal_coefficients += coefficient_width; \
 5128+    ++horizontal_contributors;                    \
 5129+    output += 3;
 5130+
 5131+#ifdef STBIR_SIMD8
 5132+
 5133+// we're loading from the XXXYYY decode by -1 to get the XXXYYY into different halves of the AVX reg fyi
 5134+#define stbir__4_coeff_start()                     \
 5135+    stbir__simdf8 tot0,tot1,c,cs; stbir__simdf t;  \
 5136+    STBIR_SIMD_NO_UNROLL(decode);                  \
 5137+    stbir__simdf8_load4b( cs, hc );                \
 5138+    stbir__simdf8_0123to00001111( c, cs );         \
 5139+    stbir__simdf8_mult_mem( tot0, c, decode - 1 ); \
 5140+    stbir__simdf8_0123to22223333( c, cs );         \
 5141+    stbir__simdf8_mult_mem( tot1, c, decode+6 - 1 );
 5142+
 5143+#define stbir__4_coeff_continue_from_4( ofs )      \
 5144+    STBIR_SIMD_NO_UNROLL(decode);                  \
 5145+    stbir__simdf8_load4b( cs, hc + (ofs) );        \
 5146+    stbir__simdf8_0123to00001111( c, cs );         \
 5147+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*3 - 1 ); \
 5148+    stbir__simdf8_0123to22223333( c, cs );         \
 5149+    stbir__simdf8_madd_mem( tot1, tot1, c, decode+(ofs)*3 + 6 - 1 );
 5150+
 5151+#define stbir__1_coeff_remnant( ofs )                          \
 5152+    STBIR_SIMD_NO_UNROLL(decode);                              \
 5153+    stbir__simdf_load1rep4( t, hc + (ofs) );                   \
 5154+    stbir__simdf8_madd_mem4( tot0, tot0, t, decode+(ofs)*3 - 1 );
 5155+
 5156+#define stbir__2_coeff_remnant( ofs )                          \
 5157+    STBIR_SIMD_NO_UNROLL(decode);                              \
 5158+    stbir__simdf8_load4b( cs, hc + (ofs) - 2 );                \
 5159+    stbir__simdf8_0123to22223333( c, cs );                     \
 5160+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*3 - 1 );
 5161+
 5162+ #define stbir__3_coeff_remnant( ofs )                           \
 5163+    STBIR_SIMD_NO_UNROLL(decode);                                \
 5164+    stbir__simdf8_load4b( cs, hc + (ofs) );                      \
 5165+    stbir__simdf8_0123to00001111( c, cs );                       \
 5166+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*3 - 1 ); \
 5167+    stbir__simdf8_0123to2222( t, cs );                           \
 5168+    stbir__simdf8_madd_mem4( tot1, tot1, t, decode+(ofs)*3 + 6 - 1 );
 5169+
 5170+#define stbir__store_output()                       \
 5171+    stbir__simdf8_add( tot0, tot0, tot1 );          \
 5172+    stbir__simdf_0123to1230( t, stbir__if_simdf8_cast_to_simdf4( tot0 ) ); \
 5173+    stbir__simdf8_add4halves( t, t, tot0 );         \
 5174+    horizontal_coefficients += coefficient_width;   \
 5175+    ++horizontal_contributors;                      \
 5176+    output += 3;                                    \
 5177+    if ( output < output_end )                      \
 5178+    {                                               \
 5179+      stbir__simdf_store( output-3, t );            \
 5180+      continue;                                     \
 5181+    }                                               \
 5182+    { stbir__simdf tt; stbir__simdf_0123to2301( tt, t ); \
 5183+    stbir__simdf_store2( output-3, t );             \
 5184+    stbir__simdf_store1( output+2-3, tt ); }        \
 5185+    break;
 5186+
 5187+
 5188+#else
 5189+
 5190+#define stbir__4_coeff_start()                  \
 5191+    stbir__simdf tot0,tot1,tot2,c,cs;           \
 5192+    STBIR_SIMD_NO_UNROLL(decode);               \
 5193+    stbir__simdf_load( cs, hc );                \
 5194+    stbir__simdf_0123to0001( c, cs );           \
 5195+    stbir__simdf_mult_mem( tot0, c, decode );   \
 5196+    stbir__simdf_0123to1122( c, cs );           \
 5197+    stbir__simdf_mult_mem( tot1, c, decode+4 ); \
 5198+    stbir__simdf_0123to2333( c, cs );           \
 5199+    stbir__simdf_mult_mem( tot2, c, decode+8 );
 5200+
 5201+#define stbir__4_coeff_continue_from_4( ofs )                 \
 5202+    STBIR_SIMD_NO_UNROLL(decode);                             \
 5203+    stbir__simdf_load( cs, hc + (ofs) );                      \
 5204+    stbir__simdf_0123to0001( c, cs );                         \
 5205+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*3 );   \
 5206+    stbir__simdf_0123to1122( c, cs );                         \
 5207+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*3+4 ); \
 5208+    stbir__simdf_0123to2333( c, cs );                         \
 5209+    stbir__simdf_madd_mem( tot2, tot2, c, decode+(ofs)*3+8 );
 5210+
 5211+#define stbir__1_coeff_remnant( ofs )         \
 5212+    STBIR_SIMD_NO_UNROLL(decode);             \
 5213+    stbir__simdf_load1z( c, hc + (ofs) );     \
 5214+    stbir__simdf_0123to0001( c, c );          \
 5215+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*3 );
 5216+
 5217+#define stbir__2_coeff_remnant( ofs )                       \
 5218+    { stbir__simdf d;                                       \
 5219+    STBIR_SIMD_NO_UNROLL(decode);                           \
 5220+    stbir__simdf_load2z( cs, hc + (ofs) );                  \
 5221+    stbir__simdf_0123to0001( c, cs );                       \
 5222+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*3 ); \
 5223+    stbir__simdf_0123to1122( c, cs );                       \
 5224+    stbir__simdf_load2z( d, decode+(ofs)*3+4 );             \
 5225+    stbir__simdf_madd( tot1, tot1, c, d ); }
 5226+
 5227+#define stbir__3_coeff_remnant( ofs )                         \
 5228+    { stbir__simdf d;                                         \
 5229+    STBIR_SIMD_NO_UNROLL(decode);                             \
 5230+    stbir__simdf_load( cs, hc + (ofs) );                      \
 5231+    stbir__simdf_0123to0001( c, cs );                         \
 5232+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*3 );   \
 5233+    stbir__simdf_0123to1122( c, cs );                         \
 5234+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*3+4 ); \
 5235+    stbir__simdf_0123to2222( c, cs );                         \
 5236+    stbir__simdf_load1z( d, decode+(ofs)*3+8 );               \
 5237+    stbir__simdf_madd( tot2, tot2, c, d );  }
 5238+
 5239+#define stbir__store_output()                       \
 5240+    stbir__simdf_0123ABCDto3ABx( c, tot0, tot1 );   \
 5241+    stbir__simdf_0123ABCDto23Ax( cs, tot1, tot2 );  \
 5242+    stbir__simdf_0123to1230( tot2, tot2 );          \
 5243+    stbir__simdf_add( tot0, tot0, cs );             \
 5244+    stbir__simdf_add( c, c, tot2 );                 \
 5245+    stbir__simdf_add( tot0, tot0, c );              \
 5246+    horizontal_coefficients += coefficient_width;   \
 5247+    ++horizontal_contributors;                      \
 5248+    output += 3;                                    \
 5249+    if ( output < output_end )                      \
 5250+    {                                               \
 5251+      stbir__simdf_store( output-3, tot0 );         \
 5252+      continue;                                     \
 5253+    }                                               \
 5254+    stbir__simdf_0123to2301( tot1, tot0 );          \
 5255+    stbir__simdf_store2( output-3, tot0 );          \
 5256+    stbir__simdf_store1( output+2-3, tot1 );        \
 5257+    break;
 5258+
 5259+#endif
 5260+
 5261+#else
 5262+
 5263+#define stbir__1_coeff_only()  \
 5264+    float tot0, tot1, tot2, c; \
 5265+    c = hc[0];                 \
 5266+    tot0 = decode[0]*c;        \
 5267+    tot1 = decode[1]*c;        \
 5268+    tot2 = decode[2]*c;
 5269+
 5270+#define stbir__2_coeff_only()  \
 5271+    float tot0, tot1, tot2, c; \
 5272+    c = hc[0];                 \
 5273+    tot0 = decode[0]*c;        \
 5274+    tot1 = decode[1]*c;        \
 5275+    tot2 = decode[2]*c;        \
 5276+    c = hc[1];                 \
 5277+    tot0 += decode[3]*c;       \
 5278+    tot1 += decode[4]*c;       \
 5279+    tot2 += decode[5]*c;
 5280+
 5281+#define stbir__3_coeff_only()  \
 5282+    float tot0, tot1, tot2, c; \
 5283+    c = hc[0];                 \
 5284+    tot0 = decode[0]*c;        \
 5285+    tot1 = decode[1]*c;        \
 5286+    tot2 = decode[2]*c;        \
 5287+    c = hc[1];                 \
 5288+    tot0 += decode[3]*c;       \
 5289+    tot1 += decode[4]*c;       \
 5290+    tot2 += decode[5]*c;       \
 5291+    c = hc[2];                 \
 5292+    tot0 += decode[6]*c;       \
 5293+    tot1 += decode[7]*c;       \
 5294+    tot2 += decode[8]*c;
 5295+
 5296+#define stbir__store_output_tiny()                \
 5297+    output[0] = tot0;                             \
 5298+    output[1] = tot1;                             \
 5299+    output[2] = tot2;                             \
 5300+    horizontal_coefficients += coefficient_width; \
 5301+    ++horizontal_contributors;                    \
 5302+    output += 3;
 5303+
 5304+#define stbir__4_coeff_start()      \
 5305+    float tota0,tota1,tota2,totb0,totb1,totb2,totc0,totc1,totc2,totd0,totd1,totd2,c;  \
 5306+    c = hc[0];                      \
 5307+    tota0 = decode[0]*c;            \
 5308+    tota1 = decode[1]*c;            \
 5309+    tota2 = decode[2]*c;            \
 5310+    c = hc[1];                      \
 5311+    totb0 = decode[3]*c;            \
 5312+    totb1 = decode[4]*c;            \
 5313+    totb2 = decode[5]*c;            \
 5314+    c = hc[2];                      \
 5315+    totc0 = decode[6]*c;            \
 5316+    totc1 = decode[7]*c;            \
 5317+    totc2 = decode[8]*c;            \
 5318+    c = hc[3];                      \
 5319+    totd0 = decode[9]*c;            \
 5320+    totd1 = decode[10]*c;           \
 5321+    totd2 = decode[11]*c;
 5322+
 5323+#define stbir__4_coeff_continue_from_4( ofs )  \
 5324+    c = hc[0+(ofs)];                           \
 5325+    tota0 += decode[0+(ofs)*3]*c;              \
 5326+    tota1 += decode[1+(ofs)*3]*c;              \
 5327+    tota2 += decode[2+(ofs)*3]*c;              \
 5328+    c = hc[1+(ofs)];                           \
 5329+    totb0 += decode[3+(ofs)*3]*c;              \
 5330+    totb1 += decode[4+(ofs)*3]*c;              \
 5331+    totb2 += decode[5+(ofs)*3]*c;              \
 5332+    c = hc[2+(ofs)];                           \
 5333+    totc0 += decode[6+(ofs)*3]*c;              \
 5334+    totc1 += decode[7+(ofs)*3]*c;              \
 5335+    totc2 += decode[8+(ofs)*3]*c;              \
 5336+    c = hc[3+(ofs)];                           \
 5337+    totd0 += decode[9+(ofs)*3]*c;              \
 5338+    totd1 += decode[10+(ofs)*3]*c;             \
 5339+    totd2 += decode[11+(ofs)*3]*c;
 5340+
 5341+#define stbir__1_coeff_remnant( ofs )  \
 5342+    c = hc[0+(ofs)];                   \
 5343+    tota0 += decode[0+(ofs)*3]*c;      \
 5344+    tota1 += decode[1+(ofs)*3]*c;      \
 5345+    tota2 += decode[2+(ofs)*3]*c;
 5346+
 5347+#define stbir__2_coeff_remnant( ofs )  \
 5348+    c = hc[0+(ofs)];                   \
 5349+    tota0 += decode[0+(ofs)*3]*c;      \
 5350+    tota1 += decode[1+(ofs)*3]*c;      \
 5351+    tota2 += decode[2+(ofs)*3]*c;      \
 5352+    c = hc[1+(ofs)];                   \
 5353+    totb0 += decode[3+(ofs)*3]*c;      \
 5354+    totb1 += decode[4+(ofs)*3]*c;      \
 5355+    totb2 += decode[5+(ofs)*3]*c;      \
 5356+
 5357+#define stbir__3_coeff_remnant( ofs )  \
 5358+    c = hc[0+(ofs)];                   \
 5359+    tota0 += decode[0+(ofs)*3]*c;      \
 5360+    tota1 += decode[1+(ofs)*3]*c;      \
 5361+    tota2 += decode[2+(ofs)*3]*c;      \
 5362+    c = hc[1+(ofs)];                   \
 5363+    totb0 += decode[3+(ofs)*3]*c;      \
 5364+    totb1 += decode[4+(ofs)*3]*c;      \
 5365+    totb2 += decode[5+(ofs)*3]*c;      \
 5366+    c = hc[2+(ofs)];                   \
 5367+    totc0 += decode[6+(ofs)*3]*c;      \
 5368+    totc1 += decode[7+(ofs)*3]*c;      \
 5369+    totc2 += decode[8+(ofs)*3]*c;
 5370+
 5371+#define stbir__store_output()                     \
 5372+    output[0] = (tota0+totc0)+(totb0+totd0);      \
 5373+    output[1] = (tota1+totc1)+(totb1+totd1);      \
 5374+    output[2] = (tota2+totc2)+(totb2+totd2);      \
 5375+    horizontal_coefficients += coefficient_width; \
 5376+    ++horizontal_contributors;                    \
 5377+    output += 3;
 5378+
 5379+#endif
 5380+
 5381+#define STBIR__horizontal_channels 3
 5382+#define STB_IMAGE_RESIZE_DO_HORIZONTALS
 5383+#include STBIR__HEADER_FILENAME
 5384+
 5385+//=================
 5386+// Do 4 channel horizontal routines
 5387+
 5388+#ifdef STBIR_SIMD
 5389+
 5390+#define stbir__1_coeff_only()             \
 5391+    stbir__simdf tot,c;                   \
 5392+    STBIR_SIMD_NO_UNROLL(decode);         \
 5393+    stbir__simdf_load1( c, hc );          \
 5394+    stbir__simdf_0123to0000( c, c );      \
 5395+    stbir__simdf_mult_mem( tot, c, decode );
 5396+
 5397+#define stbir__2_coeff_only()                       \
 5398+    stbir__simdf tot,c,cs;                          \
 5399+    STBIR_SIMD_NO_UNROLL(decode);                   \
 5400+    stbir__simdf_load2( cs, hc );                   \
 5401+    stbir__simdf_0123to0000( c, cs );               \
 5402+    stbir__simdf_mult_mem( tot, c, decode );        \
 5403+    stbir__simdf_0123to1111( c, cs );               \
 5404+    stbir__simdf_madd_mem( tot, tot, c, decode+4 );
 5405+
 5406+#define stbir__3_coeff_only()                       \
 5407+    stbir__simdf tot,c,cs;                          \
 5408+    STBIR_SIMD_NO_UNROLL(decode);                   \
 5409+    stbir__simdf_load( cs, hc );                    \
 5410+    stbir__simdf_0123to0000( c, cs );               \
 5411+    stbir__simdf_mult_mem( tot, c, decode );        \
 5412+    stbir__simdf_0123to1111( c, cs );               \
 5413+    stbir__simdf_madd_mem( tot, tot, c, decode+4 ); \
 5414+    stbir__simdf_0123to2222( c, cs );               \
 5415+    stbir__simdf_madd_mem( tot, tot, c, decode+8 );
 5416+
 5417+#define stbir__store_output_tiny()                \
 5418+    stbir__simdf_store( output, tot );            \
 5419+    horizontal_coefficients += coefficient_width; \
 5420+    ++horizontal_contributors;                    \
 5421+    output += 4;
 5422+
 5423+#ifdef STBIR_SIMD8
 5424+
 5425+#define stbir__4_coeff_start()                     \
 5426+    stbir__simdf8 tot0,c,cs; stbir__simdf t;  \
 5427+    STBIR_SIMD_NO_UNROLL(decode);                  \
 5428+    stbir__simdf8_load4b( cs, hc );                \
 5429+    stbir__simdf8_0123to00001111( c, cs );         \
 5430+    stbir__simdf8_mult_mem( tot0, c, decode );     \
 5431+    stbir__simdf8_0123to22223333( c, cs );         \
 5432+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+8 );
 5433+
 5434+#define stbir__4_coeff_continue_from_4( ofs )                  \
 5435+    STBIR_SIMD_NO_UNROLL(decode);                              \
 5436+    stbir__simdf8_load4b( cs, hc + (ofs) );                    \
 5437+    stbir__simdf8_0123to00001111( c, cs );                     \
 5438+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*4 );   \
 5439+    stbir__simdf8_0123to22223333( c, cs );                     \
 5440+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*4+8 );
 5441+
 5442+#define stbir__1_coeff_remnant( ofs )                          \
 5443+    STBIR_SIMD_NO_UNROLL(decode);                              \
 5444+    stbir__simdf_load1rep4( t, hc + (ofs) );                   \
 5445+    stbir__simdf8_madd_mem4( tot0, tot0, t, decode+(ofs)*4 );
 5446+
 5447+#define stbir__2_coeff_remnant( ofs )                          \
 5448+    STBIR_SIMD_NO_UNROLL(decode);                              \
 5449+    stbir__simdf8_load4b( cs, hc + (ofs) - 2 );                \
 5450+    stbir__simdf8_0123to22223333( c, cs );                     \
 5451+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*4 );
 5452+
 5453+ #define stbir__3_coeff_remnant( ofs )                         \
 5454+    STBIR_SIMD_NO_UNROLL(decode);                              \
 5455+    stbir__simdf8_load4b( cs, hc + (ofs) );                    \
 5456+    stbir__simdf8_0123to00001111( c, cs );                     \
 5457+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*4 );   \
 5458+    stbir__simdf8_0123to2222( t, cs );                         \
 5459+    stbir__simdf8_madd_mem4( tot0, tot0, t, decode+(ofs)*4+8 );
 5460+
 5461+#define stbir__store_output()                      \
 5462+    stbir__simdf8_add4halves( t, stbir__if_simdf8_cast_to_simdf4(tot0), tot0 );     \
 5463+    stbir__simdf_store( output, t );               \
 5464+    horizontal_coefficients += coefficient_width;  \
 5465+    ++horizontal_contributors;                     \
 5466+    output += 4;
 5467+
 5468+#else
 5469+
 5470+#define stbir__4_coeff_start()                        \
 5471+    stbir__simdf tot0,tot1,c,cs;                      \
 5472+    STBIR_SIMD_NO_UNROLL(decode);                     \
 5473+    stbir__simdf_load( cs, hc );                      \
 5474+    stbir__simdf_0123to0000( c, cs );                 \
 5475+    stbir__simdf_mult_mem( tot0, c, decode );         \
 5476+    stbir__simdf_0123to1111( c, cs );                 \
 5477+    stbir__simdf_mult_mem( tot1, c, decode+4 );       \
 5478+    stbir__simdf_0123to2222( c, cs );                 \
 5479+    stbir__simdf_madd_mem( tot0, tot0, c, decode+8 ); \
 5480+    stbir__simdf_0123to3333( c, cs );                 \
 5481+    stbir__simdf_madd_mem( tot1, tot1, c, decode+12 );
 5482+
 5483+#define stbir__4_coeff_continue_from_4( ofs )                  \
 5484+    STBIR_SIMD_NO_UNROLL(decode);                              \
 5485+    stbir__simdf_load( cs, hc + (ofs) );                       \
 5486+    stbir__simdf_0123to0000( c, cs );                          \
 5487+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*4 );    \
 5488+    stbir__simdf_0123to1111( c, cs );                          \
 5489+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*4+4 );  \
 5490+    stbir__simdf_0123to2222( c, cs );                          \
 5491+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*4+8 );  \
 5492+    stbir__simdf_0123to3333( c, cs );                          \
 5493+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*4+12 );
 5494+
 5495+#define stbir__1_coeff_remnant( ofs )                       \
 5496+    STBIR_SIMD_NO_UNROLL(decode);                           \
 5497+    stbir__simdf_load1( c, hc + (ofs) );                    \
 5498+    stbir__simdf_0123to0000( c, c );                        \
 5499+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*4 );
 5500+
 5501+#define stbir__2_coeff_remnant( ofs )                         \
 5502+    STBIR_SIMD_NO_UNROLL(decode);                             \
 5503+    stbir__simdf_load2( cs, hc + (ofs) );                     \
 5504+    stbir__simdf_0123to0000( c, cs );                         \
 5505+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*4 );   \
 5506+    stbir__simdf_0123to1111( c, cs );                         \
 5507+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*4+4 );
 5508+
 5509+#define stbir__3_coeff_remnant( ofs )                          \
 5510+    STBIR_SIMD_NO_UNROLL(decode);                              \
 5511+    stbir__simdf_load( cs, hc + (ofs) );                       \
 5512+    stbir__simdf_0123to0000( c, cs );                          \
 5513+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*4 );    \
 5514+    stbir__simdf_0123to1111( c, cs );                          \
 5515+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*4+4 );  \
 5516+    stbir__simdf_0123to2222( c, cs );                          \
 5517+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*4+8 );
 5518+
 5519+#define stbir__store_output()                     \
 5520+    stbir__simdf_add( tot0, tot0, tot1 );         \
 5521+    stbir__simdf_store( output, tot0 );           \
 5522+    horizontal_coefficients += coefficient_width; \
 5523+    ++horizontal_contributors;                    \
 5524+    output += 4;
 5525+
 5526+#endif
 5527+
 5528+#else
 5529+
 5530+#define stbir__1_coeff_only()         \
 5531+    float p0,p1,p2,p3,c;              \
 5532+    STBIR_SIMD_NO_UNROLL(decode);     \
 5533+    c = hc[0];                        \
 5534+    p0 = decode[0] * c;               \
 5535+    p1 = decode[1] * c;               \
 5536+    p2 = decode[2] * c;               \
 5537+    p3 = decode[3] * c;
 5538+
 5539+#define stbir__2_coeff_only()         \
 5540+    float p0,p1,p2,p3,c;              \
 5541+    STBIR_SIMD_NO_UNROLL(decode);     \
 5542+    c = hc[0];                        \
 5543+    p0 = decode[0] * c;               \
 5544+    p1 = decode[1] * c;               \
 5545+    p2 = decode[2] * c;               \
 5546+    p3 = decode[3] * c;               \
 5547+    c = hc[1];                        \
 5548+    p0 += decode[4] * c;              \
 5549+    p1 += decode[5] * c;              \
 5550+    p2 += decode[6] * c;              \
 5551+    p3 += decode[7] * c;
 5552+
 5553+#define stbir__3_coeff_only()         \
 5554+    float p0,p1,p2,p3,c;              \
 5555+    STBIR_SIMD_NO_UNROLL(decode);     \
 5556+    c = hc[0];                        \
 5557+    p0 = decode[0] * c;               \
 5558+    p1 = decode[1] * c;               \
 5559+    p2 = decode[2] * c;               \
 5560+    p3 = decode[3] * c;               \
 5561+    c = hc[1];                        \
 5562+    p0 += decode[4] * c;              \
 5563+    p1 += decode[5] * c;              \
 5564+    p2 += decode[6] * c;              \
 5565+    p3 += decode[7] * c;              \
 5566+    c = hc[2];                        \
 5567+    p0 += decode[8] * c;              \
 5568+    p1 += decode[9] * c;              \
 5569+    p2 += decode[10] * c;             \
 5570+    p3 += decode[11] * c;
 5571+
 5572+#define stbir__store_output_tiny()                \
 5573+    output[0] = p0;                               \
 5574+    output[1] = p1;                               \
 5575+    output[2] = p2;                               \
 5576+    output[3] = p3;                               \
 5577+    horizontal_coefficients += coefficient_width; \
 5578+    ++horizontal_contributors;                    \
 5579+    output += 4;
 5580+
 5581+#define stbir__4_coeff_start()        \
 5582+    float x0,x1,x2,x3,y0,y1,y2,y3,c;  \
 5583+    STBIR_SIMD_NO_UNROLL(decode);     \
 5584+    c = hc[0];                        \
 5585+    x0 = decode[0] * c;               \
 5586+    x1 = decode[1] * c;               \
 5587+    x2 = decode[2] * c;               \
 5588+    x3 = decode[3] * c;               \
 5589+    c = hc[1];                        \
 5590+    y0 = decode[4] * c;               \
 5591+    y1 = decode[5] * c;               \
 5592+    y2 = decode[6] * c;               \
 5593+    y3 = decode[7] * c;               \
 5594+    c = hc[2];                        \
 5595+    x0 += decode[8] * c;              \
 5596+    x1 += decode[9] * c;              \
 5597+    x2 += decode[10] * c;             \
 5598+    x3 += decode[11] * c;             \
 5599+    c = hc[3];                        \
 5600+    y0 += decode[12] * c;             \
 5601+    y1 += decode[13] * c;             \
 5602+    y2 += decode[14] * c;             \
 5603+    y3 += decode[15] * c;
 5604+
 5605+#define stbir__4_coeff_continue_from_4( ofs ) \
 5606+    STBIR_SIMD_NO_UNROLL(decode);     \
 5607+    c = hc[0+(ofs)];                  \
 5608+    x0 += decode[0+(ofs)*4] * c;      \
 5609+    x1 += decode[1+(ofs)*4] * c;      \
 5610+    x2 += decode[2+(ofs)*4] * c;      \
 5611+    x3 += decode[3+(ofs)*4] * c;      \
 5612+    c = hc[1+(ofs)];                  \
 5613+    y0 += decode[4+(ofs)*4] * c;      \
 5614+    y1 += decode[5+(ofs)*4] * c;      \
 5615+    y2 += decode[6+(ofs)*4] * c;      \
 5616+    y3 += decode[7+(ofs)*4] * c;      \
 5617+    c = hc[2+(ofs)];                  \
 5618+    x0 += decode[8+(ofs)*4] * c;      \
 5619+    x1 += decode[9+(ofs)*4] * c;      \
 5620+    x2 += decode[10+(ofs)*4] * c;     \
 5621+    x3 += decode[11+(ofs)*4] * c;     \
 5622+    c = hc[3+(ofs)];                  \
 5623+    y0 += decode[12+(ofs)*4] * c;     \
 5624+    y1 += decode[13+(ofs)*4] * c;     \
 5625+    y2 += decode[14+(ofs)*4] * c;     \
 5626+    y3 += decode[15+(ofs)*4] * c;
 5627+
 5628+#define stbir__1_coeff_remnant( ofs ) \
 5629+    STBIR_SIMD_NO_UNROLL(decode);     \
 5630+    c = hc[0+(ofs)];                  \
 5631+    x0 += decode[0+(ofs)*4] * c;      \
 5632+    x1 += decode[1+(ofs)*4] * c;      \
 5633+    x2 += decode[2+(ofs)*4] * c;      \
 5634+    x3 += decode[3+(ofs)*4] * c;
 5635+
 5636+#define stbir__2_coeff_remnant( ofs ) \
 5637+    STBIR_SIMD_NO_UNROLL(decode);     \
 5638+    c = hc[0+(ofs)];                  \
 5639+    x0 += decode[0+(ofs)*4] * c;      \
 5640+    x1 += decode[1+(ofs)*4] * c;      \
 5641+    x2 += decode[2+(ofs)*4] * c;      \
 5642+    x3 += decode[3+(ofs)*4] * c;      \
 5643+    c = hc[1+(ofs)];                  \
 5644+    y0 += decode[4+(ofs)*4] * c;      \
 5645+    y1 += decode[5+(ofs)*4] * c;      \
 5646+    y2 += decode[6+(ofs)*4] * c;      \
 5647+    y3 += decode[7+(ofs)*4] * c;
 5648+
 5649+#define stbir__3_coeff_remnant( ofs ) \
 5650+    STBIR_SIMD_NO_UNROLL(decode);     \
 5651+    c = hc[0+(ofs)];                  \
 5652+    x0 += decode[0+(ofs)*4] * c;      \
 5653+    x1 += decode[1+(ofs)*4] * c;      \
 5654+    x2 += decode[2+(ofs)*4] * c;      \
 5655+    x3 += decode[3+(ofs)*4] * c;      \
 5656+    c = hc[1+(ofs)];                  \
 5657+    y0 += decode[4+(ofs)*4] * c;      \
 5658+    y1 += decode[5+(ofs)*4] * c;      \
 5659+    y2 += decode[6+(ofs)*4] * c;      \
 5660+    y3 += decode[7+(ofs)*4] * c;      \
 5661+    c = hc[2+(ofs)];                  \
 5662+    x0 += decode[8+(ofs)*4] * c;      \
 5663+    x1 += decode[9+(ofs)*4] * c;      \
 5664+    x2 += decode[10+(ofs)*4] * c;     \
 5665+    x3 += decode[11+(ofs)*4] * c;
 5666+
 5667+#define stbir__store_output()                     \
 5668+    output[0] = x0 + y0;                          \
 5669+    output[1] = x1 + y1;                          \
 5670+    output[2] = x2 + y2;                          \
 5671+    output[3] = x3 + y3;                          \
 5672+    horizontal_coefficients += coefficient_width; \
 5673+    ++horizontal_contributors;                    \
 5674+    output += 4;
 5675+
 5676+#endif
 5677+
 5678+#define STBIR__horizontal_channels 4
 5679+#define STB_IMAGE_RESIZE_DO_HORIZONTALS
 5680+#include STBIR__HEADER_FILENAME
 5681+
 5682+
 5683+
 5684+//=================
 5685+// Do 7 channel horizontal routines
 5686+
 5687+#ifdef STBIR_SIMD
 5688+
 5689+#define stbir__1_coeff_only()                   \
 5690+    stbir__simdf tot0,tot1,c;                   \
 5691+    STBIR_SIMD_NO_UNROLL(decode);               \
 5692+    stbir__simdf_load1( c, hc );                \
 5693+    stbir__simdf_0123to0000( c, c );            \
 5694+    stbir__simdf_mult_mem( tot0, c, decode );   \
 5695+    stbir__simdf_mult_mem( tot1, c, decode+3 );
 5696+
 5697+#define stbir__2_coeff_only()                         \
 5698+    stbir__simdf tot0,tot1,c,cs;                      \
 5699+    STBIR_SIMD_NO_UNROLL(decode);                     \
 5700+    stbir__simdf_load2( cs, hc );                     \
 5701+    stbir__simdf_0123to0000( c, cs );                 \
 5702+    stbir__simdf_mult_mem( tot0, c, decode );         \
 5703+    stbir__simdf_mult_mem( tot1, c, decode+3 );       \
 5704+    stbir__simdf_0123to1111( c, cs );                 \
 5705+    stbir__simdf_madd_mem( tot0, tot0, c, decode+7 ); \
 5706+    stbir__simdf_madd_mem( tot1, tot1, c,decode+10 );
 5707+
 5708+#define stbir__3_coeff_only()                           \
 5709+    stbir__simdf tot0,tot1,c,cs;                        \
 5710+    STBIR_SIMD_NO_UNROLL(decode);                       \
 5711+    stbir__simdf_load( cs, hc );                        \
 5712+    stbir__simdf_0123to0000( c, cs );                   \
 5713+    stbir__simdf_mult_mem( tot0, c, decode );           \
 5714+    stbir__simdf_mult_mem( tot1, c, decode+3 );         \
 5715+    stbir__simdf_0123to1111( c, cs );                   \
 5716+    stbir__simdf_madd_mem( tot0, tot0, c, decode+7 );   \
 5717+    stbir__simdf_madd_mem( tot1, tot1, c, decode+10 );  \
 5718+    stbir__simdf_0123to2222( c, cs );                   \
 5719+    stbir__simdf_madd_mem( tot0, tot0, c, decode+14 );  \
 5720+    stbir__simdf_madd_mem( tot1, tot1, c, decode+17 );
 5721+
 5722+#define stbir__store_output_tiny()                \
 5723+    stbir__simdf_store( output+3, tot1 );         \
 5724+    stbir__simdf_store( output, tot0 );           \
 5725+    horizontal_coefficients += coefficient_width; \
 5726+    ++horizontal_contributors;                    \
 5727+    output += 7;
 5728+
 5729+#ifdef STBIR_SIMD8
 5730+
 5731+#define stbir__4_coeff_start()                     \
 5732+    stbir__simdf8 tot0,tot1,c,cs;                  \
 5733+    STBIR_SIMD_NO_UNROLL(decode);                  \
 5734+    stbir__simdf8_load4b( cs, hc );                \
 5735+    stbir__simdf8_0123to00000000( c, cs );         \
 5736+    stbir__simdf8_mult_mem( tot0, c, decode );     \
 5737+    stbir__simdf8_0123to11111111( c, cs );         \
 5738+    stbir__simdf8_mult_mem( tot1, c, decode+7 );   \
 5739+    stbir__simdf8_0123to22222222( c, cs );         \
 5740+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+14 );  \
 5741+    stbir__simdf8_0123to33333333( c, cs );         \
 5742+    stbir__simdf8_madd_mem( tot1, tot1, c, decode+21 );
 5743+
 5744+#define stbir__4_coeff_continue_from_4( ofs )                   \
 5745+    STBIR_SIMD_NO_UNROLL(decode);                               \
 5746+    stbir__simdf8_load4b( cs, hc + (ofs) );                     \
 5747+    stbir__simdf8_0123to00000000( c, cs );                      \
 5748+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*7 );    \
 5749+    stbir__simdf8_0123to11111111( c, cs );                      \
 5750+    stbir__simdf8_madd_mem( tot1, tot1, c, decode+(ofs)*7+7 );  \
 5751+    stbir__simdf8_0123to22222222( c, cs );                      \
 5752+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*7+14 ); \
 5753+    stbir__simdf8_0123to33333333( c, cs );                      \
 5754+    stbir__simdf8_madd_mem( tot1, tot1, c, decode+(ofs)*7+21 );
 5755+
 5756+#define stbir__1_coeff_remnant( ofs )                           \
 5757+    STBIR_SIMD_NO_UNROLL(decode);                               \
 5758+    stbir__simdf8_load1b( c, hc + (ofs) );                      \
 5759+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*7 );
 5760+
 5761+#define stbir__2_coeff_remnant( ofs )                           \
 5762+    STBIR_SIMD_NO_UNROLL(decode);                               \
 5763+    stbir__simdf8_load1b( c, hc + (ofs) );                      \
 5764+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*7 );    \
 5765+    stbir__simdf8_load1b( c, hc + (ofs)+1 );                    \
 5766+    stbir__simdf8_madd_mem( tot1, tot1, c, decode+(ofs)*7+7 );
 5767+
 5768+#define stbir__3_coeff_remnant( ofs )                           \
 5769+    STBIR_SIMD_NO_UNROLL(decode);                               \
 5770+    stbir__simdf8_load4b( cs, hc + (ofs) );                     \
 5771+    stbir__simdf8_0123to00000000( c, cs );                      \
 5772+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*7 );    \
 5773+    stbir__simdf8_0123to11111111( c, cs );                      \
 5774+    stbir__simdf8_madd_mem( tot1, tot1, c, decode+(ofs)*7+7 );  \
 5775+    stbir__simdf8_0123to22222222( c, cs );                      \
 5776+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*7+14 );
 5777+
 5778+#define stbir__store_output()                     \
 5779+    stbir__simdf8_add( tot0, tot0, tot1 );        \
 5780+    horizontal_coefficients += coefficient_width; \
 5781+    ++horizontal_contributors;                    \
 5782+    output += 7;                                  \
 5783+    if ( output < output_end )                    \
 5784+    {                                             \
 5785+      stbir__simdf8_store( output-7, tot0 );      \
 5786+      continue;                                   \
 5787+    }                                             \
 5788+    stbir__simdf_store( output-7+3, stbir__simdf_swiz(stbir__simdf8_gettop4(tot0),0,0,1,2) ); \
 5789+    stbir__simdf_store( output-7, stbir__if_simdf8_cast_to_simdf4(tot0) );           \
 5790+    break;
 5791+
 5792+#else
 5793+
 5794+#define stbir__4_coeff_start()                    \
 5795+    stbir__simdf tot0,tot1,tot2,tot3,c,cs;        \
 5796+    STBIR_SIMD_NO_UNROLL(decode);                 \
 5797+    stbir__simdf_load( cs, hc );                  \
 5798+    stbir__simdf_0123to0000( c, cs );             \
 5799+    stbir__simdf_mult_mem( tot0, c, decode );     \
 5800+    stbir__simdf_mult_mem( tot1, c, decode+3 );   \
 5801+    stbir__simdf_0123to1111( c, cs );             \
 5802+    stbir__simdf_mult_mem( tot2, c, decode+7 );   \
 5803+    stbir__simdf_mult_mem( tot3, c, decode+10 );  \
 5804+    stbir__simdf_0123to2222( c, cs );             \
 5805+    stbir__simdf_madd_mem( tot0, tot0, c, decode+14 );  \
 5806+    stbir__simdf_madd_mem( tot1, tot1, c, decode+17 );  \
 5807+    stbir__simdf_0123to3333( c, cs );                   \
 5808+    stbir__simdf_madd_mem( tot2, tot2, c, decode+21 );  \
 5809+    stbir__simdf_madd_mem( tot3, tot3, c, decode+24 );
 5810+
 5811+#define stbir__4_coeff_continue_from_4( ofs )                   \
 5812+    STBIR_SIMD_NO_UNROLL(decode);                               \
 5813+    stbir__simdf_load( cs, hc + (ofs) );                        \
 5814+    stbir__simdf_0123to0000( c, cs );                           \
 5815+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*7 );     \
 5816+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*7+3 );   \
 5817+    stbir__simdf_0123to1111( c, cs );                           \
 5818+    stbir__simdf_madd_mem( tot2, tot2, c, decode+(ofs)*7+7 );   \
 5819+    stbir__simdf_madd_mem( tot3, tot3, c, decode+(ofs)*7+10 );  \
 5820+    stbir__simdf_0123to2222( c, cs );                           \
 5821+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*7+14 );  \
 5822+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*7+17 );  \
 5823+    stbir__simdf_0123to3333( c, cs );                           \
 5824+    stbir__simdf_madd_mem( tot2, tot2, c, decode+(ofs)*7+21 );  \
 5825+    stbir__simdf_madd_mem( tot3, tot3, c, decode+(ofs)*7+24 );
 5826+
 5827+#define stbir__1_coeff_remnant( ofs )                           \
 5828+    STBIR_SIMD_NO_UNROLL(decode);                               \
 5829+    stbir__simdf_load1( c, hc + (ofs) );                        \
 5830+    stbir__simdf_0123to0000( c, c );                            \
 5831+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*7 );     \
 5832+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*7+3 );   \
 5833+
 5834+#define stbir__2_coeff_remnant( ofs )                           \
 5835+    STBIR_SIMD_NO_UNROLL(decode);                               \
 5836+    stbir__simdf_load2( cs, hc + (ofs) );                       \
 5837+    stbir__simdf_0123to0000( c, cs );                           \
 5838+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*7 );     \
 5839+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*7+3 );   \
 5840+    stbir__simdf_0123to1111( c, cs );                           \
 5841+    stbir__simdf_madd_mem( tot2, tot2, c, decode+(ofs)*7+7 );   \
 5842+    stbir__simdf_madd_mem( tot3, tot3, c, decode+(ofs)*7+10 );
 5843+
 5844+#define stbir__3_coeff_remnant( ofs )                           \
 5845+    STBIR_SIMD_NO_UNROLL(decode);                               \
 5846+    stbir__simdf_load( cs, hc + (ofs) );                        \
 5847+    stbir__simdf_0123to0000( c, cs );                           \
 5848+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*7 );     \
 5849+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*7+3 );   \
 5850+    stbir__simdf_0123to1111( c, cs );                           \
 5851+    stbir__simdf_madd_mem( tot2, tot2, c, decode+(ofs)*7+7 );   \
 5852+    stbir__simdf_madd_mem( tot3, tot3, c, decode+(ofs)*7+10 );  \
 5853+    stbir__simdf_0123to2222( c, cs );                           \
 5854+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*7+14 );  \
 5855+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*7+17 );
 5856+
 5857+#define stbir__store_output()                     \
 5858+    stbir__simdf_add( tot0, tot0, tot2 );         \
 5859+    stbir__simdf_add( tot1, tot1, tot3 );         \
 5860+    stbir__simdf_store( output+3, tot1 );         \
