commit 2f6ca91

shrub  ·  2025-12-20 20:40:26 +0000 UTC
parent 7c35d5c
this adds wallpaper scaling to the initial wallpaper implementation, so
the wallpaper scales to screen size.
2 files changed,  +10681, -1
+30, -1
 1@@ -5,11 +5,15 @@
 2 #define STBI_NO_HDR
 3 #include "../stb/stb_image.h"
 4 
 5+#define STB_IMAGE_RESIZE_IMPLEMENTATION
 6+#include "../stb/stb_image_resize2.h"
 7+
 8 #include "swc.h"
 9 #include "internal.h"
10 #include "drm.h"
11 #include "util.h"
12 #include "shm.h"
13+#include "screen.h"
14 
15 unsigned char *wallpaper = NULL;
16 struct wld_buffer *wallbuf = NULL;
17@@ -20,8 +24,33 @@ EXPORT void
18 swc_wallpaper_init(char* path)
19 {
20 	int width, height, chan;
21+	unsigned char *loaded;
22+	struct screen *screen;
23+	int target_width = 0, target_height = 0;
24+
25+	loaded = stbi_load(path, &width, &height, &chan, 4);
26+	if (!loaded)
27+		return;
28 
29-	wallpaper = stbi_load(path, &width, &height, &chan, 4);
30+	/* get screen dimensions */
31+	wl_list_for_each(screen, &swc.screens, link) {
32+		target_width = screen->base.geometry.width;
33+		target_height = screen->base.geometry.height;
34+		break;
35+	}
36+
37+	/* If we have a screen and dimensions wrong  scale  */
38+	if (target_width > 0 && target_height > 0 &&
39+	    (width != target_width || height != target_height)) {
40+		wallpaper = stbir_resize_uint8_srgb(loaded, width, height, 0,
41+		                                     NULL, target_width, target_height, 0,
42+		                                     STBIR_RGBA);
43+		stbi_image_free(loaded);
44+		width = target_width;
45+		height = target_height;
46+	} else {
47+		wallpaper = loaded;
48+	}
49 
50 	/* swap color channels to be compatible */
51 	for(int i = 0; i < width * height; i++) {
+10651, -0
    1@@ -0,0 +1,10651 @@
    2+/* stb_image_resize2 - v2.17 - 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_RX, or STBIR_XR 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_RGBA_PM, STBIR_RGBA_PM,
  151+         STBIR_RGBA_PM, STBIR_RX_PM or STBIR_XR_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+      2.17 (2025-10-25) silly format bug in easy-to-use APIs.
  334+      2.16 (2025-10-21) fixed the easy-to-use APIs to allow inverted bitmaps (negative
  335+                          strides), fix vertical filter kernel callback, fix threaded
  336+                          gather buffer priming (and assert).
  337+                          (thanks adipose, TainZerL, and Harrison Green)
  338+      2.15 (2025-07-17) fixed an assert in debug mode when using floats with input
  339+                          callbacks, work around GCC warning when adding to null ptr
  340+                          (thanks Johannes Spohr and Pyry Kovanen).
  341+      2.14 (2025-05-09) fixed a bug using downsampling gather horizontal first, and 
  342+                          scatter with vertical first.
  343+      2.13 (2025-02-27) fixed a bug when using input callbacks, turned off simd for 
  344+                          tiny-c, fixed some variables that should have been static,
  345+                          fixes a bug when calculating temp memory with resizes that
  346+                          exceed 2GB of temp memory (very large resizes).
  347+      2.12 (2024-10-18) fix incorrect use of user_data with STBIR_FREE
  348+      2.11 (2024-09-08) fix harmless asan warnings in 2-channel and 3-channel mode
  349+                          with AVX-2, fix some weird scaling edge conditions with
  350+                          point sample mode.
  351+      2.10 (2024-07-27) fix the defines GCC and mingw for loop unroll control,
  352+                          fix MSVC 32-bit arm half float routines.
  353+      2.09 (2024-06-19) fix the defines for 32-bit ARM GCC builds (was selecting
  354+                          hardware half floats).
  355+      2.08 (2024-06-10) fix for RGB->BGR three channel flips and add SIMD (thanks
  356+                          to Ryan Salsbury), fix for sub-rect resizes, use the
  357+                          pragmas to control unrolling when they are available.
  358+      2.07 (2024-05-24) fix for slow final split during threaded conversions of very 
  359+                          wide scanlines when downsampling (caused by extra input 
  360+                          converting), fix for wide scanline resamples with many 
  361+                          splits (int overflow), fix GCC warning.
  362+      2.06 (2024-02-10) fix for identical width/height 3x or more down-scaling 
  363+                          undersampling a single row on rare resize ratios (about 1%).
  364+      2.05 (2024-02-07) fix for 2 pixel to 1 pixel resizes with wrap (thanks Aras),
  365+                        fix for output callback (thanks Julien Koenen).
  366+      2.04 (2023-11-17) fix for rare AVX bug, shadowed symbol (thanks Nikola Smiljanic).
  367+      2.03 (2023-11-01) ASAN and TSAN warnings fixed, minor tweaks.
  368+      2.00 (2023-10-10) mostly new source: new api, optimizations, simd, vertical-first, etc
  369+                          2x-5x faster without simd, 4x-12x faster with simd,
  370+                          in some cases, 20x to 40x faster esp resizing large to very small.
  371+      0.96 (2019-03-04) fixed warnings
  372+      0.95 (2017-07-23) fixed warnings
  373+      0.94 (2017-03-18) fixed warnings
  374+      0.93 (2017-03-03) fixed bug with certain combinations of heights
  375+      0.92 (2017-01-02) fix integer overflow on large (>2GB) images
  376+      0.91 (2016-04-02) fix warnings; fix handling of subpixel regions
  377+      0.90 (2014-09-17) first released version
  378+
  379+   LICENSE
  380+     See end of file for license information.