 5861+    stbir__simdf_store( output, tot0 );           \
 5862+    horizontal_coefficients += coefficient_width; \
 5863+    ++horizontal_contributors;                    \
 5864+    output += 7;
 5865+
 5866+#endif
 5867+
 5868+#else
 5869+
 5870+#define stbir__1_coeff_only()        \
 5871+    float tot0, tot1, tot2, tot3, tot4, tot5, tot6, c; \
 5872+    c = hc[0];                       \
 5873+    tot0 = decode[0]*c;              \
 5874+    tot1 = decode[1]*c;              \
 5875+    tot2 = decode[2]*c;              \
 5876+    tot3 = decode[3]*c;              \
 5877+    tot4 = decode[4]*c;              \
 5878+    tot5 = decode[5]*c;              \
 5879+    tot6 = decode[6]*c;
 5880+
 5881+#define stbir__2_coeff_only()        \
 5882+    float tot0, tot1, tot2, tot3, tot4, tot5, tot6, c; \
 5883+    c = hc[0];                       \
 5884+    tot0 = decode[0]*c;              \
 5885+    tot1 = decode[1]*c;              \
 5886+    tot2 = decode[2]*c;              \
 5887+    tot3 = decode[3]*c;              \
 5888+    tot4 = decode[4]*c;              \
 5889+    tot5 = decode[5]*c;              \
 5890+    tot6 = decode[6]*c;              \
 5891+    c = hc[1];                       \
 5892+    tot0 += decode[7]*c;             \
 5893+    tot1 += decode[8]*c;             \
 5894+    tot2 += decode[9]*c;             \
 5895+    tot3 += decode[10]*c;            \
 5896+    tot4 += decode[11]*c;            \
 5897+    tot5 += decode[12]*c;            \
 5898+    tot6 += decode[13]*c;            \
 5899+
 5900+#define stbir__3_coeff_only()        \
 5901+    float tot0, tot1, tot2, tot3, tot4, tot5, tot6, c; \
 5902+    c = hc[0];                       \
 5903+    tot0 = decode[0]*c;              \
 5904+    tot1 = decode[1]*c;              \
 5905+    tot2 = decode[2]*c;              \
 5906+    tot3 = decode[3]*c;              \
 5907+    tot4 = decode[4]*c;              \
 5908+    tot5 = decode[5]*c;              \
 5909+    tot6 = decode[6]*c;              \
 5910+    c = hc[1];                       \
 5911+    tot0 += decode[7]*c;             \
 5912+    tot1 += decode[8]*c;             \
 5913+    tot2 += decode[9]*c;             \
 5914+    tot3 += decode[10]*c;            \
 5915+    tot4 += decode[11]*c;            \
 5916+    tot5 += decode[12]*c;            \
 5917+    tot6 += decode[13]*c;            \
 5918+    c = hc[2];                       \
 5919+    tot0 += decode[14]*c;            \
 5920+    tot1 += decode[15]*c;            \
 5921+    tot2 += decode[16]*c;            \
 5922+    tot3 += decode[17]*c;            \
 5923+    tot4 += decode[18]*c;            \
 5924+    tot5 += decode[19]*c;            \
 5925+    tot6 += decode[20]*c;            \
 5926+
 5927+#define stbir__store_output_tiny()                \
 5928+    output[0] = tot0;                             \
 5929+    output[1] = tot1;                             \
 5930+    output[2] = tot2;                             \
 5931+    output[3] = tot3;                             \
 5932+    output[4] = tot4;                             \
 5933+    output[5] = tot5;                             \
 5934+    output[6] = tot6;                             \
 5935+    horizontal_coefficients += coefficient_width; \
 5936+    ++horizontal_contributors;                    \
 5937+    output += 7;
 5938+
 5939+#define stbir__4_coeff_start()    \
 5940+    float x0,x1,x2,x3,x4,x5,x6,y0,y1,y2,y3,y4,y5,y6,c; \
 5941+    STBIR_SIMD_NO_UNROLL(decode); \
 5942+    c = hc[0];                    \
 5943+    x0 = decode[0] * c;           \
 5944+    x1 = decode[1] * c;           \
 5945+    x2 = decode[2] * c;           \
 5946+    x3 = decode[3] * c;           \
 5947+    x4 = decode[4] * c;           \
 5948+    x5 = decode[5] * c;           \
 5949+    x6 = decode[6] * c;           \
 5950+    c = hc[1];                    \
 5951+    y0 = decode[7] * c;           \
 5952+    y1 = decode[8] * c;           \
 5953+    y2 = decode[9] * c;           \
 5954+    y3 = decode[10] * c;          \
 5955+    y4 = decode[11] * c;          \
 5956+    y5 = decode[12] * c;          \
 5957+    y6 = decode[13] * c;          \
 5958+    c = hc[2];                    \
 5959+    x0 += decode[14] * c;         \
 5960+    x1 += decode[15] * c;         \
 5961+    x2 += decode[16] * c;         \
 5962+    x3 += decode[17] * c;         \
 5963+    x4 += decode[18] * c;         \
 5964+    x5 += decode[19] * c;         \
 5965+    x6 += decode[20] * c;         \
 5966+    c = hc[3];                    \
 5967+    y0 += decode[21] * c;         \
 5968+    y1 += decode[22] * c;         \
 5969+    y2 += decode[23] * c;         \
 5970+    y3 += decode[24] * c;         \
 5971+    y4 += decode[25] * c;         \
 5972+    y5 += decode[26] * c;         \
 5973+    y6 += decode[27] * c;
 5974+
 5975+#define stbir__4_coeff_continue_from_4( ofs ) \
 5976+    STBIR_SIMD_NO_UNROLL(decode);  \
 5977+    c = hc[0+(ofs)];               \
 5978+    x0 += decode[0+(ofs)*7] * c;   \
 5979+    x1 += decode[1+(ofs)*7] * c;   \
 5980+    x2 += decode[2+(ofs)*7] * c;   \
 5981+    x3 += decode[3+(ofs)*7] * c;   \
 5982+    x4 += decode[4+(ofs)*7] * c;   \
 5983+    x5 += decode[5+(ofs)*7] * c;   \
 5984+    x6 += decode[6+(ofs)*7] * c;   \
 5985+    c = hc[1+(ofs)];               \
 5986+    y0 += decode[7+(ofs)*7] * c;   \
 5987+    y1 += decode[8+(ofs)*7] * c;   \
 5988+    y2 += decode[9+(ofs)*7] * c;   \
 5989+    y3 += decode[10+(ofs)*7] * c;  \
 5990+    y4 += decode[11+(ofs)*7] * c;  \
 5991+    y5 += decode[12+(ofs)*7] * c;  \
 5992+    y6 += decode[13+(ofs)*7] * c;  \
 5993+    c = hc[2+(ofs)];               \
 5994+    x0 += decode[14+(ofs)*7] * c;  \
 5995+    x1 += decode[15+(ofs)*7] * c;  \
 5996+    x2 += decode[16+(ofs)*7] * c;  \
 5997+    x3 += decode[17+(ofs)*7] * c;  \
 5998+    x4 += decode[18+(ofs)*7] * c;  \
 5999+    x5 += decode[19+(ofs)*7] * c;  \
 6000+    x6 += decode[20+(ofs)*7] * c;  \
 6001+    c = hc[3+(ofs)];               \
 6002+    y0 += decode[21+(ofs)*7] * c;  \
 6003+    y1 += decode[22+(ofs)*7] * c;  \
 6004+    y2 += decode[23+(ofs)*7] * c;  \
 6005+    y3 += decode[24+(ofs)*7] * c;  \
 6006+    y4 += decode[25+(ofs)*7] * c;  \
 6007+    y5 += decode[26+(ofs)*7] * c;  \
 6008+    y6 += decode[27+(ofs)*7] * c;
 6009+
 6010+#define stbir__1_coeff_remnant( ofs ) \
 6011+    STBIR_SIMD_NO_UNROLL(decode);  \
 6012+    c = hc[0+(ofs)];               \
 6013+    x0 += decode[0+(ofs)*7] * c;   \
 6014+    x1 += decode[1+(ofs)*7] * c;   \
 6015+    x2 += decode[2+(ofs)*7] * c;   \
 6016+    x3 += decode[3+(ofs)*7] * c;   \
 6017+    x4 += decode[4+(ofs)*7] * c;   \
 6018+    x5 += decode[5+(ofs)*7] * c;   \
 6019+    x6 += decode[6+(ofs)*7] * c;   \
 6020+
 6021+#define stbir__2_coeff_remnant( ofs ) \
 6022+    STBIR_SIMD_NO_UNROLL(decode);  \
 6023+    c = hc[0+(ofs)];               \
 6024+    x0 += decode[0+(ofs)*7] * c;   \
 6025+    x1 += decode[1+(ofs)*7] * c;   \
 6026+    x2 += decode[2+(ofs)*7] * c;   \
 6027+    x3 += decode[3+(ofs)*7] * c;   \
 6028+    x4 += decode[4+(ofs)*7] * c;   \
 6029+    x5 += decode[5+(ofs)*7] * c;   \
 6030+    x6 += decode[6+(ofs)*7] * c;   \
 6031+    c = hc[1+(ofs)];               \
 6032+    y0 += decode[7+(ofs)*7] * c;   \
 6033+    y1 += decode[8+(ofs)*7] * c;   \
 6034+    y2 += decode[9+(ofs)*7] * c;   \
 6035+    y3 += decode[10+(ofs)*7] * c;  \
 6036+    y4 += decode[11+(ofs)*7] * c;  \
 6037+    y5 += decode[12+(ofs)*7] * c;  \
 6038+    y6 += decode[13+(ofs)*7] * c;  \
 6039+
 6040+#define stbir__3_coeff_remnant( ofs ) \
 6041+    STBIR_SIMD_NO_UNROLL(decode);  \
 6042+    c = hc[0+(ofs)];               \
 6043+    x0 += decode[0+(ofs)*7] * c;   \
 6044+    x1 += decode[1+(ofs)*7] * c;   \
 6045+    x2 += decode[2+(ofs)*7] * c;   \
 6046+    x3 += decode[3+(ofs)*7] * c;   \
 6047+    x4 += decode[4+(ofs)*7] * c;   \
 6048+    x5 += decode[5+(ofs)*7] * c;   \
 6049+    x6 += decode[6+(ofs)*7] * c;   \
 6050+    c = hc[1+(ofs)];               \
 6051+    y0 += decode[7+(ofs)*7] * c;   \
 6052+    y1 += decode[8+(ofs)*7] * c;   \
 6053+    y2 += decode[9+(ofs)*7] * c;   \
 6054+    y3 += decode[10+(ofs)*7] * c;  \
 6055+    y4 += decode[11+(ofs)*7] * c;  \
 6056+    y5 += decode[12+(ofs)*7] * c;  \
 6057+    y6 += decode[13+(ofs)*7] * c;  \
 6058+    c = hc[2+(ofs)];               \
 6059+    x0 += decode[14+(ofs)*7] * c;  \
 6060+    x1 += decode[15+(ofs)*7] * c;  \
 6061+    x2 += decode[16+(ofs)*7] * c;  \
 6062+    x3 += decode[17+(ofs)*7] * c;  \
 6063+    x4 += decode[18+(ofs)*7] * c;  \
 6064+    x5 += decode[19+(ofs)*7] * c;  \
 6065+    x6 += decode[20+(ofs)*7] * c;  \
 6066+
 6067+#define stbir__store_output()                     \
 6068+    output[0] = x0 + y0;                          \
 6069+    output[1] = x1 + y1;                          \
 6070+    output[2] = x2 + y2;                          \
 6071+    output[3] = x3 + y3;                          \
 6072+    output[4] = x4 + y4;                          \
 6073+    output[5] = x5 + y5;                          \
 6074+    output[6] = x6 + y6;                          \
 6075+    horizontal_coefficients += coefficient_width; \
 6076+    ++horizontal_contributors;                    \
 6077+    output += 7;
 6078+
 6079+#endif
 6080+
 6081+#define STBIR__horizontal_channels 7
 6082+#define STB_IMAGE_RESIZE_DO_HORIZONTALS
 6083+#include STBIR__HEADER_FILENAME
 6084+
 6085+
 6086+// include all of the vertical resamplers (both scatter and gather versions)
 6087+
 6088+#define STBIR__vertical_channels 1
 6089+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6090+#include STBIR__HEADER_FILENAME
 6091+
 6092+#define STBIR__vertical_channels 1
 6093+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6094+#define STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 6095+#include STBIR__HEADER_FILENAME
 6096+
 6097+#define STBIR__vertical_channels 2
 6098+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6099+#include STBIR__HEADER_FILENAME
 6100+
 6101+#define STBIR__vertical_channels 2
 6102+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6103+#define STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 6104+#include STBIR__HEADER_FILENAME
 6105+
 6106+#define STBIR__vertical_channels 3
 6107+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6108+#include STBIR__HEADER_FILENAME
 6109+
 6110+#define STBIR__vertical_channels 3
 6111+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6112+#define STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 6113+#include STBIR__HEADER_FILENAME
 6114+
 6115+#define STBIR__vertical_channels 4
 6116+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6117+#include STBIR__HEADER_FILENAME
 6118+
 6119+#define STBIR__vertical_channels 4
 6120+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6121+#define STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 6122+#include STBIR__HEADER_FILENAME
 6123+
 6124+#define STBIR__vertical_channels 5
 6125+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6126+#include STBIR__HEADER_FILENAME
 6127+
 6128+#define STBIR__vertical_channels 5
 6129+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6130+#define STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 6131+#include STBIR__HEADER_FILENAME
 6132+
 6133+#define STBIR__vertical_channels 6
 6134+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6135+#include STBIR__HEADER_FILENAME
 6136+
 6137+#define STBIR__vertical_channels 6
 6138+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6139+#define STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 6140+#include STBIR__HEADER_FILENAME
 6141+
 6142+#define STBIR__vertical_channels 7
 6143+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6144+#include STBIR__HEADER_FILENAME
 6145+
 6146+#define STBIR__vertical_channels 7
 6147+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6148+#define STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 6149+#include STBIR__HEADER_FILENAME
 6150+
 6151+#define STBIR__vertical_channels 8
 6152+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6153+#include STBIR__HEADER_FILENAME
 6154+
 6155+#define STBIR__vertical_channels 8
 6156+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6157+#define STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 6158+#include STBIR__HEADER_FILENAME
 6159+
 6160+typedef void STBIR_VERTICAL_GATHERFUNC( float * output, float const * coeffs, float const ** inputs, float const * input0_end );
 6161+
 6162+static STBIR_VERTICAL_GATHERFUNC * stbir__vertical_gathers[ 8 ] =
 6163+{
 6164+  stbir__vertical_gather_with_1_coeffs,stbir__vertical_gather_with_2_coeffs,stbir__vertical_gather_with_3_coeffs,stbir__vertical_gather_with_4_coeffs,stbir__vertical_gather_with_5_coeffs,stbir__vertical_gather_with_6_coeffs,stbir__vertical_gather_with_7_coeffs,stbir__vertical_gather_with_8_coeffs
 6165+};
 6166+
 6167+static STBIR_VERTICAL_GATHERFUNC * stbir__vertical_gathers_continues[ 8 ] =
 6168+{
 6169+  stbir__vertical_gather_with_1_coeffs_cont,stbir__vertical_gather_with_2_coeffs_cont,stbir__vertical_gather_with_3_coeffs_cont,stbir__vertical_gather_with_4_coeffs_cont,stbir__vertical_gather_with_5_coeffs_cont,stbir__vertical_gather_with_6_coeffs_cont,stbir__vertical_gather_with_7_coeffs_cont,stbir__vertical_gather_with_8_coeffs_cont
 6170+};
 6171+
 6172+typedef void STBIR_VERTICAL_SCATTERFUNC( float ** outputs, float const * coeffs, float const * input, float const * input_end );
 6173+
 6174+static STBIR_VERTICAL_SCATTERFUNC * stbir__vertical_scatter_sets[ 8 ] =
 6175+{
 6176+  stbir__vertical_scatter_with_1_coeffs,stbir__vertical_scatter_with_2_coeffs,stbir__vertical_scatter_with_3_coeffs,stbir__vertical_scatter_with_4_coeffs,stbir__vertical_scatter_with_5_coeffs,stbir__vertical_scatter_with_6_coeffs,stbir__vertical_scatter_with_7_coeffs,stbir__vertical_scatter_with_8_coeffs
 6177+};
 6178+
 6179+static STBIR_VERTICAL_SCATTERFUNC * stbir__vertical_scatter_blends[ 8 ] =
 6180+{
 6181+  stbir__vertical_scatter_with_1_coeffs_cont,stbir__vertical_scatter_with_2_coeffs_cont,stbir__vertical_scatter_with_3_coeffs_cont,stbir__vertical_scatter_with_4_coeffs_cont,stbir__vertical_scatter_with_5_coeffs_cont,stbir__vertical_scatter_with_6_coeffs_cont,stbir__vertical_scatter_with_7_coeffs_cont,stbir__vertical_scatter_with_8_coeffs_cont
 6182+};
 6183+
 6184+
 6185+static void stbir__encode_scanline( stbir__info const * stbir_info, void *output_buffer_data, float * encode_buffer, int row  STBIR_ONLY_PROFILE_GET_SPLIT_INFO )
 6186+{
 6187+  int num_pixels = stbir_info->horizontal.scale_info.output_sub_size;
 6188+  int channels = stbir_info->channels;
 6189+  int width_times_channels = num_pixels * channels;
 6190+  void * output_buffer;
 6191+
 6192+  // un-alpha weight if we need to
 6193+  if ( stbir_info->alpha_unweight )
 6194+  {
 6195+    STBIR_PROFILE_START( unalpha );
 6196+    stbir_info->alpha_unweight( encode_buffer, width_times_channels );
 6197+    STBIR_PROFILE_END( unalpha );
 6198+  }
 6199+
 6200+  // write directly into output by default
 6201+  output_buffer = output_buffer_data;
 6202+
 6203+  // if we have an output callback, we first convert the decode buffer in place (and then hand that to the callback)
 6204+  if ( stbir_info->out_pixels_cb )
 6205+    output_buffer = encode_buffer;
 6206+
 6207+  STBIR_PROFILE_START( encode );
 6208+  // convert into the output buffer
 6209+  stbir_info->encode_pixels( output_buffer, width_times_channels, encode_buffer );
 6210+  STBIR_PROFILE_END( encode );
 6211+
 6212+  // if we have an output callback, call it to send the data
 6213+  if ( stbir_info->out_pixels_cb )
 6214+    stbir_info->out_pixels_cb( output_buffer, num_pixels, row, stbir_info->user_data );
 6215+}
 6216+
 6217+
 6218+// Get the ring buffer pointer for an index
 6219+static float* stbir__get_ring_buffer_entry(stbir__info const * stbir_info, stbir__per_split_info const * split_info, int index )
 6220+{
 6221+  STBIR_ASSERT( index < stbir_info->ring_buffer_num_entries );
 6222+
 6223+  #ifdef STBIR__SEPARATE_ALLOCATIONS
 6224+    return split_info->ring_buffers[ index ];
 6225+  #else
 6226+    return (float*) ( ( (char*) split_info->ring_buffer ) + ( index * stbir_info->ring_buffer_length_bytes ) );
 6227+  #endif
 6228+}
 6229+
 6230+// Get the specified scan line from the ring buffer
 6231+static float* stbir__get_ring_buffer_scanline(stbir__info const * stbir_info, stbir__per_split_info const * split_info, int get_scanline)
 6232+{
 6233+  int ring_buffer_index = (split_info->ring_buffer_begin_index + (get_scanline - split_info->ring_buffer_first_scanline)) % stbir_info->ring_buffer_num_entries;
 6234+  return stbir__get_ring_buffer_entry( stbir_info, split_info, ring_buffer_index );
 6235+}
 6236+
 6237+static void stbir__resample_horizontal_gather(stbir__info const * stbir_info, float* output_buffer, float const * input_buffer STBIR_ONLY_PROFILE_GET_SPLIT_INFO )
 6238+{
 6239+  float const * decode_buffer = input_buffer - ( stbir_info->scanline_extents.conservative.n0 * stbir_info->effective_channels );
 6240+
 6241+  STBIR_PROFILE_START( horizontal );
 6242+  if ( ( stbir_info->horizontal.filter_enum == STBIR_FILTER_POINT_SAMPLE ) && ( stbir_info->horizontal.scale_info.scale == 1.0f ) )
 6243+    STBIR_MEMCPY( output_buffer, input_buffer, stbir_info->horizontal.scale_info.output_sub_size * sizeof( float ) * stbir_info->effective_channels );
 6244+  else
 6245+    stbir_info->horizontal_gather_channels( output_buffer, stbir_info->horizontal.scale_info.output_sub_size, decode_buffer, stbir_info->horizontal.contributors, stbir_info->horizontal.coefficients, stbir_info->horizontal.coefficient_width );
 6246+  STBIR_PROFILE_END( horizontal );
 6247+}
 6248+
 6249+static void stbir__resample_vertical_gather(stbir__info const * stbir_info, stbir__per_split_info* split_info, int n, int contrib_n0, int contrib_n1, float const * vertical_coefficients )
 6250+{
 6251+  float* encode_buffer = split_info->vertical_buffer;
 6252+  float* decode_buffer = split_info->decode_buffer;
 6253+  int vertical_first = stbir_info->vertical_first;
 6254+  int width = (vertical_first) ? ( stbir_info->scanline_extents.conservative.n1-stbir_info->scanline_extents.conservative.n0+1 ) : stbir_info->horizontal.scale_info.output_sub_size;
 6255+  int width_times_channels = stbir_info->effective_channels * width;
 6256+
 6257+  STBIR_ASSERT( stbir_info->vertical.is_gather );
 6258+
 6259+  // loop over the contributing scanlines and scale into the buffer
 6260+  STBIR_PROFILE_START( vertical );
 6261+  {
 6262+    int k = 0, total = contrib_n1 - contrib_n0 + 1;
 6263+    STBIR_ASSERT( total > 0 );
 6264+    do {
 6265+      float const * inputs[8];
 6266+      int i, cnt = total; if ( cnt > 8 ) cnt = 8;
 6267+      for( i = 0 ; i < cnt ; i++ )
 6268+        inputs[ i ] = stbir__get_ring_buffer_scanline(stbir_info, split_info, k+i+contrib_n0 );
 6269+
 6270+      // call the N scanlines at a time function (up to 8 scanlines of blending at once)
 6271+      ((k==0)?stbir__vertical_gathers:stbir__vertical_gathers_continues)[cnt-1]( (vertical_first) ? decode_buffer : encode_buffer, vertical_coefficients + k, inputs, inputs[0] + width_times_channels );
 6272+      k += cnt;
 6273+      total -= cnt;
 6274+    } while ( total );
 6275+  }
 6276+  STBIR_PROFILE_END( vertical );
 6277+
 6278+  if ( vertical_first )
 6279+  {
 6280+    // Now resample the gathered vertical data in the horizontal axis into the encode buffer
 6281+    decode_buffer[ width_times_channels ] = 0.0f; // clear two over for horizontals with a remnant of 3
 6282+    decode_buffer[ width_times_channels+1 ] = 0.0f; 
 6283+    stbir__resample_horizontal_gather(stbir_info, encode_buffer, decode_buffer  STBIR_ONLY_PROFILE_SET_SPLIT_INFO );
 6284+  }
 6285+
 6286+  stbir__encode_scanline( stbir_info, ( (char *) stbir_info->output_data ) + ((size_t)n * (size_t)stbir_info->output_stride_bytes),
 6287+                          encode_buffer, n  STBIR_ONLY_PROFILE_SET_SPLIT_INFO );
 6288+}
 6289+
 6290+static void stbir__decode_and_resample_for_vertical_gather_loop(stbir__info const * stbir_info, stbir__per_split_info* split_info, int n)
 6291+{
 6292+  int ring_buffer_index;
 6293+  float* ring_buffer;
 6294+
 6295+  // Decode the nth scanline from the source image into the decode buffer.
 6296+  stbir__decode_scanline( stbir_info, n, split_info->decode_buffer  STBIR_ONLY_PROFILE_SET_SPLIT_INFO );
 6297+
 6298+  // update new end scanline
 6299+  split_info->ring_buffer_last_scanline = n;
 6300+
 6301+  // get ring buffer
 6302+  ring_buffer_index = (split_info->ring_buffer_begin_index + (split_info->ring_buffer_last_scanline - split_info->ring_buffer_first_scanline)) % stbir_info->ring_buffer_num_entries;
 6303+  ring_buffer = stbir__get_ring_buffer_entry(stbir_info, split_info, ring_buffer_index);
 6304+
 6305+  // Now resample it into the ring buffer.
 6306+  stbir__resample_horizontal_gather( stbir_info, ring_buffer, split_info->decode_buffer  STBIR_ONLY_PROFILE_SET_SPLIT_INFO );
 6307+
 6308+  // Now it's sitting in the ring buffer ready to be used as source for the vertical sampling.
 6309+}
 6310+
 6311+static void stbir__vertical_gather_loop( stbir__info const * stbir_info, stbir__per_split_info* split_info, int split_count )
 6312+{
 6313+  int y, start_output_y, end_output_y;
 6314+  stbir__contributors* vertical_contributors = stbir_info->vertical.contributors;
 6315+  float const * vertical_coefficients = stbir_info->vertical.coefficients;
 6316+
 6317+  STBIR_ASSERT( stbir_info->vertical.is_gather );
 6318+
 6319+  start_output_y = split_info->start_output_y;
 6320+  end_output_y = split_info[split_count-1].end_output_y;
 6321+
 6322+  vertical_contributors += start_output_y;
 6323+  vertical_coefficients += start_output_y * stbir_info->vertical.coefficient_width;
 6324+
 6325+  // initialize the ring buffer for gathering
 6326+  split_info->ring_buffer_begin_index = 0;
 6327+  split_info->ring_buffer_first_scanline = vertical_contributors->n0;
 6328+  split_info->ring_buffer_last_scanline = split_info->ring_buffer_first_scanline - 1; // means "empty"
 6329+
 6330+  for (y = start_output_y; y < end_output_y; y++)
 6331+  {
 6332+    int in_first_scanline, in_last_scanline;
 6333+
 6334+    in_first_scanline = vertical_contributors->n0;
 6335+    in_last_scanline = vertical_contributors->n1;
 6336+
 6337+    // make sure the indexing hasn't broken
 6338+    STBIR_ASSERT( in_first_scanline >= split_info->ring_buffer_first_scanline );
 6339+
 6340+    // Load in new scanlines
 6341+    while (in_last_scanline > split_info->ring_buffer_last_scanline)
 6342+    {
 6343+      STBIR_ASSERT( ( split_info->ring_buffer_last_scanline - split_info->ring_buffer_first_scanline + 1 ) <= stbir_info->ring_buffer_num_entries );
 6344+
 6345+      // make sure there was room in the ring buffer when we add new scanlines
 6346+      if ( ( split_info->ring_buffer_last_scanline - split_info->ring_buffer_first_scanline + 1 ) == stbir_info->ring_buffer_num_entries )
 6347+      {
 6348+        split_info->ring_buffer_first_scanline++;
 6349+        split_info->ring_buffer_begin_index++;
 6350+      }
 6351+
 6352+      if ( stbir_info->vertical_first )
 6353+      {
 6354+        float * ring_buffer = stbir__get_ring_buffer_scanline( stbir_info, split_info, ++split_info->ring_buffer_last_scanline );
 6355+        // Decode the nth scanline from the source image into the decode buffer.
 6356+        stbir__decode_scanline( stbir_info, split_info->ring_buffer_last_scanline, ring_buffer  STBIR_ONLY_PROFILE_SET_SPLIT_INFO );
 6357+      }
 6358+      else
 6359+      {
 6360+        stbir__decode_and_resample_for_vertical_gather_loop(stbir_info, split_info, split_info->ring_buffer_last_scanline + 1);
 6361+      }
 6362+    }
 6363+
 6364+    // Now all buffers should be ready to write a row of vertical sampling, so do it.
 6365+    stbir__resample_vertical_gather(stbir_info, split_info, y, in_first_scanline, in_last_scanline, vertical_coefficients );
 6366+
 6367+    ++vertical_contributors;
 6368+    vertical_coefficients += stbir_info->vertical.coefficient_width;
 6369+  }
 6370+}
 6371+
 6372+#define STBIR__FLOAT_EMPTY_MARKER 3.0e+38F
 6373+#define STBIR__FLOAT_BUFFER_IS_EMPTY(ptr) ((ptr)[0]==STBIR__FLOAT_EMPTY_MARKER)
 6374+
 6375+static void stbir__encode_first_scanline_from_scatter(stbir__info const * stbir_info, stbir__per_split_info* split_info)
 6376+{
 6377+  // evict a scanline out into the output buffer
 6378+  float* ring_buffer_entry = stbir__get_ring_buffer_entry(stbir_info, split_info, split_info->ring_buffer_begin_index );
 6379+
 6380+  // dump the scanline out
 6381+  stbir__encode_scanline( stbir_info, ( (char *)stbir_info->output_data ) + ( (size_t)split_info->ring_buffer_first_scanline * (size_t)stbir_info->output_stride_bytes ), ring_buffer_entry, split_info->ring_buffer_first_scanline  STBIR_ONLY_PROFILE_SET_SPLIT_INFO );
 6382+
 6383+  // mark it as empty
 6384+  ring_buffer_entry[ 0 ] = STBIR__FLOAT_EMPTY_MARKER;
 6385+
 6386+  // advance the first scanline
 6387+  split_info->ring_buffer_first_scanline++;
 6388+  if ( ++split_info->ring_buffer_begin_index == stbir_info->ring_buffer_num_entries )
 6389+    split_info->ring_buffer_begin_index = 0;
 6390+}
 6391+
 6392+static void stbir__horizontal_resample_and_encode_first_scanline_from_scatter(stbir__info const * stbir_info, stbir__per_split_info* split_info)
 6393+{
 6394+  // evict a scanline out into the output buffer
 6395+
 6396+  float* ring_buffer_entry = stbir__get_ring_buffer_entry(stbir_info, split_info, split_info->ring_buffer_begin_index );
 6397+
 6398+  // Now resample it into the buffer.
 6399+  stbir__resample_horizontal_gather( stbir_info, split_info->vertical_buffer, ring_buffer_entry  STBIR_ONLY_PROFILE_SET_SPLIT_INFO );
 6400+
 6401+  // dump the scanline out
 6402+  stbir__encode_scanline( stbir_info, ( (char *)stbir_info->output_data ) + ( (size_t)split_info->ring_buffer_first_scanline * (size_t)stbir_info->output_stride_bytes ), split_info->vertical_buffer, split_info->ring_buffer_first_scanline  STBIR_ONLY_PROFILE_SET_SPLIT_INFO );
 6403+
 6404+  // mark it as empty
 6405+  ring_buffer_entry[ 0 ] = STBIR__FLOAT_EMPTY_MARKER;
 6406+
 6407+  // advance the first scanline
 6408+  split_info->ring_buffer_first_scanline++;
 6409+  if ( ++split_info->ring_buffer_begin_index == stbir_info->ring_buffer_num_entries )
 6410+    split_info->ring_buffer_begin_index = 0;
 6411+}
 6412+
 6413+static void stbir__resample_vertical_scatter(stbir__info const * stbir_info, stbir__per_split_info* split_info, int n0, int n1, float const * vertical_coefficients, float const * vertical_buffer, float const * vertical_buffer_end )
 6414+{
 6415+  STBIR_ASSERT( !stbir_info->vertical.is_gather );
 6416+
 6417+  STBIR_PROFILE_START( vertical );
 6418+  {
 6419+    int k = 0, total = n1 - n0 + 1;
 6420+    STBIR_ASSERT( total > 0 );
 6421+    do {
 6422+      float * outputs[8];
 6423+      int i, n = total; if ( n > 8 ) n = 8;
 6424+      for( i = 0 ; i < n ; i++ )
 6425+      {
 6426+        outputs[ i ] = stbir__get_ring_buffer_scanline(stbir_info, split_info, k+i+n0 );
 6427+        if ( ( i ) && ( STBIR__FLOAT_BUFFER_IS_EMPTY( outputs[i] ) != STBIR__FLOAT_BUFFER_IS_EMPTY( outputs[0] ) ) ) // make sure runs are of the same type
 6428+        {
 6429+          n = i;
 6430+          break;
 6431+        }
 6432+      }
 6433+      // call the scatter to N scanlines at a time function (up to 8 scanlines of scattering at once)
 6434+      ((STBIR__FLOAT_BUFFER_IS_EMPTY( outputs[0] ))?stbir__vertical_scatter_sets:stbir__vertical_scatter_blends)[n-1]( outputs, vertical_coefficients + k, vertical_buffer, vertical_buffer_end );
 6435+      k += n;
 6436+      total -= n;
 6437+    } while ( total );
 6438+  }
 6439+
 6440+  STBIR_PROFILE_END( vertical );
 6441+}
 6442+
 6443+typedef void stbir__handle_scanline_for_scatter_func(stbir__info const * stbir_info, stbir__per_split_info* split_info);
 6444+
 6445+static void stbir__vertical_scatter_loop( stbir__info const * stbir_info, stbir__per_split_info* split_info, int split_count )
 6446+{
 6447+  int y, start_output_y, end_output_y, start_input_y, end_input_y;
 6448+  stbir__contributors* vertical_contributors = stbir_info->vertical.contributors;
 6449+  float const * vertical_coefficients = stbir_info->vertical.coefficients;
 6450+  stbir__handle_scanline_for_scatter_func * handle_scanline_for_scatter;
 6451+  void * scanline_scatter_buffer;
 6452+  void * scanline_scatter_buffer_end;
 6453+  int on_first_input_y, last_input_y;
 6454+  int width = (stbir_info->vertical_first) ? ( stbir_info->scanline_extents.conservative.n1-stbir_info->scanline_extents.conservative.n0+1 ) : stbir_info->horizontal.scale_info.output_sub_size;
 6455+  int width_times_channels = stbir_info->effective_channels * width;
 6456+
 6457+  STBIR_ASSERT( !stbir_info->vertical.is_gather );
 6458+
 6459+  start_output_y = split_info->start_output_y;
 6460+  end_output_y = split_info[split_count-1].end_output_y;  // may do multiple split counts
 6461+
 6462+  start_input_y = split_info->start_input_y;
 6463+  end_input_y = split_info[split_count-1].end_input_y;
 6464+
 6465+  // adjust for starting offset start_input_y
 6466+  y = start_input_y + stbir_info->vertical.filter_pixel_margin;
 6467+  vertical_contributors += y ;
 6468+  vertical_coefficients += stbir_info->vertical.coefficient_width * y;
 6469+
 6470+  if ( stbir_info->vertical_first )
 6471+  {
 6472+    handle_scanline_for_scatter = stbir__horizontal_resample_and_encode_first_scanline_from_scatter;
 6473+    scanline_scatter_buffer = split_info->decode_buffer;
 6474+    scanline_scatter_buffer_end = ( (char*) scanline_scatter_buffer ) + sizeof( float ) * stbir_info->effective_channels * (stbir_info->scanline_extents.conservative.n1-stbir_info->scanline_extents.conservative.n0+1);
 6475+  }
 6476+  else
 6477+  {
 6478+    handle_scanline_for_scatter = stbir__encode_first_scanline_from_scatter;
 6479+    scanline_scatter_buffer = split_info->vertical_buffer;
 6480+    scanline_scatter_buffer_end = ( (char*) scanline_scatter_buffer ) + sizeof( float ) * stbir_info->effective_channels * stbir_info->horizontal.scale_info.output_sub_size;
 6481+  }
 6482+
 6483+  // initialize the ring buffer for scattering
 6484+  split_info->ring_buffer_first_scanline = start_output_y;
 6485+  split_info->ring_buffer_last_scanline = -1;
 6486+  split_info->ring_buffer_begin_index = -1;
 6487+
 6488+  // mark all the buffers as empty to start
 6489+  for( y = 0 ; y < stbir_info->ring_buffer_num_entries ; y++ )
 6490+  {
 6491+    float * decode_buffer = stbir__get_ring_buffer_entry( stbir_info, split_info, y );
 6492+    decode_buffer[ width_times_channels ] = 0.0f; // clear two over for horizontals with a remnant of 3
 6493+    decode_buffer[ width_times_channels+1 ] = 0.0f; 
 6494+    decode_buffer[0] = STBIR__FLOAT_EMPTY_MARKER; // only used on scatter
 6495+  }
 6496+
 6497+  // do the loop in input space
 6498+  on_first_input_y = 1; last_input_y = start_input_y;
 6499+  for (y = start_input_y ; y < end_input_y; y++)
 6500+  {
 6501+    int out_first_scanline, out_last_scanline;
 6502+
 6503+    out_first_scanline = vertical_contributors->n0;
 6504+    out_last_scanline = vertical_contributors->n1;
 6505+
 6506+    STBIR_ASSERT(out_last_scanline - out_first_scanline + 1 <= stbir_info->ring_buffer_num_entries);
 6507+
 6508+    if ( ( out_last_scanline >= out_first_scanline ) && ( ( ( out_first_scanline >= start_output_y ) && ( out_first_scanline < end_output_y ) ) || ( ( out_last_scanline >= start_output_y ) && ( out_last_scanline < end_output_y ) ) ) )
 6509+    {
 6510+      float const * vc = vertical_coefficients;
 6511+
 6512+      // keep track of the range actually seen for the next resize
 6513+      last_input_y = y;
 6514+      if ( ( on_first_input_y ) && ( y > start_input_y ) )
 6515+        split_info->start_input_y = y;
 6516+      on_first_input_y = 0;
 6517+
 6518+      // clip the region
 6519+      if ( out_first_scanline < start_output_y )
 6520+      {
 6521+        vc += start_output_y - out_first_scanline;
 6522+        out_first_scanline = start_output_y;
 6523+      }
 6524+
 6525+      if ( out_last_scanline >= end_output_y )
 6526+        out_last_scanline = end_output_y - 1;
 6527+
 6528+      // if very first scanline, init the index
 6529+      if (split_info->ring_buffer_begin_index < 0)
 6530+        split_info->ring_buffer_begin_index = out_first_scanline - start_output_y;
 6531+
 6532+      STBIR_ASSERT( split_info->ring_buffer_begin_index <= out_first_scanline );
 6533+
 6534+      // Decode the nth scanline from the source image into the decode buffer.
 6535+      stbir__decode_scanline( stbir_info, y, split_info->decode_buffer  STBIR_ONLY_PROFILE_SET_SPLIT_INFO );
 6536+
 6537+      // When horizontal first, we resample horizontally into the vertical buffer before we scatter it out
 6538+      if ( !stbir_info->vertical_first )
 6539+        stbir__resample_horizontal_gather( stbir_info, split_info->vertical_buffer, split_info->decode_buffer  STBIR_ONLY_PROFILE_SET_SPLIT_INFO );
 6540+
 6541+      // Now it's sitting in the buffer ready to be distributed into the ring buffers.
 6542+
 6543+      // evict from the ringbuffer, if we need are full
 6544+      if ( ( ( split_info->ring_buffer_last_scanline - split_info->ring_buffer_first_scanline + 1 ) == stbir_info->ring_buffer_num_entries ) &&
 6545+           ( out_last_scanline > split_info->ring_buffer_last_scanline ) )
 6546+        handle_scanline_for_scatter( stbir_info, split_info );
 6547+
 6548+      // Now the horizontal buffer is ready to write to all ring buffer rows, so do it.
 6549+      stbir__resample_vertical_scatter(stbir_info, split_info, out_first_scanline, out_last_scanline, vc, (float*)scanline_scatter_buffer, (float*)scanline_scatter_buffer_end );
 6550+
 6551+      // update the end of the buffer
 6552+      if ( out_last_scanline > split_info->ring_buffer_last_scanline )
 6553+        split_info->ring_buffer_last_scanline = out_last_scanline;
 6554+    }
 6555+    ++vertical_contributors;
 6556+    vertical_coefficients += stbir_info->vertical.coefficient_width;
 6557+  }
 6558+
 6559+  // now evict the scanlines that are left over in the ring buffer
 6560+  while ( split_info->ring_buffer_first_scanline < end_output_y )
 6561+    handle_scanline_for_scatter(stbir_info, split_info);
 6562+
 6563+  // update the end_input_y if we do multiple resizes with the same data
 6564+  ++last_input_y;
 6565+  for( y = 0 ; y < split_count; y++ )
 6566+    if ( split_info[y].end_input_y > last_input_y )
 6567+      split_info[y].end_input_y = last_input_y;
 6568+}
 6569+
 6570+
 6571+static stbir__kernel_callback * stbir__builtin_kernels[] =   { 0, stbir__filter_trapezoid,  stbir__filter_triangle, stbir__filter_cubic, stbir__filter_catmullrom, stbir__filter_mitchell, stbir__filter_point };
 6572+static stbir__support_callback * stbir__builtin_supports[] = { 0, stbir__support_trapezoid, stbir__support_one,     stbir__support_two,  stbir__support_two,       stbir__support_two,     stbir__support_zeropoint5 };
 6573+
 6574+static void stbir__set_sampler(stbir__sampler * samp, stbir_filter filter, stbir__kernel_callback * kernel, stbir__support_callback * support, stbir_edge edge, stbir__scale_info * scale_info, int always_gather, void * user_data )
 6575+{
 6576+  // set filter
 6577+  if (filter == 0)
 6578+  {
 6579+    filter = STBIR_DEFAULT_FILTER_DOWNSAMPLE; // default to downsample
 6580+    if (scale_info->scale >= ( 1.0f - stbir__small_float ) )
 6581+    {
 6582+      if ( (scale_info->scale <= ( 1.0f + stbir__small_float ) ) && ( STBIR_CEILF(scale_info->pixel_shift) == scale_info->pixel_shift ) )
 6583+        filter = STBIR_FILTER_POINT_SAMPLE;
 6584+      else
 6585+        filter = STBIR_DEFAULT_FILTER_UPSAMPLE;
 6586+    }
 6587+  }
 6588+  samp->filter_enum = filter;
 6589+
 6590+  STBIR_ASSERT(samp->filter_enum != 0);
 6591+  STBIR_ASSERT((unsigned)samp->filter_enum < STBIR_FILTER_OTHER);
 6592+  samp->filter_kernel = stbir__builtin_kernels[ filter ];
 6593+  samp->filter_support = stbir__builtin_supports[ filter ];
 6594+
 6595+  if ( kernel && support )
 6596+  {
 6597+    samp->filter_kernel = kernel;
 6598+    samp->filter_support = support;
 6599+    samp->filter_enum = STBIR_FILTER_OTHER;
 6600+  }
 6601+
 6602+  samp->edge = edge;
 6603+  samp->filter_pixel_width  = stbir__get_filter_pixel_width (samp->filter_support, scale_info->scale, user_data );
 6604+  // Gather is always better, but in extreme downsamples, you have to most or all of the data in memory
 6605+  //    For horizontal, we always have all the pixels, so we always use gather here (always_gather==1).
 6606+  //    For vertical, we use gather if scaling up (which means we will have samp->filter_pixel_width
 6607+  //    scanlines in memory at once).
 6608+  samp->is_gather = 0;
 6609+  if ( scale_info->scale >= ( 1.0f - stbir__small_float ) )
 6610+    samp->is_gather = 1;
 6611+  else if ( ( always_gather ) || ( samp->filter_pixel_width <= STBIR_FORCE_GATHER_FILTER_SCANLINES_AMOUNT ) )
 6612+    samp->is_gather = 2;
 6613+
 6614+  // pre calculate stuff based on the above
 6615+  samp->coefficient_width = stbir__get_coefficient_width(samp, samp->is_gather, user_data);
 6616+
 6617+  // filter_pixel_width is the conservative size in pixels of input that affect an output pixel.
 6618+  //   In rare cases (only with 2 pix to 1 pix with the default filters), it's possible that the 
 6619+  //   filter will extend before or after the scanline beyond just one extra entire copy of the 
 6620+  //   scanline (we would hit the edge twice). We don't let you do that, so we clamp the total 
 6621+  //   width to 3x the total of input pixel (once for the scanline, once for the left side 
 6622+  //   overhang, and once for the right side). We only do this for edge mode, since the other 
 6623+  //   modes can just re-edge clamp back in again.
 6624+  if ( edge == STBIR_EDGE_WRAP )
 6625+    if ( samp->filter_pixel_width > ( scale_info->input_full_size * 3 ) )
 6626+      samp->filter_pixel_width = scale_info->input_full_size * 3;
 6627+
 6628+  // This is how much to expand buffers to account for filters seeking outside
 6629+  // the image boundaries.
 6630+  samp->filter_pixel_margin = samp->filter_pixel_width / 2;
 6631+  
 6632+  // filter_pixel_margin is the amount that this filter can overhang on just one side of either 
 6633+  //   end of the scanline (left or the right). Since we only allow you to overhang 1 scanline's 
 6634+  //   worth of pixels, we clamp this one side of overhang to the input scanline size. Again, 
 6635+  //   this clamping only happens in rare cases with the default filters (2 pix to 1 pix). 
 6636+  if ( edge == STBIR_EDGE_WRAP )
 6637+    if ( samp->filter_pixel_margin > scale_info->input_full_size )
 6638+      samp->filter_pixel_margin = scale_info->input_full_size;
 6639+
 6640+  samp->num_contributors = stbir__get_contributors(samp, samp->is_gather);
 6641+
 6642+  samp->contributors_size = samp->num_contributors * sizeof(stbir__contributors);
 6643+  samp->coefficients_size = samp->num_contributors * samp->coefficient_width * sizeof(float) + sizeof(float)*STBIR_INPUT_CALLBACK_PADDING; // extra sizeof(float) is padding
 6644+
 6645+  samp->gather_prescatter_contributors = 0;
 6646+  samp->gather_prescatter_coefficients = 0;
 6647+  if ( samp->is_gather == 0 )
 6648+  {
 6649+    samp->gather_prescatter_coefficient_width = samp->filter_pixel_width;
 6650+    samp->gather_prescatter_num_contributors  = stbir__get_contributors(samp, 2);
 6651+    samp->gather_prescatter_contributors_size = samp->gather_prescatter_num_contributors * sizeof(stbir__contributors);
 6652+    samp->gather_prescatter_coefficients_size = samp->gather_prescatter_num_contributors * samp->gather_prescatter_coefficient_width * sizeof(float);
 6653+  }
 6654+}
 6655+
 6656+static void stbir__get_conservative_extents( stbir__sampler * samp, stbir__contributors * range, void * user_data )
 6657+{
 6658+  float scale = samp->scale_info.scale;
 6659+  float out_shift = samp->scale_info.pixel_shift;
 6660+  stbir__support_callback * support = samp->filter_support;
 6661+  int input_full_size = samp->scale_info.input_full_size;
 6662+  stbir_edge edge = samp->edge;
 6663+  float inv_scale = samp->scale_info.inv_scale;
 6664+
 6665+  STBIR_ASSERT( samp->is_gather != 0 );
 6666+
 6667+  if ( samp->is_gather == 1 )
 6668+  {
 6669+    int in_first_pixel, in_last_pixel;
 6670+    float out_filter_radius = support(inv_scale, user_data) * scale;
 6671+
 6672+    stbir__calculate_in_pixel_range( &in_first_pixel, &in_last_pixel, 0.5, out_filter_radius, inv_scale, out_shift, input_full_size, edge );
 6673+    range->n0 = in_first_pixel;
 6674+    stbir__calculate_in_pixel_range( &in_first_pixel, &in_last_pixel, ( (float)(samp->scale_info.output_sub_size-1) ) + 0.5f, out_filter_radius, inv_scale, out_shift, input_full_size, edge );
 6675+    range->n1 = in_last_pixel;
 6676+  }
 6677+  else if ( samp->is_gather == 2 ) // downsample gather, refine
 6678+  {
 6679+    float in_pixels_radius = support(scale, user_data) * inv_scale;
 6680+    int filter_pixel_margin = samp->filter_pixel_margin;
 6681+    int output_sub_size = samp->scale_info.output_sub_size;
 6682+    int input_end;
 6683+    int n;
 6684+    int in_first_pixel, in_last_pixel;
 6685+
 6686+    // get a conservative area of the input range
 6687+    stbir__calculate_in_pixel_range( &in_first_pixel, &in_last_pixel, 0, 0, inv_scale, out_shift, input_full_size, edge );
 6688+    range->n0 = in_first_pixel;
 6689+    stbir__calculate_in_pixel_range( &in_first_pixel, &in_last_pixel, (float)output_sub_size, 0, inv_scale, out_shift, input_full_size, edge );
 6690+    range->n1 = in_last_pixel;
 6691+
 6692+    // now go through the margin to the start of area to find bottom
 6693+    n = range->n0 + 1;
 6694+    input_end = -filter_pixel_margin;
 6695+    while( n >= input_end )
 6696+    {
 6697+      int out_first_pixel, out_last_pixel;
 6698+      stbir__calculate_out_pixel_range( &out_first_pixel, &out_last_pixel, ((float)n)+0.5f, in_pixels_radius, scale, out_shift, output_sub_size );
 6699+      if ( out_first_pixel > out_last_pixel )
 6700+        break;
 6701+
 6702+      if ( ( out_first_pixel < output_sub_size ) || ( out_last_pixel >= 0 ) )
 6703+        range->n0 = n;
 6704+      --n;
 6705+    }
 6706+
 6707+    // now go through the end of the area through the margin to find top
 6708+    n = range->n1 - 1;
 6709+    input_end = n + 1 + filter_pixel_margin;
 6710+    while( n <= input_end )
 6711+    {
 6712+      int out_first_pixel, out_last_pixel;
 6713+      stbir__calculate_out_pixel_range( &out_first_pixel, &out_last_pixel, ((float)n)+0.5f, in_pixels_radius, scale, out_shift, output_sub_size );
 6714+      if ( out_first_pixel > out_last_pixel )
 6715+        break;
 6716+      if ( ( out_first_pixel < output_sub_size ) || ( out_last_pixel >= 0 ) )
 6717+        range->n1 = n;
 6718+      ++n;
 6719+    }
 6720+  }
 6721+
 6722+  if ( samp->edge == STBIR_EDGE_WRAP )
 6723+  {
 6724+    // if we are wrapping, and we are very close to the image size (so the edges might merge), just use the scanline up to the edge
 6725+    if ( ( range->n0 > 0 ) && ( range->n1 >= input_full_size ) )
 6726+    {
 6727+      int marg = range->n1 - input_full_size + 1;
 6728+      if ( ( marg + STBIR__MERGE_RUNS_PIXEL_THRESHOLD ) >= range->n0 )
 6729+        range->n0 = 0;
 6730+    }
 6731+    if ( ( range->n0 < 0 ) && ( range->n1 < (input_full_size-1) ) )
 6732+    {
 6733+      int marg = -range->n0;
 6734+      if ( ( input_full_size - marg - STBIR__MERGE_RUNS_PIXEL_THRESHOLD - 1 ) <= range->n1 )
 6735+        range->n1 = input_full_size - 1;
 6736+    }
 6737+  }
 6738+  else
 6739+  {
 6740+    // for non-edge-wrap modes, we never read over the edge, so clamp
 6741+    if ( range->n0 < 0 )
 6742+      range->n0 = 0;
 6743+    if ( range->n1 >= input_full_size )
 6744+      range->n1 = input_full_size - 1;
 6745+  }
 6746+}
 6747+
 6748+static void stbir__get_split_info( stbir__per_split_info* split_info, int splits, int output_height, int vertical_pixel_margin, int input_full_height, int is_gather, stbir__contributors * contribs )
 6749+{
 6750+  int i, cur;
 6751+  int left = output_height;
 6752+
 6753+  cur = 0;
 6754+  for( i = 0 ; i < splits ; i++ )
 6755+  {
 6756+    int each;
 6757+
 6758+    split_info[i].start_output_y = cur;
 6759+    each = left / ( splits - i );
 6760+    split_info[i].end_output_y = cur + each;
 6761+
 6762+    // ok, when we are gathering, we need to make sure we are starting on a y offset that doesn't have
 6763+    //   a "special" set of coefficients. Basically, with exactly the right filter at exactly the right
 6764+    //   resize at exactly the right phase, some of the coefficents can be zero. When they are zero, we
 6765+    //   don't process them at all.  But this leads to a tricky thing with the thread splits, where we
 6766+    //   might have a set of two coeffs like this for example: (4,4) and (3,6).  The 4,4 means there was
 6767+    //   just one single coeff because things worked out perfectly (normally, they all have 4 coeffs
 6768+    //   like the range 3,6.  The problem is that if we start right on the (4,4) on a brand new thread,
 6769+    //   then when we get to (3,6), we don't have the "3" sample in memory (because we didn't load
 6770+    //   it on the initial (4,4) range because it didn't have a 3 (we only add new samples that are 
 6771+    //   larger than our existing samples - it's just how the eviction works). So, our solution here
 6772+    //   is pretty simple, if we start right on a range that has samples that start earlier, then we 
 6773+    //   simply bump up our previous thread split range to include it, and then start this threads
 6774+    //   range with the smaller sample. It just moves one scanline from one thread split to another,
 6775+    //   so that we end with the unusual one, instead of start with it. To do this, we check 2-4 
 6776+    //   sample at each thread split start and then occassionally move them.
 6777+    
 6778+    if ( ( is_gather ) && ( i ) )
 6779+    {
 6780+      stbir__contributors * small_contribs;
 6781+      int j, smallest, stop, start_n0;
 6782+      stbir__contributors * split_contribs = contribs + cur;
 6783+
 6784+      // scan for a max of 3x the filter width or until the next thread split
 6785+      stop = vertical_pixel_margin * 3;
 6786+      if ( each < stop )
 6787+        stop = each;
 6788+
 6789+      // loops a few times before early out
 6790+      smallest = 0;
 6791+      small_contribs = split_contribs;
 6792+      start_n0 = small_contribs->n0;
 6793+      for( j = 1 ; j <= stop ; j++ )
 6794+      {
 6795+        ++split_contribs;
 6796+        if ( split_contribs->n0 > start_n0 )
 6797+          break;
 6798+        if ( split_contribs->n0 < small_contribs->n0 )
 6799+        {
 6800+          small_contribs = split_contribs;
 6801+          smallest = j;
 6802+        }
 6803+      }
 6804+
 6805+      split_info[i-1].end_output_y += smallest;
 6806+      split_info[i].start_output_y += smallest;
 6807+    }
 6808+
 6809+    cur += each;
 6810+    left -= each;
 6811+
 6812+    // scatter range (updated to minimum as you run it)
 6813+    split_info[i].start_input_y = -vertical_pixel_margin;
 6814+    split_info[i].end_input_y = input_full_height + vertical_pixel_margin;
 6815+  }
 6816+}
 6817+
 6818+static void stbir__free_internal_mem( stbir__info *info )
 6819+{
 6820+  #define STBIR__FREE_AND_CLEAR( ptr ) { if ( ptr ) { void * p = (ptr); (ptr) = 0; STBIR_FREE( p, info->user_data); } }
 6821+
 6822+  if ( info )
 6823+  {
 6824+  #ifndef STBIR__SEPARATE_ALLOCATIONS
 6825+    STBIR__FREE_AND_CLEAR( info->alloced_mem );
 6826+  #else
 6827+    int i,j;
 6828+
 6829+    if ( ( info->vertical.gather_prescatter_contributors ) && ( (void*)info->vertical.gather_prescatter_contributors != (void*)info->split_info[0].decode_buffer ) )
 6830+    {
 6831+      STBIR__FREE_AND_CLEAR( info->vertical.gather_prescatter_coefficients );
 6832+      STBIR__FREE_AND_CLEAR( info->vertical.gather_prescatter_contributors );
 6833+    }
 6834+    for( i = 0 ; i < info->splits ; i++ )
 6835+    {
 6836+      for( j = 0 ; j < info->alloc_ring_buffer_num_entries ; j++ )
 6837+      {
 6838+        #ifdef STBIR_SIMD8
 6839+        if ( info->effective_channels == 3 )
 6840+          --info->split_info[i].ring_buffers[j]; // avx in 3 channel mode needs one float at the start of the buffer
 6841+        #endif
 6842+        STBIR__FREE_AND_CLEAR( info->split_info[i].ring_buffers[j] );
 6843+      }
 6844+
 6845+      #ifdef STBIR_SIMD8
 6846+      if ( info->effective_channels == 3 )
 6847+        --info->split_info[i].decode_buffer; // avx in 3 channel mode needs one float at the start of the buffer
 6848+      #endif
 6849+      STBIR__FREE_AND_CLEAR( info->split_info[i].decode_buffer );
 6850+      STBIR__FREE_AND_CLEAR( info->split_info[i].ring_buffers );
 6851+      STBIR__FREE_AND_CLEAR( info->split_info[i].vertical_buffer );
 6852+    }
 6853+    STBIR__FREE_AND_CLEAR( info->split_info );
 6854+    if ( info->vertical.coefficients != info->horizontal.coefficients )
 6855+    {
 6856+      STBIR__FREE_AND_CLEAR( info->vertical.coefficients );
 6857+      STBIR__FREE_AND_CLEAR( info->vertical.contributors );
 6858+    }
 6859+    STBIR__FREE_AND_CLEAR( info->horizontal.coefficients );
 6860+    STBIR__FREE_AND_CLEAR( info->horizontal.contributors );
 6861+    STBIR__FREE_AND_CLEAR( info->alloced_mem );
 6862+    STBIR_FREE( info, info->user_data );
 6863+  #endif
 6864+  }
 6865+
 6866+  #undef STBIR__FREE_AND_CLEAR
 6867+}
 6868+
 6869+static int stbir__get_max_split( int splits, int height )
 6870+{
 6871+  int i;
 6872+  int max = 0;
 6873+
 6874+  for( i = 0 ; i < splits ; i++ )
 6875+  {
 6876+    int each = height / ( splits - i );
 6877+    if ( each > max )
 6878+      max = each;
 6879+    height -= each;
 6880+  }
 6881+  return max;
 6882+}
 6883+
 6884+static stbir__horizontal_gather_channels_func ** stbir__horizontal_gather_n_coeffs_funcs[8] =
 6885+{
 6886+  0, stbir__horizontal_gather_1_channels_with_n_coeffs_funcs, stbir__horizontal_gather_2_channels_with_n_coeffs_funcs, stbir__horizontal_gather_3_channels_with_n_coeffs_funcs, stbir__horizontal_gather_4_channels_with_n_coeffs_funcs, 0,0, stbir__horizontal_gather_7_channels_with_n_coeffs_funcs
 6887+};
 6888+
 6889+static stbir__horizontal_gather_channels_func ** stbir__horizontal_gather_channels_funcs[8] =
 6890+{
 6891+  0, stbir__horizontal_gather_1_channels_funcs, stbir__horizontal_gather_2_channels_funcs, stbir__horizontal_gather_3_channels_funcs, stbir__horizontal_gather_4_channels_funcs, 0,0, stbir__horizontal_gather_7_channels_funcs
 6892+};
 6893+
 6894+// there are six resize classifications: 0 == vertical scatter, 1 == vertical gather < 1x scale, 2 == vertical gather 1x-2x scale, 4 == vertical gather < 3x scale, 4 == vertical gather > 3x scale, 5 == <=4 pixel height, 6 == <=4 pixel wide column
 6895+#define STBIR_RESIZE_CLASSIFICATIONS 8
 6896+
 6897+static float stbir__compute_weights[5][STBIR_RESIZE_CLASSIFICATIONS][4]=  // 5 = 0=1chan, 1=2chan, 2=3chan, 3=4chan, 4=7chan
 6898+{
 6899+  {
 6900+    { 1.00000f, 1.00000f, 0.31250f, 1.00000f },
 6901+    { 0.56250f, 0.59375f, 0.00000f, 0.96875f },
 6902+    { 1.00000f, 0.06250f, 0.00000f, 1.00000f },
 6903+    { 0.00000f, 0.09375f, 1.00000f, 1.00000f },
 6904+    { 1.00000f, 1.00000f, 0.31250f, 1.00000f },
 6905+    { 0.03125f, 0.12500f, 1.00000f, 1.00000f },
 6906+    { 1.00000f, 1.00000f, 0.06250f, 1.00000f },
 6907+    { 0.00000f, 1.00000f, 0.00000f, 0.03125f },
 6908+  }, {
 6909+    { 0.00000f, 0.84375f, 0.00000f, 0.03125f },
 6910+    { 0.09375f, 0.93750f, 0.00000f, 0.78125f },
 6911+    { 0.87500f, 0.21875f, 0.00000f, 0.96875f },
 6912+    { 0.09375f, 0.09375f, 1.00000f, 1.00000f },
 6913+    { 0.00000f, 0.84375f, 0.00000f, 0.03125f },
 6914+    { 0.03125f, 0.12500f, 1.00000f, 1.00000f },
 6915+    { 1.00000f, 1.00000f, 0.06250f, 1.00000f },
 6916+    { 0.00000f, 1.00000f, 0.00000f, 0.53125f },
 6917+  }, {
 6918+    { 0.00000f, 0.53125f, 0.00000f, 0.03125f },
 6919+    { 0.06250f, 0.96875f, 0.00000f, 0.53125f },
 6920+    { 0.87500f, 0.18750f, 0.00000f, 0.93750f },
 6921+    { 0.00000f, 0.09375f, 1.00000f, 1.00000f },
 6922+    { 0.00000f, 0.53125f, 0.00000f, 0.03125f },
 6923+    { 0.03125f, 0.12500f, 1.00000f, 1.00000f },
 6924+    { 1.00000f, 1.00000f, 0.06250f, 1.00000f },
 6925+    { 0.00000f, 1.00000f, 0.00000f, 0.56250f },
 6926+  }, {
 6927+    { 0.00000f, 0.50000f, 0.00000f, 0.71875f },
 6928+    { 0.06250f, 0.84375f, 0.00000f, 0.87500f },
 6929+    { 1.00000f, 0.50000f, 0.50000f, 0.96875f },
 6930+    { 1.00000f, 0.09375f, 0.31250f, 0.50000f },
 6931+    { 0.00000f, 0.50000f, 0.00000f, 0.71875f },
 6932+    { 1.00000f, 0.03125f, 0.03125f, 0.53125f },
 6933+    { 1.00000f, 1.00000f, 0.06250f, 1.00000f },
 6934+    { 0.00000f, 1.00000f, 0.03125f, 0.18750f },
 6935+  }, {
 6936+    { 0.00000f, 0.59375f, 0.00000f, 0.96875f },
 6937+    { 0.06250f, 0.81250f, 0.06250f, 0.59375f },
 6938+    { 0.75000f, 0.43750f, 0.12500f, 0.96875f },
 6939+    { 0.87500f, 0.06250f, 0.18750f, 0.43750f },
 6940+    { 0.00000f, 0.59375f, 0.00000f, 0.96875f },
 6941+    { 0.15625f, 0.12500f, 1.00000f, 1.00000f },
 6942+    { 1.00000f, 1.00000f, 0.06250f, 1.00000f },
 6943+    { 0.00000f, 1.00000f, 0.03125f, 0.34375f },
 6944+  }
 6945+};
 6946+
 6947+// structure that allow us to query and override info for training the costs
 6948+typedef struct STBIR__V_FIRST_INFO
 6949+{
 6950+  double v_cost, h_cost;
 6951+  int control_v_first; // 0 = no control, 1 = force hori, 2 = force vert
 6952+  int v_first;
 6953+  int v_resize_classification;
 6954+  int is_gather;
 6955+} STBIR__V_FIRST_INFO;
 6956+
 6957+#ifdef STBIR__V_FIRST_INFO_BUFFER
 6958+static STBIR__V_FIRST_INFO STBIR__V_FIRST_INFO_BUFFER = {0};
 6959+#define STBIR__V_FIRST_INFO_POINTER &STBIR__V_FIRST_INFO_BUFFER
 6960+#else
 6961+#define STBIR__V_FIRST_INFO_POINTER 0
 6962+#endif
 6963+
 6964+// Figure out whether to scale along the horizontal or vertical first.
 6965+//   This only *super* important when you are scaling by a massively
 6966+//   different amount in the vertical vs the horizontal (for example, if
 6967+//   you are scaling by 2x in the width, and 0.5x in the height, then you
 6968+//   want to do the vertical scale first, because it's around 3x faster
 6969+//   in that order.
 6970+//
 6971+//   In more normal circumstances, this makes a 20-40% differences, so
 6972+//     it's good to get right, but not critical. The normal way that you
 6973+//     decide which direction goes first is just figuring out which
 6974+//     direction does more multiplies. But with modern CPUs with their
 6975+//     fancy caches and SIMD and high IPC abilities, so there's just a lot
 6976+//     more that goes into it.
 6977+//
 6978+//   My handwavy sort of solution is to have an app that does a whole
 6979+//     bunch of timing for both vertical and horizontal first modes,
 6980+//     and then another app that can read lots of these timing files
 6981+//     and try to search for the best weights to use. Dotimings.c
 6982+//     is the app that does a bunch of timings, and vf_train.c is the
 6983+//     app that solves for the best weights (and shows how well it
 6984+//     does currently).
 6985+
 6986+static int stbir__should_do_vertical_first( float weights_table[STBIR_RESIZE_CLASSIFICATIONS][4], int horizontal_filter_pixel_width, float horizontal_scale, int horizontal_output_size, int vertical_filter_pixel_width, float vertical_scale, int vertical_output_size, int is_gather, STBIR__V_FIRST_INFO * info )
 6987+{
 6988+  double v_cost, h_cost;
 6989+  float * weights;
 6990+  int vertical_first;
 6991+  int v_classification;
 6992+
 6993+  // categorize the resize into buckets
 6994+  if ( ( vertical_output_size <= 4 ) || ( horizontal_output_size <= 4 ) )
 6995+    v_classification = ( vertical_output_size < horizontal_output_size ) ? 6 : 7;
 6996+  else if ( ( !is_gather ) && ( ( vertical_output_size <= 16 ) || ( horizontal_output_size <= 16 ) ) )
 6997+    v_classification = 4;
 6998+  else if ( vertical_scale <= 1.0f )
 6999+    v_classification = ( is_gather ) ? 1 : 0;
 7000+  else if ( vertical_scale <= 2.0f)
 7001+    v_classification = 2;
 7002+  else if ( vertical_scale <= 3.0f)
 7003+    v_classification = 3;
 7004+  else 
 7005+    v_classification = 5; // everything bigger than 3x
 7006+
 7007+  // use the right weights
 7008+  weights = weights_table[ v_classification ];
 7009+
 7010+  // this is the costs when you don't take into account modern CPUs with high ipc and simd and caches - wish we had a better estimate
 7011+  h_cost = (float)horizontal_filter_pixel_width * weights[0] + horizontal_scale * (float)vertical_filter_pixel_width * weights[1];
 7012+  v_cost = (float)vertical_filter_pixel_width  * weights[2] + vertical_scale * (float)horizontal_filter_pixel_width * weights[3];
 7013+
 7014+  // use computation estimate to decide vertical first or not
 7015+  vertical_first = ( v_cost <= h_cost ) ? 1 : 0;
 7016+
 7017+  // save these, if requested
 7018+  if ( info )
 7019+  {
 7020+    info->h_cost = h_cost;
 7021+    info->v_cost = v_cost;
 7022+    info->v_resize_classification = v_classification;
 7023+    info->v_first = vertical_first;
 7024+    info->is_gather = is_gather;
 7025+  }
 7026+
 7027+  // and this allows us to override everything for testing (see dotiming.c)
 7028+  if ( ( info ) && ( info->control_v_first ) )
 7029+    vertical_first = ( info->control_v_first == 2 ) ? 1 : 0;
 7030+
 7031+  return vertical_first;
 7032+}
 7033+
 7034+// layout lookups - must match stbir_internal_pixel_layout
 7035+static unsigned char stbir__pixel_channels[] = {
 7036+  1,2,3,3,4,   // 1ch, 2ch, rgb, bgr, 4ch
 7037+  4,4,4,4,2,2, // RGBA,BGRA,ARGB,ABGR,RA,AR
 7038+  4,4,4,4,2,2, // RGBA_PM,BGRA_PM,ARGB_PM,ABGR_PM,RA_PM,AR_PM
 7039+};
 7040+
 7041+// the internal pixel layout enums are in a different order, so we can easily do range comparisons of types
 7042+//   the public pixel layout is ordered in a way that if you cast num_channels (1-4) to the enum, you get something sensible
 7043+static stbir_internal_pixel_layout stbir__pixel_layout_convert_public_to_internal[] = {
 7044+  STBIRI_BGR, STBIRI_1CHANNEL, STBIRI_2CHANNEL, STBIRI_RGB, STBIRI_RGBA,
 7045+  STBIRI_4CHANNEL, STBIRI_BGRA, STBIRI_ARGB, STBIRI_ABGR, STBIRI_RA, STBIRI_AR,
 7046+  STBIRI_RGBA_PM, STBIRI_BGRA_PM, STBIRI_ARGB_PM, STBIRI_ABGR_PM, STBIRI_RA_PM, STBIRI_AR_PM,
 7047+};
 7048+
 7049+static stbir__info * stbir__alloc_internal_mem_and_build_samplers( stbir__sampler * horizontal, stbir__sampler * vertical, stbir__contributors * conservative, stbir_pixel_layout input_pixel_layout_public, stbir_pixel_layout output_pixel_layout_public, int splits, int new_x, int new_y, int fast_alpha, void * user_data STBIR_ONLY_PROFILE_BUILD_GET_INFO )
 7050+{
 7051+  static char stbir_channel_count_index[8]={ 9,0,1,2, 3,9,9,4 };
 7052+
 7053+  stbir__info * info = 0;
 7054+  void * alloced = 0;
 7055+  size_t alloced_total = 0;
 7056+  int vertical_first;
 7057+  size_t decode_buffer_size, ring_buffer_length_bytes, ring_buffer_size, vertical_buffer_size;
 7058+  int alloc_ring_buffer_num_entries;
 7059+
 7060+  int alpha_weighting_type = 0; // 0=none, 1=simple, 2=fancy
 7061+  int conservative_split_output_size = stbir__get_max_split( splits, vertical->scale_info.output_sub_size );
 7062+  stbir_internal_pixel_layout input_pixel_layout = stbir__pixel_layout_convert_public_to_internal[ input_pixel_layout_public ];
 7063+  stbir_internal_pixel_layout output_pixel_layout = stbir__pixel_layout_convert_public_to_internal[ output_pixel_layout_public ];
 7064+  int channels = stbir__pixel_channels[ input_pixel_layout ];
 7065+  int effective_channels = channels;
 7066+
 7067+  // first figure out what type of alpha weighting to use (if any)
 7068+  if ( ( horizontal->filter_enum != STBIR_FILTER_POINT_SAMPLE ) || ( vertical->filter_enum != STBIR_FILTER_POINT_SAMPLE ) ) // no alpha weighting on point sampling
 7069+  {
 7070+    if ( ( input_pixel_layout >= STBIRI_RGBA ) && ( input_pixel_layout <= STBIRI_AR ) && ( output_pixel_layout >= STBIRI_RGBA ) && ( output_pixel_layout <= STBIRI_AR ) )
 7071+    {
 7072+      if ( fast_alpha )
 7073+      {
 7074+        alpha_weighting_type = 4;
 7075+      }
 7076+      else
 7077+      {
 7078+        static int fancy_alpha_effective_cnts[6] = { 7, 7, 7, 7, 3, 3 };
 7079+        alpha_weighting_type = 2;
 7080+        effective_channels = fancy_alpha_effective_cnts[ input_pixel_layout - STBIRI_RGBA ];
 7081+      }
 7082+    }
 7083+    else if ( ( input_pixel_layout >= STBIRI_RGBA_PM ) && ( input_pixel_layout <= STBIRI_AR_PM ) && ( output_pixel_layout >= STBIRI_RGBA ) && ( output_pixel_layout <= STBIRI_AR ) )
 7084+    {
 7085+      // input premult, output non-premult
 7086+      alpha_weighting_type = 3;
 7087+    }
 7088+    else if ( ( input_pixel_layout >= STBIRI_RGBA ) && ( input_pixel_layout <= STBIRI_AR ) && ( output_pixel_layout >= STBIRI_RGBA_PM ) && ( output_pixel_layout <= STBIRI_AR_PM ) )
 7089+    {
 7090+      // input non-premult, output premult
 7091+      alpha_weighting_type = 1;
 7092+    }
 7093+  }
 7094+
 7095+  // channel in and out count must match currently
 7096+  if ( channels != stbir__pixel_channels[ output_pixel_layout ] )
 7097+    return 0;
 7098+
 7099+  // get vertical first
 7100+  vertical_first = stbir__should_do_vertical_first( stbir__compute_weights[ (int)stbir_channel_count_index[ effective_channels ] ], horizontal->filter_pixel_width, horizontal->scale_info.scale, horizontal->scale_info.output_sub_size, vertical->filter_pixel_width, vertical->scale_info.scale, vertical->scale_info.output_sub_size, vertical->is_gather, STBIR__V_FIRST_INFO_POINTER );
 7101+
 7102+  // sometimes read one float off in some of the unrolled loops (with a weight of zero coeff, so it doesn't have an effect)
 7103+  //   we use a few extra floats instead of just 1, so that input callback buffer can overlap with the decode buffer without
 7104+  //   the conversion routines overwriting the callback input data.
 7105+  decode_buffer_size = ( conservative->n1 - conservative->n0 + 1 ) * effective_channels * sizeof(float) + sizeof(float)*STBIR_INPUT_CALLBACK_PADDING; // extra floats for input callback stagger
 7106+
 7107+#if defined( STBIR__SEPARATE_ALLOCATIONS ) && defined(STBIR_SIMD8)
 7108+  if ( effective_channels == 3 )
 7109+    decode_buffer_size += sizeof(float); // avx in 3 channel mode needs one float at the start of the buffer (only with separate allocations)
 7110+#endif
 7111+
 7112+  ring_buffer_length_bytes = (size_t)horizontal->scale_info.output_sub_size * (size_t)effective_channels * sizeof(float) + sizeof(float)*STBIR_INPUT_CALLBACK_PADDING; // extra floats for padding
 7113+
 7114+  // if we do vertical first, the ring buffer holds a whole decoded line
 7115+  if ( vertical_first )
 7116+    ring_buffer_length_bytes = ( decode_buffer_size + 15 ) & ~15;
 7117+
 7118+  if ( ( ring_buffer_length_bytes & 4095 ) == 0 ) ring_buffer_length_bytes += 64*3; // avoid 4k alias
 7119+
 7120+  // One extra entry because floating point precision problems sometimes cause an extra to be necessary.
 7121+  alloc_ring_buffer_num_entries = vertical->filter_pixel_width + 1;
 7122+
 7123+  // we never need more ring buffer entries than the scanlines we're outputting when in scatter mode
 7124+  if ( ( !vertical->is_gather ) && ( alloc_ring_buffer_num_entries > conservative_split_output_size ) )
 7125+    alloc_ring_buffer_num_entries = conservative_split_output_size;
 7126+
 7127+  ring_buffer_size = (size_t)alloc_ring_buffer_num_entries * (size_t)ring_buffer_length_bytes;
 7128+
 7129+  // The vertical buffer is used differently, depending on whether we are scattering
 7130+  //   the vertical scanlines, or gathering them.
 7131+  //   If scattering, it's used at the temp buffer to accumulate each output.
 7132+  //   If gathering, it's just the output buffer.
 7133+  vertical_buffer_size = (size_t)horizontal->scale_info.output_sub_size * (size_t)effective_channels * sizeof(float) + sizeof(float);  // extra float for padding
 7134+
 7135+  // we make two passes through this loop, 1st to add everything up, 2nd to allocate and init
 7136+  for(;;)
 7137+  {
 7138+    int i;
 7139+    void * advance_mem = alloced;
 7140+    int copy_horizontal = 0;
 7141+    stbir__sampler * possibly_use_horizontal_for_pivot = 0;
 7142+
 7143+#ifdef STBIR__SEPARATE_ALLOCATIONS
 7144+    #define STBIR__NEXT_PTR( ptr, size, ntype ) if ( alloced ) { void * p = STBIR_MALLOC( size, user_data); if ( p == 0 ) { stbir__free_internal_mem( info ); return 0; } (ptr) = (ntype*)p; }
 7145+#else
 7146+    #define STBIR__NEXT_PTR( ptr, size, ntype ) advance_mem = (void*) ( ( ((size_t)advance_mem) + 15 ) & ~15 ); if ( alloced ) ptr = (ntype*)advance_mem; advance_mem = (char*)(((size_t)advance_mem) + (size));
 7147+#endif
 7148+
 7149+    STBIR__NEXT_PTR( info, sizeof( stbir__info ), stbir__info );
 7150+
 7151+    STBIR__NEXT_PTR( info->split_info, sizeof( stbir__per_split_info ) * splits, stbir__per_split_info );
 7152+
 7153+    if ( info )
 7154+    {
 7155+      static stbir__alpha_weight_func * fancy_alpha_weights[6]  =    { stbir__fancy_alpha_weight_4ch,   stbir__fancy_alpha_weight_4ch,   stbir__fancy_alpha_weight_4ch,   stbir__fancy_alpha_weight_4ch,   stbir__fancy_alpha_weight_2ch,   stbir__fancy_alpha_weight_2ch };
 7156+      static stbir__alpha_unweight_func * fancy_alpha_unweights[6] = { stbir__fancy_alpha_unweight_4ch, stbir__fancy_alpha_unweight_4ch, stbir__fancy_alpha_unweight_4ch, stbir__fancy_alpha_unweight_4ch, stbir__fancy_alpha_unweight_2ch, stbir__fancy_alpha_unweight_2ch };
 7157+      static stbir__alpha_weight_func * simple_alpha_weights[6] = { stbir__simple_alpha_weight_4ch, stbir__simple_alpha_weight_4ch, stbir__simple_alpha_weight_4ch, stbir__simple_alpha_weight_4ch, stbir__simple_alpha_weight_2ch, stbir__simple_alpha_weight_2ch };
 7158+      static stbir__alpha_unweight_func * simple_alpha_unweights[6] = { stbir__simple_alpha_unweight_4ch, stbir__simple_alpha_unweight_4ch, stbir__simple_alpha_unweight_4ch, stbir__simple_alpha_unweight_4ch, stbir__simple_alpha_unweight_2ch, stbir__simple_alpha_unweight_2ch };
 7159+
 7160+      // initialize info fields
 7161+      info->alloced_mem = alloced;
 7162+      info->alloced_total = alloced_total;
 7163+
 7164+      info->channels = channels;
 7165+      info->effective_channels = effective_channels;
 7166+
 7167+      info->offset_x = new_x;
 7168+      info->offset_y = new_y;
 7169+      info->alloc_ring_buffer_num_entries = (int)alloc_ring_buffer_num_entries;
 7170+      info->ring_buffer_num_entries = 0;
 7171+      info->ring_buffer_length_bytes = (int)ring_buffer_length_bytes;
 7172+      info->splits = splits;
 7173+      info->vertical_first = vertical_first;
 7174+
 7175+      info->input_pixel_layout_internal = input_pixel_layout;
 7176+      info->output_pixel_layout_internal = output_pixel_layout;
 7177+
 7178+      // setup alpha weight functions
 7179+      info->alpha_weight = 0;
 7180+      info->alpha_unweight = 0;
 7181+
 7182+      // handle alpha weighting functions and overrides
 7183+      if ( alpha_weighting_type == 2 )
 7184+      {
 7185+        // high quality alpha multiplying on the way in, dividing on the way out
 7186+        info->alpha_weight = fancy_alpha_weights[ input_pixel_layout - STBIRI_RGBA ];
 7187+        info->alpha_unweight = fancy_alpha_unweights[ output_pixel_layout - STBIRI_RGBA ];
 7188+      }
 7189+      else if ( alpha_weighting_type == 4 )
 7190+      {
 7191+        // fast alpha multiplying on the way in, dividing on the way out
 7192+        info->alpha_weight = simple_alpha_weights[ input_pixel_layout - STBIRI_RGBA ];
 7193+        info->alpha_unweight = simple_alpha_unweights[ output_pixel_layout - STBIRI_RGBA ];
 7194+      }
 7195+      else if ( alpha_weighting_type == 1 )
 7196+      {
 7197+        // fast alpha on the way in, leave in premultiplied form on way out
 7198+        info->alpha_weight = simple_alpha_weights[ input_pixel_layout - STBIRI_RGBA ];
 7199+      }
 7200+      else if ( alpha_weighting_type == 3 )
 7201+      {
 7202+        // incoming is premultiplied, fast alpha dividing on the way out - non-premultiplied output
 7203+        info->alpha_unweight = simple_alpha_unweights[ output_pixel_layout - STBIRI_RGBA ];
 7204+      }
 7205+
 7206+      // handle 3-chan color flipping, using the alpha weight path
 7207+      if ( ( ( input_pixel_layout == STBIRI_RGB ) && ( output_pixel_layout == STBIRI_BGR ) ) ||
 7208+           ( ( input_pixel_layout == STBIRI_BGR ) && ( output_pixel_layout == STBIRI_RGB ) ) )
 7209+      {
 7210+        // do the flipping on the smaller of the two ends
 7211+        if ( horizontal->scale_info.scale < 1.0f )
 7212+          info->alpha_unweight = stbir__simple_flip_3ch;
 7213+        else
 7214+          info->alpha_weight = stbir__simple_flip_3ch;
 7215+      }
 7216+
 7217+    }
 7218+
 7219+    // get all the per-split buffers
 7220+    for( i = 0 ; i < splits ; i++ )
 7221+    {
 7222+      STBIR__NEXT_PTR( info->split_info[i].decode_buffer, decode_buffer_size, float );
 7223+
 7224+#ifdef STBIR__SEPARATE_ALLOCATIONS
 7225+
 7226+      #ifdef STBIR_SIMD8
 7227+      if ( ( info ) && ( effective_channels == 3 ) )
 7228+        ++info->split_info[i].decode_buffer; // avx in 3 channel mode needs one float at the start of the buffer
 7229+      #endif
 7230+
 7231+      STBIR__NEXT_PTR( info->split_info[i].ring_buffers, alloc_ring_buffer_num_entries * sizeof(float*), float* );
 7232+      {
 7233+        int j;
 7234+        for( j = 0 ; j < alloc_ring_buffer_num_entries ; j++ )
 7235+        {
 7236+          STBIR__NEXT_PTR( info->split_info[i].ring_buffers[j], ring_buffer_length_bytes, float );
 7237+          #ifdef STBIR_SIMD8
 7238+          if ( ( info ) && ( effective_channels == 3 ) )
 7239+            ++info->split_info[i].ring_buffers[j]; // avx in 3 channel mode needs one float at the start of the buffer
 7240+          #endif
 7241+        }
 7242+      }
 7243+#else
 7244+      STBIR__NEXT_PTR( info->split_info[i].ring_buffer, ring_buffer_size, float );
 7245+#endif
 7246+      STBIR__NEXT_PTR( info->split_info[i].vertical_buffer, vertical_buffer_size, float );
 7247+    }
 7248+
 7249+    // alloc memory for to-be-pivoted coeffs (if necessary)
 7250+    if ( vertical->is_gather == 0 )
 7251+    {
 7252+      size_t both;
 7253+      size_t temp_mem_amt;
 7254+
 7255+      // when in vertical scatter mode, we first build the coefficients in gather mode, and then pivot after,
 7256+      //   that means we need two buffers, so we try to use the decode buffer and ring buffer for this. if that
 7257+      //   is too small, we just allocate extra memory to use as this temp.