  381+*/
  382+
  383+#if !defined(STB_IMAGE_RESIZE_DO_HORIZONTALS) && !defined(STB_IMAGE_RESIZE_DO_VERTICALS) && !defined(STB_IMAGE_RESIZE_DO_CODERS)   // for internal re-includes
  384+
  385+#ifndef STBIR_INCLUDE_STB_IMAGE_RESIZE2_H
  386+#define STBIR_INCLUDE_STB_IMAGE_RESIZE2_H
  387+
  388+#include <stddef.h>
  389+#ifdef _MSC_VER
  390+typedef unsigned char    stbir_uint8;
  391+typedef unsigned short   stbir_uint16;
  392+typedef unsigned int     stbir_uint32;
  393+typedef unsigned __int64 stbir_uint64;
  394+#else
  395+#include <stdint.h>
  396+typedef uint8_t  stbir_uint8;
  397+typedef uint16_t stbir_uint16;
  398+typedef uint32_t stbir_uint32;
  399+typedef uint64_t stbir_uint64;
  400+#endif
  401+
  402+#ifndef STBIRDEF
  403+#ifdef STB_IMAGE_RESIZE_STATIC
  404+#define STBIRDEF static
  405+#else
  406+#ifdef __cplusplus
  407+#define STBIRDEF extern "C"
  408+#else
  409+#define STBIRDEF extern
  410+#endif
  411+#endif
  412+#endif
  413+
  414+//////////////////////////////////////////////////////////////////////////////
  415+////   start "header file" ///////////////////////////////////////////////////
  416+//
  417+// Easy-to-use API:
  418+//
  419+//     * stride is the offset between successive rows of image data
  420+//        in memory, in bytes. specify 0 for packed continuously in memory
  421+//     * colorspace is linear or sRGB as specified by function name
  422+//     * Uses the default filters
  423+//     * Uses edge mode clamped
  424+//     * returned result is 1 for success or 0 in case of an error.
  425+
  426+
  427+// stbir_pixel_layout specifies:
  428+//   number of channels
  429+//   order of channels
  430+//   whether color is premultiplied by alpha
  431+// for back compatibility, you can cast the old channel count to an stbir_pixel_layout
  432+typedef enum
  433+{
  434+  STBIR_1CHANNEL = 1,
  435+  STBIR_2CHANNEL = 2,
  436+  STBIR_RGB      = 3,               // 3-chan, with order specified (for channel flipping)
  437+  STBIR_BGR      = 0,               // 3-chan, with order specified (for channel flipping)
  438+  STBIR_4CHANNEL = 5,
  439+
  440+  STBIR_RGBA = 4,                   // alpha formats, where alpha is NOT premultiplied into color channels
  441+  STBIR_BGRA = 6,
  442+  STBIR_ARGB = 7,
  443+  STBIR_ABGR = 8,
  444+  STBIR_RA   = 9,
  445+  STBIR_AR   = 10,
  446+
  447+  STBIR_RGBA_PM = 11,               // alpha formats, where alpha is premultiplied into color channels
  448+  STBIR_BGRA_PM = 12,
  449+  STBIR_ARGB_PM = 13,
  450+  STBIR_ABGR_PM = 14,
  451+  STBIR_RA_PM   = 15,
  452+  STBIR_AR_PM   = 16,
  453+
  454+  STBIR_RGBA_NO_AW = 11,            // alpha formats, where NO alpha weighting is applied at all!
  455+  STBIR_BGRA_NO_AW = 12,            //   these are just synonyms for the _PM flags (which also do
  456+  STBIR_ARGB_NO_AW = 13,            //   no alpha weighting). These names just make it more clear
  457+  STBIR_ABGR_NO_AW = 14,            //   for some folks).
  458+  STBIR_RA_NO_AW   = 15,
  459+  STBIR_AR_NO_AW   = 16,
  460+
  461+} stbir_pixel_layout;
  462+
  463+//===============================================================
  464+//  Simple-complexity API
  465+//
  466+//    If output_pixels is NULL (0), then we will allocate the buffer and return it to you.
  467+//--------------------------------
  468+
  469+STBIRDEF unsigned char * stbir_resize_uint8_srgb( const unsigned char *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
  470+                                                        unsigned char *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
  471+                                                        stbir_pixel_layout pixel_type );
  472+
  473+STBIRDEF unsigned char * stbir_resize_uint8_linear( 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 float * stbir_resize_float_linear( const float *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
  478+                                                  float *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
  479+                                                  stbir_pixel_layout pixel_type );
  480+//===============================================================
  481+
  482+//===============================================================
  483+// Medium-complexity API
  484+//
  485+// This extends the easy-to-use API as follows:
  486+//
  487+//     * Can specify the datatype - U8, U8_SRGB, U16, FLOAT, HALF_FLOAT
  488+//     * Edge wrap can selected explicitly
  489+//     * Filter can be selected explicitly
  490+//--------------------------------
  491+
  492+typedef enum
  493+{
  494+  STBIR_EDGE_CLAMP   = 0,
  495+  STBIR_EDGE_REFLECT = 1,
  496+  STBIR_EDGE_WRAP    = 2,  // this edge mode is slower and uses more memory
  497+  STBIR_EDGE_ZERO    = 3,
  498+} stbir_edge;
  499+
  500+typedef enum
  501+{
  502+  STBIR_FILTER_DEFAULT      = 0,  // use same filter type that easy-to-use API chooses
  503+  STBIR_FILTER_BOX          = 1,  // A trapezoid w/1-pixel wide ramps, same result as box for integer scale ratios
  504+  STBIR_FILTER_TRIANGLE     = 2,  // On upsampling, produces same results as bilinear texture filtering
  505+  STBIR_FILTER_CUBICBSPLINE = 3,  // The cubic b-spline (aka Mitchell-Netrevalli with B=1,C=0), gaussian-esque
  506+  STBIR_FILTER_CATMULLROM   = 4,  // An interpolating cubic spline
  507+  STBIR_FILTER_MITCHELL     = 5,  // Mitchell-Netrevalli filter with B=1/3, C=1/3
  508+  STBIR_FILTER_POINT_SAMPLE = 6,  // Simple point sampling
  509+  STBIR_FILTER_OTHER        = 7,  // User callback specified
  510+} stbir_filter;
  511+
  512+typedef enum
  513+{
  514+  STBIR_TYPE_UINT8            = 0,
  515+  STBIR_TYPE_UINT8_SRGB       = 1,
  516+  STBIR_TYPE_UINT8_SRGB_ALPHA = 2,  // alpha channel, when present, should also be SRGB (this is very unusual)
  517+  STBIR_TYPE_UINT16           = 3,
  518+  STBIR_TYPE_FLOAT            = 4,
  519+  STBIR_TYPE_HALF_FLOAT       = 5
  520+} stbir_datatype;
  521+
  522+// medium api
  523+STBIRDEF void *  stbir_resize( const void *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
  524+                                     void *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
  525+                               stbir_pixel_layout pixel_layout, stbir_datatype data_type,
  526+                               stbir_edge edge, stbir_filter filter );
  527+//===============================================================
  528+
  529+
  530+
  531+//===============================================================
  532+// Extended-complexity API
  533+//
  534+// This API exposes all resize functionality.