 7258+
 7259+      both = (size_t)vertical->gather_prescatter_contributors_size + (size_t)vertical->gather_prescatter_coefficients_size;
 7260+
 7261+#ifdef STBIR__SEPARATE_ALLOCATIONS
 7262+      temp_mem_amt = decode_buffer_size;
 7263+
 7264+      #ifdef STBIR_SIMD8
 7265+      if ( effective_channels == 3 )
 7266+        --temp_mem_amt; // avx in 3 channel mode needs one float at the start of the buffer
 7267+      #endif
 7268+#else
 7269+      temp_mem_amt = (size_t)( decode_buffer_size + ring_buffer_size + vertical_buffer_size ) * (size_t)splits;
 7270+#endif
 7271+      if ( temp_mem_amt >= both )
 7272+      {
 7273+        if ( info )
 7274+        {
 7275+          vertical->gather_prescatter_contributors = (stbir__contributors*)info->split_info[0].decode_buffer;
 7276+          vertical->gather_prescatter_coefficients = (float*) ( ( (char*)info->split_info[0].decode_buffer ) + vertical->gather_prescatter_contributors_size );
 7277+        }
 7278+      }
 7279+      else
 7280+      {
 7281+        // ring+decode memory is too small, so allocate temp memory
 7282+        STBIR__NEXT_PTR( vertical->gather_prescatter_contributors, vertical->gather_prescatter_contributors_size, stbir__contributors );
 7283+        STBIR__NEXT_PTR( vertical->gather_prescatter_coefficients, vertical->gather_prescatter_coefficients_size, float );
 7284+      }
 7285+    }
 7286+
 7287+    STBIR__NEXT_PTR( horizontal->contributors, horizontal->contributors_size, stbir__contributors );
 7288+    STBIR__NEXT_PTR( horizontal->coefficients, horizontal->coefficients_size, float );
 7289+
 7290+    // are the two filters identical?? (happens a lot with mipmap generation)
 7291+    if ( ( horizontal->filter_kernel == vertical->filter_kernel ) && ( horizontal->filter_support == vertical->filter_support ) && ( horizontal->edge == vertical->edge ) && ( horizontal->scale_info.output_sub_size == vertical->scale_info.output_sub_size ) )
 7292+    {
 7293+      float diff_scale = horizontal->scale_info.scale - vertical->scale_info.scale;
 7294+      float diff_shift = horizontal->scale_info.pixel_shift - vertical->scale_info.pixel_shift;
 7295+      if ( diff_scale < 0.0f ) diff_scale = -diff_scale;
 7296+      if ( diff_shift < 0.0f ) diff_shift = -diff_shift;
 7297+      if ( ( diff_scale <= stbir__small_float ) && ( diff_shift <= stbir__small_float ) )
 7298+      {
 7299+        if ( horizontal->is_gather == vertical->is_gather )
 7300+        {
 7301+          copy_horizontal = 1;
 7302+          goto no_vert_alloc;
 7303+        }
 7304+        // everything matches, but vertical is scatter, horizontal is gather, use horizontal coeffs for vertical pivot coeffs
 7305+        possibly_use_horizontal_for_pivot = horizontal;
 7306+      }
 7307+    }
 7308+
 7309+    STBIR__NEXT_PTR( vertical->contributors, vertical->contributors_size, stbir__contributors );
 7310+    STBIR__NEXT_PTR( vertical->coefficients, vertical->coefficients_size, float );
 7311+
 7312+   no_vert_alloc:
 7313+
 7314+    if ( info )
 7315+    {
 7316+      STBIR_PROFILE_BUILD_START( horizontal );
 7317+
 7318+      stbir__calculate_filters( horizontal, 0, user_data STBIR_ONLY_PROFILE_BUILD_SET_INFO );
 7319+
 7320+      // setup the horizontal gather functions
 7321+      // start with defaulting to the n_coeffs functions (specialized on channels and remnant leftover)
 7322+      info->horizontal_gather_channels = stbir__horizontal_gather_n_coeffs_funcs[ effective_channels ][ horizontal->extent_info.widest & 3 ];
 7323+      // but if the number of coeffs <= 12, use another set of special cases. <=12 coeffs is any enlarging resize, or shrinking resize down to about 1/3 size
 7324+      if ( horizontal->extent_info.widest <= 12 )
 7325+        info->horizontal_gather_channels = stbir__horizontal_gather_channels_funcs[ effective_channels ][ horizontal->extent_info.widest - 1 ];
 7326+
 7327+      info->scanline_extents.conservative.n0 = conservative->n0;
 7328+      info->scanline_extents.conservative.n1 = conservative->n1;
 7329+
 7330+      // get exact extents
 7331+      stbir__get_extents( horizontal, &info->scanline_extents );
 7332+
 7333+      // pack the horizontal coeffs
 7334+      horizontal->coefficient_width = stbir__pack_coefficients(horizontal->num_contributors, horizontal->contributors, horizontal->coefficients, horizontal->coefficient_width, horizontal->extent_info.widest, info->scanline_extents.conservative.n0, info->scanline_extents.conservative.n1 );
 7335+
 7336+      STBIR_MEMCPY( &info->horizontal, horizontal, sizeof( stbir__sampler ) );
 7337+
 7338+      STBIR_PROFILE_BUILD_END( horizontal );
 7339+
 7340+      if ( copy_horizontal )
 7341+      {
 7342+        STBIR_MEMCPY( &info->vertical, horizontal, sizeof( stbir__sampler ) );
 7343+      }
 7344+      else
 7345+      {
 7346+        STBIR_PROFILE_BUILD_START( vertical );
 7347+
 7348+        stbir__calculate_filters( vertical, possibly_use_horizontal_for_pivot, user_data STBIR_ONLY_PROFILE_BUILD_SET_INFO );
 7349+        STBIR_MEMCPY( &info->vertical, vertical, sizeof( stbir__sampler ) );
 7350+
 7351+        STBIR_PROFILE_BUILD_END( vertical );
 7352+      }
 7353+
 7354+      // setup the vertical split ranges
 7355+      stbir__get_split_info( info->split_info, info->splits, info->vertical.scale_info.output_sub_size, info->vertical.filter_pixel_margin, info->vertical.scale_info.input_full_size, info->vertical.is_gather, info->vertical.contributors );
 7356+
 7357+      // now we know precisely how many entries we need
 7358+      info->ring_buffer_num_entries = info->vertical.extent_info.widest;
 7359+
 7360+      // we never need more ring buffer entries than the scanlines we're outputting
 7361+      if ( ( !info->vertical.is_gather ) && ( info->ring_buffer_num_entries > conservative_split_output_size ) )
 7362+        info->ring_buffer_num_entries = conservative_split_output_size;
 7363+      STBIR_ASSERT( info->ring_buffer_num_entries <= info->alloc_ring_buffer_num_entries );
 7364+    }
 7365+    #undef STBIR__NEXT_PTR
 7366+
 7367+
 7368+    // is this the first time through loop?
 7369+    if ( info == 0 )
 7370+    {
 7371+      alloced_total = ( 15 + (size_t)advance_mem );
 7372+      alloced = STBIR_MALLOC( alloced_total, user_data );
 7373+      if ( alloced == 0 )
 7374+        return 0;
 7375+    }
 7376+    else
 7377+      return info;  // success
 7378+  }
 7379+}
 7380+
 7381+static int stbir__perform_resize( stbir__info const * info, int split_start, int split_count )
 7382+{
 7383+  stbir__per_split_info * split_info = info->split_info + split_start;
 7384+
 7385+  STBIR_PROFILE_CLEAR_EXTRAS();
 7386+
 7387+  STBIR_PROFILE_FIRST_START( looping );
 7388+  if (info->vertical.is_gather)
 7389+    stbir__vertical_gather_loop( info, split_info, split_count );
 7390+  else
 7391+    stbir__vertical_scatter_loop( info, split_info, split_count );
 7392+  STBIR_PROFILE_END( looping );
 7393+
 7394+  return 1;
 7395+}
 7396+
 7397+static void stbir__update_info_from_resize( stbir__info * info, STBIR_RESIZE * resize )
 7398+{
 7399+  static stbir__decode_pixels_func * decode_simple[STBIR_TYPE_HALF_FLOAT-STBIR_TYPE_UINT8_SRGB+1]=
 7400+  {
 7401+    /* 1ch-4ch */ stbir__decode_uint8_srgb, stbir__decode_uint8_srgb, 0, stbir__decode_float_linear, stbir__decode_half_float_linear,
 7402+  };
 7403+
 7404+  static stbir__decode_pixels_func * decode_alphas[STBIRI_AR-STBIRI_RGBA+1][STBIR_TYPE_HALF_FLOAT-STBIR_TYPE_UINT8_SRGB+1]=
 7405+  {
 7406+    { /* RGBA */ stbir__decode_uint8_srgb4_linearalpha,      stbir__decode_uint8_srgb,      0, stbir__decode_float_linear,      stbir__decode_half_float_linear },
 7407+    { /* BGRA */ stbir__decode_uint8_srgb4_linearalpha_BGRA, stbir__decode_uint8_srgb_BGRA, 0, stbir__decode_float_linear_BGRA, stbir__decode_half_float_linear_BGRA },
 7408+    { /* ARGB */ stbir__decode_uint8_srgb4_linearalpha_ARGB, stbir__decode_uint8_srgb_ARGB, 0, stbir__decode_float_linear_ARGB, stbir__decode_half_float_linear_ARGB },
 7409+    { /* ABGR */ stbir__decode_uint8_srgb4_linearalpha_ABGR, stbir__decode_uint8_srgb_ABGR, 0, stbir__decode_float_linear_ABGR, stbir__decode_half_float_linear_ABGR },
 7410+    { /* RA   */ stbir__decode_uint8_srgb2_linearalpha,      stbir__decode_uint8_srgb,      0, stbir__decode_float_linear,      stbir__decode_half_float_linear },
 7411+    { /* AR   */ stbir__decode_uint8_srgb2_linearalpha_AR,   stbir__decode_uint8_srgb_AR,   0, stbir__decode_float_linear_AR,   stbir__decode_half_float_linear_AR },
 7412+  };
 7413+
 7414+  static stbir__decode_pixels_func * decode_simple_scaled_or_not[2][2]=
 7415+  {
 7416+    { stbir__decode_uint8_linear_scaled,  stbir__decode_uint8_linear }, { stbir__decode_uint16_linear_scaled, stbir__decode_uint16_linear },
 7417+  };
 7418+
 7419+  static stbir__decode_pixels_func * decode_alphas_scaled_or_not[STBIRI_AR-STBIRI_RGBA+1][2][2]=
 7420+  {
 7421+    { /* RGBA */ { stbir__decode_uint8_linear_scaled,       stbir__decode_uint8_linear },      { stbir__decode_uint16_linear_scaled,      stbir__decode_uint16_linear } },
 7422+    { /* BGRA */ { stbir__decode_uint8_linear_scaled_BGRA,  stbir__decode_uint8_linear_BGRA }, { stbir__decode_uint16_linear_scaled_BGRA, stbir__decode_uint16_linear_BGRA } },
 7423+    { /* ARGB */ { stbir__decode_uint8_linear_scaled_ARGB,  stbir__decode_uint8_linear_ARGB }, { stbir__decode_uint16_linear_scaled_ARGB, stbir__decode_uint16_linear_ARGB } },
 7424+    { /* ABGR */ { stbir__decode_uint8_linear_scaled_ABGR,  stbir__decode_uint8_linear_ABGR }, { stbir__decode_uint16_linear_scaled_ABGR, stbir__decode_uint16_linear_ABGR } },
 7425+    { /* RA   */ { stbir__decode_uint8_linear_scaled,       stbir__decode_uint8_linear },      { stbir__decode_uint16_linear_scaled,      stbir__decode_uint16_linear } },
 7426+    { /* AR   */ { stbir__decode_uint8_linear_scaled_AR,    stbir__decode_uint8_linear_AR },   { stbir__decode_uint16_linear_scaled_AR,   stbir__decode_uint16_linear_AR } }
 7427+  };
 7428+
 7429+  static stbir__encode_pixels_func * encode_simple[STBIR_TYPE_HALF_FLOAT-STBIR_TYPE_UINT8_SRGB+1]=
 7430+  {
 7431+    /* 1ch-4ch */ stbir__encode_uint8_srgb, stbir__encode_uint8_srgb, 0, stbir__encode_float_linear, stbir__encode_half_float_linear,
 7432+  };
 7433+
 7434+  static stbir__encode_pixels_func * encode_alphas[STBIRI_AR-STBIRI_RGBA+1][STBIR_TYPE_HALF_FLOAT-STBIR_TYPE_UINT8_SRGB+1]=
 7435+  {
 7436+    { /* RGBA */ stbir__encode_uint8_srgb4_linearalpha,      stbir__encode_uint8_srgb,      0, stbir__encode_float_linear,      stbir__encode_half_float_linear },
 7437+    { /* BGRA */ stbir__encode_uint8_srgb4_linearalpha_BGRA, stbir__encode_uint8_srgb_BGRA, 0, stbir__encode_float_linear_BGRA, stbir__encode_half_float_linear_BGRA },
 7438+    { /* ARGB */ stbir__encode_uint8_srgb4_linearalpha_ARGB, stbir__encode_uint8_srgb_ARGB, 0, stbir__encode_float_linear_ARGB, stbir__encode_half_float_linear_ARGB },
 7439+    { /* ABGR */ stbir__encode_uint8_srgb4_linearalpha_ABGR, stbir__encode_uint8_srgb_ABGR, 0, stbir__encode_float_linear_ABGR, stbir__encode_half_float_linear_ABGR },
 7440+    { /* RA   */ stbir__encode_uint8_srgb2_linearalpha,      stbir__encode_uint8_srgb,      0, stbir__encode_float_linear,      stbir__encode_half_float_linear },
 7441+    { /* AR   */ stbir__encode_uint8_srgb2_linearalpha_AR,   stbir__encode_uint8_srgb_AR,   0, stbir__encode_float_linear_AR,   stbir__encode_half_float_linear_AR }
 7442+  };
 7443+
 7444+  static stbir__encode_pixels_func * encode_simple_scaled_or_not[2][2]=
 7445+  {
 7446+    { stbir__encode_uint8_linear_scaled,  stbir__encode_uint8_linear }, { stbir__encode_uint16_linear_scaled, stbir__encode_uint16_linear },
 7447+  };
 7448+
 7449+  static stbir__encode_pixels_func * encode_alphas_scaled_or_not[STBIRI_AR-STBIRI_RGBA+1][2][2]=
 7450+  {
 7451+    { /* RGBA */ { stbir__encode_uint8_linear_scaled,       stbir__encode_uint8_linear },       { stbir__encode_uint16_linear_scaled,      stbir__encode_uint16_linear } },
 7452+    { /* BGRA */ { stbir__encode_uint8_linear_scaled_BGRA,  stbir__encode_uint8_linear_BGRA },  { stbir__encode_uint16_linear_scaled_BGRA, stbir__encode_uint16_linear_BGRA } },
 7453+    { /* ARGB */ { stbir__encode_uint8_linear_scaled_ARGB,  stbir__encode_uint8_linear_ARGB },  { stbir__encode_uint16_linear_scaled_ARGB, stbir__encode_uint16_linear_ARGB } },
 7454+    { /* ABGR */ { stbir__encode_uint8_linear_scaled_ABGR,  stbir__encode_uint8_linear_ABGR },  { stbir__encode_uint16_linear_scaled_ABGR, stbir__encode_uint16_linear_ABGR } },
 7455+    { /* RA   */ { stbir__encode_uint8_linear_scaled,       stbir__encode_uint8_linear },       { stbir__encode_uint16_linear_scaled,      stbir__encode_uint16_linear } },
 7456+    { /* AR   */ { stbir__encode_uint8_linear_scaled_AR,    stbir__encode_uint8_linear_AR },    { stbir__encode_uint16_linear_scaled_AR,   stbir__encode_uint16_linear_AR } }
 7457+  };
 7458+
 7459+  stbir__decode_pixels_func * decode_pixels = 0;
 7460+  stbir__encode_pixels_func * encode_pixels = 0;
 7461+  stbir_datatype input_type, output_type;
 7462+
 7463+  input_type = resize->input_data_type;
 7464+  output_type = resize->output_data_type;
 7465+  info->input_data = resize->input_pixels;
 7466+  info->input_stride_bytes = resize->input_stride_in_bytes;
 7467+  info->output_stride_bytes = resize->output_stride_in_bytes;
 7468+
 7469+  // if we're completely point sampling, then we can turn off SRGB
 7470+  if ( ( info->horizontal.filter_enum == STBIR_FILTER_POINT_SAMPLE ) && ( info->vertical.filter_enum == STBIR_FILTER_POINT_SAMPLE ) )
 7471+  {
 7472+    if ( ( ( input_type  == STBIR_TYPE_UINT8_SRGB ) || ( input_type  == STBIR_TYPE_UINT8_SRGB_ALPHA ) ) &&
 7473+         ( ( output_type == STBIR_TYPE_UINT8_SRGB ) || ( output_type == STBIR_TYPE_UINT8_SRGB_ALPHA ) ) )
 7474+    {
 7475+      input_type = STBIR_TYPE_UINT8;
 7476+      output_type = STBIR_TYPE_UINT8;
 7477+    }
 7478+  }
 7479+
 7480+  // recalc the output and input strides
 7481+  if ( info->input_stride_bytes == 0 )
 7482+    info->input_stride_bytes = info->channels * info->horizontal.scale_info.input_full_size * stbir__type_size[input_type];
 7483+
 7484+  if ( info->output_stride_bytes == 0 )
 7485+    info->output_stride_bytes = info->channels * info->horizontal.scale_info.output_sub_size * stbir__type_size[output_type];
 7486+
 7487+  // calc offset
 7488+  info->output_data = ( (char*) resize->output_pixels ) + ( (size_t) info->offset_y * (size_t) resize->output_stride_in_bytes ) + ( info->offset_x * info->channels * stbir__type_size[output_type] );
 7489+
 7490+  info->in_pixels_cb = resize->input_cb;
 7491+  info->user_data = resize->user_data;
 7492+  info->out_pixels_cb = resize->output_cb;
 7493+
 7494+  // setup the input format converters
 7495+  if ( ( input_type == STBIR_TYPE_UINT8 ) || ( input_type == STBIR_TYPE_UINT16 ) )
 7496+  {
 7497+    int non_scaled = 0;
 7498+
 7499+    // check if we can run unscaled - 0-255.0/0-65535.0 instead of 0-1.0 (which is a tiny bit faster when doing linear 8->8 or 16->16)
 7500+    if ( ( !info->alpha_weight ) && ( !info->alpha_unweight )  ) // don't short circuit when alpha weighting (get everything to 0-1.0 as usual)
 7501+      if ( ( ( input_type == STBIR_TYPE_UINT8 ) && ( output_type == STBIR_TYPE_UINT8 ) ) || ( ( input_type == STBIR_TYPE_UINT16 ) && ( output_type == STBIR_TYPE_UINT16 ) ) )
 7502+        non_scaled = 1;
 7503+
 7504+    if ( info->input_pixel_layout_internal <= STBIRI_4CHANNEL )
 7505+      decode_pixels = decode_simple_scaled_or_not[ input_type == STBIR_TYPE_UINT16 ][ non_scaled ];
 7506+    else
 7507+      decode_pixels = decode_alphas_scaled_or_not[ ( info->input_pixel_layout_internal - STBIRI_RGBA ) % ( STBIRI_AR-STBIRI_RGBA+1 ) ][ input_type == STBIR_TYPE_UINT16 ][ non_scaled ];
 7508+  }
 7509+  else
 7510+  {
 7511+    if ( info->input_pixel_layout_internal <= STBIRI_4CHANNEL )
 7512+      decode_pixels = decode_simple[ input_type - STBIR_TYPE_UINT8_SRGB ];
 7513+    else
 7514+      decode_pixels = decode_alphas[ ( info->input_pixel_layout_internal - STBIRI_RGBA ) % ( STBIRI_AR-STBIRI_RGBA+1 ) ][ input_type - STBIR_TYPE_UINT8_SRGB ];
 7515+  }
 7516+
 7517+  // setup the output format converters
 7518+  if ( ( output_type == STBIR_TYPE_UINT8 ) || ( output_type == STBIR_TYPE_UINT16 ) )
 7519+  {
 7520+    int non_scaled = 0;
 7521+
 7522+    // check if we can run unscaled - 0-255.0/0-65535.0 instead of 0-1.0 (which is a tiny bit faster when doing linear 8->8 or 16->16)
 7523+    if ( ( !info->alpha_weight ) && ( !info->alpha_unweight ) ) // don't short circuit when alpha weighting (get everything to 0-1.0 as usual)
 7524+      if ( ( ( input_type == STBIR_TYPE_UINT8 ) && ( output_type == STBIR_TYPE_UINT8 ) ) || ( ( input_type == STBIR_TYPE_UINT16 ) && ( output_type == STBIR_TYPE_UINT16 ) ) )
 7525+        non_scaled = 1;
 7526+
 7527+    if ( info->output_pixel_layout_internal <= STBIRI_4CHANNEL )
 7528+      encode_pixels = encode_simple_scaled_or_not[ output_type == STBIR_TYPE_UINT16 ][ non_scaled ];
 7529+    else
 7530+      encode_pixels = encode_alphas_scaled_or_not[ ( info->output_pixel_layout_internal - STBIRI_RGBA ) % ( STBIRI_AR-STBIRI_RGBA+1 ) ][ output_type == STBIR_TYPE_UINT16 ][ non_scaled ];
 7531+  }
 7532+  else
 7533+  {
 7534+    if ( info->output_pixel_layout_internal <= STBIRI_4CHANNEL )
 7535+      encode_pixels = encode_simple[ output_type - STBIR_TYPE_UINT8_SRGB ];
 7536+    else
 7537+      encode_pixels = encode_alphas[ ( info->output_pixel_layout_internal - STBIRI_RGBA ) % ( STBIRI_AR-STBIRI_RGBA+1 ) ][ output_type - STBIR_TYPE_UINT8_SRGB ];
 7538+  }
 7539+
 7540+  info->input_type = input_type;
 7541+  info->output_type = output_type;
 7542+  info->decode_pixels = decode_pixels;
 7543+  info->encode_pixels = encode_pixels;
 7544+}
 7545+
 7546+static void stbir__clip( int * outx, int * outsubw, int outw, double * u0, double * u1 )
 7547+{
 7548+  double per, adj;
 7549+  int over;
 7550+
 7551+  // do left/top edge
 7552+  if ( *outx < 0 )
 7553+  {
 7554+    per = ( (double)*outx ) / ( (double)*outsubw ); // is negative
 7555+    adj = per * ( *u1 - *u0 );
 7556+    *u0 -= adj; // increases u0
 7557+    *outx = 0;
 7558+  }
 7559+
 7560+  // do right/bot edge
 7561+  over = outw - ( *outx + *outsubw );
 7562+  if ( over < 0 )
 7563+  {
 7564+    per = ( (double)over ) / ( (double)*outsubw ); // is negative
 7565+    adj = per * ( *u1 - *u0 );
 7566+    *u1 += adj; // decrease u1
 7567+    *outsubw = outw - *outx;
 7568+  }
 7569+}
 7570+
 7571+// converts a double to a rational that has less than one float bit of error (returns 0 if unable to do so)
 7572+static int stbir__double_to_rational(double f, stbir_uint32 limit, stbir_uint32 *numer, stbir_uint32 *denom, int limit_denom ) // limit_denom (1) or limit numer (0)
 7573+{
 7574+  double err;
 7575+  stbir_uint64 top, bot;
 7576+  stbir_uint64 numer_last = 0;
 7577+  stbir_uint64 denom_last = 1;
 7578+  stbir_uint64 numer_estimate = 1;
 7579+  stbir_uint64 denom_estimate = 0;
 7580+
 7581+  // scale to past float error range
 7582+  top = (stbir_uint64)( f * (double)(1 << 25) );
 7583+  bot = 1 << 25;
 7584+
 7585+  // keep refining, but usually stops in a few loops - usually 5 for bad cases
 7586+  for(;;)
 7587+  {
 7588+    stbir_uint64 est, temp;
 7589+
 7590+    // hit limit, break out and do best full range estimate
 7591+    if ( ( ( limit_denom ) ? denom_estimate : numer_estimate ) >= limit )
 7592+      break;
 7593+
 7594+    // is the current error less than 1 bit of a float? if so, we're done
 7595+    if ( denom_estimate )
 7596+    {
 7597+      err = ( (double)numer_estimate / (double)denom_estimate ) - f;
 7598+      if ( err < 0.0 ) err = -err;
 7599+      if ( err < ( 1.0 / (double)(1<<24) ) )
 7600+      {
 7601+        // yup, found it
 7602+        *numer = (stbir_uint32) numer_estimate;
 7603+        *denom = (stbir_uint32) denom_estimate;
 7604+        return 1;
 7605+      }
 7606+    }
 7607+
 7608+    // no more refinement bits left? break out and do full range estimate
 7609+    if ( bot == 0 )
 7610+      break;
 7611+
 7612+    // gcd the estimate bits
 7613+    est = top / bot;
 7614+    temp = top % bot;
 7615+    top = bot;
 7616+    bot = temp;
 7617+
 7618+    // move remainders
 7619+    temp = est * denom_estimate + denom_last;
 7620+    denom_last = denom_estimate;
 7621+    denom_estimate = temp;
 7622+
 7623+    // move remainders
 7624+    temp = est * numer_estimate + numer_last;
 7625+    numer_last = numer_estimate;
 7626+    numer_estimate = temp;
 7627+  }
 7628+
 7629+  // we didn't find anything good enough for float, use a full range estimate
 7630+  if ( limit_denom )
 7631+  {
 7632+    numer_estimate= (stbir_uint64)( f * (double)limit + 0.5 );
 7633+    denom_estimate = limit;
 7634+  }
 7635+  else
 7636+  {
 7637+    numer_estimate = limit;
 7638+    denom_estimate = (stbir_uint64)( ( (double)limit / f ) + 0.5 );
 7639+  }
 7640+
 7641+  *numer = (stbir_uint32) numer_estimate;
 7642+  *denom = (stbir_uint32) denom_estimate;
 7643+
 7644+  err = ( denom_estimate ) ? ( ( (double)(stbir_uint32)numer_estimate / (double)(stbir_uint32)denom_estimate ) - f ) : 1.0;
 7645+  if ( err < 0.0 ) err = -err;
 7646+  return ( err < ( 1.0 / (double)(1<<24) ) ) ? 1 : 0;
 7647+}
 7648+
 7649+static int stbir__calculate_region_transform( stbir__scale_info * scale_info, int output_full_range, int * output_offset, int output_sub_range, int input_full_range, double input_s0, double input_s1 )
 7650+{
 7651+  double output_range, input_range, output_s, input_s, ratio, scale;
 7652+
 7653+  input_s = input_s1 - input_s0;
 7654+
 7655+  // null area
 7656+  if ( ( output_full_range == 0 ) || ( input_full_range == 0 ) ||
 7657+       ( output_sub_range == 0 ) || ( input_s <= stbir__small_float ) )
 7658+    return 0;
 7659+
 7660+  // are either of the ranges completely out of bounds?
 7661+  if ( ( *output_offset >= output_full_range ) || ( ( *output_offset + output_sub_range ) <= 0 ) || ( input_s0 >= (1.0f-stbir__small_float) ) || ( input_s1 <= stbir__small_float ) )
 7662+    return 0;
 7663+
 7664+  output_range = (double)output_full_range;
 7665+  input_range = (double)input_full_range;
 7666+
 7667+  output_s = ( (double)output_sub_range) / output_range;
 7668+
 7669+  // figure out the scaling to use
 7670+  ratio = output_s / input_s;
 7671+
 7672+  // save scale before clipping
 7673+  scale = ( output_range / input_range ) * ratio;
 7674+  scale_info->scale = (float)scale;
 7675+  scale_info->inv_scale = (float)( 1.0 / scale );
 7676+
 7677+  // clip output area to left/right output edges (and adjust input area)
 7678+  stbir__clip( output_offset, &output_sub_range, output_full_range, &input_s0, &input_s1 );
 7679+
 7680+  // recalc input area
 7681+  input_s = input_s1 - input_s0;
 7682+
 7683+  // after clipping do we have zero input area?
 7684+  if ( input_s <= stbir__small_float )
 7685+    return 0;
 7686+
 7687+  // calculate and store the starting source offsets in output pixel space
 7688+  scale_info->pixel_shift = (float) ( input_s0 * ratio * output_range );
 7689+
 7690+  scale_info->scale_is_rational = stbir__double_to_rational( scale, ( scale <= 1.0 ) ? output_full_range : input_full_range, &scale_info->scale_numerator, &scale_info->scale_denominator, ( scale >= 1.0 ) );
 7691+
 7692+  scale_info->input_full_size = input_full_range;
 7693+  scale_info->output_sub_size = output_sub_range;
 7694+
 7695+  return 1;
 7696+}
 7697+
 7698+
 7699+static void stbir__init_and_set_layout( STBIR_RESIZE * resize, stbir_pixel_layout pixel_layout, stbir_datatype data_type )
 7700+{
 7701+  resize->input_cb = 0;
 7702+  resize->output_cb = 0;
 7703+  resize->user_data = resize;
 7704+  resize->samplers = 0;
 7705+  resize->called_alloc = 0;
 7706+  resize->horizontal_filter = STBIR_FILTER_DEFAULT;
 7707+  resize->horizontal_filter_kernel = 0; resize->horizontal_filter_support = 0;
 7708+  resize->vertical_filter = STBIR_FILTER_DEFAULT;
 7709+  resize->vertical_filter_kernel = 0; resize->vertical_filter_support = 0;
 7710+  resize->horizontal_edge = STBIR_EDGE_CLAMP;
 7711+  resize->vertical_edge = STBIR_EDGE_CLAMP;
 7712+  resize->input_s0 = 0; resize->input_t0 = 0; resize->input_s1 = 1; resize->input_t1 = 1;
 7713+  resize->output_subx = 0; resize->output_suby = 0; resize->output_subw = resize->output_w; resize->output_subh = resize->output_h;
 7714+  resize->input_data_type = data_type;
 7715+  resize->output_data_type = data_type;
 7716+  resize->input_pixel_layout_public = pixel_layout;
 7717+  resize->output_pixel_layout_public = pixel_layout;
 7718+  resize->needs_rebuild = 1;
 7719+}
 7720+
 7721+STBIRDEF void stbir_resize_init( STBIR_RESIZE * resize,
 7722+                                 const void *input_pixels,  int input_w,  int input_h, int input_stride_in_bytes, // stride can be zero
 7723+                                       void *output_pixels, int output_w, int output_h, int output_stride_in_bytes, // stride can be zero
 7724+                                 stbir_pixel_layout pixel_layout, stbir_datatype data_type )
 7725+{
 7726+  resize->input_pixels = input_pixels;
 7727+  resize->input_w = input_w;
 7728+  resize->input_h = input_h;
 7729+  resize->input_stride_in_bytes = input_stride_in_bytes;
 7730+  resize->output_pixels = output_pixels;
 7731+  resize->output_w = output_w;
 7732+  resize->output_h = output_h;
 7733+  resize->output_stride_in_bytes = output_stride_in_bytes;
 7734+  resize->fast_alpha = 0;
 7735+
 7736+  stbir__init_and_set_layout( resize, pixel_layout, data_type );
 7737+}
 7738+
 7739+// You can update parameters any time after resize_init
 7740+STBIRDEF void stbir_set_datatypes( STBIR_RESIZE * resize, stbir_datatype input_type, stbir_datatype output_type )  // by default, datatype from resize_init
 7741+{
 7742+  resize->input_data_type = input_type;
 7743+  resize->output_data_type = output_type;
 7744+  if ( ( resize->samplers ) && ( !resize->needs_rebuild ) )
 7745+    stbir__update_info_from_resize( resize->samplers, resize );
 7746+}
 7747+
 7748+STBIRDEF void stbir_set_pixel_callbacks( STBIR_RESIZE * resize, stbir_input_callback * input_cb, stbir_output_callback * output_cb )   // no callbacks by default
 7749+{
 7750+  resize->input_cb = input_cb;
 7751+  resize->output_cb = output_cb;
 7752+
 7753+  if ( ( resize->samplers ) && ( !resize->needs_rebuild ) )
 7754+  {
 7755+    resize->samplers->in_pixels_cb = input_cb;
 7756+    resize->samplers->out_pixels_cb = output_cb;
 7757+  }
 7758+}
 7759+
 7760+STBIRDEF void stbir_set_user_data( STBIR_RESIZE * resize, void * user_data )                                     // pass back STBIR_RESIZE* by default
 7761+{
 7762+  resize->user_data = user_data;
 7763+  if ( ( resize->samplers ) && ( !resize->needs_rebuild ) )
 7764+    resize->samplers->user_data = user_data;
 7765+}
 7766+
 7767+STBIRDEF void stbir_set_buffer_ptrs( STBIR_RESIZE * resize, const void * input_pixels, int input_stride_in_bytes, void * output_pixels, int output_stride_in_bytes )
 7768+{
 7769+  resize->input_pixels = input_pixels;
 7770+  resize->input_stride_in_bytes = input_stride_in_bytes;
 7771+  resize->output_pixels = output_pixels;
 7772+  resize->output_stride_in_bytes = output_stride_in_bytes;
 7773+  if ( ( resize->samplers ) && ( !resize->needs_rebuild ) )
 7774+    stbir__update_info_from_resize( resize->samplers, resize );
 7775+}
 7776+
 7777+
 7778+STBIRDEF int stbir_set_edgemodes( STBIR_RESIZE * resize, stbir_edge horizontal_edge, stbir_edge vertical_edge )       // CLAMP by default
 7779+{
 7780+  resize->horizontal_edge = horizontal_edge;
 7781+  resize->vertical_edge = vertical_edge;
 7782+  resize->needs_rebuild = 1;
 7783+  return 1;
 7784+}
 7785+
 7786+STBIRDEF int stbir_set_filters( STBIR_RESIZE * resize, stbir_filter horizontal_filter, stbir_filter vertical_filter ) // STBIR_DEFAULT_FILTER_UPSAMPLE/DOWNSAMPLE by default
 7787+{
 7788+  resize->horizontal_filter = horizontal_filter;
 7789+  resize->vertical_filter = vertical_filter;
 7790+  resize->needs_rebuild = 1;
 7791+  return 1;
 7792+}
 7793+
 7794+STBIRDEF int stbir_set_filter_callbacks( STBIR_RESIZE * resize, stbir__kernel_callback * horizontal_filter, stbir__support_callback * horizontal_support, stbir__kernel_callback * vertical_filter, stbir__support_callback * vertical_support )
 7795+{
 7796+  resize->horizontal_filter_kernel = horizontal_filter; resize->horizontal_filter_support = horizontal_support;
 7797+  resize->vertical_filter_kernel = vertical_filter; resize->vertical_filter_support = vertical_support;
 7798+  resize->needs_rebuild = 1;
 7799+  return 1;
 7800+}
 7801+
 7802+STBIRDEF int stbir_set_pixel_layouts( STBIR_RESIZE * resize, stbir_pixel_layout input_pixel_layout, stbir_pixel_layout output_pixel_layout )   // sets new pixel layouts
 7803+{
 7804+  resize->input_pixel_layout_public = input_pixel_layout;
 7805+  resize->output_pixel_layout_public = output_pixel_layout;
 7806+  resize->needs_rebuild = 1;
 7807+  return 1;
 7808+}
 7809+
 7810+
 7811+STBIRDEF int stbir_set_non_pm_alpha_speed_over_quality( STBIR_RESIZE * resize, int non_pma_alpha_speed_over_quality )   // sets alpha speed
 7812+{
 7813+  resize->fast_alpha = non_pma_alpha_speed_over_quality;
 7814+  resize->needs_rebuild = 1;
 7815+  return 1;
 7816+}
 7817+
 7818+STBIRDEF int stbir_set_input_subrect( STBIR_RESIZE * resize, double s0, double t0, double s1, double t1 )                 // sets input region (full region by default)
 7819+{
 7820+  resize->input_s0 = s0;
 7821+  resize->input_t0 = t0;
 7822+  resize->input_s1 = s1;
 7823+  resize->input_t1 = t1;
 7824+  resize->needs_rebuild = 1;
 7825+
 7826+  // are we inbounds?
 7827+  if ( ( s1 < stbir__small_float ) || ( (s1-s0) < stbir__small_float ) ||
 7828+       ( t1 < stbir__small_float ) || ( (t1-t0) < stbir__small_float ) ||
 7829+       ( s0 > (1.0f-stbir__small_float) ) ||
 7830+       ( t0 > (1.0f-stbir__small_float) ) )
 7831+    return 0;
 7832+
 7833+  return 1;
 7834+}
 7835+
 7836+STBIRDEF int stbir_set_output_pixel_subrect( STBIR_RESIZE * resize, int subx, int suby, int subw, int subh )          // sets input region (full region by default)
 7837+{
 7838+  resize->output_subx = subx;
 7839+  resize->output_suby = suby;
 7840+  resize->output_subw = subw;
 7841+  resize->output_subh = subh;
 7842+  resize->needs_rebuild = 1;
 7843+
 7844+  // are we inbounds?
 7845+  if ( ( subx >= resize->output_w ) || ( ( subx + subw ) <= 0 ) || ( suby >= resize->output_h ) || ( ( suby + subh ) <= 0 ) || ( subw == 0 ) || ( subh == 0 ) )
 7846+    return 0;
 7847+
 7848+  return 1;
 7849+}
 7850+
 7851+STBIRDEF int stbir_set_pixel_subrect( STBIR_RESIZE * resize, int subx, int suby, int subw, int subh )                 // sets both regions (full regions by default)
 7852+{
 7853+  double s0, t0, s1, t1;
 7854+
 7855+  s0 = ( (double)subx ) / ( (double)resize->output_w );
 7856+  t0 = ( (double)suby ) / ( (double)resize->output_h );
 7857+  s1 = ( (double)(subx+subw) ) / ( (double)resize->output_w );
 7858+  t1 = ( (double)(suby+subh) ) / ( (double)resize->output_h );
 7859+
 7860+  resize->input_s0 = s0;
 7861+  resize->input_t0 = t0;
 7862+  resize->input_s1 = s1;
 7863+  resize->input_t1 = t1;
 7864+  resize->output_subx = subx;
 7865+  resize->output_suby = suby;
 7866+  resize->output_subw = subw;
 7867+  resize->output_subh = subh;
 7868+  resize->needs_rebuild = 1;
 7869+
 7870+  // are we inbounds?
 7871+  if ( ( subx >= resize->output_w ) || ( ( subx + subw ) <= 0 ) || ( suby >= resize->output_h ) || ( ( suby + subh ) <= 0 ) || ( subw == 0 ) || ( subh == 0 ) )
 7872+    return 0;
 7873+
 7874+  return 1;
 7875+}
 7876+
 7877+static int stbir__perform_build( STBIR_RESIZE * resize, int splits )
 7878+{
 7879+  stbir__contributors conservative = { 0, 0 };
 7880+  stbir__sampler horizontal, vertical;
 7881+  int new_output_subx, new_output_suby;
 7882+  stbir__info * out_info;
 7883+  #ifdef STBIR_PROFILE
 7884+  stbir__info profile_infod;  // used to contain building profile info before everything is allocated
 7885+  stbir__info * profile_info = &profile_infod;
 7886+  #endif
 7887+
 7888+  // have we already built the samplers?