  535+//
  536+//     * Separate filter types for each axis
  537+//     * Separate edge modes for each axis
  538+//     * Separate input and output data types
  539+//     * Can specify regions with subpixel correctness
  540+//     * Can specify alpha flags
  541+//     * Can specify a memory callback
  542+//     * Can specify a callback data type for pixel input and output
  543+//     * Can be threaded for a single resize
  544+//     * Can be used to resize many frames without recalculating the sampler info
  545+//
  546+//  Use this API as follows:
  547+//     1) Call the stbir_resize_init function on a local STBIR_RESIZE structure
  548+//     2) Call any of the stbir_set functions
  549+//     3) Optionally call stbir_build_samplers() if you are going to resample multiple times
  550+//        with the same input and output dimensions (like resizing video frames)
  551+//     4) Resample by calling stbir_resize_extended().
  552+//     5) Call stbir_free_samplers() if you called stbir_build_samplers()
  553+//--------------------------------
  554+
  555+
  556+// Types:
  557+
  558+// INPUT CALLBACK: this callback is used for input scanlines
  559+typedef void const * stbir_input_callback( void * optional_output, void const * input_ptr, int num_pixels, int x, int y, void * context );
  560+
  561+// OUTPUT CALLBACK: this callback is used for output scanlines
  562+typedef void stbir_output_callback( void const * output_ptr, int num_pixels, int y, void * context );
  563+
  564+// callbacks for user installed filters
  565+typedef float stbir__kernel_callback( float x, float scale, void * user_data ); // centered at zero
  566+typedef float stbir__support_callback( float scale, void * user_data );
  567+
  568+// internal structure with precomputed scaling
  569+typedef struct stbir__info stbir__info;
  570+
  571+typedef struct STBIR_RESIZE  // use the stbir_resize_init and stbir_override functions to set these values for future compatibility
  572+{
  573+  void * user_data;
  574+  void const * input_pixels;
  575+  int input_w, input_h;
  576+  double input_s0, input_t0, input_s1, input_t1;
  577+  stbir_input_callback * input_cb;
  578+  void * output_pixels;
  579+  int output_w, output_h;
  580+  int output_subx, output_suby, output_subw, output_subh;
  581+  stbir_output_callback * output_cb;
  582+  int input_stride_in_bytes;
  583+  int output_stride_in_bytes;
  584+  int splits;
  585+  int fast_alpha;
  586+  int needs_rebuild;
  587+  int called_alloc;
  588+  stbir_pixel_layout input_pixel_layout_public;
  589+  stbir_pixel_layout output_pixel_layout_public;
  590+  stbir_datatype input_data_type;
  591+  stbir_datatype output_data_type;
  592+  stbir_filter horizontal_filter, vertical_filter;
  593+  stbir_edge horizontal_edge, vertical_edge;
  594+  stbir__kernel_callback * horizontal_filter_kernel; stbir__support_callback * horizontal_filter_support;
  595+  stbir__kernel_callback * vertical_filter_kernel; stbir__support_callback * vertical_filter_support;
  596+  stbir__info * samplers;
  597+} STBIR_RESIZE;
  598+
  599+// extended complexity api
  600+
  601+
  602+// First off, you must ALWAYS call stbir_resize_init on your resize structure before any of the other calls!
  603+STBIRDEF void stbir_resize_init( STBIR_RESIZE * resize,
  604+                                 const void *input_pixels,  int input_w,  int input_h, int input_stride_in_bytes, // stride can be zero
  605+                                       void *output_pixels, int output_w, int output_h, int output_stride_in_bytes, // stride can be zero
  606+                                 stbir_pixel_layout pixel_layout, stbir_datatype data_type );
  607+
  608+//===============================================================
  609+// You can update these parameters any time after resize_init and there is no cost
  610+//--------------------------------
  611+
  612+STBIRDEF void stbir_set_datatypes( STBIR_RESIZE * resize, stbir_datatype input_type, stbir_datatype output_type );
  613+STBIRDEF void stbir_set_pixel_callbacks( STBIR_RESIZE * resize, stbir_input_callback * input_cb, stbir_output_callback * output_cb );   // no callbacks by default
  614+STBIRDEF void stbir_set_user_data( STBIR_RESIZE * resize, void * user_data );                                               // pass back STBIR_RESIZE* by default
  615+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 );
  616+
  617+//===============================================================
  618+
  619+
  620+//===============================================================
  621+// If you call any of these functions, you will trigger a sampler rebuild!
  622+//--------------------------------
  623+
  624+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
  625+STBIRDEF int stbir_set_edgemodes( STBIR_RESIZE * resize, stbir_edge horizontal_edge, stbir_edge vertical_edge );       // CLAMP by default
  626+
  627+STBIRDEF int stbir_set_filters( STBIR_RESIZE * resize, stbir_filter horizontal_filter, stbir_filter vertical_filter ); // STBIR_DEFAULT_FILTER_UPSAMPLE/DOWNSAMPLE by default
  628+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 );
  629+
  630+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)
  631+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)
  632+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)
  633+
  634+// when inputting AND outputting non-premultiplied alpha pixels, we use a slower but higher quality technique
  635+//   that fills the zero alpha pixel's RGB values with something plausible.  If you don't care about areas of
  636+//   zero alpha, you can call this function to get about a 25% speed improvement for STBIR_RGBA to STBIR_RGBA
  637+//   types of resizes.
  638+STBIRDEF int stbir_set_non_pm_alpha_speed_over_quality( STBIR_RESIZE * resize, int non_pma_alpha_speed_over_quality );
  639+//===============================================================
  640+
  641+
  642+//===============================================================
  643+// You can call build_samplers to prebuild all the internal data we need to resample.