 7889+  if ( resize->samplers )
 7890+    return 0;
 7891+
 7892+  #define STBIR_RETURN_ERROR_AND_ASSERT( exp )  STBIR_ASSERT( !(exp) ); if (exp) return 0;
 7893+  STBIR_RETURN_ERROR_AND_ASSERT( (unsigned)resize->horizontal_filter >= STBIR_FILTER_OTHER)
 7894+  STBIR_RETURN_ERROR_AND_ASSERT( (unsigned)resize->vertical_filter >= STBIR_FILTER_OTHER)
 7895+  #undef STBIR_RETURN_ERROR_AND_ASSERT
 7896+
 7897+  if ( splits <= 0 )
 7898+    return 0;
 7899+
 7900+  STBIR_PROFILE_BUILD_FIRST_START( build );
 7901+
 7902+  new_output_subx = resize->output_subx;
 7903+  new_output_suby = resize->output_suby;
 7904+
 7905+  // do horizontal clip and scale calcs
 7906+  if ( !stbir__calculate_region_transform( &horizontal.scale_info, resize->output_w, &new_output_subx, resize->output_subw, resize->input_w, resize->input_s0, resize->input_s1 ) )
 7907+    return 0;
 7908+
 7909+  // do vertical clip and scale calcs
 7910+  if ( !stbir__calculate_region_transform( &vertical.scale_info, resize->output_h, &new_output_suby, resize->output_subh, resize->input_h, resize->input_t0, resize->input_t1 ) )
 7911+    return 0;
 7912+
 7913+  // if nothing to do, just return
 7914+  if ( ( horizontal.scale_info.output_sub_size == 0 ) || ( vertical.scale_info.output_sub_size == 0 ) )
 7915+    return 0;
 7916+
 7917+  stbir__set_sampler(&horizontal, resize->horizontal_filter, resize->horizontal_filter_kernel, resize->horizontal_filter_support, resize->horizontal_edge, &horizontal.scale_info, 1, resize->user_data );
 7918+  stbir__get_conservative_extents( &horizontal, &conservative, resize->user_data );
 7919+  stbir__set_sampler(&vertical, resize->vertical_filter, resize->vertical_filter_kernel, resize->vertical_filter_support, resize->vertical_edge, &vertical.scale_info, 0, resize->user_data );
 7920+
 7921+  if ( ( vertical.scale_info.output_sub_size / splits ) < STBIR_FORCE_MINIMUM_SCANLINES_FOR_SPLITS ) // each split should be a minimum of 4 scanlines (handwavey choice)
 7922+  {
 7923+    splits = vertical.scale_info.output_sub_size / STBIR_FORCE_MINIMUM_SCANLINES_FOR_SPLITS;
 7924+    if ( splits == 0 ) splits = 1;
 7925+  }
 7926+
 7927+  STBIR_PROFILE_BUILD_START( alloc );
 7928+  out_info = stbir__alloc_internal_mem_and_build_samplers( &horizontal, &vertical, &conservative, resize->input_pixel_layout_public, resize->output_pixel_layout_public, splits, new_output_subx, new_output_suby, resize->fast_alpha, resize->user_data STBIR_ONLY_PROFILE_BUILD_SET_INFO );
 7929+  STBIR_PROFILE_BUILD_END( alloc );
 7930+  STBIR_PROFILE_BUILD_END( build );
 7931+
 7932+  if ( out_info )
 7933+  {
 7934+    resize->splits = splits;
 7935+    resize->samplers = out_info;
 7936+    resize->needs_rebuild = 0;
 7937+    #ifdef STBIR_PROFILE
 7938+      STBIR_MEMCPY( &out_info->profile, &profile_infod.profile, sizeof( out_info->profile ) );
 7939+    #endif
 7940+
 7941+    // update anything that can be changed without recalcing samplers
 7942+    stbir__update_info_from_resize( out_info, resize );
 7943+
 7944+    return splits;
 7945+  }
 7946+
 7947+  return 0;
 7948+}
 7949+
 7950+STBIRDEF void stbir_free_samplers( STBIR_RESIZE * resize )
 7951+{
 7952+  if ( resize->samplers )
 7953+  {
 7954+    stbir__free_internal_mem( resize->samplers );
 7955+    resize->samplers = 0;
 7956+    resize->called_alloc = 0;
 7957+  }
 7958+}
 7959+
 7960+STBIRDEF int stbir_build_samplers_with_splits( STBIR_RESIZE * resize, int splits )
 7961+{
 7962+  if ( ( resize->samplers == 0 ) || ( resize->needs_rebuild ) )
 7963+  {
 7964+    if ( resize->samplers )
 7965+      stbir_free_samplers( resize );
 7966+
 7967+    resize->called_alloc = 1;
 7968+    return stbir__perform_build( resize, splits );
 7969+  }
 7970+
 7971+  STBIR_PROFILE_BUILD_CLEAR( resize->samplers );
 7972+
 7973+  return 1;
 7974+}
 7975+
 7976+STBIRDEF int stbir_build_samplers( STBIR_RESIZE * resize )
 7977+{
 7978+  return stbir_build_samplers_with_splits( resize, 1 );
 7979+}
 7980+
 7981+STBIRDEF int stbir_resize_extended( STBIR_RESIZE * resize )
 7982+{
 7983+  int result;
 7984+
 7985+  if ( ( resize->samplers == 0 ) || ( resize->needs_rebuild ) )
 7986+  {
 7987+    int alloc_state = resize->called_alloc;  // remember allocated state
 7988+
 7989+    if ( resize->samplers )
 7990+    {
 7991+      stbir__free_internal_mem( resize->samplers );
 7992+      resize->samplers = 0;
 7993+    }
 7994+
 7995+    if ( !stbir_build_samplers( resize ) )
 7996+      return 0;
 7997+
 7998+    resize->called_alloc = alloc_state;
 7999+
 8000+    // if build_samplers succeeded (above), but there are no samplers set, then
 8001+    //   the area to stretch into was zero pixels, so don't do anything and return
 8002+    //   success
 8003+    if ( resize->samplers == 0 )
 8004+      return 1;
 8005+  }
 8006+  else
 8007+  {
 8008+    // didn't build anything - clear it
 8009+    STBIR_PROFILE_BUILD_CLEAR( resize->samplers );
 8010+  }
 8011+
 8012+  // do resize
 8013+  result = stbir__perform_resize( resize->samplers, 0, resize->splits );
 8014+
 8015+  // if we alloced, then free
 8016+  if ( !resize->called_alloc )
 8017+  {
 8018+    stbir_free_samplers( resize );
 8019+    resize->samplers = 0;
 8020+  }
 8021+
 8022+  return result;
 8023+}
 8024+
 8025+STBIRDEF int stbir_resize_extended_split( STBIR_RESIZE * resize, int split_start, int split_count )
 8026+{
 8027+  STBIR_ASSERT( resize->samplers );
 8028+
 8029+  // if we're just doing the whole thing, call full
 8030+  if ( ( split_start == -1 ) || ( ( split_start == 0 ) && ( split_count == resize->splits ) ) )
 8031+    return stbir_resize_extended( resize );
 8032+
 8033+  // you **must** build samplers first when using split resize
 8034+  if ( ( resize->samplers == 0 ) || ( resize->needs_rebuild ) )
 8035+    return 0;
 8036+
 8037+  if ( ( split_start >= resize->splits ) || ( split_start < 0 ) || ( ( split_start + split_count ) > resize->splits ) || ( split_count <= 0 ) )
 8038+    return 0;
 8039+
 8040+  // do resize
 8041+  return stbir__perform_resize( resize->samplers, split_start, split_count );
 8042+}
 8043+
 8044+
 8045+static void * stbir_quick_resize_helper( const void *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
 8046+                                               void *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
 8047+                                               stbir_pixel_layout pixel_layout, stbir_datatype data_type, stbir_edge edge, stbir_filter filter )
 8048+{
 8049+  STBIR_RESIZE resize;
 8050+  int scanline_output_in_bytes;
 8051+  int positive_output_stride_in_bytes;
 8052+  void * start_ptr;
 8053+  void * free_ptr;
 8054+
 8055+  scanline_output_in_bytes = output_w * stbir__type_size[ data_type ] * stbir__pixel_channels[ stbir__pixel_layout_convert_public_to_internal[ pixel_layout ] ];
 8056+  if ( scanline_output_in_bytes == 0 )
 8057+    return 0;
 8058+
 8059+  // if zero stride, use scanline output
 8060+  if ( output_stride_in_bytes == 0 )
 8061+    output_stride_in_bytes = scanline_output_in_bytes;
 8062+
 8063+  // abs value for inverted images (negative pitches)
 8064+  positive_output_stride_in_bytes = output_stride_in_bytes;
 8065+  if ( positive_output_stride_in_bytes < 0 )
 8066+    positive_output_stride_in_bytes = -positive_output_stride_in_bytes;
 8067+
 8068+  // is the requested stride smaller than the scanline output? if so, just fail
 8069+  if ( positive_output_stride_in_bytes < scanline_output_in_bytes )
 8070+    return 0;
 8071+
 8072+  start_ptr = output_pixels;
 8073+  free_ptr = 0;  // no free pointer, since they passed buffer to use
 8074+
 8075+  // did they pass a zero for the dest? if so, allocate the buffer
 8076+  if ( output_pixels == 0 )
 8077+  {
 8078+    size_t size;
 8079+    char * ptr;
 8080+  
 8081+    size = (size_t)positive_output_stride_in_bytes * (size_t)output_h;
 8082+    if ( size == 0 )
 8083+      return 0;
 8084+
 8085+    ptr = (char*) STBIR_MALLOC( size, 0 );
 8086+    if ( ptr == 0 )
 8087+      return 0;
 8088+
 8089+    free_ptr = ptr;
 8090+
 8091+    // point at the last scanline, if they requested a flipped image
 8092+    if ( output_stride_in_bytes < 0 )
 8093+      start_ptr = ptr + ( (size_t)positive_output_stride_in_bytes * (size_t)( output_h - 1 ) );
 8094+    else
 8095+      start_ptr = ptr;
 8096+  }
 8097+
 8098+  // ok, now do the resize
 8099+  stbir_resize_init( &resize,
 8100+                     input_pixels,  input_w,  input_h,  input_stride_in_bytes,
 8101+                     start_ptr, output_w, output_h, output_stride_in_bytes,
 8102+                     pixel_layout, data_type );
 8103+
 8104+  resize.horizontal_edge = edge;
 8105+  resize.vertical_edge = edge;
 8106+  resize.horizontal_filter = filter;
 8107+  resize.vertical_filter = filter;
 8108+
 8109+  if ( !stbir_resize_extended( &resize ) )
 8110+  {
 8111+    if ( free_ptr )
 8112+      STBIR_FREE( free_ptr, 0 );
 8113+    return 0;
 8114+  }
 8115+
 8116+  return (free_ptr) ? free_ptr : start_ptr;
 8117+}
 8118+
 8119+
 8120+
 8121+STBIRDEF unsigned char * stbir_resize_uint8_linear( const unsigned char *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
 8122+                                                          unsigned char *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
 8123+                                                          stbir_pixel_layout pixel_layout )
 8124+{
 8125+  return (unsigned char *) stbir_quick_resize_helper( input_pixels , input_w , input_h, input_stride_in_bytes, 
 8126+                                                      output_pixels, output_w, output_h, output_stride_in_bytes, 
 8127+                                                      pixel_layout, STBIR_TYPE_UINT8, STBIR_EDGE_CLAMP, STBIR_FILTER_DEFAULT );
 8128+}
 8129+
 8130+STBIRDEF unsigned char * stbir_resize_uint8_srgb( const unsigned char *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
 8131+                                                        unsigned char *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
 8132+                                                        stbir_pixel_layout pixel_layout )
 8133+{
 8134+  return (unsigned char *) stbir_quick_resize_helper( input_pixels , input_w , input_h, input_stride_in_bytes, 
 8135+                                                      output_pixels, output_w, output_h, output_stride_in_bytes, 
 8136+                                                      pixel_layout, STBIR_TYPE_UINT8_SRGB, STBIR_EDGE_CLAMP, STBIR_FILTER_DEFAULT );
 8137+}
 8138+
 8139+
 8140+STBIRDEF float * stbir_resize_float_linear( const float *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
 8141+                                                  float *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
 8142+                                                  stbir_pixel_layout pixel_layout )
 8143+{
 8144+  return (float *) stbir_quick_resize_helper( input_pixels , input_w , input_h, input_stride_in_bytes, 
 8145+                                              output_pixels, output_w, output_h, output_stride_in_bytes, 
 8146+                                              pixel_layout, STBIR_TYPE_FLOAT, STBIR_EDGE_CLAMP, STBIR_FILTER_DEFAULT  );
 8147+}
 8148+
 8149+
 8150+STBIRDEF void * stbir_resize( const void *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
 8151+                                    void *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
 8152+                                    stbir_pixel_layout pixel_layout, stbir_datatype data_type,
 8153+                                    stbir_edge edge, stbir_filter filter )
 8154+{
 8155+  return (void *) stbir_quick_resize_helper( input_pixels , input_w , input_h, input_stride_in_bytes, 
 8156+                                             output_pixels, output_w, output_h, output_stride_in_bytes, 
 8157+                                             pixel_layout, data_type, edge, filter  );
 8158+}
 8159+
 8160+#ifdef STBIR_PROFILE
 8161+
 8162+STBIRDEF void stbir_resize_build_profile_info( STBIR_PROFILE_INFO * info, STBIR_RESIZE const * resize )
 8163+{
 8164+  static char const * bdescriptions[6] = { "Building", "Allocating", "Horizontal sampler", "Vertical sampler", "Coefficient cleanup", "Coefficient pivot" } ;
 8165+  stbir__info* samp = resize->samplers;
 8166+  int i;
 8167+
 8168+  typedef int testa[ (STBIR__ARRAY_SIZE( bdescriptions ) == (STBIR__ARRAY_SIZE( samp->profile.array )-1) )?1:-1];
 8169+  typedef int testb[ (sizeof( samp->profile.array ) == (sizeof(samp->profile.named)) )?1:-1];
 8170+  typedef int testc[ (sizeof( info->clocks ) >= (sizeof(samp->profile.named)) )?1:-1];
 8171+
 8172+  for( i = 0 ; i < STBIR__ARRAY_SIZE( bdescriptions ) ; i++)
 8173+    info->clocks[i] = samp->profile.array[i+1];
 8174+
 8175+  info->total_clocks = samp->profile.named.total;
 8176+  info->descriptions = bdescriptions;
 8177+  info->count = STBIR__ARRAY_SIZE( bdescriptions );
 8178+}
 8179+
 8180+STBIRDEF void stbir_resize_split_profile_info( STBIR_PROFILE_INFO * info, STBIR_RESIZE const * resize, int split_start, int split_count )
 8181+{
 8182+  static char const * descriptions[7] = { "Looping", "Vertical sampling", "Horizontal sampling", "Scanline input", "Scanline output", "Alpha weighting", "Alpha unweighting" };
 8183+  stbir__per_split_info * split_info;
 8184+  int s, i;
 8185+
 8186+  typedef int testa[ (STBIR__ARRAY_SIZE( descriptions ) == (STBIR__ARRAY_SIZE( split_info->profile.array )-1) )?1:-1];
 8187+  typedef int testb[ (sizeof( split_info->profile.array ) == (sizeof(split_info->profile.named)) )?1:-1];
 8188+  typedef int testc[ (sizeof( info->clocks ) >= (sizeof(split_info->profile.named)) )?1:-1];
 8189+
 8190+  if ( split_start == -1 )
 8191+  {
 8192+    split_start = 0;
 8193+    split_count = resize->samplers->splits;
 8194+  }
 8195+
 8196+  if ( ( split_start >= resize->splits ) || ( split_start < 0 ) || ( ( split_start + split_count ) > resize->splits ) || ( split_count <= 0 ) )
 8197+  {
 8198+    info->total_clocks = 0;
 8199+    info->descriptions = 0;
 8200+    info->count = 0;
 8201+    return;
 8202+  }
 8203+
 8204+  split_info = resize->samplers->split_info + split_start;
 8205+
 8206+  // sum up the profile from all the splits
 8207+  for( i = 0 ; i < STBIR__ARRAY_SIZE( descriptions ) ; i++ )
 8208+  {
 8209+    stbir_uint64 sum = 0;
 8210+    for( s = 0 ; s < split_count ; s++ )
 8211+      sum += split_info[s].profile.array[i+1];
 8212+    info->clocks[i] = sum;
 8213+  }
 8214+
 8215+  info->total_clocks = split_info->profile.named.total; 
 8216+  info->descriptions = descriptions;
 8217+  info->count = STBIR__ARRAY_SIZE( descriptions );
 8218+}
 8219+
 8220+STBIRDEF void stbir_resize_extended_profile_info( STBIR_PROFILE_INFO * info, STBIR_RESIZE const * resize )
 8221+{
 8222+  stbir_resize_split_profile_info( info, resize, -1, 0 );
 8223+}
 8224+
 8225+#endif // STBIR_PROFILE
 8226+
 8227+#undef STBIR_BGR
 8228+#undef STBIR_1CHANNEL
 8229+#undef STBIR_2CHANNEL
 8230+#undef STBIR_RGB
 8231+#undef STBIR_RGBA
 8232+#undef STBIR_4CHANNEL
 8233+#undef STBIR_BGRA
 8234+#undef STBIR_ARGB
 8235+#undef STBIR_ABGR
 8236+#undef STBIR_RA
 8237+#undef STBIR_AR
 8238+#undef STBIR_RGBA_PM
 8239+#undef STBIR_BGRA_PM
 8240+#undef STBIR_ARGB_PM
 8241+#undef STBIR_ABGR_PM
 8242+#undef STBIR_RA_PM
 8243+#undef STBIR_AR_PM
 8244+
 8245+#endif // STB_IMAGE_RESIZE_IMPLEMENTATION
 8246+
 8247+#else  // STB_IMAGE_RESIZE_HORIZONTALS&STB_IMAGE_RESIZE_DO_VERTICALS
 8248+
 8249+// we reinclude the header file to define all the horizontal functions
 8250+//   specializing each function for the number of coeffs is 20-40% faster *OVERALL*
 8251+
 8252+// by including the header file again this way, we can still debug the functions
 8253+
 8254+#define STBIR_strs_join2( start, mid, end ) start##mid##end
 8255+#define STBIR_strs_join1( start, mid, end ) STBIR_strs_join2( start, mid, end )
 8256+
 8257+#define STBIR_strs_join24( start, mid1, mid2, end ) start##mid1##mid2##end
 8258+#define STBIR_strs_join14( start, mid1, mid2, end ) STBIR_strs_join24( start, mid1, mid2, end )
 8259+
 8260+#ifdef STB_IMAGE_RESIZE_DO_CODERS
 8261+
 8262+#ifdef stbir__decode_suffix
 8263+#define STBIR__CODER_NAME( name ) STBIR_strs_join1( name, _, stbir__decode_suffix )
 8264+#else
 8265+#define STBIR__CODER_NAME( name ) name
 8266+#endif
 8267+
 8268+#ifdef stbir__decode_swizzle
 8269+#define stbir__decode_simdf8_flip(reg) STBIR_strs_join1( STBIR_strs_join1( STBIR_strs_join1( STBIR_strs_join1( stbir__simdf8_0123to,stbir__decode_order0,stbir__decode_order1),stbir__decode_order2,stbir__decode_order3),stbir__decode_order0,stbir__decode_order1),stbir__decode_order2,stbir__decode_order3)(reg, reg)
 8270+#define stbir__decode_simdf4_flip(reg) STBIR_strs_join1( STBIR_strs_join1( stbir__simdf_0123to,stbir__decode_order0,stbir__decode_order1),stbir__decode_order2,stbir__decode_order3)(reg, reg)
 8271+#define stbir__encode_simdf8_unflip(reg) STBIR_strs_join1( STBIR_strs_join1( STBIR_strs_join1( STBIR_strs_join1( stbir__simdf8_0123to,stbir__encode_order0,stbir__encode_order1),stbir__encode_order2,stbir__encode_order3),stbir__encode_order0,stbir__encode_order1),stbir__encode_order2,stbir__encode_order3)(reg, reg)
 8272+#define stbir__encode_simdf4_unflip(reg) STBIR_strs_join1( STBIR_strs_join1( stbir__simdf_0123to,stbir__encode_order0,stbir__encode_order1),stbir__encode_order2,stbir__encode_order3)(reg, reg)
 8273+#else
 8274+#define stbir__decode_order0 0
 8275+#define stbir__decode_order1 1
 8276+#define stbir__decode_order2 2
 8277+#define stbir__decode_order3 3
 8278+#define stbir__encode_order0 0
 8279+#define stbir__encode_order1 1
 8280+#define stbir__encode_order2 2
 8281+#define stbir__encode_order3 3
 8282+#define stbir__decode_simdf8_flip(reg)
 8283+#define stbir__decode_simdf4_flip(reg)
 8284+#define stbir__encode_simdf8_unflip(reg)
 8285+#define stbir__encode_simdf4_unflip(reg)
 8286+#endif
 8287+
 8288+#ifdef STBIR_SIMD8
 8289+#define stbir__encode_simdfX_unflip  stbir__encode_simdf8_unflip
 8290+#else
 8291+#define stbir__encode_simdfX_unflip  stbir__encode_simdf4_unflip
 8292+#endif
 8293+
 8294+static float * STBIR__CODER_NAME( stbir__decode_uint8_linear_scaled )( float * decodep, int width_times_channels, void const * inputp )
 8295+{
 8296+  float STBIR_STREAMOUT_PTR( * ) decode = decodep;
 8297+  float * decode_end = (float*) decode + width_times_channels;
 8298+  unsigned char const * input = (unsigned char const*)inputp;
 8299+
 8300+  #ifdef STBIR_SIMD
 8301+  unsigned char const * end_input_m16 = input + width_times_channels - 16;
 8302+  if ( width_times_channels >= 16 )
 8303+  {
 8304+    decode_end -= 16;
 8305+    STBIR_NO_UNROLL_LOOP_START_INF_FOR
 8306+    for(;;)
 8307+    {
 8308+      #ifdef STBIR_SIMD8
 8309+      stbir__simdi i; stbir__simdi8 o0,o1;
 8310+      stbir__simdf8 of0, of1;
 8311+      STBIR_NO_UNROLL(decode);
 8312+      stbir__simdi_load( i, input );
 8313+      stbir__simdi8_expand_u8_to_u32( o0, o1, i );
 8314+      stbir__simdi8_convert_i32_to_float( of0, o0 );
 8315+      stbir__simdi8_convert_i32_to_float( of1, o1 );
 8316+      stbir__simdf8_mult( of0, of0, STBIR_max_uint8_as_float_inverted8);
 8317+      stbir__simdf8_mult( of1, of1, STBIR_max_uint8_as_float_inverted8);
 8318+      stbir__decode_simdf8_flip( of0 );
 8319+      stbir__decode_simdf8_flip( of1 );
 8320+      stbir__simdf8_store( decode + 0, of0 );
 8321+      stbir__simdf8_store( decode + 8, of1 );
 8322+      #else
 8323+      stbir__simdi i, o0, o1, o2, o3;
 8324+      stbir__simdf of0, of1, of2, of3;
 8325+      STBIR_NO_UNROLL(decode);
 8326+      stbir__simdi_load( i, input );
 8327+      stbir__simdi_expand_u8_to_u32( o0,o1,o2,o3,i);
 8328+      stbir__simdi_convert_i32_to_float( of0, o0 );
 8329+      stbir__simdi_convert_i32_to_float( of1, o1 );
 8330+      stbir__simdi_convert_i32_to_float( of2, o2 );
 8331+      stbir__simdi_convert_i32_to_float( of3, o3 );
 8332+      stbir__simdf_mult( of0, of0, STBIR__CONSTF(STBIR_max_uint8_as_float_inverted) );
 8333+      stbir__simdf_mult( of1, of1, STBIR__CONSTF(STBIR_max_uint8_as_float_inverted) );
 8334+      stbir__simdf_mult( of2, of2, STBIR__CONSTF(STBIR_max_uint8_as_float_inverted) );
 8335+      stbir__simdf_mult( of3, of3, STBIR__CONSTF(STBIR_max_uint8_as_float_inverted) );
 8336+      stbir__decode_simdf4_flip( of0 );
 8337+      stbir__decode_simdf4_flip( of1 );
 8338+      stbir__decode_simdf4_flip( of2 );
 8339+      stbir__decode_simdf4_flip( of3 );
 8340+      stbir__simdf_store( decode + 0,  of0 );
 8341+      stbir__simdf_store( decode + 4,  of1 );
 8342+      stbir__simdf_store( decode + 8,  of2 );
 8343+      stbir__simdf_store( decode + 12, of3 );
 8344+      #endif
 8345+      decode += 16;
 8346+      input += 16;
 8347+      if ( decode <= decode_end )
 8348+        continue;
 8349+      if ( decode == ( decode_end + 16 ) )
 8350+        break;
 8351+      decode = decode_end; // backup and do last couple
 8352+      input = end_input_m16;
 8353+    }
 8354+    return decode_end + 16;
 8355+  }
 8356+  #endif
 8357+
 8358+  // try to do blocks of 4 when you can
 8359+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 8360+  decode += 4;
 8361+  STBIR_SIMD_NO_UNROLL_LOOP_START
 8362+  while( decode <= decode_end )
 8363+  {
 8364+    STBIR_SIMD_NO_UNROLL(decode);
 8365+    decode[0-4] = ((float)(input[stbir__decode_order0])) * stbir__max_uint8_as_float_inverted;
 8366+    decode[1-4] = ((float)(input[stbir__decode_order1])) * stbir__max_uint8_as_float_inverted;
 8367+    decode[2-4] = ((float)(input[stbir__decode_order2])) * stbir__max_uint8_as_float_inverted;
 8368+    decode[3-4] = ((float)(input[stbir__decode_order3])) * stbir__max_uint8_as_float_inverted;
 8369+    decode += 4;
 8370+    input += 4;
 8371+  }
 8372+  decode -= 4;
 8373+  #endif
 8374+
 8375+  // do the remnants
 8376+  #if stbir__coder_min_num < 4
 8377+  STBIR_NO_UNROLL_LOOP_START
 8378+  while( decode < decode_end )
 8379+  {
 8380+    STBIR_NO_UNROLL(decode);
 8381+    decode[0] = ((float)(input[stbir__decode_order0])) * stbir__max_uint8_as_float_inverted;
 8382+    #if stbir__coder_min_num >= 2
 8383+    decode[1] = ((float)(input[stbir__decode_order1])) * stbir__max_uint8_as_float_inverted;
 8384+    #endif
 8385+    #if stbir__coder_min_num >= 3
 8386+    decode[2] = ((float)(input[stbir__decode_order2])) * stbir__max_uint8_as_float_inverted;
 8387+    #endif
 8388+    decode += stbir__coder_min_num;
 8389+    input += stbir__coder_min_num;
 8390+  }
 8391+  #endif
 8392+
 8393+  return decode_end;
 8394+}
 8395+
 8396+static void STBIR__CODER_NAME( stbir__encode_uint8_linear_scaled )( void * outputp, int width_times_channels, float const * encode )
 8397+{
 8398+  unsigned char STBIR_SIMD_STREAMOUT_PTR( * ) output = (unsigned char *) outputp;
 8399+  unsigned char * end_output = ( (unsigned char *) output ) + width_times_channels;
 8400+
 8401+  #ifdef STBIR_SIMD
 8402+  if ( width_times_channels >= stbir__simdfX_float_count*2 )
 8403+  {
 8404+    float const * end_encode_m8 = encode + width_times_channels - stbir__simdfX_float_count*2;
 8405+    end_output -= stbir__simdfX_float_count*2;
 8406+    STBIR_NO_UNROLL_LOOP_START_INF_FOR
 8407+    for(;;)
 8408+    {
 8409+      stbir__simdfX e0, e1;
 8410+      stbir__simdi i;
 8411+      STBIR_SIMD_NO_UNROLL(encode);
 8412+      stbir__simdfX_madd_mem( e0, STBIR_simd_point5X, STBIR_max_uint8_as_floatX, encode );
 8413+      stbir__simdfX_madd_mem( e1, STBIR_simd_point5X, STBIR_max_uint8_as_floatX, encode+stbir__simdfX_float_count );
 8414+      stbir__encode_simdfX_unflip( e0 );
 8415+      stbir__encode_simdfX_unflip( e1 );
 8416+      #ifdef STBIR_SIMD8
 8417+      stbir__simdf8_pack_to_16bytes( i, e0, e1 );
 8418+      stbir__simdi_store( output, i );
 8419+      #else
 8420+      stbir__simdf_pack_to_8bytes( i, e0, e1 );
 8421+      stbir__simdi_store2( output, i );
 8422+      #endif
 8423+      encode += stbir__simdfX_float_count*2;
 8424+      output += stbir__simdfX_float_count*2;
 8425+      if ( output <= end_output )
 8426+        continue;
 8427+      if ( output == ( end_output + stbir__simdfX_float_count*2 ) )
 8428+        break;
 8429+      output = end_output; // backup and do last couple
 8430+      encode = end_encode_m8;
 8431+    }
 8432+    return;
 8433+  }
 8434+
 8435+  // try to do blocks of 4 when you can
 8436+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 8437+  output += 4;
 8438+  STBIR_NO_UNROLL_LOOP_START
 8439+  while( output <= end_output )
 8440+  {
 8441+    stbir__simdf e0;
 8442+    stbir__simdi i0;
 8443+    STBIR_NO_UNROLL(encode);
 8444+    stbir__simdf_load( e0, encode );
 8445+    stbir__simdf_madd( e0, STBIR__CONSTF(STBIR_simd_point5), STBIR__CONSTF(STBIR_max_uint8_as_float), e0 );
 8446+    stbir__encode_simdf4_unflip( e0 );
 8447+    stbir__simdf_pack_to_8bytes( i0, e0, e0 );  // only use first 4
 8448+    *(int*)(output-4) = stbir__simdi_to_int( i0 );
 8449+    output += 4;
 8450+    encode += 4;
 8451+  }
 8452+  output -= 4;
 8453+  #endif
 8454+
 8455+  // do the remnants
 8456+  #if stbir__coder_min_num < 4
 8457+  STBIR_NO_UNROLL_LOOP_START
 8458+  while( output < end_output )
 8459+  {
 8460+    stbir__simdf e0;
 8461+    STBIR_NO_UNROLL(encode);
 8462+    stbir__simdf_madd1_mem( e0, STBIR__CONSTF(STBIR_simd_point5), STBIR__CONSTF(STBIR_max_uint8_as_float), encode+stbir__encode_order0 ); output[0] = stbir__simdf_convert_float_to_uint8( e0 );
 8463+    #if stbir__coder_min_num >= 2
 8464+    stbir__simdf_madd1_mem( e0, STBIR__CONSTF(STBIR_simd_point5), STBIR__CONSTF(STBIR_max_uint8_as_float), encode+stbir__encode_order1 ); output[1] = stbir__simdf_convert_float_to_uint8( e0 );
 8465+    #endif
 8466+    #if stbir__coder_min_num >= 3
 8467+    stbir__simdf_madd1_mem( e0, STBIR__CONSTF(STBIR_simd_point5), STBIR__CONSTF(STBIR_max_uint8_as_float), encode+stbir__encode_order2 ); output[2] = stbir__simdf_convert_float_to_uint8( e0 );
 8468+    #endif
 8469+    output += stbir__coder_min_num;
 8470+    encode += stbir__coder_min_num;
 8471+  }
 8472+  #endif
 8473+
 8474+  #else
 8475+
 8476+  // try to do blocks of 4 when you can
 8477+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 8478+  output += 4;
 8479+  while( output <= end_output )
 8480+  {
 8481+    float f;
 8482+    f = encode[stbir__encode_order0] * stbir__max_uint8_as_float + 0.5f; STBIR_CLAMP(f, 0, 255); output[0-4] = (unsigned char)f;
 8483+    f = encode[stbir__encode_order1] * stbir__max_uint8_as_float + 0.5f; STBIR_CLAMP(f, 0, 255); output[1-4] = (unsigned char)f;
 8484+    f = encode[stbir__encode_order2] * stbir__max_uint8_as_float + 0.5f; STBIR_CLAMP(f, 0, 255); output[2-4] = (unsigned char)f;
 8485+    f = encode[stbir__encode_order3] * stbir__max_uint8_as_float + 0.5f; STBIR_CLAMP(f, 0, 255); output[3-4] = (unsigned char)f;
 8486+    output += 4;
 8487+    encode += 4;
 8488+  }
 8489+  output -= 4;
 8490+  #endif
 8491+
 8492+  // do the remnants
 8493+  #if stbir__coder_min_num < 4
 8494+  STBIR_NO_UNROLL_LOOP_START
 8495+  while( output < end_output )
 8496+  {
 8497+    float f;
 8498+    STBIR_NO_UNROLL(encode);
 8499+    f = encode[stbir__encode_order0] * stbir__max_uint8_as_float + 0.5f; STBIR_CLAMP(f, 0, 255); output[0] = (unsigned char)f;
 8500+    #if stbir__coder_min_num >= 2
 8501+    f = encode[stbir__encode_order1] * stbir__max_uint8_as_float + 0.5f; STBIR_CLAMP(f, 0, 255); output[1] = (unsigned char)f;
 8502+    #endif
 8503+    #if stbir__coder_min_num >= 3
 8504+    f = encode[stbir__encode_order2] * stbir__max_uint8_as_float + 0.5f; STBIR_CLAMP(f, 0, 255); output[2] = (unsigned char)f;
 8505+    #endif
 8506+    output += stbir__coder_min_num;
 8507+    encode += stbir__coder_min_num;
 8508+  }
 8509+  #endif
 8510+  #endif
 8511+}
 8512+
 8513+static float * STBIR__CODER_NAME(stbir__decode_uint8_linear)( float * decodep, int width_times_channels, void const * inputp )
 8514+{
 8515+  float STBIR_STREAMOUT_PTR( * ) decode = decodep;
 8516+  float * decode_end = (float*) decode + width_times_channels;
 8517+  unsigned char const * input = (unsigned char const*)inputp;
 8518+
 8519+  #ifdef STBIR_SIMD
 8520+  unsigned char const * end_input_m16 = input + width_times_channels - 16;
 8521+  if ( width_times_channels >= 16 )
 8522+  {
 8523+    decode_end -= 16;
 8524+    STBIR_NO_UNROLL_LOOP_START_INF_FOR
 8525+    for(;;)
 8526+    {
 8527+      #ifdef STBIR_SIMD8
 8528+      stbir__simdi i; stbir__simdi8 o0,o1;
 8529+      stbir__simdf8 of0, of1;
 8530+      STBIR_NO_UNROLL(decode);
 8531+      stbir__simdi_load( i, input );
 8532+      stbir__simdi8_expand_u8_to_u32( o0, o1, i );
 8533+      stbir__simdi8_convert_i32_to_float( of0, o0 );
 8534+      stbir__simdi8_convert_i32_to_float( of1, o1 );
 8535+      stbir__decode_simdf8_flip( of0 );
 8536+      stbir__decode_simdf8_flip( of1 );
 8537+      stbir__simdf8_store( decode + 0, of0 );
 8538+      stbir__simdf8_store( decode + 8, of1 );
 8539+      #else
 8540+      stbir__simdi i, o0, o1, o2, o3;
 8541+      stbir__simdf of0, of1, of2, of3;
 8542+      STBIR_NO_UNROLL(decode);
 8543+      stbir__simdi_load( i, input );
 8544+      stbir__simdi_expand_u8_to_u32( o0,o1,o2,o3,i);
 8545+      stbir__simdi_convert_i32_to_float( of0, o0 );
 8546+      stbir__simdi_convert_i32_to_float( of1, o1 );
 8547+      stbir__simdi_convert_i32_to_float( of2, o2 );
 8548+      stbir__simdi_convert_i32_to_float( of3, o3 );
 8549+      stbir__decode_simdf4_flip( of0 );
 8550+      stbir__decode_simdf4_flip( of1 );
 8551+      stbir__decode_simdf4_flip( of2 );
 8552+      stbir__decode_simdf4_flip( of3 );
 8553+      stbir__simdf_store( decode + 0,  of0 );
 8554+      stbir__simdf_store( decode + 4,  of1 );
 8555+      stbir__simdf_store( decode + 8,  of2 );
 8556+      stbir__simdf_store( decode + 12, of3 );
 8557+#endif
 8558+      decode += 16;
 8559+      input += 16;
 8560+      if ( decode <= decode_end )
 8561+        continue;
 8562+      if ( decode == ( decode_end + 16 ) )
 8563+        break;
 8564+      decode = decode_end; // backup and do last couple
 8565+      input = end_input_m16;
 8566+    }
 8567+    return decode_end + 16;
 8568+  }
 8569+  #endif
 8570+
 8571+  // try to do blocks of 4 when you can
 8572+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 8573+  decode += 4;
 8574+  STBIR_SIMD_NO_UNROLL_LOOP_START
 8575+  while( decode <= decode_end )
 8576+  {
 8577+    STBIR_SIMD_NO_UNROLL(decode);
 8578+    decode[0-4] = ((float)(input[stbir__decode_order0]));
 8579+    decode[1-4] = ((float)(input[stbir__decode_order1]));
 8580+    decode[2-4] = ((float)(input[stbir__decode_order2]));
 8581+    decode[3-4] = ((float)(input[stbir__decode_order3]));
 8582+    decode += 4;
 8583+    input += 4;
 8584+  }
 8585+  decode -= 4;
 8586+  #endif
 8587+
 8588+  // do the remnants
 8589+  #if stbir__coder_min_num < 4
 8590+  STBIR_NO_UNROLL_LOOP_START
 8591+  while( decode < decode_end )
 8592+  {
 8593+    STBIR_NO_UNROLL(decode);
 8594+    decode[0] = ((float)(input[stbir__decode_order0]));
 8595+    #if stbir__coder_min_num >= 2
 8596+    decode[1] = ((float)(input[stbir__decode_order1]));
 8597+    #endif
 8598+    #if stbir__coder_min_num >= 3
 8599+    decode[2] = ((float)(input[stbir__decode_order2]));
 8600+    #endif
 8601+    decode += stbir__coder_min_num;
 8602+    input += stbir__coder_min_num;
 8603+  }
 8604+  #endif
 8605+  return decode_end;
 8606+}
 8607+
 8608+static void STBIR__CODER_NAME( stbir__encode_uint8_linear )( void * outputp, int width_times_channels, float const * encode )
 8609+{
 8610+  unsigned char STBIR_SIMD_STREAMOUT_PTR( * ) output = (unsigned char *) outputp;
 8611+  unsigned char * end_output = ( (unsigned char *) output ) + width_times_channels;
 8612+
 8613+  #ifdef STBIR_SIMD
 8614+  if ( width_times_channels >= stbir__simdfX_float_count*2 )
 8615+  {
 8616+    float const * end_encode_m8 = encode + width_times_channels - stbir__simdfX_float_count*2;
 8617+    end_output -= stbir__simdfX_float_count*2;
 8618+    STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 8619+    for(;;)
 8620+    {
 8621+      stbir__simdfX e0, e1;
 8622+      stbir__simdi i;
 8623+      STBIR_SIMD_NO_UNROLL(encode);
 8624+      stbir__simdfX_add_mem( e0, STBIR_simd_point5X, encode );
 8625+      stbir__simdfX_add_mem( e1, STBIR_simd_point5X, encode+stbir__simdfX_float_count );
 8626+      stbir__encode_simdfX_unflip( e0 );
 8627+      stbir__encode_simdfX_unflip( e1 );
 8628+      #ifdef STBIR_SIMD8
 8629+      stbir__simdf8_pack_to_16bytes( i, e0, e1 );
 8630+      stbir__simdi_store( output, i );
 8631+      #else
 8632+      stbir__simdf_pack_to_8bytes( i, e0, e1 );
 8633+      stbir__simdi_store2( output, i );
 8634+      #endif
 8635+      encode += stbir__simdfX_float_count*2;
 8636+      output += stbir__simdfX_float_count*2;
 8637+      if ( output <= end_output )
 8638+        continue;
 8639+      if ( output == ( end_output + stbir__simdfX_float_count*2 ) )
 8640+        break;
 8641+      output = end_output; // backup and do last couple
 8642+      encode = end_encode_m8;
 8643+    }
 8644+    return;
 8645+  }
 8646+
 8647+  // try to do blocks of 4 when you can
 8648+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 8649+  output += 4;
 8650+  STBIR_NO_UNROLL_LOOP_START
 8651+  while( output <= end_output )
 8652+  {
 8653+    stbir__simdf e0;
 8654+    stbir__simdi i0;
 8655+    STBIR_NO_UNROLL(encode);
 8656+    stbir__simdf_load( e0, encode );
 8657+    stbir__simdf_add( e0, STBIR__CONSTF(STBIR_simd_point5), e0 );
 8658+    stbir__encode_simdf4_unflip( e0 );
 8659+    stbir__simdf_pack_to_8bytes( i0, e0, e0 );  // only use first 4
 8660+    *(int*)(output-4) = stbir__simdi_to_int( i0 );
 8661+    output += 4;
 8662+    encode += 4;
 8663+  }
 8664+  output -= 4;
 8665+  #endif
 8666+
 8667+  #else
 8668+
 8669+  // try to do blocks of 4 when you can
 8670+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 8671+  output += 4;
 8672+  while( output <= end_output )
 8673+  {
 8674+    float f;
 8675+    f = encode[stbir__encode_order0] + 0.5f; STBIR_CLAMP(f, 0, 255); output[0-4] = (unsigned char)f;