  644+//   Then, if you call resize_extended many times with the same resize, you only pay the
  645+//   cost once.
  646+// If you do call build_samplers, you MUST call free_samplers eventually.
  647+//--------------------------------
  648+
  649+// This builds the samplers and does one allocation
  650+STBIRDEF int stbir_build_samplers( STBIR_RESIZE * resize );
  651+
  652+// You MUST call this, if you call stbir_build_samplers or stbir_build_samplers_with_splits
  653+STBIRDEF void stbir_free_samplers( STBIR_RESIZE * resize );
  654+//===============================================================
  655+
  656+
  657+// And this is the main function to perform the resize synchronously on one thread.
  658+STBIRDEF int stbir_resize_extended( STBIR_RESIZE * resize );
  659+
  660+
  661+//===============================================================
  662+// Use these functions for multithreading.
  663+//   1) You call stbir_build_samplers_with_splits first on the main thread
  664+//   2) Then stbir_resize_with_split on each thread
  665+//   3) stbir_free_samplers when done on the main thread
  666+//--------------------------------
  667+
  668+// This will build samplers for threading.
  669+//   You can pass in the number of threads you'd like to use (try_splits).
  670+//   It returns the number of splits (threads) that you can call it with.
  671+///  It might be less if the image resize can't be split up that many ways.
  672+
  673+STBIRDEF int stbir_build_samplers_with_splits( STBIR_RESIZE * resize, int try_splits );
  674+
  675+// This function does a split of the resizing (you call this fuction for each
  676+// split, on multiple threads). A split is a piece of the output resize pixel space.
  677+
  678+// Note that you MUST call stbir_build_samplers_with_splits before stbir_resize_extended_split!
  679+
  680+// Usually, you will always call stbir_resize_split with split_start as the thread_index
  681+//   and "1" for the split_count.
  682+// But, if you have a weird situation where you MIGHT want 8 threads, but sometimes
  683+//   only 4 threads, you can use 0,2,4,6 for the split_start's and use "2" for the
  684+//   split_count each time to turn in into a 4 thread resize. (This is unusual).
  685+
  686+STBIRDEF int stbir_resize_extended_split( STBIR_RESIZE * resize, int split_start, int split_count );
  687+//===============================================================
  688+
  689+
  690+//===============================================================
  691+// Pixel Callbacks info:
  692+//--------------------------------
  693+
  694+//   The input callback is super flexible - it calls you with the input address
  695+//   (based on the stride and base pointer), it gives you an optional_output
  696+//   pointer that you can fill, or you can just return your own pointer into
  697+//   your own data.
  698+//
  699+//   You can also do conversion from non-supported data types if necessary - in
  700+//   this case, you ignore the input_ptr and just use the x and y parameters to
  701+//   calculate your own input_ptr based on the size of each non-supported pixel.
  702+//   (Something like the third example below.)
  703+//
  704+//   You can also install just an input or just an output callback by setting the
  705+//   callback that you don't want to zero.
  706+//
  707+//     First example, progress: (getting a callback that you can monitor the progress):
  708+//        void const * my_callback( void * optional_output, void const * input_ptr, int num_pixels, int x, int y, void * context )
  709+//        {
  710+//           percentage_done = y / input_height;
  711+//           return input_ptr;  // use buffer from call
  712+//        }
  713+//
  714+//     Next example, copying: (copy from some other buffer or stream):
  715+//        void const * my_callback( void * optional_output, void const * input_ptr, int num_pixels, int x, int y, void * context )
  716+//        {
  717+//           CopyOrStreamData( optional_output, other_data_src, num_pixels * pixel_width_in_bytes );
  718+//           return optional_output;  // return the optional buffer that we filled
  719+//        }
  720+//
  721+//     Third example, input another buffer without copying: (zero-copy from other buffer):
  722+//        void const * my_callback( void * optional_output, void const * input_ptr, int num_pixels, int x, int y, void * context )
  723+//        {
  724+//           void * pixels = ( (char*) other_image_base ) + ( y * other_image_stride ) + ( x * other_pixel_width_in_bytes );
  725+//           return pixels;       // return pointer to your data without copying
  726+//        }
  727+//
  728+//
  729+//   The output callback is considerably simpler - it just calls you so that you can dump
  730+//   out each scanline. You could even directly copy out to disk if you have a simple format
  731+//   like TGA or BMP. You can also convert to other output types here if you want.
  732+//
  733+//   Simple example:
  734+//        void const * my_output( void * output_ptr, int num_pixels, int y, void * context )
  735+//        {
  736+//           percentage_done = y / output_height;
  737+//           fwrite( output_ptr, pixel_width_in_bytes, num_pixels, output_file );
  738+//        }
  739+//===============================================================
  740+
  741+
  742+
  743+
  744+//===============================================================
  745+// optional built-in profiling API
  746+//--------------------------------
  747+
  748+#ifdef STBIR_PROFILE
  749+
  750+typedef struct STBIR_PROFILE_INFO
  751+{
  752+  stbir_uint64 total_clocks;
  753+
  754+  // how many clocks spent (of total_clocks) in the various resize routines, along with a string description
  755+  //    there are "resize_count" number of zones
  756+  stbir_uint64 clocks[ 8 ];
  757+  char const ** descriptions;
  758+
  759+  // count of clocks and descriptions
  760+  stbir_uint32 count;
  761+} STBIR_PROFILE_INFO;
  762+
  763+// use after calling stbir_resize_extended (or stbir_build_samplers or stbir_build_samplers_with_splits)
  764+STBIRDEF void stbir_resize_build_profile_info( STBIR_PROFILE_INFO * out_info, STBIR_RESIZE const * resize );
  765+
  766+// use after calling stbir_resize_extended
  767+STBIRDEF void stbir_resize_extended_profile_info( STBIR_PROFILE_INFO * out_info, STBIR_RESIZE const * resize );
  768+
  769+// use after calling stbir_resize_extended_split
  770+STBIRDEF void stbir_resize_split_profile_info( STBIR_PROFILE_INFO * out_info, STBIR_RESIZE const * resize, int split_start, int split_num );
  771+
  772+//===============================================================
  773+
  774+#endif
  775+
  776+
  777+////   end header file   /////////////////////////////////////////////////////
  778+#endif // STBIR_INCLUDE_STB_IMAGE_RESIZE2_H
  779+
  780+#if defined(STB_IMAGE_RESIZE_IMPLEMENTATION) || defined(STB_IMAGE_RESIZE2_IMPLEMENTATION)
  781+
  782+#ifndef STBIR_ASSERT
  783+#include <assert.h>
  784+#define STBIR_ASSERT(x) assert(x)
  785+#endif
  786+
  787+#ifndef STBIR_MALLOC
  788+#include <stdlib.h>
  789+#define STBIR_MALLOC(size,user_data) ((void)(user_data), malloc(size))
  790+#define STBIR_FREE(ptr,user_data)    ((void)(user_data), free(ptr))
  791+// (we used the comma operator to evaluate user_data, to avoid "unused parameter" warnings)
  792+#endif
  793+
  794+#ifdef _MSC_VER
  795+
  796+#define stbir__inline __forceinline
  797+
  798+#else
  799+
  800+#define stbir__inline __inline__
  801+
  802+// Clang address sanitizer
  803+#if defined(__has_feature)
  804+  #if __has_feature(address_sanitizer) || __has_feature(memory_sanitizer)
  805+    #ifndef STBIR__SEPARATE_ALLOCATIONS
  806+      #define STBIR__SEPARATE_ALLOCATIONS
  807+    #endif
  808+  #endif
  809+#endif
  810+
  811+#endif
  812+
  813+// GCC and MSVC
  814+#if defined(__SANITIZE_ADDRESS__)
  815+  #ifndef STBIR__SEPARATE_ALLOCATIONS
  816+    #define STBIR__SEPARATE_ALLOCATIONS
  817+  #endif
  818+#endif
  819+
  820+// Always turn off automatic FMA use - use STBIR_USE_FMA if you want.
  821+// Otherwise, this is a determinism disaster.
  822+#ifndef STBIR_DONT_CHANGE_FP_CONTRACT  // override in case you don't want this behavior
  823+#if defined(_MSC_VER) && !defined(__clang__)
  824+#if _MSC_VER > 1200
  825+#pragma fp_contract(off)
  826+#endif
  827+#elif defined(__GNUC__) &&  !defined(__clang__)
  828+#pragma GCC optimize("fp-contract=off")
  829+#else
  830+#pragma STDC FP_CONTRACT OFF
  831+#endif
  832+#endif
  833+
  834+#ifdef _MSC_VER
  835+#define STBIR__UNUSED(v)  (void)(v)
  836+#else
  837+#define STBIR__UNUSED(v)  (void)sizeof(v)
  838+#endif
  839+
  840+#define STBIR__ARRAY_SIZE(a) (sizeof((a))/sizeof((a)[0]))
  841+
  842+
  843+#ifndef STBIR_DEFAULT_FILTER_UPSAMPLE
  844+#define STBIR_DEFAULT_FILTER_UPSAMPLE    STBIR_FILTER_CATMULLROM
  845+#endif
  846+
  847+#ifndef STBIR_DEFAULT_FILTER_DOWNSAMPLE
  848+#define STBIR_DEFAULT_FILTER_DOWNSAMPLE  STBIR_FILTER_MITCHELL
  849+#endif
  850+
  851+
  852+#ifndef STBIR__HEADER_FILENAME
  853+#define STBIR__HEADER_FILENAME "stb_image_resize2.h"
  854+#endif
  855+
  856+// the internal pixel layout enums are in a different order, so we can easily do range comparisons of types
  857+//   the public pixel layout is ordered in a way that if you cast num_channels (1-4) to the enum, you get something sensible
  858+typedef enum
  859+{
  860+  STBIRI_1CHANNEL = 0,
  861+  STBIRI_2CHANNEL = 1,
  862+  STBIRI_RGB      = 2,
  863+  STBIRI_BGR      = 3,
  864+  STBIRI_4CHANNEL = 4,
  865+
  866+  STBIRI_RGBA = 5,
  867+  STBIRI_BGRA = 6,
  868+  STBIRI_ARGB = 7,
  869+  STBIRI_ABGR = 8,
  870+  STBIRI_RA   = 9,
  871+  STBIRI_AR   = 10,
  872+
  873+  STBIRI_RGBA_PM = 11,
  874+  STBIRI_BGRA_PM = 12,
  875+  STBIRI_ARGB_PM = 13,
  876+  STBIRI_ABGR_PM = 14,
  877+  STBIRI_RA_PM   = 15,
  878+  STBIRI_AR_PM   = 16,
  879+} stbir_internal_pixel_layout;
  880+
  881+// define the public pixel layouts to not compile inside the implementation (to avoid accidental use)
  882+#define STBIR_BGR bad_dont_use_in_implementation
  883+#define STBIR_1CHANNEL STBIR_BGR
  884+#define STBIR_2CHANNEL STBIR_BGR
  885+#define STBIR_RGB STBIR_BGR
  886+#define STBIR_RGBA STBIR_BGR
  887+#define STBIR_4CHANNEL STBIR_BGR
  888+#define STBIR_BGRA STBIR_BGR
  889+#define STBIR_ARGB STBIR_BGR
  890+#define STBIR_ABGR STBIR_BGR
  891+#define STBIR_RA STBIR_BGR
  892+#define STBIR_AR STBIR_BGR
  893+#define STBIR_RGBA_PM STBIR_BGR
  894+#define STBIR_BGRA_PM STBIR_BGR
  895+#define STBIR_ARGB_PM STBIR_BGR
  896+#define STBIR_ABGR_PM STBIR_BGR
  897+#define STBIR_RA_PM STBIR_BGR
  898+#define STBIR_AR_PM STBIR_BGR
  899+
  900+// must match stbir_datatype
  901+static unsigned char stbir__type_size[] = {
  902+  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
  903+};
  904+
  905+// When gathering, the contributors are which source pixels contribute.
  906+// When scattering, the contributors are which destination pixels are contributed to.