 8676+    f = encode[stbir__encode_order1] + 0.5f; STBIR_CLAMP(f, 0, 255); output[1-4] = (unsigned char)f;
 8677+    f = encode[stbir__encode_order2] + 0.5f; STBIR_CLAMP(f, 0, 255); output[2-4] = (unsigned char)f;
 8678+    f = encode[stbir__encode_order3] + 0.5f; STBIR_CLAMP(f, 0, 255); output[3-4] = (unsigned char)f;
 8679+    output += 4;
 8680+    encode += 4;
 8681+  }
 8682+  output -= 4;
 8683+  #endif
 8684+
 8685+  #endif
 8686+
 8687+  // do the remnants
 8688+  #if stbir__coder_min_num < 4
 8689+  STBIR_NO_UNROLL_LOOP_START
 8690+  while( output < end_output )
 8691+  {
 8692+    float f;
 8693+    STBIR_NO_UNROLL(encode);
 8694+    f = encode[stbir__encode_order0] + 0.5f; STBIR_CLAMP(f, 0, 255); output[0] = (unsigned char)f;
 8695+    #if stbir__coder_min_num >= 2
 8696+    f = encode[stbir__encode_order1] + 0.5f; STBIR_CLAMP(f, 0, 255); output[1] = (unsigned char)f;
 8697+    #endif
 8698+    #if stbir__coder_min_num >= 3
 8699+    f = encode[stbir__encode_order2] + 0.5f; STBIR_CLAMP(f, 0, 255); output[2] = (unsigned char)f;
 8700+    #endif
 8701+    output += stbir__coder_min_num;
 8702+    encode += stbir__coder_min_num;
 8703+  }
 8704+  #endif
 8705+}
 8706+
 8707+static float * STBIR__CODER_NAME(stbir__decode_uint8_srgb)( float * decodep, int width_times_channels, void const * inputp )
 8708+{
 8709+  float STBIR_STREAMOUT_PTR( * ) decode = decodep;
 8710+  float * decode_end = (float*) decode + width_times_channels;
 8711+  unsigned char const * input = (unsigned char const *)inputp;
 8712+
 8713+  // try to do blocks of 4 when you can
 8714+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 8715+  decode += 4;
 8716+  while( decode <= decode_end )
 8717+  {
 8718+    decode[0-4] = stbir__srgb_uchar_to_linear_float[ input[ stbir__decode_order0 ] ];
 8719+    decode[1-4] = stbir__srgb_uchar_to_linear_float[ input[ stbir__decode_order1 ] ];
 8720+    decode[2-4] = stbir__srgb_uchar_to_linear_float[ input[ stbir__decode_order2 ] ];
 8721+    decode[3-4] = stbir__srgb_uchar_to_linear_float[ input[ stbir__decode_order3 ] ];
 8722+    decode += 4;
 8723+    input += 4;
 8724+  }
 8725+  decode -= 4;
 8726+  #endif
 8727+
 8728+  // do the remnants
 8729+  #if stbir__coder_min_num < 4
 8730+  STBIR_NO_UNROLL_LOOP_START
 8731+  while( decode < decode_end )
 8732+  {
 8733+    STBIR_NO_UNROLL(decode);
 8734+    decode[0] = stbir__srgb_uchar_to_linear_float[ input[ stbir__decode_order0 ] ];
 8735+    #if stbir__coder_min_num >= 2
 8736+    decode[1] = stbir__srgb_uchar_to_linear_float[ input[ stbir__decode_order1 ] ];
 8737+    #endif
 8738+    #if stbir__coder_min_num >= 3
 8739+    decode[2] = stbir__srgb_uchar_to_linear_float[ input[ stbir__decode_order2 ] ];
 8740+    #endif
 8741+    decode += stbir__coder_min_num;
 8742+    input += stbir__coder_min_num;
 8743+  }
 8744+  #endif
 8745+  return decode_end;
 8746+}
 8747+
 8748+#define stbir__min_max_shift20( i, f ) \
 8749+    stbir__simdf_max( f, f, stbir_simdf_casti(STBIR__CONSTI( STBIR_almost_zero )) ); \
 8750+    stbir__simdf_min( f, f, stbir_simdf_casti(STBIR__CONSTI( STBIR_almost_one  )) ); \
 8751+    stbir__simdi_32shr( i, stbir_simdi_castf( f ), 20 );
 8752+
 8753+#define stbir__scale_and_convert( i, f ) \
 8754+    stbir__simdf_madd( f, STBIR__CONSTF( STBIR_simd_point5 ), STBIR__CONSTF( STBIR_max_uint8_as_float ), f ); \
 8755+    stbir__simdf_max( f, f, stbir__simdf_zeroP() ); \
 8756+    stbir__simdf_min( f, f, STBIR__CONSTF( STBIR_max_uint8_as_float ) ); \
 8757+    stbir__simdf_convert_float_to_i32( i, f );
 8758+
 8759+#define stbir__linear_to_srgb_finish( i, f ) \
 8760+{ \
 8761+    stbir__simdi temp;  \
 8762+    stbir__simdi_32shr( temp, stbir_simdi_castf( f ), 12 ) ; \
 8763+    stbir__simdi_and( temp, temp, STBIR__CONSTI(STBIR_mantissa_mask) ); \
 8764+    stbir__simdi_or( temp, temp, STBIR__CONSTI(STBIR_topscale) ); \
 8765+    stbir__simdi_16madd( i, i, temp ); \
 8766+    stbir__simdi_32shr( i, i, 16 ); \
 8767+}
 8768+
 8769+#define stbir__simdi_table_lookup2( v0,v1, table ) \
 8770+{ \
 8771+  stbir__simdi_u32 temp0,temp1; \
 8772+  temp0.m128i_i128 = v0; \
 8773+  temp1.m128i_i128 = v1; \
 8774+  temp0.m128i_u32[0] = table[temp0.m128i_i32[0]]; temp0.m128i_u32[1] = table[temp0.m128i_i32[1]]; temp0.m128i_u32[2] = table[temp0.m128i_i32[2]]; temp0.m128i_u32[3] = table[temp0.m128i_i32[3]]; \
 8775+  temp1.m128i_u32[0] = table[temp1.m128i_i32[0]]; temp1.m128i_u32[1] = table[temp1.m128i_i32[1]]; temp1.m128i_u32[2] = table[temp1.m128i_i32[2]]; temp1.m128i_u32[3] = table[temp1.m128i_i32[3]]; \
 8776+  v0 = temp0.m128i_i128; \
 8777+  v1 = temp1.m128i_i128; \
 8778+}
 8779+
 8780+#define stbir__simdi_table_lookup3( v0,v1,v2, table ) \
 8781+{ \
 8782+  stbir__simdi_u32 temp0,temp1,temp2; \
 8783+  temp0.m128i_i128 = v0; \
 8784+  temp1.m128i_i128 = v1; \
 8785+  temp2.m128i_i128 = v2; \
 8786+  temp0.m128i_u32[0] = table[temp0.m128i_i32[0]]; temp0.m128i_u32[1] = table[temp0.m128i_i32[1]]; temp0.m128i_u32[2] = table[temp0.m128i_i32[2]]; temp0.m128i_u32[3] = table[temp0.m128i_i32[3]]; \
 8787+  temp1.m128i_u32[0] = table[temp1.m128i_i32[0]]; temp1.m128i_u32[1] = table[temp1.m128i_i32[1]]; temp1.m128i_u32[2] = table[temp1.m128i_i32[2]]; temp1.m128i_u32[3] = table[temp1.m128i_i32[3]]; \
 8788+  temp2.m128i_u32[0] = table[temp2.m128i_i32[0]]; temp2.m128i_u32[1] = table[temp2.m128i_i32[1]]; temp2.m128i_u32[2] = table[temp2.m128i_i32[2]]; temp2.m128i_u32[3] = table[temp2.m128i_i32[3]]; \
 8789+  v0 = temp0.m128i_i128; \
 8790+  v1 = temp1.m128i_i128; \
 8791+  v2 = temp2.m128i_i128; \
 8792+}
 8793+
 8794+#define stbir__simdi_table_lookup4( v0,v1,v2,v3, table ) \
 8795+{ \
 8796+  stbir__simdi_u32 temp0,temp1,temp2,temp3; \
 8797+  temp0.m128i_i128 = v0; \
 8798+  temp1.m128i_i128 = v1; \
 8799+  temp2.m128i_i128 = v2; \
 8800+  temp3.m128i_i128 = v3; \
 8801+  temp0.m128i_u32[0] = table[temp0.m128i_i32[0]]; temp0.m128i_u32[1] = table[temp0.m128i_i32[1]]; temp0.m128i_u32[2] = table[temp0.m128i_i32[2]]; temp0.m128i_u32[3] = table[temp0.m128i_i32[3]]; \
 8802+  temp1.m128i_u32[0] = table[temp1.m128i_i32[0]]; temp1.m128i_u32[1] = table[temp1.m128i_i32[1]]; temp1.m128i_u32[2] = table[temp1.m128i_i32[2]]; temp1.m128i_u32[3] = table[temp1.m128i_i32[3]]; \
 8803+  temp2.m128i_u32[0] = table[temp2.m128i_i32[0]]; temp2.m128i_u32[1] = table[temp2.m128i_i32[1]]; temp2.m128i_u32[2] = table[temp2.m128i_i32[2]]; temp2.m128i_u32[3] = table[temp2.m128i_i32[3]]; \
 8804+  temp3.m128i_u32[0] = table[temp3.m128i_i32[0]]; temp3.m128i_u32[1] = table[temp3.m128i_i32[1]]; temp3.m128i_u32[2] = table[temp3.m128i_i32[2]]; temp3.m128i_u32[3] = table[temp3.m128i_i32[3]]; \
 8805+  v0 = temp0.m128i_i128; \
 8806+  v1 = temp1.m128i_i128; \
 8807+  v2 = temp2.m128i_i128; \
 8808+  v3 = temp3.m128i_i128; \
 8809+}
 8810+
 8811+static void STBIR__CODER_NAME( stbir__encode_uint8_srgb )( void * outputp, int width_times_channels, float const * encode )
 8812+{
 8813+  unsigned char STBIR_SIMD_STREAMOUT_PTR( * ) output = (unsigned char*) outputp;
 8814+  unsigned char * end_output = ( (unsigned char*) output ) + width_times_channels;
 8815+
 8816+  #ifdef STBIR_SIMD
 8817+
 8818+  if ( width_times_channels >= 16 )
 8819+  {
 8820+    float const * end_encode_m16 = encode + width_times_channels - 16;
 8821+    end_output -= 16;
 8822+    STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 8823+    for(;;)
 8824+    {
 8825+      stbir__simdf f0, f1, f2, f3;
 8826+      stbir__simdi i0, i1, i2, i3;
 8827+      STBIR_SIMD_NO_UNROLL(encode);
 8828+
 8829+      stbir__simdf_load4_transposed( f0, f1, f2, f3, encode );
 8830+
 8831+      stbir__min_max_shift20( i0, f0 );
 8832+      stbir__min_max_shift20( i1, f1 );
 8833+      stbir__min_max_shift20( i2, f2 );
 8834+      stbir__min_max_shift20( i3, f3 );
 8835+
 8836+      stbir__simdi_table_lookup4( i0, i1, i2, i3, ( fp32_to_srgb8_tab4 - (127-13)*8 ) );
 8837+
 8838+      stbir__linear_to_srgb_finish( i0, f0 );
 8839+      stbir__linear_to_srgb_finish( i1, f1 );
 8840+      stbir__linear_to_srgb_finish( i2, f2 );
 8841+      stbir__linear_to_srgb_finish( i3, f3 );
 8842+
 8843+      stbir__interleave_pack_and_store_16_u8( output,  STBIR_strs_join1(i, ,stbir__encode_order0), STBIR_strs_join1(i, ,stbir__encode_order1), STBIR_strs_join1(i, ,stbir__encode_order2), STBIR_strs_join1(i, ,stbir__encode_order3) );
 8844+
 8845+      encode += 16;
 8846+      output += 16;
 8847+      if ( output <= end_output )
 8848+        continue;
 8849+      if ( output == ( end_output + 16 ) )
 8850+        break;
 8851+      output = end_output; // backup and do last couple
 8852+      encode = end_encode_m16;
 8853+    }
 8854+    return;
 8855+  }
 8856+  #endif
 8857+
 8858+  // try to do blocks of 4 when you can
 8859+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 8860+  output += 4;
 8861+  STBIR_SIMD_NO_UNROLL_LOOP_START
 8862+  while ( output <= end_output )
 8863+  {
 8864+    STBIR_SIMD_NO_UNROLL(encode);
 8865+
 8866+    output[0-4] = stbir__linear_to_srgb_uchar( encode[stbir__encode_order0] );
 8867+    output[1-4] = stbir__linear_to_srgb_uchar( encode[stbir__encode_order1] );
 8868+    output[2-4] = stbir__linear_to_srgb_uchar( encode[stbir__encode_order2] );
 8869+    output[3-4] = stbir__linear_to_srgb_uchar( encode[stbir__encode_order3] );
 8870+
 8871+    output += 4;
 8872+    encode += 4;
 8873+  }
 8874+  output -= 4;
 8875+  #endif
 8876+
 8877+  // do the remnants
 8878+  #if stbir__coder_min_num < 4
 8879+  STBIR_NO_UNROLL_LOOP_START
 8880+  while( output < end_output )
 8881+  {
 8882+    STBIR_NO_UNROLL(encode);
 8883+    output[0] = stbir__linear_to_srgb_uchar( encode[stbir__encode_order0] );
 8884+    #if stbir__coder_min_num >= 2
 8885+    output[1] = stbir__linear_to_srgb_uchar( encode[stbir__encode_order1] );
 8886+    #endif
 8887+    #if stbir__coder_min_num >= 3
 8888+    output[2] = stbir__linear_to_srgb_uchar( encode[stbir__encode_order2] );
 8889+    #endif
 8890+    output += stbir__coder_min_num;
 8891+    encode += stbir__coder_min_num;
 8892+  }
 8893+  #endif
 8894+}
 8895+
 8896+#if ( stbir__coder_min_num == 4 ) || ( ( stbir__coder_min_num == 1 ) && ( !defined(stbir__decode_swizzle) ) )
 8897+
 8898+static float * STBIR__CODER_NAME(stbir__decode_uint8_srgb4_linearalpha)( float * decodep, int width_times_channels, void const * inputp )
 8899+{
 8900+  float STBIR_STREAMOUT_PTR( * ) decode = decodep;
 8901+  float * decode_end = (float*) decode + width_times_channels;
 8902+  unsigned char const * input = (unsigned char const *)inputp;
 8903+
 8904+  do {
 8905+    decode[0] = stbir__srgb_uchar_to_linear_float[ input[stbir__decode_order0] ];
 8906+    decode[1] = stbir__srgb_uchar_to_linear_float[ input[stbir__decode_order1] ];
 8907+    decode[2] = stbir__srgb_uchar_to_linear_float[ input[stbir__decode_order2] ];
 8908+    decode[3] = ( (float) input[stbir__decode_order3] ) * stbir__max_uint8_as_float_inverted;
 8909+    input += 4;
 8910+    decode += 4;
 8911+  } while( decode < decode_end );
 8912+  return decode_end;
 8913+}
 8914+
 8915+
 8916+static void STBIR__CODER_NAME( stbir__encode_uint8_srgb4_linearalpha )( void * outputp, int width_times_channels, float const * encode )
 8917+{
 8918+  unsigned char STBIR_SIMD_STREAMOUT_PTR( * ) output = (unsigned char*) outputp;
 8919+  unsigned char * end_output = ( (unsigned char*) output ) + width_times_channels;
 8920+
 8921+  #ifdef STBIR_SIMD
 8922+
 8923+  if ( width_times_channels >= 16 )
 8924+  {
 8925+    float const * end_encode_m16 = encode + width_times_channels - 16;
 8926+    end_output -= 16;
 8927+    STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 8928+    for(;;)
 8929+    {
 8930+      stbir__simdf f0, f1, f2, f3;
 8931+      stbir__simdi i0, i1, i2, i3;
 8932+
 8933+      STBIR_SIMD_NO_UNROLL(encode);
 8934+      stbir__simdf_load4_transposed( f0, f1, f2, f3, encode );
 8935+
 8936+      stbir__min_max_shift20( i0, f0 );
 8937+      stbir__min_max_shift20( i1, f1 );
 8938+      stbir__min_max_shift20( i2, f2 );
 8939+      stbir__scale_and_convert( i3, f3 );
 8940+
 8941+      stbir__simdi_table_lookup3( i0, i1, i2, ( fp32_to_srgb8_tab4 - (127-13)*8 ) );
 8942+
 8943+      stbir__linear_to_srgb_finish( i0, f0 );
 8944+      stbir__linear_to_srgb_finish( i1, f1 );
 8945+      stbir__linear_to_srgb_finish( i2, f2 );
 8946+
 8947+      stbir__interleave_pack_and_store_16_u8( output,  STBIR_strs_join1(i, ,stbir__encode_order0), STBIR_strs_join1(i, ,stbir__encode_order1), STBIR_strs_join1(i, ,stbir__encode_order2), STBIR_strs_join1(i, ,stbir__encode_order3) );
 8948+
 8949+      output += 16;
 8950+      encode += 16;
 8951+
 8952+      if ( output <= end_output )
 8953+        continue;
 8954+      if ( output == ( end_output + 16 ) )
 8955+        break;
 8956+      output = end_output; // backup and do last couple
 8957+      encode = end_encode_m16;
 8958+    }
 8959+    return;
 8960+  }
 8961+  #endif
 8962+
 8963+  STBIR_SIMD_NO_UNROLL_LOOP_START
 8964+  do {
 8965+    float f;
 8966+    STBIR_SIMD_NO_UNROLL(encode);
 8967+
 8968+    output[stbir__decode_order0] = stbir__linear_to_srgb_uchar( encode[0] );
 8969+    output[stbir__decode_order1] = stbir__linear_to_srgb_uchar( encode[1] );
 8970+    output[stbir__decode_order2] = stbir__linear_to_srgb_uchar( encode[2] );
 8971+
 8972+    f = encode[3] * stbir__max_uint8_as_float + 0.5f;
 8973+    STBIR_CLAMP(f, 0, 255);
 8974+    output[stbir__decode_order3] = (unsigned char) f;
 8975+
 8976+    output += 4;
 8977+    encode += 4;
 8978+  } while( output < end_output );
 8979+}
 8980+
 8981+#endif
 8982+
 8983+#if ( stbir__coder_min_num == 2 ) || ( ( stbir__coder_min_num == 1 ) && ( !defined(stbir__decode_swizzle) ) )
 8984+
 8985+static float * STBIR__CODER_NAME(stbir__decode_uint8_srgb2_linearalpha)( float * decodep, int width_times_channels, void const * inputp )
 8986+{
 8987+  float STBIR_STREAMOUT_PTR( * ) decode = decodep;
 8988+  float * decode_end = (float*) decode + width_times_channels;
 8989+  unsigned char const * input = (unsigned char const *)inputp;
 8990+
 8991+  decode += 4;
 8992+  while( decode <= decode_end )
 8993+  {
 8994+    decode[0-4] = stbir__srgb_uchar_to_linear_float[ input[stbir__decode_order0] ];
 8995+    decode[1-4] = ( (float) input[stbir__decode_order1] ) * stbir__max_uint8_as_float_inverted;
 8996+    decode[2-4] = stbir__srgb_uchar_to_linear_float[ input[stbir__decode_order0+2] ];
 8997+    decode[3-4] = ( (float) input[stbir__decode_order1+2] ) * stbir__max_uint8_as_float_inverted;
 8998+    input += 4;
 8999+    decode += 4;
 9000+  }
 9001+  decode -= 4;
 9002+  if( decode < decode_end )
 9003+  {
 9004+    decode[0] = stbir__srgb_uchar_to_linear_float[ input[stbir__decode_order0] ];
 9005+    decode[1] = ( (float) input[stbir__decode_order1] ) * stbir__max_uint8_as_float_inverted;
 9006+  }
 9007+  return decode_end;
 9008+}
 9009+
 9010+static void STBIR__CODER_NAME( stbir__encode_uint8_srgb2_linearalpha )( void * outputp, int width_times_channels, float const * encode )
 9011+{
 9012+  unsigned char STBIR_SIMD_STREAMOUT_PTR( * ) output = (unsigned char*) outputp;
 9013+  unsigned char * end_output = ( (unsigned char*) output ) + width_times_channels;
 9014+
 9015+  #ifdef STBIR_SIMD
 9016+
 9017+  if ( width_times_channels >= 16 )
 9018+  {
 9019+    float const * end_encode_m16 = encode + width_times_channels - 16;
 9020+    end_output -= 16;
 9021+    STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 9022+    for(;;)
 9023+    {
 9024+      stbir__simdf f0, f1, f2, f3;
 9025+      stbir__simdi i0, i1, i2, i3;
 9026+
 9027+      STBIR_SIMD_NO_UNROLL(encode);
 9028+      stbir__simdf_load4_transposed( f0, f1, f2, f3, encode );
 9029+
 9030+      stbir__min_max_shift20( i0, f0 );
 9031+      stbir__scale_and_convert( i1, f1 );
 9032+      stbir__min_max_shift20( i2, f2 );
 9033+      stbir__scale_and_convert( i3, f3 );
 9034+
 9035+      stbir__simdi_table_lookup2( i0, i2, ( fp32_to_srgb8_tab4 - (127-13)*8 ) );
 9036+
 9037+      stbir__linear_to_srgb_finish( i0, f0 );
 9038+      stbir__linear_to_srgb_finish( i2, f2 );
 9039+
 9040+      stbir__interleave_pack_and_store_16_u8( output,  STBIR_strs_join1(i, ,stbir__encode_order0), STBIR_strs_join1(i, ,stbir__encode_order1), STBIR_strs_join1(i, ,stbir__encode_order2), STBIR_strs_join1(i, ,stbir__encode_order3) );
 9041+
 9042+      output += 16;
 9043+      encode += 16;
 9044+      if ( output <= end_output )
 9045+        continue;
 9046+      if ( output == ( end_output + 16 ) )
 9047+        break;
 9048+      output = end_output; // backup and do last couple
 9049+      encode = end_encode_m16;
 9050+    }
 9051+    return;
 9052+  }
 9053+  #endif
 9054+
 9055+  STBIR_SIMD_NO_UNROLL_LOOP_START
 9056+  do {
 9057+    float f;
 9058+    STBIR_SIMD_NO_UNROLL(encode);
 9059+
 9060+    output[stbir__decode_order0] = stbir__linear_to_srgb_uchar( encode[0] );
 9061+
 9062+    f = encode[1] * stbir__max_uint8_as_float + 0.5f;
 9063+    STBIR_CLAMP(f, 0, 255);
 9064+    output[stbir__decode_order1] = (unsigned char) f;
 9065+
 9066+    output += 2;
 9067+    encode += 2;
 9068+  } while( output < end_output );
 9069+}
 9070+
 9071+#endif
 9072+
 9073+static float * STBIR__CODER_NAME(stbir__decode_uint16_linear_scaled)( float * decodep, int width_times_channels, void const * inputp )
 9074+{
 9075+  float STBIR_STREAMOUT_PTR( * ) decode = decodep;
 9076+  float * decode_end = (float*) decode + width_times_channels;
 9077+  unsigned short const * input = (unsigned short const *)inputp;
 9078+
 9079+  #ifdef STBIR_SIMD
 9080+  unsigned short const * end_input_m8 = input + width_times_channels - 8;
 9081+  if ( width_times_channels >= 8 )
 9082+  {
 9083+    decode_end -= 8;
 9084+    STBIR_NO_UNROLL_LOOP_START_INF_FOR
 9085+    for(;;)
 9086+    {
 9087+      #ifdef STBIR_SIMD8
 9088+      stbir__simdi i; stbir__simdi8 o;
 9089+      stbir__simdf8 of;
 9090+      STBIR_NO_UNROLL(decode);
 9091+      stbir__simdi_load( i, input );
 9092+      stbir__simdi8_expand_u16_to_u32( o, i );
 9093+      stbir__simdi8_convert_i32_to_float( of, o );
 9094+      stbir__simdf8_mult( of, of, STBIR_max_uint16_as_float_inverted8);
 9095+      stbir__decode_simdf8_flip( of );
 9096+      stbir__simdf8_store( decode + 0, of );
 9097+      #else
 9098+      stbir__simdi i, o0, o1;
 9099+      stbir__simdf of0, of1;
 9100+      STBIR_NO_UNROLL(decode);
 9101+      stbir__simdi_load( i, input );
 9102+      stbir__simdi_expand_u16_to_u32( o0,o1,i );
 9103+      stbir__simdi_convert_i32_to_float( of0, o0 );
 9104+      stbir__simdi_convert_i32_to_float( of1, o1 );
 9105+      stbir__simdf_mult( of0, of0, STBIR__CONSTF(STBIR_max_uint16_as_float_inverted) );
 9106+      stbir__simdf_mult( of1, of1, STBIR__CONSTF(STBIR_max_uint16_as_float_inverted));
 9107+      stbir__decode_simdf4_flip( of0 );
 9108+      stbir__decode_simdf4_flip( of1 );
 9109+      stbir__simdf_store( decode + 0,  of0 );
 9110+      stbir__simdf_store( decode + 4,  of1 );
 9111+      #endif
 9112+      decode += 8;
 9113+      input += 8;
 9114+      if ( decode <= decode_end )
 9115+        continue;
 9116+      if ( decode == ( decode_end + 8 ) )
 9117+        break;
 9118+      decode = decode_end; // backup and do last couple
 9119+      input = end_input_m8;
 9120+    }
 9121+    return decode_end + 8;
 9122+  }
 9123+  #endif
 9124+
 9125+  // try to do blocks of 4 when you can
 9126+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9127+  decode += 4;
 9128+  STBIR_SIMD_NO_UNROLL_LOOP_START
 9129+  while( decode <= decode_end )
 9130+  {
 9131+    STBIR_SIMD_NO_UNROLL(decode);
 9132+    decode[0-4] = ((float)(input[stbir__decode_order0])) * stbir__max_uint16_as_float_inverted;
 9133+    decode[1-4] = ((float)(input[stbir__decode_order1])) * stbir__max_uint16_as_float_inverted;
 9134+    decode[2-4] = ((float)(input[stbir__decode_order2])) * stbir__max_uint16_as_float_inverted;
 9135+    decode[3-4] = ((float)(input[stbir__decode_order3])) * stbir__max_uint16_as_float_inverted;
 9136+    decode += 4;
 9137+    input += 4;
 9138+  }
 9139+  decode -= 4;
 9140+  #endif
 9141+
 9142+  // do the remnants
 9143+  #if stbir__coder_min_num < 4
 9144+  STBIR_NO_UNROLL_LOOP_START
 9145+  while( decode < decode_end )
 9146+  {
 9147+    STBIR_NO_UNROLL(decode);
 9148+    decode[0] = ((float)(input[stbir__decode_order0])) * stbir__max_uint16_as_float_inverted;
 9149+    #if stbir__coder_min_num >= 2
 9150+    decode[1] = ((float)(input[stbir__decode_order1])) * stbir__max_uint16_as_float_inverted;
 9151+    #endif
 9152+    #if stbir__coder_min_num >= 3
 9153+    decode[2] = ((float)(input[stbir__decode_order2])) * stbir__max_uint16_as_float_inverted;
 9154+    #endif
 9155+    decode += stbir__coder_min_num;
 9156+    input += stbir__coder_min_num;
 9157+  }
 9158+  #endif
 9159+  return decode_end;
 9160+}
 9161+
 9162+
 9163+static void STBIR__CODER_NAME(stbir__encode_uint16_linear_scaled)( void * outputp, int width_times_channels, float const * encode )
 9164+{
 9165+  unsigned short STBIR_SIMD_STREAMOUT_PTR( * ) output = (unsigned short*) outputp;
 9166+  unsigned short * end_output = ( (unsigned short*) output ) + width_times_channels;
 9167+
 9168+  #ifdef STBIR_SIMD
 9169+  {
 9170+    if ( width_times_channels >= stbir__simdfX_float_count*2 )
 9171+    {
 9172+      float const * end_encode_m8 = encode + width_times_channels - stbir__simdfX_float_count*2;
 9173+      end_output -= stbir__simdfX_float_count*2;
 9174+      STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 9175+      for(;;)
 9176+      {
 9177+        stbir__simdfX e0, e1;
 9178+        stbir__simdiX i;
 9179+        STBIR_SIMD_NO_UNROLL(encode);
 9180+        stbir__simdfX_madd_mem( e0, STBIR_simd_point5X, STBIR_max_uint16_as_floatX, encode );
 9181+        stbir__simdfX_madd_mem( e1, STBIR_simd_point5X, STBIR_max_uint16_as_floatX, encode+stbir__simdfX_float_count );
 9182+        stbir__encode_simdfX_unflip( e0 );
 9183+        stbir__encode_simdfX_unflip( e1 );
 9184+        stbir__simdfX_pack_to_words( i, e0, e1 );
 9185+        stbir__simdiX_store( output, i );
 9186+        encode += stbir__simdfX_float_count*2;
 9187+        output += stbir__simdfX_float_count*2;
 9188+        if ( output <= end_output )
 9189+          continue;
 9190+        if ( output == ( end_output + stbir__simdfX_float_count*2 ) )
 9191+          break;
 9192+        output = end_output;     // backup and do last couple
 9193+        encode = end_encode_m8;
 9194+      }
 9195+      return;
 9196+    }
 9197+  }
 9198+
 9199+  // try to do blocks of 4 when you can
 9200+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9201+  output += 4;
 9202+  STBIR_NO_UNROLL_LOOP_START
 9203+  while( output <= end_output )
 9204+  {
 9205+    stbir__simdf e;
 9206+    stbir__simdi i;
 9207+    STBIR_NO_UNROLL(encode);
 9208+    stbir__simdf_load( e, encode );
 9209+    stbir__simdf_madd( e, STBIR__CONSTF(STBIR_simd_point5), STBIR__CONSTF(STBIR_max_uint16_as_float), e );
 9210+    stbir__encode_simdf4_unflip( e );
 9211+    stbir__simdf_pack_to_8words( i, e, e );  // only use first 4
 9212+    stbir__simdi_store2( output-4, i );
 9213+    output += 4;
 9214+    encode += 4;
 9215+  }
 9216+  output -= 4;
 9217+  #endif
 9218+
 9219+  // do the remnants
 9220+  #if stbir__coder_min_num < 4
 9221+  STBIR_NO_UNROLL_LOOP_START
 9222+  while( output < end_output )
 9223+  {
 9224+    stbir__simdf e;
 9225+    STBIR_NO_UNROLL(encode);
 9226+    stbir__simdf_madd1_mem( e, STBIR__CONSTF(STBIR_simd_point5), STBIR__CONSTF(STBIR_max_uint16_as_float), encode+stbir__encode_order0 ); output[0] = stbir__simdf_convert_float_to_short( e );
 9227+    #if stbir__coder_min_num >= 2
 9228+    stbir__simdf_madd1_mem( e, STBIR__CONSTF(STBIR_simd_point5), STBIR__CONSTF(STBIR_max_uint16_as_float), encode+stbir__encode_order1 ); output[1] = stbir__simdf_convert_float_to_short( e );
 9229+    #endif
 9230+    #if stbir__coder_min_num >= 3
 9231+    stbir__simdf_madd1_mem( e, STBIR__CONSTF(STBIR_simd_point5), STBIR__CONSTF(STBIR_max_uint16_as_float), encode+stbir__encode_order2 ); output[2] = stbir__simdf_convert_float_to_short( e );
 9232+    #endif
 9233+    output += stbir__coder_min_num;
 9234+    encode += stbir__coder_min_num;
 9235+  }
 9236+  #endif
 9237+
 9238+  #else
 9239+
 9240+  // try to do blocks of 4 when you can
 9241+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9242+  output += 4;
 9243+  STBIR_SIMD_NO_UNROLL_LOOP_START
 9244+  while( output <= end_output )
 9245+  {
 9246+    float f;
 9247+    STBIR_SIMD_NO_UNROLL(encode);
 9248+    f = encode[stbir__encode_order0] * stbir__max_uint16_as_float + 0.5f; STBIR_CLAMP(f, 0, 65535); output[0-4] = (unsigned short)f;
 9249+    f = encode[stbir__encode_order1] * stbir__max_uint16_as_float + 0.5f; STBIR_CLAMP(f, 0, 65535); output[1-4] = (unsigned short)f;
 9250+    f = encode[stbir__encode_order2] * stbir__max_uint16_as_float + 0.5f; STBIR_CLAMP(f, 0, 65535); output[2-4] = (unsigned short)f;
 9251+    f = encode[stbir__encode_order3] * stbir__max_uint16_as_float + 0.5f; STBIR_CLAMP(f, 0, 65535); output[3-4] = (unsigned short)f;
 9252+    output += 4;
 9253+    encode += 4;
 9254+  }
 9255+  output -= 4;
 9256+  #endif
 9257+
 9258+  // do the remnants
 9259+  #if stbir__coder_min_num < 4
 9260+  STBIR_NO_UNROLL_LOOP_START
 9261+  while( output < end_output )
 9262+  {
 9263+    float f;
 9264+    STBIR_NO_UNROLL(encode);
 9265+    f = encode[stbir__encode_order0] * stbir__max_uint16_as_float + 0.5f; STBIR_CLAMP(f, 0, 65535); output[0] = (unsigned short)f;
 9266+    #if stbir__coder_min_num >= 2
 9267+    f = encode[stbir__encode_order1] * stbir__max_uint16_as_float + 0.5f; STBIR_CLAMP(f, 0, 65535); output[1] = (unsigned short)f;
 9268+    #endif
 9269+    #if stbir__coder_min_num >= 3
 9270+    f = encode[stbir__encode_order2] * stbir__max_uint16_as_float + 0.5f; STBIR_CLAMP(f, 0, 65535); output[2] = (unsigned short)f;
 9271+    #endif
 9272+    output += stbir__coder_min_num;
 9273+    encode += stbir__coder_min_num;
 9274+  }
 9275+  #endif
 9276+  #endif
 9277+}
 9278+
 9279+static float * STBIR__CODER_NAME(stbir__decode_uint16_linear)( float * decodep, int width_times_channels, void const * inputp )
 9280+{
 9281+  float STBIR_STREAMOUT_PTR( * ) decode = decodep;
 9282+  float * decode_end = (float*) decode + width_times_channels;
 9283+  unsigned short const * input = (unsigned short const *)inputp;
 9284+
 9285+  #ifdef STBIR_SIMD
 9286+  unsigned short const * end_input_m8 = input + width_times_channels - 8;
 9287+  if ( width_times_channels >= 8 )
 9288+  {
 9289+    decode_end -= 8;
 9290+    STBIR_NO_UNROLL_LOOP_START_INF_FOR
 9291+    for(;;)
 9292+    {
 9293+      #ifdef STBIR_SIMD8
 9294+      stbir__simdi i; stbir__simdi8 o;
 9295+      stbir__simdf8 of;
 9296+      STBIR_NO_UNROLL(decode);
 9297+      stbir__simdi_load( i, input );
 9298+      stbir__simdi8_expand_u16_to_u32( o, i );
 9299+      stbir__simdi8_convert_i32_to_float( of, o );
 9300+      stbir__decode_simdf8_flip( of );
 9301+      stbir__simdf8_store( decode + 0, of );
 9302+      #else
 9303+      stbir__simdi i, o0, o1;
 9304+      stbir__simdf of0, of1;
 9305+      STBIR_NO_UNROLL(decode);
 9306+      stbir__simdi_load( i, input );
 9307+      stbir__simdi_expand_u16_to_u32( o0, o1, i );
 9308+      stbir__simdi_convert_i32_to_float( of0, o0 );
 9309+      stbir__simdi_convert_i32_to_float( of1, o1 );
 9310+      stbir__decode_simdf4_flip( of0 );
 9311+      stbir__decode_simdf4_flip( of1 );
 9312+      stbir__simdf_store( decode + 0,  of0 );
 9313+      stbir__simdf_store( decode + 4,  of1 );
 9314+      #endif
 9315+      decode += 8;
 9316+      input += 8;
 9317+      if ( decode <= decode_end )
 9318+        continue;
 9319+      if ( decode == ( decode_end + 8 ) )
 9320+        break;
 9321+      decode = decode_end; // backup and do last couple
 9322+      input = end_input_m8;
 9323+    }
 9324+    return decode_end + 8;
 9325+  }
 9326+  #endif
 9327+
 9328+  // try to do blocks of 4 when you can
 9329+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9330+  decode += 4;
 9331+  STBIR_SIMD_NO_UNROLL_LOOP_START
 9332+  while( decode <= decode_end )
 9333+  {
 9334+    STBIR_SIMD_NO_UNROLL(decode);
 9335+    decode[0-4] = ((float)(input[stbir__decode_order0]));
 9336+    decode[1-4] = ((float)(input[stbir__decode_order1]));
 9337+    decode[2-4] = ((float)(input[stbir__decode_order2]));
 9338+    decode[3-4] = ((float)(input[stbir__decode_order3]));
 9339+    decode += 4;
 9340+    input += 4;
 9341+  }
 9342+  decode -= 4;
 9343+  #endif
 9344+
 9345+  // do the remnants
 9346+  #if stbir__coder_min_num < 4
 9347+  STBIR_NO_UNROLL_LOOP_START
 9348+  while( decode < decode_end )
 9349+  {
 9350+    STBIR_NO_UNROLL(decode);
 9351+    decode[0] = ((float)(input[stbir__decode_order0]));
 9352+    #if stbir__coder_min_num >= 2
 9353+    decode[1] = ((float)(input[stbir__decode_order1]));
 9354+    #endif
 9355+    #if stbir__coder_min_num >= 3
 9356+    decode[2] = ((float)(input[stbir__decode_order2]));
 9357+    #endif
 9358+    decode += stbir__coder_min_num;
 9359+    input += stbir__coder_min_num;
 9360+  }
 9361+  #endif
 9362+  return decode_end;
 9363+}
 9364+
 9365+static void STBIR__CODER_NAME(stbir__encode_uint16_linear)( void * outputp, int width_times_channels, float const * encode )
 9366+{
 9367+  unsigned short STBIR_SIMD_STREAMOUT_PTR( * ) output = (unsigned short*) outputp;
 9368+  unsigned short * end_output = ( (unsigned short*) output ) + width_times_channels;
 9369+
 9370+  #ifdef STBIR_SIMD
 9371+  {
 9372+    if ( width_times_channels >= stbir__simdfX_float_count*2 )
 9373+    {
 9374+      float const * end_encode_m8 = encode + width_times_channels - stbir__simdfX_float_count*2;
 9375+      end_output -= stbir__simdfX_float_count*2;
 9376+      STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 9377+      for(;;)
 9378+      {
 9379+        stbir__simdfX e0, e1;
 9380+        stbir__simdiX i;
 9381+        STBIR_SIMD_NO_UNROLL(encode);
 9382+        stbir__simdfX_add_mem( e0, STBIR_simd_point5X, encode );
 9383+        stbir__simdfX_add_mem( e1, STBIR_simd_point5X, encode+stbir__simdfX_float_count );
 9384+        stbir__encode_simdfX_unflip( e0 );
 9385+        stbir__encode_simdfX_unflip( e1 );
 9386+        stbir__simdfX_pack_to_words( i, e0, e1 );
 9387+        stbir__simdiX_store( output, i );
 9388+        encode += stbir__simdfX_float_count*2;
 9389+        output += stbir__simdfX_float_count*2;
 9390+        if ( output <= end_output )
 9391+          continue;
 9392+        if ( output == ( end_output + stbir__simdfX_float_count*2 ) )
 9393+          break;
 9394+        output = end_output; // backup and do last couple
 9395+        encode = end_encode_m8;
 9396+      }
 9397+      return;
 9398+    }
 9399+  }
 9400+
 9401+  // try to do blocks of 4 when you can
 9402+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9403+  output += 4;
 9404+  STBIR_NO_UNROLL_LOOP_START
 9405+  while( output <= end_output )
 9406+  {
 9407+    stbir__simdf e;
 9408+    stbir__simdi i;
 9409+    STBIR_NO_UNROLL(encode);
 9410+    stbir__simdf_load( e, encode );
 9411+    stbir__simdf_add( e, STBIR__CONSTF(STBIR_simd_point5), e );
 9412+    stbir__encode_simdf4_unflip( e );
 9413+    stbir__simdf_pack_to_8words( i, e, e );  // only use first 4
 9414+    stbir__simdi_store2( output-4, i );
 9415+    output += 4;
 9416+    encode += 4;
 9417+  }
 9418+  output -= 4;
 9419+  #endif
 9420+
 9421+  #else
 9422+
 9423+  // try to do blocks of 4 when you can
 9424+  #if  stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9425+  output += 4;
 9426+  STBIR_SIMD_NO_UNROLL_LOOP_START
 9427+  while( output <= end_output )
 9428+  {
 9429+    float f;
 9430+    STBIR_SIMD_NO_UNROLL(encode);
 9431+    f = encode[stbir__encode_order0] + 0.5f; STBIR_CLAMP(f, 0, 65535); output[0-4] = (unsigned short)f;
 9432+    f = encode[stbir__encode_order1] + 0.5f; STBIR_CLAMP(f, 0, 65535); output[1-4] = (unsigned short)f;
 9433+    f = encode[stbir__encode_order2] + 0.5f; STBIR_CLAMP(f, 0, 65535); output[2-4] = (unsigned short)f;
 9434+    f = encode[stbir__encode_order3] + 0.5f; STBIR_CLAMP(f, 0, 65535); output[3-4] = (unsigned short)f;
 9435+    output += 4;
 9436+    encode += 4;
 9437+  }
 9438+  output -= 4;
 9439+  #endif
 9440+
 9441+  #endif
 9442+
 9443+  // do the remnants
 9444+  #if stbir__coder_min_num < 4
 9445+  STBIR_NO_UNROLL_LOOP_START
 9446+  while( output < end_output )
 9447+  {
 9448+    float f;
 9449+    STBIR_NO_UNROLL(encode);
 9450+    f = encode[stbir__encode_order0] + 0.5f; STBIR_CLAMP(f, 0, 65535); output[0] = (unsigned short)f;
 9451+    #if stbir__coder_min_num >= 2
 9452+    f = encode[stbir__encode_order1] + 0.5f; STBIR_CLAMP(f, 0, 65535); output[1] = (unsigned short)f;
 9453+    #endif
 9454+    #if stbir__coder_min_num >= 3
 9455+    f = encode[stbir__encode_order2] + 0.5f; STBIR_CLAMP(f, 0, 65535); output[2] = (unsigned short)f;
 9456+    #endif
 9457+    output += stbir__coder_min_num;
 9458+    encode += stbir__coder_min_num;
 9459+  }
 9460+  #endif
 9461+}
 9462+
 9463+static float * STBIR__CODER_NAME(stbir__decode_half_float_linear)( float * decodep, int width_times_channels, void const * inputp )
 9464+{
 9465+  float STBIR_STREAMOUT_PTR( * ) decode = decodep;
 9466+  float * decode_end = (float*) decode + width_times_channels;
 9467+  stbir__FP16 const * input = (stbir__FP16 const *)inputp;
 9468+
 9469+  #ifdef STBIR_SIMD
 9470+  if ( width_times_channels >= 8 )
 9471+  {
 9472+    stbir__FP16 const * end_input_m8 = input + width_times_channels - 8;
 9473+    decode_end -= 8;
 9474+    STBIR_NO_UNROLL_LOOP_START_INF_FOR
 9475+    for(;;)
 9476+    {
 9477+      STBIR_NO_UNROLL(decode);
 9478+
 9479+      stbir__half_to_float_SIMD( decode, input );
 9480+      #ifdef stbir__decode_swizzle
 9481+      #ifdef STBIR_SIMD8
 9482+      {
 9483+        stbir__simdf8 of;
 9484+        stbir__simdf8_load( of, decode );
 9485+        stbir__decode_simdf8_flip( of );
 9486+        stbir__simdf8_store( decode, of );
 9487+      }
 9488+      #else
 9489+      {
 9490+        stbir__simdf of0,of1;