  907+typedef struct
  908+{
  909+  int n0; // First contributing pixel
  910+  int n1; // Last contributing pixel
  911+} stbir__contributors;
  912+
  913+typedef struct
  914+{
  915+  int lowest;    // First sample index for whole filter
  916+  int highest;   // Last sample index for whole filter
  917+  int widest;    // widest single set of samples for an output
  918+} stbir__filter_extent_info;
  919+
  920+typedef struct
  921+{
  922+  int n0; // First pixel of decode buffer to write to
  923+  int n1; // Last pixel of decode that will be written to
  924+  int pixel_offset_for_input;  // Pixel offset into input_scanline
  925+} stbir__span;
  926+
  927+typedef struct stbir__scale_info
  928+{
  929+  int input_full_size;
  930+  int output_sub_size;
  931+  float scale;
  932+  float inv_scale;
  933+  float pixel_shift; // starting shift in output pixel space (in pixels)
  934+  int scale_is_rational;
  935+  stbir_uint32 scale_numerator, scale_denominator;
  936+} stbir__scale_info;
  937+
  938+typedef struct
  939+{
  940+  stbir__contributors * contributors;
  941+  float* coefficients;
  942+  stbir__contributors * gather_prescatter_contributors;
  943+  float * gather_prescatter_coefficients;
  944+  stbir__scale_info scale_info;
  945+  float support;
  946+  stbir_filter filter_enum;
  947+  stbir__kernel_callback * filter_kernel;
  948+  stbir__support_callback * filter_support;
  949+  stbir_edge edge;
  950+  int coefficient_width;
  951+  int filter_pixel_width;
  952+  int filter_pixel_margin;
  953+  int num_contributors;
  954+  int contributors_size;
  955+  int coefficients_size;
  956+  stbir__filter_extent_info extent_info;
  957+  int is_gather;  // 0 = scatter, 1 = gather with scale >= 1, 2 = gather with scale < 1
  958+  int gather_prescatter_num_contributors;
  959+  int gather_prescatter_coefficient_width;
  960+  int gather_prescatter_contributors_size;
  961+  int gather_prescatter_coefficients_size;
  962+} stbir__sampler;
  963+
  964+typedef struct
  965+{
  966+  stbir__contributors conservative;
  967+  int edge_sizes[2];    // this can be less than filter_pixel_margin, if the filter and scaling falls off
  968+  stbir__span spans[2]; // can be two spans, if doing input subrect with clamp mode WRAP
  969+} stbir__extents;
  970+
  971+typedef struct
  972+{
  973+#ifdef STBIR_PROFILE
  974+  union
  975+  {
  976+    struct { stbir_uint64 total, looping, vertical, horizontal, decode, encode, alpha, unalpha; } named;
  977+    stbir_uint64 array[8];
  978+  } profile;
  979+  stbir_uint64 * current_zone_excluded_ptr;
  980+#endif
  981+  float* decode_buffer;
  982+
  983+  int ring_buffer_first_scanline;
  984+  int ring_buffer_last_scanline;
  985+  int ring_buffer_begin_index;    // first_scanline is at this index in the ring buffer
  986+  int start_output_y, end_output_y;
  987+  int start_input_y, end_input_y;  // used in scatter only
  988+
  989+  #ifdef STBIR__SEPARATE_ALLOCATIONS
  990+    float** ring_buffers; // one pointer for each ring buffer
  991+  #else
  992+    float* ring_buffer;  // one big buffer that we index into
  993+  #endif
  994+
  995+  float* vertical_buffer;
  996+
  997+  char no_cache_straddle[64];
  998+} stbir__per_split_info;
  999+
 1000+typedef float * stbir__decode_pixels_func( float * decode, int width_times_channels, void const * input );
 1001+typedef void stbir__alpha_weight_func( float * decode_buffer, int width_times_channels );
 1002+typedef void stbir__horizontal_gather_channels_func( float * output_buffer, unsigned int output_sub_size, float const * decode_buffer,
 1003+  stbir__contributors const * horizontal_contributors, float const * horizontal_coefficients, int coefficient_width );
 1004+typedef void stbir__alpha_unweight_func(float * encode_buffer, int width_times_channels );
 1005+typedef void stbir__encode_pixels_func( void * output, int width_times_channels, float const * encode );
 1006+
 1007+struct stbir__info
 1008+{
 1009+#ifdef STBIR_PROFILE
 1010+  union
 1011+  {
 1012+    struct { stbir_uint64 total, build, alloc, horizontal, vertical, cleanup, pivot; } named;
 1013+    stbir_uint64 array[7];
 1014+  } profile;
 1015+  stbir_uint64 * current_zone_excluded_ptr;
 1016+#endif
 1017+  stbir__sampler horizontal;
 1018+  stbir__sampler vertical;
 1019+
 1020+  void const * input_data;
 1021+  void * output_data;
 1022+
 1023+  int input_stride_bytes;
 1024+  int output_stride_bytes;
 1025+  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)
 1026+  int ring_buffer_num_entries;    // Total number of entries in the ring buffer.
 1027+
 1028+  stbir_datatype input_type;
 1029+  stbir_datatype output_type;
 1030+
 1031+  stbir_input_callback * in_pixels_cb;
 1032+  void * user_data;
 1033+  stbir_output_callback * out_pixels_cb;
 1034+
 1035+  stbir__extents scanline_extents;
 1036+
 1037+  void * alloced_mem;
 1038+  stbir__per_split_info * split_info;  // by default 1, but there will be N of these allocated based on the thread init you did
 1039+
 1040+  stbir__decode_pixels_func * decode_pixels;
 1041+  stbir__alpha_weight_func * alpha_weight;
 1042+  stbir__horizontal_gather_channels_func * horizontal_gather_channels;
 1043+  stbir__alpha_unweight_func * alpha_unweight;
 1044+  stbir__encode_pixels_func * encode_pixels;
 1045+
 1046+  int alloc_ring_buffer_num_entries;    // Number of entries in the ring buffer that will be allocated
 1047+  int splits; // count of splits
 1048+
 1049+  stbir_internal_pixel_layout input_pixel_layout_internal;
 1050+  stbir_internal_pixel_layout output_pixel_layout_internal;
 1051+
 1052+  int input_color_and_type;
 1053+  int offset_x, offset_y; // offset within output_data
 1054+  int vertical_first;
 1055+  int channels;
 1056+  int effective_channels; // same as channels, except on RGBA/ARGB (7), or XA/AX (3)
 1057+  size_t alloced_total;
 1058+};
 1059+
 1060+
 1061+#define stbir__max_uint8_as_float             255.0f
 1062+#define stbir__max_uint16_as_float            65535.0f
 1063+#define stbir__max_uint8_as_float_inverted    3.9215689e-03f     // (1.0f/255.0f)
 1064+#define stbir__max_uint16_as_float_inverted   1.5259022e-05f     // (1.0f/65535.0f)
 1065+#define stbir__small_float ((float)1 / (1 << 20) / (1 << 20) / (1 << 20) / (1 << 20) / (1 << 20) / (1 << 20))
 1066+
 1067+// min/max friendly
 1068+#define STBIR_CLAMP(x, xmin, xmax) for(;;) { \
 1069+  if ( (x) < (xmin) ) (x) = (xmin);     \
 1070+  if ( (x) > (xmax) ) (x) = (xmax);     \
 1071+  break;                                \
 1072+}
 1073+
 1074+static stbir__inline int stbir__min(int a, int b)
 1075+{
 1076+  return a < b ? a : b;
 1077+}
 1078+
 1079+static stbir__inline int stbir__max(int a, int b)
 1080+{
 1081+  return a > b ? a : b;
 1082+}
 1083+
 1084+static float stbir__srgb_uchar_to_linear_float[256] = {
 1085+  0.000000f, 0.000304f, 0.000607f, 0.000911f, 0.001214f, 0.001518f, 0.001821f, 0.002125f, 0.002428f, 0.002732f, 0.003035f,
 1086+  0.003347f, 0.003677f, 0.004025f, 0.004391f, 0.004777f, 0.005182f, 0.005605f, 0.006049f, 0.006512f, 0.006995f, 0.007499f,
 1087+  0.008023f, 0.008568f, 0.009134f, 0.009721f, 0.010330f, 0.010960f, 0.011612f, 0.012286f, 0.012983f, 0.013702f, 0.014444f,
 1088+  0.015209f, 0.015996f, 0.016807f, 0.017642f, 0.018500f, 0.019382f, 0.020289f, 0.021219f, 0.022174f, 0.023153f, 0.024158f,
 1089+  0.025187f, 0.026241f, 0.027321f, 0.028426f, 0.029557f, 0.030713f, 0.031896f, 0.033105f, 0.034340f, 0.035601f, 0.036889f,
 1090+  0.038204f, 0.039546f, 0.040915f, 0.042311f, 0.043735f, 0.045186f, 0.046665f, 0.048172f, 0.049707f, 0.051269f, 0.052861f,
 1091+  0.054480f, 0.056128f, 0.057805f, 0.059511f, 0.061246f, 0.063010f, 0.064803f, 0.066626f, 0.068478f, 0.070360f, 0.072272f,
 1092+  0.074214f, 0.076185f, 0.078187f, 0.080220f, 0.082283f, 0.084376f, 0.086500f, 0.088656f, 0.090842f, 0.093059f, 0.095307f,
 1093+  0.097587f, 0.099899f, 0.102242f, 0.104616f, 0.107023f, 0.109462f, 0.111932f, 0.114435f, 0.116971f, 0.119538f, 0.122139f,
 1094+  0.124772f, 0.127438f, 0.130136f, 0.132868f, 0.135633f, 0.138432f, 0.141263f, 0.144128f, 0.147027f, 0.149960f, 0.152926f,
 1095+  0.155926f, 0.158961f, 0.162029f, 0.165132f, 0.168269f, 0.171441f, 0.174647f, 0.177888f, 0.181164f, 0.184475f, 0.187821f,
 1096+  0.191202f, 0.194618f, 0.198069f, 0.201556f, 0.205079f, 0.208637f, 0.212231f, 0.215861f, 0.219526f, 0.223228f, 0.226966f,
 1097+  0.230740f, 0.234551f, 0.238398f, 0.242281f, 0.246201f, 0.250158f, 0.254152f, 0.258183f, 0.262251f, 0.266356f, 0.270498f,
 1098+  0.274677f, 0.278894f, 0.283149f, 0.287441f, 0.291771f, 0.296138f, 0.300544f, 0.304987f, 0.309469f, 0.313989f, 0.318547f,
 1099+  0.323143f, 0.327778f, 0.332452f, 0.337164f, 0.341914f, 0.346704f, 0.351533f, 0.356400f, 0.361307f, 0.366253f, 0.371238f,
 1100+  0.376262f, 0.381326f, 0.386430f, 0.391573f, 0.396755f, 0.401978f, 0.407240f, 0.412543f, 0.417885f, 0.423268f, 0.428691f,
 1101+  0.434154f, 0.439657f, 0.445201f, 0.450786f, 0.456411f, 0.462077f, 0.467784f, 0.473532f, 0.479320f, 0.485150f, 0.491021f,
 1102+  0.496933f, 0.502887f, 0.508881f, 0.514918f, 0.520996f, 0.527115f, 0.533276f, 0.539480f, 0.545725f, 0.552011f, 0.558340f,
 1103+  0.564712f, 0.571125f, 0.577581f, 0.584078f, 0.590619f, 0.597202f, 0.603827f, 0.610496f, 0.617207f, 0.623960f, 0.630757f,
 1104+  0.637597f, 0.644480f, 0.651406f, 0.658375f, 0.665387f, 0.672443f, 0.679543f, 0.686685f, 0.693872f, 0.701102f, 0.708376f,
 1105+  0.715694f, 0.723055f, 0.730461f, 0.737911f, 0.745404f, 0.752942f, 0.760525f, 0.768151f, 0.775822f, 0.783538f, 0.791298f,