 9491+        stbir__simdf_load( of0, decode );
 9492+        stbir__simdf_load( of1, decode+4 );
 9493+        stbir__decode_simdf4_flip( of0 );
 9494+        stbir__decode_simdf4_flip( of1 );
 9495+        stbir__simdf_store( decode, of0 );
 9496+        stbir__simdf_store( decode+4, of1 );
 9497+      }
 9498+      #endif
 9499+      #endif
 9500+      decode += 8;
 9501+      input += 8;
 9502+      if ( decode <= decode_end )
 9503+        continue;
 9504+      if ( decode == ( decode_end + 8 ) )
 9505+        break;
 9506+      decode = decode_end; // backup and do last couple
 9507+      input = end_input_m8;
 9508+    }
 9509+    return decode_end + 8;
 9510+  }
 9511+  #endif
 9512+
 9513+  // try to do blocks of 4 when you can
 9514+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9515+  decode += 4;
 9516+  STBIR_SIMD_NO_UNROLL_LOOP_START
 9517+  while( decode <= decode_end )
 9518+  {
 9519+    STBIR_SIMD_NO_UNROLL(decode);
 9520+    decode[0-4] = stbir__half_to_float(input[stbir__decode_order0]);
 9521+    decode[1-4] = stbir__half_to_float(input[stbir__decode_order1]);
 9522+    decode[2-4] = stbir__half_to_float(input[stbir__decode_order2]);
 9523+    decode[3-4] = stbir__half_to_float(input[stbir__decode_order3]);
 9524+    decode += 4;
 9525+    input += 4;
 9526+  }
 9527+  decode -= 4;
 9528+  #endif
 9529+
 9530+  // do the remnants
 9531+  #if stbir__coder_min_num < 4
 9532+  STBIR_NO_UNROLL_LOOP_START
 9533+  while( decode < decode_end )
 9534+  {
 9535+    STBIR_NO_UNROLL(decode);
 9536+    decode[0] = stbir__half_to_float(input[stbir__decode_order0]);
 9537+    #if stbir__coder_min_num >= 2
 9538+    decode[1] = stbir__half_to_float(input[stbir__decode_order1]);
 9539+    #endif
 9540+    #if stbir__coder_min_num >= 3
 9541+    decode[2] = stbir__half_to_float(input[stbir__decode_order2]);
 9542+    #endif
 9543+    decode += stbir__coder_min_num;
 9544+    input += stbir__coder_min_num;
 9545+  }
 9546+  #endif
 9547+  return decode_end;
 9548+}
 9549+
 9550+static void STBIR__CODER_NAME( stbir__encode_half_float_linear )( void * outputp, int width_times_channels, float const * encode )
 9551+{
 9552+  stbir__FP16 STBIR_SIMD_STREAMOUT_PTR( * ) output = (stbir__FP16*) outputp;
 9553+  stbir__FP16 * end_output = ( (stbir__FP16*) output ) + width_times_channels;
 9554+
 9555+  #ifdef STBIR_SIMD
 9556+  if ( width_times_channels >= 8 )
 9557+  {
 9558+    float const * end_encode_m8 = encode + width_times_channels - 8;
 9559+    end_output -= 8;
 9560+    STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 9561+    for(;;)
 9562+    {
 9563+      STBIR_SIMD_NO_UNROLL(encode);
 9564+      #ifdef stbir__decode_swizzle
 9565+      #ifdef STBIR_SIMD8
 9566+      {
 9567+        stbir__simdf8 of;
 9568+        stbir__simdf8_load( of, encode );
 9569+        stbir__encode_simdf8_unflip( of );
 9570+        stbir__float_to_half_SIMD( output, (float*)&of );
 9571+      }
 9572+      #else
 9573+      {
 9574+        stbir__simdf of[2];
 9575+        stbir__simdf_load( of[0], encode );
 9576+        stbir__simdf_load( of[1], encode+4 );
 9577+        stbir__encode_simdf4_unflip( of[0] );
 9578+        stbir__encode_simdf4_unflip( of[1] );
 9579+        stbir__float_to_half_SIMD( output, (float*)of );
 9580+      }
 9581+      #endif
 9582+      #else
 9583+      stbir__float_to_half_SIMD( output, encode );
 9584+      #endif
 9585+      encode += 8;
 9586+      output += 8;
 9587+      if ( output <= end_output )
 9588+        continue;
 9589+      if ( output == ( end_output + 8 ) )
 9590+        break;
 9591+      output = end_output; // backup and do last couple
 9592+      encode = end_encode_m8;
 9593+    }
 9594+    return;
 9595+  }
 9596+  #endif
 9597+
 9598+  // try to do blocks of 4 when you can
 9599+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9600+  output += 4;
 9601+  STBIR_SIMD_NO_UNROLL_LOOP_START
 9602+  while( output <= end_output )
 9603+  {
 9604+    STBIR_SIMD_NO_UNROLL(output);
 9605+    output[0-4] = stbir__float_to_half(encode[stbir__encode_order0]);
 9606+    output[1-4] = stbir__float_to_half(encode[stbir__encode_order1]);
 9607+    output[2-4] = stbir__float_to_half(encode[stbir__encode_order2]);
 9608+    output[3-4] = stbir__float_to_half(encode[stbir__encode_order3]);
 9609+    output += 4;
 9610+    encode += 4;
 9611+  }
 9612+  output -= 4;
 9613+  #endif
 9614+
 9615+  // do the remnants
 9616+  #if stbir__coder_min_num < 4
 9617+  STBIR_NO_UNROLL_LOOP_START
 9618+  while( output < end_output )
 9619+  {
 9620+    STBIR_NO_UNROLL(output);
 9621+    output[0] = stbir__float_to_half(encode[stbir__encode_order0]);
 9622+    #if stbir__coder_min_num >= 2
 9623+    output[1] = stbir__float_to_half(encode[stbir__encode_order1]);
 9624+    #endif
 9625+    #if stbir__coder_min_num >= 3
 9626+    output[2] = stbir__float_to_half(encode[stbir__encode_order2]);
 9627+    #endif
 9628+    output += stbir__coder_min_num;
 9629+    encode += stbir__coder_min_num;
 9630+  }
 9631+  #endif
 9632+}
 9633+
 9634+static float * STBIR__CODER_NAME(stbir__decode_float_linear)( float * decodep, int width_times_channels, void const * inputp )
 9635+{
 9636+  #ifdef stbir__decode_swizzle
 9637+  float STBIR_STREAMOUT_PTR( * ) decode = decodep;
 9638+  float * decode_end = (float*) decode + width_times_channels;
 9639+  float const * input = (float const *)inputp;
 9640+
 9641+  #ifdef STBIR_SIMD
 9642+  if ( width_times_channels >= 16 )
 9643+  {
 9644+    float const * end_input_m16 = input + width_times_channels - 16;
 9645+    decode_end -= 16;
 9646+    STBIR_NO_UNROLL_LOOP_START_INF_FOR
 9647+    for(;;)
 9648+    {
 9649+      STBIR_NO_UNROLL(decode);
 9650+      #ifdef stbir__decode_swizzle
 9651+      #ifdef STBIR_SIMD8
 9652+      {
 9653+        stbir__simdf8 of0,of1;
 9654+        stbir__simdf8_load( of0, input );
 9655+        stbir__simdf8_load( of1, input+8 );
 9656+        stbir__decode_simdf8_flip( of0 );
 9657+        stbir__decode_simdf8_flip( of1 );
 9658+        stbir__simdf8_store( decode, of0 );
 9659+        stbir__simdf8_store( decode+8, of1 );
 9660+      }
 9661+      #else
 9662+      {
 9663+        stbir__simdf of0,of1,of2,of3;
 9664+        stbir__simdf_load( of0, input );
 9665+        stbir__simdf_load( of1, input+4 );
 9666+        stbir__simdf_load( of2, input+8 );
 9667+        stbir__simdf_load( of3, input+12 );
 9668+        stbir__decode_simdf4_flip( of0 );
 9669+        stbir__decode_simdf4_flip( of1 );
 9670+        stbir__decode_simdf4_flip( of2 );
 9671+        stbir__decode_simdf4_flip( of3 );
 9672+        stbir__simdf_store( decode, of0 );
 9673+        stbir__simdf_store( decode+4, of1 );
 9674+        stbir__simdf_store( decode+8, of2 );
 9675+        stbir__simdf_store( decode+12, of3 );
 9676+      }
 9677+      #endif
 9678+      #endif
 9679+      decode += 16;
 9680+      input += 16;
 9681+      if ( decode <= decode_end )
 9682+        continue;
 9683+      if ( decode == ( decode_end + 16 ) )
 9684+        break;
 9685+      decode = decode_end; // backup and do last couple
 9686+      input = end_input_m16;
 9687+    }
 9688+    return decode_end + 16;
 9689+  }
 9690+  #endif
 9691+
 9692+  // try to do blocks of 4 when you can
 9693+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9694+  decode += 4;
 9695+  STBIR_SIMD_NO_UNROLL_LOOP_START
 9696+  while( decode <= decode_end )
 9697+  {
 9698+    STBIR_SIMD_NO_UNROLL(decode);
 9699+    decode[0-4] = input[stbir__decode_order0];
 9700+    decode[1-4] = input[stbir__decode_order1];
 9701+    decode[2-4] = input[stbir__decode_order2];
 9702+    decode[3-4] = input[stbir__decode_order3];
 9703+    decode += 4;
 9704+    input += 4;
 9705+  }
 9706+  decode -= 4;
 9707+  #endif
 9708+
 9709+  // do the remnants
 9710+  #if stbir__coder_min_num < 4
 9711+  STBIR_NO_UNROLL_LOOP_START
 9712+  while( decode < decode_end )
 9713+  {
 9714+    STBIR_NO_UNROLL(decode);
 9715+    decode[0] = input[stbir__decode_order0];
 9716+    #if stbir__coder_min_num >= 2
 9717+    decode[1] = input[stbir__decode_order1];
 9718+    #endif
 9719+    #if stbir__coder_min_num >= 3
 9720+    decode[2] = input[stbir__decode_order2];
 9721+    #endif
 9722+    decode += stbir__coder_min_num;
 9723+    input += stbir__coder_min_num;
 9724+  }
 9725+  #endif
 9726+  return decode_end;
 9727+
 9728+  #else
 9729+
 9730+  if ( (void*)decodep != inputp )
 9731+    STBIR_MEMCPY( decodep, inputp, width_times_channels * sizeof( float ) );
 9732+
 9733+  return decodep + width_times_channels;
 9734+
 9735+  #endif
 9736+}
 9737+
 9738+static void STBIR__CODER_NAME( stbir__encode_float_linear )( void * outputp, int width_times_channels, float const * encode )
 9739+{
 9740+  #if !defined( STBIR_FLOAT_HIGH_CLAMP ) && !defined(STBIR_FLOAT_LOW_CLAMP) && !defined(stbir__decode_swizzle)
 9741+
 9742+  if ( (void*)outputp != (void*) encode )
 9743+    STBIR_MEMCPY( outputp, encode, width_times_channels * sizeof( float ) );
 9744+
 9745+  #else
 9746+
 9747+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = (float*) outputp;
 9748+  float * end_output = ( (float*) output ) + width_times_channels;
 9749+
 9750+  #ifdef STBIR_FLOAT_HIGH_CLAMP
 9751+  #define stbir_scalar_hi_clamp( v ) if ( v > STBIR_FLOAT_HIGH_CLAMP ) v = STBIR_FLOAT_HIGH_CLAMP;
 9752+  #else
 9753+  #define stbir_scalar_hi_clamp( v )
 9754+  #endif
 9755+  #ifdef STBIR_FLOAT_LOW_CLAMP
 9756+  #define stbir_scalar_lo_clamp( v ) if ( v < STBIR_FLOAT_LOW_CLAMP ) v = STBIR_FLOAT_LOW_CLAMP;
 9757+  #else
 9758+  #define stbir_scalar_lo_clamp( v )
 9759+  #endif
 9760+
 9761+  #ifdef STBIR_SIMD
 9762+
 9763+  #ifdef STBIR_FLOAT_HIGH_CLAMP
 9764+  const stbir__simdfX high_clamp = stbir__simdf_frepX(STBIR_FLOAT_HIGH_CLAMP);
 9765+  #endif
 9766+  #ifdef STBIR_FLOAT_LOW_CLAMP
 9767+  const stbir__simdfX low_clamp = stbir__simdf_frepX(STBIR_FLOAT_LOW_CLAMP);
 9768+  #endif
 9769+
 9770+  if ( width_times_channels >= ( stbir__simdfX_float_count * 2 ) )
 9771+  {
 9772+    float const * end_encode_m8 = encode + width_times_channels - ( stbir__simdfX_float_count * 2 );
 9773+    end_output -= ( stbir__simdfX_float_count * 2 );
 9774+    STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 9775+    for(;;)
 9776+    {
 9777+      stbir__simdfX e0, e1;
 9778+      STBIR_SIMD_NO_UNROLL(encode);
 9779+      stbir__simdfX_load( e0, encode );
 9780+      stbir__simdfX_load( e1, encode+stbir__simdfX_float_count );
 9781+#ifdef STBIR_FLOAT_HIGH_CLAMP
 9782+      stbir__simdfX_min( e0, e0, high_clamp );
 9783+      stbir__simdfX_min( e1, e1, high_clamp );
 9784+#endif
 9785+#ifdef STBIR_FLOAT_LOW_CLAMP
 9786+      stbir__simdfX_max( e0, e0, low_clamp );
 9787+      stbir__simdfX_max( e1, e1, low_clamp );
 9788+#endif
 9789+      stbir__encode_simdfX_unflip( e0 );
 9790+      stbir__encode_simdfX_unflip( e1 );
 9791+      stbir__simdfX_store( output, e0 );
 9792+      stbir__simdfX_store( output+stbir__simdfX_float_count, e1 );
 9793+      encode += stbir__simdfX_float_count * 2;
 9794+      output += stbir__simdfX_float_count * 2;
 9795+      if ( output <= end_output )
 9796+        continue;
 9797+      if ( output == ( end_output + ( stbir__simdfX_float_count * 2 ) ) )
 9798+        break;
 9799+      output = end_output; // backup and do last couple
 9800+      encode = end_encode_m8;
 9801+    }
 9802+    return;
 9803+  }
 9804+
 9805+  // try to do blocks of 4 when you can
 9806+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9807+  output += 4;
 9808+  STBIR_NO_UNROLL_LOOP_START
 9809+  while( output <= end_output )
 9810+  {
 9811+    stbir__simdf e0;
 9812+    STBIR_NO_UNROLL(encode);
 9813+    stbir__simdf_load( e0, encode );
 9814+#ifdef STBIR_FLOAT_HIGH_CLAMP
 9815+    stbir__simdf_min( e0, e0, high_clamp );
 9816+#endif
 9817+#ifdef STBIR_FLOAT_LOW_CLAMP
 9818+    stbir__simdf_max( e0, e0, low_clamp );
 9819+#endif
 9820+    stbir__encode_simdf4_unflip( e0 );
 9821+    stbir__simdf_store( output-4, e0 );
 9822+    output += 4;
 9823+    encode += 4;
 9824+  }
 9825+  output -= 4;
 9826+  #endif
 9827+
 9828+  #else
 9829+
 9830+  // try to do blocks of 4 when you can
 9831+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9832+  output += 4;
 9833+  STBIR_SIMD_NO_UNROLL_LOOP_START
 9834+  while( output <= end_output )
 9835+  {
 9836+    float e;
 9837+    STBIR_SIMD_NO_UNROLL(encode);
 9838+    e = encode[ stbir__encode_order0 ]; stbir_scalar_hi_clamp( e ); stbir_scalar_lo_clamp( e ); output[0-4] = e;
 9839+    e = encode[ stbir__encode_order1 ]; stbir_scalar_hi_clamp( e ); stbir_scalar_lo_clamp( e ); output[1-4] = e;
 9840+    e = encode[ stbir__encode_order2 ]; stbir_scalar_hi_clamp( e ); stbir_scalar_lo_clamp( e ); output[2-4] = e;
 9841+    e = encode[ stbir__encode_order3 ]; stbir_scalar_hi_clamp( e ); stbir_scalar_lo_clamp( e ); output[3-4] = e;
 9842+    output += 4;
 9843+    encode += 4;
 9844+  }
 9845+  output -= 4;
 9846+
 9847+  #endif
 9848+
 9849+  #endif
 9850+
 9851+  // do the remnants
 9852+  #if stbir__coder_min_num < 4
 9853+  STBIR_NO_UNROLL_LOOP_START
 9854+  while( output < end_output )
 9855+  {
 9856+    float e;
 9857+    STBIR_NO_UNROLL(encode);
 9858+    e = encode[ stbir__encode_order0 ]; stbir_scalar_hi_clamp( e ); stbir_scalar_lo_clamp( e ); output[0] = e;
 9859+    #if stbir__coder_min_num >= 2
 9860+    e = encode[ stbir__encode_order1 ]; stbir_scalar_hi_clamp( e ); stbir_scalar_lo_clamp( e ); output[1] = e;
 9861+    #endif
 9862+    #if stbir__coder_min_num >= 3
 9863+    e = encode[ stbir__encode_order2 ]; stbir_scalar_hi_clamp( e ); stbir_scalar_lo_clamp( e ); output[2] = e;
 9864+    #endif
 9865+    output += stbir__coder_min_num;
 9866+    encode += stbir__coder_min_num;
 9867+  }
 9868+  #endif
 9869+
 9870+  #endif
 9871+}
 9872+
 9873+#undef stbir__decode_suffix
 9874+#undef stbir__decode_simdf8_flip
 9875+#undef stbir__decode_simdf4_flip
 9876+#undef stbir__decode_order0
 9877+#undef stbir__decode_order1
 9878+#undef stbir__decode_order2
 9879+#undef stbir__decode_order3
 9880+#undef stbir__encode_order0
 9881+#undef stbir__encode_order1
 9882+#undef stbir__encode_order2
 9883+#undef stbir__encode_order3
 9884+#undef stbir__encode_simdf8_unflip
 9885+#undef stbir__encode_simdf4_unflip
 9886+#undef stbir__encode_simdfX_unflip
 9887+#undef STBIR__CODER_NAME
 9888+#undef stbir__coder_min_num
 9889+#undef stbir__decode_swizzle
 9890+#undef stbir_scalar_hi_clamp
 9891+#undef stbir_scalar_lo_clamp
 9892+#undef STB_IMAGE_RESIZE_DO_CODERS
 9893+
 9894+#elif defined( STB_IMAGE_RESIZE_DO_VERTICALS)
 9895+
 9896+#ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 9897+#define STBIR_chans( start, end ) STBIR_strs_join14(start,STBIR__vertical_channels,end,_cont)
 9898+#else
 9899+#define STBIR_chans( start, end ) STBIR_strs_join1(start,STBIR__vertical_channels,end)
 9900+#endif
 9901+
 9902+#if STBIR__vertical_channels >= 1
 9903+#define stbIF0( code ) code
 9904+#else
 9905+#define stbIF0( code )
 9906+#endif
 9907+#if STBIR__vertical_channels >= 2
 9908+#define stbIF1( code ) code
 9909+#else
 9910+#define stbIF1( code )
 9911+#endif
 9912+#if STBIR__vertical_channels >= 3
 9913+#define stbIF2( code ) code
 9914+#else
 9915+#define stbIF2( code )
 9916+#endif
 9917+#if STBIR__vertical_channels >= 4
 9918+#define stbIF3( code ) code
 9919+#else
 9920+#define stbIF3( code )
 9921+#endif
 9922+#if STBIR__vertical_channels >= 5
 9923+#define stbIF4( code ) code
 9924+#else
 9925+#define stbIF4( code )
 9926+#endif
 9927+#if STBIR__vertical_channels >= 6
 9928+#define stbIF5( code ) code
 9929+#else
 9930+#define stbIF5( code )
 9931+#endif
 9932+#if STBIR__vertical_channels >= 7
 9933+#define stbIF6( code ) code
 9934+#else
 9935+#define stbIF6( code )
 9936+#endif
 9937+#if STBIR__vertical_channels >= 8
 9938+#define stbIF7( code ) code
 9939+#else
 9940+#define stbIF7( code )
 9941+#endif
 9942+
 9943+static void STBIR_chans( stbir__vertical_scatter_with_,_coeffs)( float ** outputs, float const * vertical_coefficients, float const * input, float const * input_end )
 9944+{
 9945+  stbIF0( float STBIR_SIMD_STREAMOUT_PTR( * ) output0 = outputs[0]; float c0s = vertical_coefficients[0]; )
 9946+  stbIF1( float STBIR_SIMD_STREAMOUT_PTR( * ) output1 = outputs[1]; float c1s = vertical_coefficients[1]; )
 9947+  stbIF2( float STBIR_SIMD_STREAMOUT_PTR( * ) output2 = outputs[2]; float c2s = vertical_coefficients[2]; )
 9948+  stbIF3( float STBIR_SIMD_STREAMOUT_PTR( * ) output3 = outputs[3]; float c3s = vertical_coefficients[3]; )
 9949+  stbIF4( float STBIR_SIMD_STREAMOUT_PTR( * ) output4 = outputs[4]; float c4s = vertical_coefficients[4]; )
 9950+  stbIF5( float STBIR_SIMD_STREAMOUT_PTR( * ) output5 = outputs[5]; float c5s = vertical_coefficients[5]; )
 9951+  stbIF6( float STBIR_SIMD_STREAMOUT_PTR( * ) output6 = outputs[6]; float c6s = vertical_coefficients[6]; )
 9952+  stbIF7( float STBIR_SIMD_STREAMOUT_PTR( * ) output7 = outputs[7]; float c7s = vertical_coefficients[7]; )
 9953+
 9954+  #ifdef STBIR_SIMD
 9955+  {
 9956+    stbIF0(stbir__simdfX c0 = stbir__simdf_frepX( c0s ); )
 9957+    stbIF1(stbir__simdfX c1 = stbir__simdf_frepX( c1s ); )
 9958+    stbIF2(stbir__simdfX c2 = stbir__simdf_frepX( c2s ); )
 9959+    stbIF3(stbir__simdfX c3 = stbir__simdf_frepX( c3s ); )
 9960+    stbIF4(stbir__simdfX c4 = stbir__simdf_frepX( c4s ); )
 9961+    stbIF5(stbir__simdfX c5 = stbir__simdf_frepX( c5s ); )
 9962+    stbIF6(stbir__simdfX c6 = stbir__simdf_frepX( c6s ); )
 9963+    stbIF7(stbir__simdfX c7 = stbir__simdf_frepX( c7s ); )
 9964+    STBIR_SIMD_NO_UNROLL_LOOP_START
 9965+    while ( ( (char*)input_end - (char*) input ) >= (16*stbir__simdfX_float_count) )
 9966+    {
 9967+      stbir__simdfX o0, o1, o2, o3, r0, r1, r2, r3;
 9968+      STBIR_SIMD_NO_UNROLL(output0);
 9969+
 9970+      stbir__simdfX_load( r0, input );               stbir__simdfX_load( r1, input+stbir__simdfX_float_count );     stbir__simdfX_load( r2, input+(2*stbir__simdfX_float_count) );      stbir__simdfX_load( r3, input+(3*stbir__simdfX_float_count) );
 9971+
 9972+      #ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 9973+      stbIF0( stbir__simdfX_load( o0, output0 );     stbir__simdfX_load( o1, output0+stbir__simdfX_float_count );   stbir__simdfX_load( o2, output0+(2*stbir__simdfX_float_count) );    stbir__simdfX_load( o3, output0+(3*stbir__simdfX_float_count) );
 9974+              stbir__simdfX_madd( o0, o0, r0, c0 );  stbir__simdfX_madd( o1, o1, r1, c0 );  stbir__simdfX_madd( o2, o2, r2, c0 );   stbir__simdfX_madd( o3, o3, r3, c0 );
 9975+              stbir__simdfX_store( output0, o0 );    stbir__simdfX_store( output0+stbir__simdfX_float_count, o1 );  stbir__simdfX_store( output0+(2*stbir__simdfX_float_count), o2 );   stbir__simdfX_store( output0+(3*stbir__simdfX_float_count), o3 ); )
 9976+      stbIF1( stbir__simdfX_load( o0, output1 );     stbir__simdfX_load( o1, output1+stbir__simdfX_float_count );   stbir__simdfX_load( o2, output1+(2*stbir__simdfX_float_count) );    stbir__simdfX_load( o3, output1+(3*stbir__simdfX_float_count) );
 9977+              stbir__simdfX_madd( o0, o0, r0, c1 );  stbir__simdfX_madd( o1, o1, r1, c1 );  stbir__simdfX_madd( o2, o2, r2, c1 );   stbir__simdfX_madd( o3, o3, r3, c1 );
 9978+              stbir__simdfX_store( output1, o0 );    stbir__simdfX_store( output1+stbir__simdfX_float_count, o1 );  stbir__simdfX_store( output1+(2*stbir__simdfX_float_count), o2 );   stbir__simdfX_store( output1+(3*stbir__simdfX_float_count), o3 ); )
 9979+      stbIF2( stbir__simdfX_load( o0, output2 );     stbir__simdfX_load( o1, output2+stbir__simdfX_float_count );   stbir__simdfX_load( o2, output2+(2*stbir__simdfX_float_count) );    stbir__simdfX_load( o3, output2+(3*stbir__simdfX_float_count) );
 9980+              stbir__simdfX_madd( o0, o0, r0, c2 );  stbir__simdfX_madd( o1, o1, r1, c2 );  stbir__simdfX_madd( o2, o2, r2, c2 );   stbir__simdfX_madd( o3, o3, r3, c2 );
 9981+              stbir__simdfX_store( output2, o0 );    stbir__simdfX_store( output2+stbir__simdfX_float_count, o1 );  stbir__simdfX_store( output2+(2*stbir__simdfX_float_count), o2 );   stbir__simdfX_store( output2+(3*stbir__simdfX_float_count), o3 ); )
 9982+      stbIF3( stbir__simdfX_load( o0, output3 );     stbir__simdfX_load( o1, output3+stbir__simdfX_float_count );   stbir__simdfX_load( o2, output3+(2*stbir__simdfX_float_count) );    stbir__simdfX_load( o3, output3+(3*stbir__simdfX_float_count) );
 9983+              stbir__simdfX_madd( o0, o0, r0, c3 );  stbir__simdfX_madd( o1, o1, r1, c3 );  stbir__simdfX_madd( o2, o2, r2, c3 );   stbir__simdfX_madd( o3, o3, r3, c3 );
 9984+              stbir__simdfX_store( output3, o0 );    stbir__simdfX_store( output3+stbir__simdfX_float_count, o1 );  stbir__simdfX_store( output3+(2*stbir__simdfX_float_count), o2 );   stbir__simdfX_store( output3+(3*stbir__simdfX_float_count), o3 ); )
 9985+      stbIF4( stbir__simdfX_load( o0, output4 );     stbir__simdfX_load( o1, output4+stbir__simdfX_float_count );   stbir__simdfX_load( o2, output4+(2*stbir__simdfX_float_count) );    stbir__simdfX_load( o3, output4+(3*stbir__simdfX_float_count) );
 9986+              stbir__simdfX_madd( o0, o0, r0, c4 );  stbir__simdfX_madd( o1, o1, r1, c4 );  stbir__simdfX_madd( o2, o2, r2, c4 );   stbir__simdfX_madd( o3, o3, r3, c4 );
 9987+              stbir__simdfX_store( output4, o0 );    stbir__simdfX_store( output4+stbir__simdfX_float_count, o1 );  stbir__simdfX_store( output4+(2*stbir__simdfX_float_count), o2 );   stbir__simdfX_store( output4+(3*stbir__simdfX_float_count), o3 ); )
 9988+      stbIF5( stbir__simdfX_load( o0, output5 );     stbir__simdfX_load( o1, output5+stbir__simdfX_float_count );   stbir__simdfX_load( o2, output5+(2*stbir__simdfX_float_count));    stbir__simdfX_load( o3, output5+(3*stbir__simdfX_float_count) );
 9989+              stbir__simdfX_madd( o0, o0, r0, c5 );  stbir__simdfX_madd( o1, o1, r1, c5 );  stbir__simdfX_madd( o2, o2, r2, c5 );   stbir__simdfX_madd( o3, o3, r3, c5 );
 9990+              stbir__simdfX_store( output5, o0 );    stbir__simdfX_store( output5+stbir__simdfX_float_count, o1 );  stbir__simdfX_store( output5+(2*stbir__simdfX_float_count), o2 );   stbir__simdfX_store( output5+(3*stbir__simdfX_float_count), o3 ); )
 9991+      stbIF6( stbir__simdfX_load( o0, output6 );     stbir__simdfX_load( o1, output6+stbir__simdfX_float_count );   stbir__simdfX_load( o2, output6+(2*stbir__simdfX_float_count) );    stbir__simdfX_load( o3, output6+(3*stbir__simdfX_float_count) );
 9992+              stbir__simdfX_madd( o0, o0, r0, c6 );  stbir__simdfX_madd( o1, o1, r1, c6 );  stbir__simdfX_madd( o2, o2, r2, c6 );   stbir__simdfX_madd( o3, o3, r3, c6 );
 9993+              stbir__simdfX_store( output6, o0 );    stbir__simdfX_store( output6+stbir__simdfX_float_count, o1 );  stbir__simdfX_store( output6+(2*stbir__simdfX_float_count), o2 );   stbir__simdfX_store( output6+(3*stbir__simdfX_float_count), o3 ); )
 9994+      stbIF7( stbir__simdfX_load( o0, output7 );     stbir__simdfX_load( o1, output7+stbir__simdfX_float_count );   stbir__simdfX_load( o2, output7+(2*stbir__simdfX_float_count) );    stbir__simdfX_load( o3, output7+(3*stbir__simdfX_float_count) );
 9995+              stbir__simdfX_madd( o0, o0, r0, c7 );  stbir__simdfX_madd( o1, o1, r1, c7 );  stbir__simdfX_madd( o2, o2, r2, c7 );   stbir__simdfX_madd( o3, o3, r3, c7 );
 9996+              stbir__simdfX_store( output7, o0 );    stbir__simdfX_store( output7+stbir__simdfX_float_count, o1 );  stbir__simdfX_store( output7+(2*stbir__simdfX_float_count), o2 );   stbir__simdfX_store( output7+(3*stbir__simdfX_float_count), o3 ); )
 9997+      #else
 9998+      stbIF0( stbir__simdfX_mult( o0, r0, c0 );      stbir__simdfX_mult( o1, r1, c0 );      stbir__simdfX_mult( o2, r2, c0 );       stbir__simdfX_mult( o3, r3, c0 );
 9999+              stbir__simdfX_store( output0, o0 );    stbir__simdfX_store( output0+stbir__simdfX_float_count, o1 );  stbir__simdfX_store( output0+(2*stbir__simdfX_float_count), o2 );   stbir__simdfX_store( output0+(3*stbir__simdfX_float_count), o3 ); )
10000+      stbIF1( stbir__simdfX_mult( o0, r0, c1 );      stbir__simdfX_mult( o1, r1, c1 );      stbir__simdfX_mult( o2, r2, c1 );       stbir__simdfX_mult( o3, r3, c1 );
10001+              stbir__simdfX_store( output1, o0 );    stbir__simdfX_store( output1+stbir__simdfX_float_count, o1 );  stbir__simdfX_store( output1+(2*stbir__simdfX_float_count), o2 );   stbir__simdfX_store( output1+(3*stbir__simdfX_float_count), o3 ); )
10002+      stbIF2( stbir__simdfX_mult( o0, r0, c2 );      stbir__simdfX_mult( o1, r1, c2 );      stbir__simdfX_mult( o2, r2, c2 );       stbir__simdfX_mult( o3, r3, c2 );
10003+              stbir__simdfX_store( output2, o0 );    stbir__simdfX_store( output2+stbir__simdfX_float_count, o1 );  stbir__simdfX_store( output2+(2*stbir__simdfX_float_count), o2 );   stbir__simdfX_store( output2+(3*stbir__simdfX_float_count), o3 ); )
10004+      stbIF3( stbir__simdfX_mult( o0, r0, c3 );      stbir__simdfX_mult( o1, r1, c3 );      stbir__simdfX_mult( o2, r2, c3 );       stbir__simdfX_mult( o3, r3, c3 );
10005+              stbir__simdfX_store( output3, o0 );    stbir__simdfX_store( output3+stbir__simdfX_float_count, o1 );  stbir__simdfX_store( output3+(2*stbir__simdfX_float_count), o2 );   stbir__simdfX_store( output3+(3*stbir__simdfX_float_count), o3 ); )
10006+      stbIF4( stbir__simdfX_mult( o0, r0, c4 );      stbir__simdfX_mult( o1, r1, c4 );      stbir__simdfX_mult( o2, r2, c4 );       stbir__simdfX_mult( o3, r3, c4 );
10007+              stbir__simdfX_store( output4, o0 );    stbir__simdfX_store( output4+stbir__simdfX_float_count, o1 );  stbir__simdfX_store( output4+(2*stbir__simdfX_float_count), o2 );   stbir__simdfX_store( output4+(3*stbir__simdfX_float_count), o3 ); )
10008+      stbIF5( stbir__simdfX_mult( o0, r0, c5 );      stbir__simdfX_mult( o1, r1, c5 );      stbir__simdfX_mult( o2, r2, c5 );       stbir__simdfX_mult( o3, r3, c5 );
10009+              stbir__simdfX_store( output5, o0 );    stbir__simdfX_store( output5+stbir__simdfX_float_count, o1 );  stbir__simdfX_store( output5+(2*stbir__simdfX_float_count), o2 );   stbir__simdfX_store( output5+(3*stbir__simdfX_float_count), o3 ); )
10010+      stbIF6( stbir__simdfX_mult( o0, r0, c6 );      stbir__simdfX_mult( o1, r1, c6 );      stbir__simdfX_mult( o2, r2, c6 );       stbir__simdfX_mult( o3, r3, c6 );
10011+              stbir__simdfX_store( output6, o0 );    stbir__simdfX_store( output6+stbir__simdfX_float_count, o1 );  stbir__simdfX_store( output6+(2*stbir__simdfX_float_count), o2 );   stbir__simdfX_store( output6+(3*stbir__simdfX_float_count), o3 ); )
10012+      stbIF7( stbir__simdfX_mult( o0, r0, c7 );      stbir__simdfX_mult( o1, r1, c7 );      stbir__simdfX_mult( o2, r2, c7 );       stbir__simdfX_mult( o3, r3, c7 );
10013+              stbir__simdfX_store( output7, o0 );    stbir__simdfX_store( output7+stbir__simdfX_float_count, o1 );  stbir__simdfX_store( output7+(2*stbir__simdfX_float_count), o2 );   stbir__simdfX_store( output7+(3*stbir__simdfX_float_count), o3 ); )
10014+      #endif
10015+
10016+      input += (4*stbir__simdfX_float_count);
10017+      stbIF0( output0 += (4*stbir__simdfX_float_count); ) stbIF1( output1 += (4*stbir__simdfX_float_count); ) stbIF2( output2 += (4*stbir__simdfX_float_count); ) stbIF3( output3 += (4*stbir__simdfX_float_count); ) stbIF4( output4 += (4*stbir__simdfX_float_count); ) stbIF5( output5 += (4*stbir__simdfX_float_count); ) stbIF6( output6 += (4*stbir__simdfX_float_count); ) stbIF7( output7 += (4*stbir__simdfX_float_count); )
10018+    }
10019+    STBIR_SIMD_NO_UNROLL_LOOP_START
10020+    while ( ( (char*)input_end - (char*) input ) >= 16 )
10021+    {
10022+      stbir__simdf o0, r0;
10023+      STBIR_SIMD_NO_UNROLL(output0);
10024+
10025+      stbir__simdf_load( r0, input );
10026+
10027+      #ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
10028+      stbIF0( stbir__simdf_load( o0, output0 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c0 ) );  stbir__simdf_store( output0, o0 ); )
10029+      stbIF1( stbir__simdf_load( o0, output1 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c1 ) );  stbir__simdf_store( output1, o0 ); )
10030+      stbIF2( stbir__simdf_load( o0, output2 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c2 ) );  stbir__simdf_store( output2, o0 ); )
10031+      stbIF3( stbir__simdf_load( o0, output3 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c3 ) );  stbir__simdf_store( output3, o0 ); )
10032+      stbIF4( stbir__simdf_load( o0, output4 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c4 ) );  stbir__simdf_store( output4, o0 ); )
10033+      stbIF5( stbir__simdf_load( o0, output5 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c5 ) );  stbir__simdf_store( output5, o0 ); )
10034+      stbIF6( stbir__simdf_load( o0, output6 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c6 ) );  stbir__simdf_store( output6, o0 ); )
10035+      stbIF7( stbir__simdf_load( o0, output7 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c7 ) );  stbir__simdf_store( output7, o0 ); )
10036+      #else
10037+      stbIF0( stbir__simdf_mult( o0, r0, stbir__if_simdf8_cast_to_simdf4( c0 ) );   stbir__simdf_store( output0, o0 ); )
10038+      stbIF1( stbir__simdf_mult( o0, r0, stbir__if_simdf8_cast_to_simdf4( c1 ) );   stbir__simdf_store( output1, o0 ); )
10039+      stbIF2( stbir__simdf_mult( o0, r0, stbir__if_simdf8_cast_to_simdf4( c2 ) );   stbir__simdf_store( output2, o0 ); )
10040+      stbIF3( stbir__simdf_mult( o0, r0, stbir__if_simdf8_cast_to_simdf4( c3 ) );   stbir__simdf_store( output3, o0 ); )
10041+      stbIF4( stbir__simdf_mult( o0, r0, stbir__if_simdf8_cast_to_simdf4( c4 ) );   stbir__simdf_store( output4, o0 ); )
10042+      stbIF5( stbir__simdf_mult( o0, r0, stbir__if_simdf8_cast_to_simdf4( c5 ) );   stbir__simdf_store( output5, o0 ); )
10043+      stbIF6( stbir__simdf_mult( o0, r0, stbir__if_simdf8_cast_to_simdf4( c6 ) );   stbir__simdf_store( output6, o0 ); )
10044+      stbIF7( stbir__simdf_mult( o0, r0, stbir__if_simdf8_cast_to_simdf4( c7 ) );   stbir__simdf_store( output7, o0 ); )
10045+      #endif
10046+
10047+      input += 4;
10048+      stbIF0( output0 += 4; ) stbIF1( output1 += 4; ) stbIF2( output2 += 4; ) stbIF3( output3 += 4; ) stbIF4( output4 += 4; ) stbIF5( output5 += 4; ) stbIF6( output6 += 4; ) stbIF7( output7 += 4; )
10049+    }
10050+  }
10051+  #else
10052+  STBIR_NO_UNROLL_LOOP_START
10053+  while ( ( (char*)input_end - (char*) input ) >= 16 )
10054+  {
10055+    float r0, r1, r2, r3;
10056+    STBIR_NO_UNROLL(input);
10057+
10058+    r0 = input[0], r1 = input[1], r2 = input[2], r3 = input[3];
10059+
10060+    #ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
10061+    stbIF0( output0[0] += ( r0 * c0s ); output0[1] += ( r1 * c0s ); output0[2] += ( r2 * c0s ); output0[3] += ( r3 * c0s ); )
10062+    stbIF1( output1[0] += ( r0 * c1s ); output1[1] += ( r1 * c1s ); output1[2] += ( r2 * c1s ); output1[3] += ( r3 * c1s ); )
10063+    stbIF2( output2[0] += ( r0 * c2s ); output2[1] += ( r1 * c2s ); output2[2] += ( r2 * c2s ); output2[3] += ( r3 * c2s ); )
10064+    stbIF3( output3[0] += ( r0 * c3s ); output3[1] += ( r1 * c3s ); output3[2] += ( r2 * c3s ); output3[3] += ( r3 * c3s ); )
10065+    stbIF4( output4[0] += ( r0 * c4s ); output4[1] += ( r1 * c4s ); output4[2] += ( r2 * c4s ); output4[3] += ( r3 * c4s ); )
10066+    stbIF5( output5[0] += ( r0 * c5s ); output5[1] += ( r1 * c5s ); output5[2] += ( r2 * c5s ); output5[3] += ( r3 * c5s ); )
10067+    stbIF6( output6[0] += ( r0 * c6s ); output6[1] += ( r1 * c6s ); output6[2] += ( r2 * c6s ); output6[3] += ( r3 * c6s ); )
10068+    stbIF7( output7[0] += ( r0 * c7s ); output7[1] += ( r1 * c7s ); output7[2] += ( r2 * c7s ); output7[3] += ( r3 * c7s ); )
10069+    #else
10070+    stbIF0( output0[0]  = ( r0 * c0s ); output0[1]  = ( r1 * c0s ); output0[2]  = ( r2 * c0s ); output0[3]  = ( r3 * c0s ); )
10071+    stbIF1( output1[0]  = ( r0 * c1s ); output1[1]  = ( r1 * c1s ); output1[2]  = ( r2 * c1s ); output1[3]  = ( r3 * c1s ); )
10072+    stbIF2( output2[0]  = ( r0 * c2s ); output2[1]  = ( r1 * c2s ); output2[2]  = ( r2 * c2s ); output2[3]  = ( r3 * c2s ); )
10073+    stbIF3( output3[0]  = ( r0 * c3s ); output3[1]  = ( r1 * c3s ); output3[2]  = ( r2 * c3s ); output3[3]  = ( r3 * c3s ); )
10074+    stbIF4( output4[0]  = ( r0 * c4s ); output4[1]  = ( r1 * c4s ); output4[2]  = ( r2 * c4s ); output4[3]  = ( r3 * c4s ); )
10075+    stbIF5( output5[0]  = ( r0 * c5s ); output5[1]  = ( r1 * c5s ); output5[2]  = ( r2 * c5s ); output5[3]  = ( r3 * c5s ); )
10076+    stbIF6( output6[0]  = ( r0 * c6s ); output6[1]  = ( r1 * c6s ); output6[2]  = ( r2 * c6s ); output6[3]  = ( r3 * c6s ); )
10077+    stbIF7( output7[0]  = ( r0 * c7s ); output7[1]  = ( r1 * c7s ); output7[2]  = ( r2 * c7s ); output7[3]  = ( r3 * c7s ); )
10078+    #endif
10079+
10080+    input += 4;
10081+    stbIF0( output0 += 4; ) stbIF1( output1 += 4; ) stbIF2( output2 += 4; ) stbIF3( output3 += 4; ) stbIF4( output4 += 4; ) stbIF5( output5 += 4; ) stbIF6( output6 += 4; ) stbIF7( output7 += 4; )
10082+  }
10083+  #endif
10084+  STBIR_NO_UNROLL_LOOP_START
10085+  while ( input < input_end )
10086+  {
10087+    float r = input[0];
10088+    STBIR_NO_UNROLL(output0);
10089+
10090+    #ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
10091+    stbIF0( output0[0] += ( r * c0s ); )
10092+    stbIF1( output1[0] += ( r * c1s ); )
10093+    stbIF2( output2[0] += ( r * c2s ); )
10094+    stbIF3( output3[0] += ( r * c3s ); )
10095+    stbIF4( output4[0] += ( r * c4s ); )
10096+    stbIF5( output5[0] += ( r * c5s ); )
10097+    stbIF6( output6[0] += ( r * c6s ); )
10098+    stbIF7( output7[0] += ( r * c7s ); )
10099+    #else
10100+    stbIF0( output0[0]  = ( r * c0s ); )
10101+    stbIF1( output1[0]  = ( r * c1s ); )
10102+    stbIF2( output2[0]  = ( r * c2s ); )
10103+    stbIF3( output3[0]  = ( r * c3s ); )
10104+    stbIF4( output4[0]  = ( r * c4s ); )
10105+    stbIF5( output5[0]  = ( r * c5s ); )
10106+    stbIF6( output6[0]  = ( r * c6s ); )
10107+    stbIF7( output7[0]  = ( r * c7s ); )
10108+    #endif
10109+
10110+    ++input;
10111+    stbIF0( ++output0; ) stbIF1( ++output1; ) stbIF2( ++output2; ) stbIF3( ++output3; ) stbIF4( ++output4; ) stbIF5( ++output5; ) stbIF6( ++output6; ) stbIF7( ++output7; )
10112+  }
10113+}
10114+
10115+static void STBIR_chans( stbir__vertical_gather_with_,_coeffs)( float * outputp, float const * vertical_coefficients, float const ** inputs, float const * input0_end )
10116+{
10117+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = outputp;
10118+
10119+  stbIF0( float const * input0 = inputs[0]; float c0s = vertical_coefficients[0]; )
10120+  stbIF1( float const * input1 = inputs[1]; float c1s = vertical_coefficients[1]; )
10121+  stbIF2( float const * input2 = inputs[2]; float c2s = vertical_coefficients[2]; )
10122+  stbIF3( float const * input3 = inputs[3]; float c3s = vertical_coefficients[3]; )
10123+  stbIF4( float const * input4 = inputs[4]; float c4s = vertical_coefficients[4]; )
10124+  stbIF5( float const * input5 = inputs[5]; float c5s = vertical_coefficients[5]; )
10125+  stbIF6( float const * input6 = inputs[6]; float c6s = vertical_coefficients[6]; )
10126+  stbIF7( float const * input7 = inputs[7]; float c7s = vertical_coefficients[7]; )
10127+
10128+#if ( STBIR__vertical_channels == 1 ) && !defined(STB_IMAGE_RESIZE_VERTICAL_CONTINUE)
10129+  // check single channel one weight
10130+  if ( ( c0s >= (1.0f-0.000001f) ) && ( c0s <= (1.0f+0.000001f) ) )
10131+  {
10132+    STBIR_MEMCPY( output, input0, (char*)input0_end - (char*)input0 );
10133+    return;
10134+  }
10135+#endif
10136+
10137+  #ifdef STBIR_SIMD
10138+  {
10139+    stbIF0(stbir__simdfX c0 = stbir__simdf_frepX( c0s ); )
10140+    stbIF1(stbir__simdfX c1 = stbir__simdf_frepX( c1s ); )
10141+    stbIF2(stbir__simdfX c2 = stbir__simdf_frepX( c2s ); )
10142+    stbIF3(stbir__simdfX c3 = stbir__simdf_frepX( c3s ); )
10143+    stbIF4(stbir__simdfX c4 = stbir__simdf_frepX( c4s ); )
10144+    stbIF5(stbir__simdfX c5 = stbir__simdf_frepX( c5s ); )
10145+    stbIF6(stbir__simdfX c6 = stbir__simdf_frepX( c6s ); )
10146+    stbIF7(stbir__simdfX c7 = stbir__simdf_frepX( c7s ); )
10147+
10148+    STBIR_SIMD_NO_UNROLL_LOOP_START
10149+    while ( ( (char*)input0_end - (char*) input0 ) >= (16*stbir__simdfX_float_count) )
10150+    {
10151+      stbir__simdfX o0, o1, o2, o3, r0, r1, r2, r3;
10152+      STBIR_SIMD_NO_UNROLL(output);
10153+
10154+      // prefetch four loop iterations ahead (doesn't affect much for small resizes, but helps with big ones)
10155+      stbIF0( stbir__prefetch( input0 + (16*stbir__simdfX_float_count) ); )
10156+      stbIF1( stbir__prefetch( input1 + (16*stbir__simdfX_float_count) ); )
10157+      stbIF2( stbir__prefetch( input2 + (16*stbir__simdfX_float_count) ); )
10158+      stbIF3( stbir__prefetch( input3 + (16*stbir__simdfX_float_count) ); )
10159+      stbIF4( stbir__prefetch( input4 + (16*stbir__simdfX_float_count) ); )
10160+      stbIF5( stbir__prefetch( input5 + (16*stbir__simdfX_float_count) ); )
10161+      stbIF6( stbir__prefetch( input6 + (16*stbir__simdfX_float_count) ); )
10162+      stbIF7( stbir__prefetch( input7 + (16*stbir__simdfX_float_count) ); )
10163+
10164+      #ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
10165+      stbIF0( stbir__simdfX_load( o0, output );      stbir__simdfX_load( o1, output+stbir__simdfX_float_count );   stbir__simdfX_load( o2, output+(2*stbir__simdfX_float_count) );   stbir__simdfX_load( o3, output+(3*stbir__simdfX_float_count) );
10166+              stbir__simdfX_load( r0, input0 );      stbir__simdfX_load( r1, input0+stbir__simdfX_float_count );   stbir__simdfX_load( r2, input0+(2*stbir__simdfX_float_count) );   stbir__simdfX_load( r3, input0+(3*stbir__simdfX_float_count) );
10167+              stbir__simdfX_madd( o0, o0, r0, c0 );  stbir__simdfX_madd( o1, o1, r1, c0 );                         stbir__simdfX_madd( o2, o2, r2, c0 );                             stbir__simdfX_madd( o3, o3, r3, c0 ); )
10168+      #else
10169+      stbIF0( stbir__simdfX_load( r0, input0 );      stbir__simdfX_load( r1, input0+stbir__simdfX_float_count );   stbir__simdfX_load( r2, input0+(2*stbir__simdfX_float_count) );   stbir__simdfX_load( r3, input0+(3*stbir__simdfX_float_count) );
10170+              stbir__simdfX_mult( o0, r0, c0 );      stbir__simdfX_mult( o1, r1, c0 );                             stbir__simdfX_mult( o2, r2, c0 );                                 stbir__simdfX_mult( o3, r3, c0 );  )
10171+      #endif
10172+
10173+      stbIF1( stbir__simdfX_load( r0, input1 );      stbir__simdfX_load( r1, input1+stbir__simdfX_float_count );   stbir__simdfX_load( r2, input1+(2*stbir__simdfX_float_count) );   stbir__simdfX_load( r3, input1+(3*stbir__simdfX_float_count) );
10174+              stbir__simdfX_madd( o0, o0, r0, c1 );  stbir__simdfX_madd( o1, o1, r1, c1 );                         stbir__simdfX_madd( o2, o2, r2, c1 );                             stbir__simdfX_madd( o3, o3, r3, c1 ); )
10175+      stbIF2( stbir__simdfX_load( r0, input2 );      stbir__simdfX_load( r1, input2+stbir__simdfX_float_count );   stbir__simdfX_load( r2, input2+(2*stbir__simdfX_float_count) );   stbir__simdfX_load( r3, input2+(3*stbir__simdfX_float_count) );
10176+              stbir__simdfX_madd( o0, o0, r0, c2 );  stbir__simdfX_madd( o1, o1, r1, c2 );                         stbir__simdfX_madd( o2, o2, r2, c2 );                             stbir__simdfX_madd( o3, o3, r3, c2 ); )
10177+      stbIF3( stbir__simdfX_load( r0, input3 );      stbir__simdfX_load( r1, input3+stbir__simdfX_float_count );   stbir__simdfX_load( r2, input3+(2*stbir__simdfX_float_count) );   stbir__simdfX_load( r3, input3+(3*stbir__simdfX_float_count) );
10178+              stbir__simdfX_madd( o0, o0, r0, c3 );  stbir__simdfX_madd( o1, o1, r1, c3 );                         stbir__simdfX_madd( o2, o2, r2, c3 );                             stbir__simdfX_madd( o3, o3, r3, c3 ); )
10179+      stbIF4( stbir__simdfX_load( r0, input4 );      stbir__simdfX_load( r1, input4+stbir__simdfX_float_count );   stbir__simdfX_load( r2, input4+(2*stbir__simdfX_float_count) );   stbir__simdfX_load( r3, input4+(3*stbir__simdfX_float_count) );
10180+              stbir__simdfX_madd( o0, o0, r0, c4 );  stbir__simdfX_madd( o1, o1, r1, c4 );                         stbir__simdfX_madd( o2, o2, r2, c4 );                             stbir__simdfX_madd( o3, o3, r3, c4 ); )
10181+      stbIF5( stbir__simdfX_load( r0, input5 );      stbir__simdfX_load( r1, input5+stbir__simdfX_float_count );   stbir__simdfX_load( r2, input5+(2*stbir__simdfX_float_count) );   stbir__simdfX_load( r3, input5+(3*stbir__simdfX_float_count) );
10182+              stbir__simdfX_madd( o0, o0, r0, c5 );  stbir__simdfX_madd( o1, o1, r1, c5 );                         stbir__simdfX_madd( o2, o2, r2, c5 );                             stbir__simdfX_madd( o3, o3, r3, c5 ); )
10183+      stbIF6( stbir__simdfX_load( r0, input6 );      stbir__simdfX_load( r1, input6+stbir__simdfX_float_count );   stbir__simdfX_load( r2, input6+(2*stbir__simdfX_float_count) );   stbir__simdfX_load( r3, input6+(3*stbir__simdfX_float_count) );
10184+              stbir__simdfX_madd( o0, o0, r0, c6 );  stbir__simdfX_madd( o1, o1, r1, c6 );                         stbir__simdfX_madd( o2, o2, r2, c6 );                             stbir__simdfX_madd( o3, o3, r3, c6 ); )
10185+      stbIF7( stbir__simdfX_load( r0, input7 );      stbir__simdfX_load( r1, input7+stbir__simdfX_float_count );   stbir__simdfX_load( r2, input7+(2*stbir__simdfX_float_count) );   stbir__simdfX_load( r3, input7+(3*stbir__simdfX_float_count) );
10186+              stbir__simdfX_madd( o0, o0, r0, c7 );  stbir__simdfX_madd( o1, o1, r1, c7 );                         stbir__simdfX_madd( o2, o2, r2, c7 );                             stbir__simdfX_madd( o3, o3, r3, c7 ); )
10187+
10188+      stbir__simdfX_store( output, o0 );             stbir__simdfX_store( output+stbir__simdfX_float_count, o1 );  stbir__simdfX_store( output+(2*stbir__simdfX_float_count), o2 );  stbir__simdfX_store( output+(3*stbir__simdfX_float_count), o3 );
10189+      output += (4*stbir__simdfX_float_count);
10190+      stbIF0( input0 += (4*stbir__simdfX_float_count); ) stbIF1( input1 += (4*stbir__simdfX_float_count); ) stbIF2( input2 += (4*stbir__simdfX_float_count); ) stbIF3( input3 += (4*stbir__simdfX_float_count); ) stbIF4( input4 += (4*stbir__simdfX_float_count); ) stbIF5( input5 += (4*stbir__simdfX_float_count); ) stbIF6( input6 += (4*stbir__simdfX_float_count); ) stbIF7( input7 += (4*stbir__simdfX_float_count); )
10191+    }
10192+
10193+    STBIR_SIMD_NO_UNROLL_LOOP_START
10194+    while ( ( (char*)input0_end - (char*) input0 ) >= 16 )
10195+    {
10196+      stbir__simdf o0, r0;
10197+      STBIR_SIMD_NO_UNROLL(output);
10198+
10199+      #ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
10200+      stbIF0( stbir__simdf_load( o0, output );   stbir__simdf_load( r0, input0 ); stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c0 ) ); )
10201+      #else
10202+      stbIF0( stbir__simdf_load( r0, input0 );  stbir__simdf_mult( o0, r0, stbir__if_simdf8_cast_to_simdf4( c0 ) ); )
10203+      #endif
10204+      stbIF1( stbir__simdf_load( r0, input1 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c1 ) ); )
10205+      stbIF2( stbir__simdf_load( r0, input2 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c2 ) ); )
10206+      stbIF3( stbir__simdf_load( r0, input3 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c3 ) ); )
10207+      stbIF4( stbir__simdf_load( r0, input4 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c4 ) ); )
10208+      stbIF5( stbir__simdf_load( r0, input5 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c5 ) ); )
10209+      stbIF6( stbir__simdf_load( r0, input6 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c6 ) ); )
10210+      stbIF7( stbir__simdf_load( r0, input7 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c7 ) ); )
10211+
10212+      stbir__simdf_store( output, o0 );
10213+      output += 4;
10214+      stbIF0( input0 += 4; ) stbIF1( input1 += 4; ) stbIF2( input2 += 4; ) stbIF3( input3 += 4; ) stbIF4( input4 += 4; ) stbIF5( input5 += 4; ) stbIF6( input6 += 4; ) stbIF7( input7 += 4; )
10215+    }
10216+  }
10217+  #else
10218+  STBIR_NO_UNROLL_LOOP_START
10219+  while ( ( (char*)input0_end - (char*) input0 ) >= 16 )
10220+  {
10221+    float o0, o1, o2, o3;
10222+    STBIR_NO_UNROLL(output);
10223+    #ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
10224+    stbIF0( o0 = output[0] + input0[0] * c0s; o1 = output[1] + input0[1] * c0s; o2 = output[2] + input0[2] * c0s; o3 = output[3] + input0[3] * c0s; )
10225+    #else
10226+    stbIF0( o0  = input0[0] * c0s; o1  = input0[1] * c0s; o2  = input0[2] * c0s; o3  = input0[3] * c0s; )
10227+    #endif
10228+    stbIF1( o0 += input1[0] * c1s; o1 += input1[1] * c1s; o2 += input1[2] * c1s; o3 += input1[3] * c1s; )
10229+    stbIF2( o0 += input2[0] * c2s; o1 += input2[1] * c2s; o2 += input2[2] * c2s; o3 += input2[3] * c2s; )
10230+    stbIF3( o0 += input3[0] * c3s; o1 += input3[1] * c3s; o2 += input3[2] * c3s; o3 += input3[3] * c3s; )
10231+    stbIF4( o0 += input4[0] * c4s; o1 += input4[1] * c4s; o2 += input4[2] * c4s; o3 += input4[3] * c4s; )
10232+    stbIF5( o0 += input5[0] * c5s; o1 += input5[1] * c5s; o2 += input5[2] * c5s; o3 += input5[3] * c5s; )
10233+    stbIF6( o0 += input6[0] * c6s; o1 += input6[1] * c6s; o2 += input6[2] * c6s; o3 += input6[3] * c6s; )
10234+    stbIF7( o0 += input7[0] * c7s; o1 += input7[1] * c7s; o2 += input7[2] * c7s; o3 += input7[3] * c7s; )
10235+    output[0] = o0; output[1] = o1; output[2] = o2; output[3] = o3;
10236+    output += 4;
10237+    stbIF0( input0 += 4; ) stbIF1( input1 += 4; ) stbIF2( input2 += 4; ) stbIF3( input3 += 4; ) stbIF4( input4 += 4; ) stbIF5( input5 += 4; ) stbIF6( input6 += 4; ) stbIF7( input7 += 4; )
10238+  }
10239+  #endif
10240+  STBIR_NO_UNROLL_LOOP_START
10241+  while ( input0 < input0_end )
10242+  {
10243+    float o0;
10244+    STBIR_NO_UNROLL(output);
10245+    #ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
10246+    stbIF0( o0 = output[0] + input0[0] * c0s; )
10247+    #else
10248+    stbIF0( o0  = input0[0] * c0s; )
10249+    #endif
10250+    stbIF1( o0 += input1[0] * c1s; )
10251+    stbIF2( o0 += input2[0] * c2s; )
10252+    stbIF3( o0 += input3[0] * c3s; )
10253+    stbIF4( o0 += input4[0] * c4s; )
10254+    stbIF5( o0 += input5[0] * c5s; )
10255+    stbIF6( o0 += input6[0] * c6s; )
10256+    stbIF7( o0 += input7[0] * c7s; )
10257+    output[0] = o0;
10258+    ++output;
10259+    stbIF0( ++input0; ) stbIF1( ++input1; ) stbIF2( ++input2; ) stbIF3( ++input3; ) stbIF4( ++input4; ) stbIF5( ++input5; ) stbIF6( ++input6; ) stbIF7( ++input7; )
10260+  }
10261+}
10262+
10263+#undef stbIF0
10264+#undef stbIF1
10265+#undef stbIF2
10266+#undef stbIF3
10267+#undef stbIF4
10268+#undef stbIF5
10269+#undef stbIF6
10270+#undef stbIF7
10271+#undef STB_IMAGE_RESIZE_DO_VERTICALS
10272+#undef STBIR__vertical_channels
10273+#undef STB_IMAGE_RESIZE_DO_HORIZONTALS
10274+#undef STBIR_strs_join24
10275+#undef STBIR_strs_join14
10276+#undef STBIR_chans
10277+#ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
10278+#undef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
10279+#endif
10280+
10281+#else // !STB_IMAGE_RESIZE_DO_VERTICALS
10282+
10283+#define STBIR_chans( start, end ) STBIR_strs_join1(start,STBIR__horizontal_channels,end)
10284+
10285+#ifndef stbir__2_coeff_only
10286+#define stbir__2_coeff_only()             \
10287+    stbir__1_coeff_only();                \
10288+    stbir__1_coeff_remnant(1);
10289+#endif
10290+
10291+#ifndef stbir__2_coeff_remnant
10292+#define stbir__2_coeff_remnant( ofs )     \
10293+    stbir__1_coeff_remnant(ofs);          \
10294+    stbir__1_coeff_remnant((ofs)+1);
10295+#endif
10296+
10297+#ifndef stbir__3_coeff_only
10298+#define stbir__3_coeff_only()             \
10299+    stbir__2_coeff_only();                \
10300+    stbir__1_coeff_remnant(2);
10301+#endif
10302+
10303+#ifndef stbir__3_coeff_remnant
10304+#define stbir__3_coeff_remnant( ofs )     \
10305+    stbir__2_coeff_remnant(ofs);          \
10306+    stbir__1_coeff_remnant((ofs)+2);
10307+#endif
10308+
10309+#ifndef stbir__3_coeff_setup
10310+#define stbir__3_coeff_setup()
10311+#endif
10312+
10313+#ifndef stbir__4_coeff_start
10314+#define stbir__4_coeff_start()            \
10315+    stbir__2_coeff_only();                \
10316+    stbir__2_coeff_remnant(2);
10317+#endif
10318+
10319+#ifndef stbir__4_coeff_continue_from_4
10320+#define stbir__4_coeff_continue_from_4( ofs )     \
10321+    stbir__2_coeff_remnant(ofs);                  \
10322+    stbir__2_coeff_remnant((ofs)+2);
10323+#endif
10324+
10325+#ifndef stbir__store_output_tiny
10326+#define stbir__store_output_tiny stbir__store_output
10327+#endif
10328+
10329+static void STBIR_chans( stbir__horizontal_gather_,_channels_with_1_coeff)( float * output_buffer, unsigned int output_sub_size, float const * decode_buffer, stbir__contributors const * horizontal_contributors, float const * horizontal_coefficients, int coefficient_width )
10330+{
10331+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10332+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10333+  STBIR_SIMD_NO_UNROLL_LOOP_START
10334+  do {
10335+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10336+    float const * hc = horizontal_coefficients;
10337+    stbir__1_coeff_only();
10338+    stbir__store_output_tiny();
10339+  } while ( output < output_end );
10340+}
10341+
10342+static void STBIR_chans( stbir__horizontal_gather_,_channels_with_2_coeffs)( float * output_buffer, unsigned int output_sub_size, float const * decode_buffer, stbir__contributors const * horizontal_contributors, float const * horizontal_coefficients, int coefficient_width )
10343+{
10344+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10345+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10346+  STBIR_SIMD_NO_UNROLL_LOOP_START
10347+  do {
10348+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10349+    float const * hc = horizontal_coefficients;
10350+    stbir__2_coeff_only();
10351+    stbir__store_output_tiny();
10352+  } while ( output < output_end );
10353+}
10354+
10355+static void STBIR_chans( stbir__horizontal_gather_,_channels_with_3_coeffs)( float * output_buffer, unsigned int output_sub_size, float const * decode_buffer, stbir__contributors const * horizontal_contributors, float const * horizontal_coefficients, int coefficient_width )
10356+{
10357+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10358+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10359+  STBIR_SIMD_NO_UNROLL_LOOP_START
10360+  do {
10361+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10362+    float const * hc = horizontal_coefficients;
10363+    stbir__3_coeff_only();
10364+    stbir__store_output_tiny();
10365+  } while ( output < output_end );
10366+}
10367+
10368+static void STBIR_chans( stbir__horizontal_gather_,_channels_with_4_coeffs)( float * output_buffer, unsigned int output_sub_size, float const * decode_buffer, stbir__contributors const * horizontal_contributors, float const * horizontal_coefficients, int coefficient_width )
10369+{
10370+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10371+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10372+  STBIR_SIMD_NO_UNROLL_LOOP_START
10373+  do {
10374+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10375+    float const * hc = horizontal_coefficients;
10376+    stbir__4_coeff_start();
10377+    stbir__store_output();
10378+  } while ( output < output_end );
10379+}
10380+
10381+static void STBIR_chans( stbir__horizontal_gather_,_channels_with_5_coeffs)( float * output_buffer, unsigned int output_sub_size, float const * decode_buffer, stbir__contributors const * horizontal_contributors, float const * horizontal_coefficients, int coefficient_width )
10382+{
10383+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10384+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10385+  STBIR_SIMD_NO_UNROLL_LOOP_START
10386+  do {
10387+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10388+    float const * hc = horizontal_coefficients;
10389+    stbir__4_coeff_start();
10390+    stbir__1_coeff_remnant(4);
10391+    stbir__store_output();
10392+  } while ( output < output_end );
10393+}
10394+
10395+static void STBIR_chans( stbir__horizontal_gather_,_channels_with_6_coeffs)( float * output_buffer, unsigned int output_sub_size, float const * decode_buffer, stbir__contributors const * horizontal_contributors, float const * horizontal_coefficients, int coefficient_width )
10396+{
10397+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10398+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10399+  STBIR_SIMD_NO_UNROLL_LOOP_START
10400+  do {
10401+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10402+    float const * hc = horizontal_coefficients;
10403+    stbir__4_coeff_start();
10404+    stbir__2_coeff_remnant(4);
10405+    stbir__store_output();
10406+  } while ( output < output_end );
10407+}
10408+
10409+static void STBIR_chans( stbir__horizontal_gather_,_channels_with_7_coeffs)( float * output_buffer, unsigned int output_sub_size, float const * decode_buffer, stbir__contributors const * horizontal_contributors, float const * horizontal_coefficients, int coefficient_width )
10410+{
10411+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10412+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10413+  stbir__3_coeff_setup();
10414+  STBIR_SIMD_NO_UNROLL_LOOP_START
10415+  do {
10416+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10417+    float const * hc = horizontal_coefficients;
10418+
10419+    stbir__4_coeff_start();
10420+    stbir__3_coeff_remnant(4);
10421+    stbir__store_output();
10422+  } while ( output < output_end );
10423+}
10424+
10425+static void STBIR_chans( stbir__horizontal_gather_,_channels_with_8_coeffs)( float * output_buffer, unsigned int output_sub_size, float const * decode_buffer, stbir__contributors const * horizontal_contributors, float const * horizontal_coefficients, int coefficient_width )
10426+{
10427+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10428+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10429+  STBIR_SIMD_NO_UNROLL_LOOP_START
10430+  do {
10431+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10432+    float const * hc = horizontal_coefficients;
10433+    stbir__4_coeff_start();
10434+    stbir__4_coeff_continue_from_4(4);
10435+    stbir__store_output();
10436+  } while ( output < output_end );
10437+}
10438+
10439+static void STBIR_chans( stbir__horizontal_gather_,_channels_with_9_coeffs)( float * output_buffer, unsigned int output_sub_size, float const * decode_buffer, stbir__contributors const * horizontal_contributors, float const * horizontal_coefficients, int coefficient_width )
10440+{
10441+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10442+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10443+  STBIR_SIMD_NO_UNROLL_LOOP_START
10444+  do {
10445+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10446+    float const * hc = horizontal_coefficients;
10447+    stbir__4_coeff_start();
10448+    stbir__4_coeff_continue_from_4(4);
10449+    stbir__1_coeff_remnant(8);
10450+    stbir__store_output();
10451+  } while ( output < output_end );
10452+}
10453+
10454+static void STBIR_chans( stbir__horizontal_gather_,_channels_with_10_coeffs)( float * output_buffer, unsigned int output_sub_size, float const * decode_buffer, stbir__contributors const * horizontal_contributors, float const * horizontal_coefficients, int coefficient_width )
10455+{
10456+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10457+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10458+  STBIR_SIMD_NO_UNROLL_LOOP_START
10459+  do {
10460+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10461+    float const * hc = horizontal_coefficients;
10462+    stbir__4_coeff_start();
10463+    stbir__4_coeff_continue_from_4(4);
10464+    stbir__2_coeff_remnant(8);
10465+    stbir__store_output();
10466+  } while ( output < output_end );
10467+}
10468+
10469+static void STBIR_chans( stbir__horizontal_gather_,_channels_with_11_coeffs)( float * output_buffer, unsigned int output_sub_size, float const * decode_buffer, stbir__contributors const * horizontal_contributors, float const * horizontal_coefficients, int coefficient_width )
10470+{
10471+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10472+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10473+  stbir__3_coeff_setup();
10474+  STBIR_SIMD_NO_UNROLL_LOOP_START
10475+  do {
10476+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10477+    float const * hc = horizontal_coefficients;
10478+    stbir__4_coeff_start();
10479+    stbir__4_coeff_continue_from_4(4);
10480+    stbir__3_coeff_remnant(8);
10481+    stbir__store_output();
10482+  } while ( output < output_end );
10483+}
10484+
10485+static void STBIR_chans( stbir__horizontal_gather_,_channels_with_12_coeffs)( float * output_buffer, unsigned int output_sub_size, float const * decode_buffer, stbir__contributors const * horizontal_contributors, float const * horizontal_coefficients, int coefficient_width )
10486+{
10487+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10488+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10489+  STBIR_SIMD_NO_UNROLL_LOOP_START
10490+  do {
10491+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10492+    float const * hc = horizontal_coefficients;
10493+    stbir__4_coeff_start();
10494+    stbir__4_coeff_continue_from_4(4);
10495+    stbir__4_coeff_continue_from_4(8);
10496+    stbir__store_output();
10497+  } while ( output < output_end );
10498+}
10499+
10500+static void STBIR_chans( stbir__horizontal_gather_,_channels_with_n_coeffs_mod0 )( float * output_buffer, unsigned int output_sub_size, float const * decode_buffer, stbir__contributors const * horizontal_contributors, float const * horizontal_coefficients, int coefficient_width )
10501+{
10502+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10503+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10504+  STBIR_SIMD_NO_UNROLL_LOOP_START
10505+  do {
10506+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10507+    int n = ( ( horizontal_contributors->n1 - horizontal_contributors->n0 + 1 ) - 4 + 3 ) >> 2;
10508+    float const * hc = horizontal_coefficients;
10509+
10510+    stbir__4_coeff_start();
10511+    STBIR_SIMD_NO_UNROLL_LOOP_START
10512+    do {
10513+      hc += 4;
10514+      decode += STBIR__horizontal_channels * 4;
10515+      stbir__4_coeff_continue_from_4( 0 );
10516+      --n;
10517+    } while ( n > 0 );
10518+    stbir__store_output();
10519+  } while ( output < output_end );
10520+}
10521+
10522+static void STBIR_chans( stbir__horizontal_gather_,_channels_with_n_coeffs_mod1 )( float * output_buffer, unsigned int output_sub_size, float const * decode_buffer, stbir__contributors const * horizontal_contributors, float const * horizontal_coefficients, int coefficient_width )
10523+{
10524+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10525+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10526+  STBIR_SIMD_NO_UNROLL_LOOP_START
10527+  do {
10528+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10529+    int n = ( ( horizontal_contributors->n1 - horizontal_contributors->n0 + 1 ) - 5 + 3 ) >> 2;
10530+    float const * hc = horizontal_coefficients;
10531+
10532+    stbir__4_coeff_start();
10533+    STBIR_SIMD_NO_UNROLL_LOOP_START
10534+    do {
10535+      hc += 4;
10536+      decode += STBIR__horizontal_channels * 4;
10537+      stbir__4_coeff_continue_from_4( 0 );
10538+      --n;
10539+    } while ( n > 0 );
10540+    stbir__1_coeff_remnant( 4 );
10541+    stbir__store_output();
10542+  } while ( output < output_end );
10543+}
10544+
10545+static void STBIR_chans( stbir__horizontal_gather_,_channels_with_n_coeffs_mod2 )( float * output_buffer, unsigned int output_sub_size, float const * decode_buffer, stbir__contributors const * horizontal_contributors, float const * horizontal_coefficients, int coefficient_width )
10546+{
10547+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10548+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10549+  STBIR_SIMD_NO_UNROLL_LOOP_START
10550+  do {
10551+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10552+    int n = ( ( horizontal_contributors->n1 - horizontal_contributors->n0 + 1 ) - 6 + 3 ) >> 2;
10553+    float const * hc = horizontal_coefficients;
10554+
10555+    stbir__4_coeff_start();
10556+    STBIR_SIMD_NO_UNROLL_LOOP_START
10557+    do {
10558+      hc += 4;
10559+      decode += STBIR__horizontal_channels * 4;
10560+      stbir__4_coeff_continue_from_4( 0 );
10561+      --n;
10562+    } while ( n > 0 );
10563+    stbir__2_coeff_remnant( 4 );
10564+
10565+    stbir__store_output();
10566+  } while ( output < output_end );
10567+}
10568+
10569+static void STBIR_chans( stbir__horizontal_gather_,_channels_with_n_coeffs_mod3 )( float * output_buffer, unsigned int output_sub_size, float const * decode_buffer, stbir__contributors const * horizontal_contributors, float const * horizontal_coefficients, int coefficient_width )
10570+{
10571+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10572+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10573+  stbir__3_coeff_setup();
10574+  STBIR_SIMD_NO_UNROLL_LOOP_START
10575+  do {
10576+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10577+    int n = ( ( horizontal_contributors->n1 - horizontal_contributors->n0 + 1 ) - 7 + 3 ) >> 2;
10578+    float const * hc = horizontal_coefficients;
10579+
10580+    stbir__4_coeff_start();
10581+    STBIR_SIMD_NO_UNROLL_LOOP_START
10582+    do {
10583+      hc += 4;
10584+      decode += STBIR__horizontal_channels * 4;
10585+      stbir__4_coeff_continue_from_4( 0 );
10586+      --n;
10587+    } while ( n > 0 );
10588+    stbir__3_coeff_remnant( 4 );
10589+
10590+    stbir__store_output();
10591+  } while ( output < output_end );
10592+}
10593+
10594+static stbir__horizontal_gather_channels_func * STBIR_chans(stbir__horizontal_gather_,_channels_with_n_coeffs_funcs)[4]=
10595+{
10596+  STBIR_chans(stbir__horizontal_gather_,_channels_with_n_coeffs_mod0),
10597+  STBIR_chans(stbir__horizontal_gather_,_channels_with_n_coeffs_mod1),
10598+  STBIR_chans(stbir__horizontal_gather_,_channels_with_n_coeffs_mod2),
10599+  STBIR_chans(stbir__horizontal_gather_,_channels_with_n_coeffs_mod3),
10600+};
10601+
10602+static stbir__horizontal_gather_channels_func * STBIR_chans(stbir__horizontal_gather_,_channels_funcs)[12]=
10603+{
10604+  STBIR_chans(stbir__horizontal_gather_,_channels_with_1_coeff),
10605+  STBIR_chans(stbir__horizontal_gather_,_channels_with_2_coeffs),
10606+  STBIR_chans(stbir__horizontal_gather_,_channels_with_3_coeffs),
10607+  STBIR_chans(stbir__horizontal_gather_,_channels_with_4_coeffs),
10608+  STBIR_chans(stbir__horizontal_gather_,_channels_with_5_coeffs),
10609+  STBIR_chans(stbir__horizontal_gather_,_channels_with_6_coeffs),
10610+  STBIR_chans(stbir__horizontal_gather_,_channels_with_7_coeffs),
10611+  STBIR_chans(stbir__horizontal_gather_,_channels_with_8_coeffs),
10612+  STBIR_chans(stbir__horizontal_gather_,_channels_with_9_coeffs),
10613+  STBIR_chans(stbir__horizontal_gather_,_channels_with_10_coeffs),
10614+  STBIR_chans(stbir__horizontal_gather_,_channels_with_11_coeffs),
10615+  STBIR_chans(stbir__horizontal_gather_,_channels_with_12_coeffs),
10616+};
10617+
10618+#undef STBIR__horizontal_channels
10619+#undef STB_IMAGE_RESIZE_DO_HORIZONTALS
10620+#undef stbir__1_coeff_only
10621+#undef stbir__1_coeff_remnant
10622+#undef stbir__2_coeff_only
10623+#undef stbir__2_coeff_remnant
10624+#undef stbir__3_coeff_only
10625+#undef stbir__3_coeff_remnant
10626+#undef stbir__3_coeff_setup
10627+#undef stbir__4_coeff_start
10628+#undef stbir__4_coeff_continue_from_4
10629+#undef stbir__store_output
10630+#undef stbir__store_output_tiny
10631+#undef STBIR_chans
10632+
10633+#endif  // HORIZONALS
10634+
10635+#undef STBIR_strs_join2
10636+#undef STBIR_strs_join1
10637+
10638+#endif // STB_IMAGE_RESIZE_DO_HORIZONTALS/VERTICALS/CODERS
10639+
10640+/*
10641+------------------------------------------------------------------------------
10642+This software is available under 2 licenses -- choose whichever you prefer.
10643+------------------------------------------------------------------------------
10644+ALTERNATIVE A - MIT License
10645+Copyright (c) 2017 Sean Barrett
10646+Permission is hereby granted, free of charge, to any person obtaining a copy of
10647+this software and associated documentation files (the "Software"), to deal in
10648+the Software without restriction, including without limitation the rights to
10649+use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
10650+of the Software, and to permit persons to whom the Software is furnished to do
10651+so, subject to the following conditions:
10652+The above copyright notice and this permission notice shall be included in all
10653+copies or substantial portions of the Software.
10654+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
10655+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
10656+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
10657+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
10658+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
10659+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
10660+SOFTWARE.
10661+------------------------------------------------------------------------------
10662+ALTERNATIVE B - Public Domain (www.unlicense.org)
10663+This is free and unencumbered software released into the public domain.
10664+Anyone is free to copy, modify, publish, use, compile, sell, or distribute this
10665+software, either in source code form or as a compiled binary, for any purpose,
10666+commercial or non-commercial, and by any means.
10667+In jurisdictions that recognize copyright laws, the author or authors of this
10668+software dedicate any and all copyright interest in the software to the public
10669+domain. We make this dedication for the benefit of the public at large and to
10670+the detriment of our heirs and successors. We intend this dedication to be an
10671+overt act of relinquishment in perpetuity of all present and future rights to
10672+this software under copyright law.
10673+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
10674+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
10675+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
10676+AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
10677+ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
10678+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
10679+------------------------------------------------------------------------------
10680+*/
+51, -0
 1@@ -0,0 +1,51 @@
 2+#ifndef STBI_ALLOC_H
 3+#define STBI_ALLOC_H
 4+
 5+#include <sys/mman.h>
 6+#include <unistd.h>
 7+#include <string.h>
 8+#include <stdint.h>
 9+
10+#define STBI_MALLOC(sz)  stbi_mmap_alloc(sz)
11+#define STBI_FREE(p)     stbi_mmap_free(p)
12+#define STBI_REALLOC(p,sz) stbi_mmap_realloc(p,sz)
13+
14+static void *stbi_mmap_alloc(size_t sz)
15+{
16+	if (sz == 0) return NULL;
17+	long page = sysconf(_SC_PAGESIZE);
18+	size_t mapsz = ((sz + page) & ~(page - 1)) + page;
19+	void *p = mmap(NULL, mapsz, PROT_READ|PROT_WRITE, MAP_PRIVATE|MAP_ANONYMOUS, -1, 0);
20+	if (p == MAP_FAILED) return NULL;
21+	*(size_t*)p = mapsz;
22+	return (char*)p + page;
23+}
24+
25+static void stbi_mmap_free(void *p)
26+{
27+	if (!p) return;
28+	long page = sysconf(_SC_PAGESIZE);
29+	void *base = (char*)p - page;
30+	size_t mapsz = *(size_t*)base;
31+	munmap(base, mapsz);
32+}
33+
34+static void *stbi_mmap_realloc(void *p, size_t newsz)
35+{
36+	if (!p) return stbi_mmap_alloc(newsz);
37+	if (newsz == 0) {
38+		stbi_mmap_free(p);
39+		return NULL;
40+	}
41+	long page = sysconf(_SC_PAGESIZE);
42+	void *base = (char*)p - page;
43+	size_t old_mapsz = *(size_t*)base;
44+	size_t oldsz = old_mapsz - page;
45+	void *newp = stbi_mmap_alloc(newsz);
46+	if (!newp) return p;
47+	memcpy(newp, p, oldsz < newsz ? oldsz : newsz);
48+	munmap(base, old_mapsz);
49+	return newp;
50+}
51+
52+#endif