 1106+  0.799103f, 0.806952f, 0.814847f, 0.822786f, 0.830770f, 0.838799f, 0.846873f, 0.854993f, 0.863157f, 0.871367f, 0.879622f,
 1107+  0.887923f, 0.896269f, 0.904661f, 0.913099f, 0.921582f, 0.930111f, 0.938686f, 0.947307f, 0.955974f, 0.964686f, 0.973445f,
 1108+  0.982251f, 0.991102f, 1.0f
 1109+};
 1110+
 1111+typedef union
 1112+{
 1113+  unsigned int u;
 1114+  float f;
 1115+} stbir__FP32;
 1116+
 1117+// From https://gist.github.com/rygorous/2203834
 1118+
 1119+static const stbir_uint32 fp32_to_srgb8_tab4[104] = {
 1120+  0x0073000d, 0x007a000d, 0x0080000d, 0x0087000d, 0x008d000d, 0x0094000d, 0x009a000d, 0x00a1000d,
 1121+  0x00a7001a, 0x00b4001a, 0x00c1001a, 0x00ce001a, 0x00da001a, 0x00e7001a, 0x00f4001a, 0x0101001a,
 1122+  0x010e0033, 0x01280033, 0x01410033, 0x015b0033, 0x01750033, 0x018f0033, 0x01a80033, 0x01c20033,
 1123+  0x01dc0067, 0x020f0067, 0x02430067, 0x02760067, 0x02aa0067, 0x02dd0067, 0x03110067, 0x03440067,
 1124+  0x037800ce, 0x03df00ce, 0x044600ce, 0x04ad00ce, 0x051400ce, 0x057b00c5, 0x05dd00bc, 0x063b00b5,
 1125+  0x06970158, 0x07420142, 0x07e30130, 0x087b0120, 0x090b0112, 0x09940106, 0x0a1700fc, 0x0a9500f2,
 1126+  0x0b0f01cb, 0x0bf401ae, 0x0ccb0195, 0x0d950180, 0x0e56016e, 0x0f0d015e, 0x0fbc0150, 0x10630143,
 1127+  0x11070264, 0x1238023e, 0x1357021d, 0x14660201, 0x156601e9, 0x165a01d3, 0x174401c0, 0x182401af,
 1128+  0x18fe0331, 0x1a9602fe, 0x1c1502d2, 0x1d7e02ad, 0x1ed4028d, 0x201a0270, 0x21520256, 0x227d0240,
 1129+  0x239f0443, 0x25c003fe, 0x27bf03c4, 0x29a10392, 0x2b6a0367, 0x2d1d0341, 0x2ebe031f, 0x304d0300,
 1130+  0x31d105b0, 0x34a80555, 0x37520507, 0x39d504c5, 0x3c37048b, 0x3e7c0458, 0x40a8042a, 0x42bd0401,
 1131+  0x44c20798, 0x488e071e, 0x4c1c06b6, 0x4f76065d, 0x52a50610, 0x55ac05cc, 0x5892058f, 0x5b590559,
 1132+  0x5e0c0a23, 0x631c0980, 0x67db08f6, 0x6c55087f, 0x70940818, 0x74a007bd, 0x787d076c, 0x7c330723,
 1133+};
 1134+
 1135+static stbir__inline stbir_uint8 stbir__linear_to_srgb_uchar(float in)
 1136+{
 1137+  static const stbir__FP32 almostone = { 0x3f7fffff }; // 1-eps
 1138+  static const stbir__FP32 minval = { (127-13) << 23 };
 1139+  stbir_uint32 tab,bias,scale,t;
 1140+  stbir__FP32 f;
 1141+
 1142+  // Clamp to [2^(-13), 1-eps]; these two values map to 0 and 1, respectively.
 1143+  // The tests are carefully written so that NaNs map to 0, same as in the reference
 1144+  // implementation.
 1145+  if (!(in > minval.f)) // written this way to catch NaNs
 1146+      return 0;
 1147+  if (in > almostone.f)
 1148+      return 255;
 1149+
 1150+  // Do the table lookup and unpack bias, scale
 1151+  f.f = in;
 1152+  tab = fp32_to_srgb8_tab4[(f.u - minval.u) >> 20];
 1153+  bias = (tab >> 16) << 9;
 1154+  scale = tab & 0xffff;
 1155+
 1156+  // Grab next-highest mantissa bits and perform linear interpolation
 1157+  t = (f.u >> 12) & 0xff;
 1158+  return (unsigned char) ((bias + scale*t) >> 16);
 1159+}
 1160+
 1161+#ifndef STBIR_FORCE_GATHER_FILTER_SCANLINES_AMOUNT
 1162+#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.
 1163+#endif
 1164+
 1165+#ifndef STBIR_FORCE_MINIMUM_SCANLINES_FOR_SPLITS
 1166+#define STBIR_FORCE_MINIMUM_SCANLINES_FOR_SPLITS 4 // when threading, what is the minimum number of scanlines for a split?
 1167+#endif
 1168+
 1169+#define STBIR_INPUT_CALLBACK_PADDING 3
 1170+
 1171+#ifdef _M_IX86_FP
 1172+#if ( _M_IX86_FP >= 1 )
 1173+#ifndef STBIR_SSE
 1174+#define STBIR_SSE
 1175+#endif
 1176+#endif
 1177+#endif
 1178+
 1179+#ifdef __TINYC__
 1180+  // tiny c has no intrinsics yet - this can become a version check if they add them
 1181+  #define STBIR_NO_SIMD
 1182+#endif
 1183+
 1184+#if defined(_x86_64) || defined( __x86_64__ ) || defined( _M_X64 ) || defined(__x86_64) || defined(_M_AMD64) || defined(__SSE2__) || defined(STBIR_SSE) || defined(STBIR_SSE2)
 1185+  #ifndef STBIR_SSE2
 1186+    #define STBIR_SSE2
 1187+  #endif
 1188+  #if defined(__AVX__) || defined(STBIR_AVX2)
 1189+    #ifndef STBIR_AVX
 1190+      #ifndef STBIR_NO_AVX
 1191+        #define STBIR_AVX
 1192+      #endif
 1193+    #endif
 1194+  #endif
 1195+  #if defined(__AVX2__) || defined(STBIR_AVX2)
 1196+    #ifndef STBIR_NO_AVX2
 1197+      #ifndef STBIR_AVX2
 1198+        #define STBIR_AVX2
 1199+      #endif
 1200+      #if defined( _MSC_VER ) && !defined(__clang__)
 1201+        #ifndef STBIR_FP16C  // FP16C instructions are on all AVX2 cpus, so we can autoselect it here on microsoft - clang needs -m16c
 1202+          #define STBIR_FP16C
 1203+        #endif
 1204+      #endif
 1205+    #endif
 1206+  #endif
 1207+  #ifdef __F16C__
 1208+    #ifndef STBIR_FP16C  // turn on FP16C instructions if the define is set (for clang and gcc)
 1209+      #define STBIR_FP16C
 1210+    #endif
 1211+  #endif
 1212+#endif
 1213+
 1214+#if defined( _M_ARM64 ) || defined( __aarch64__ ) || defined( __arm64__ ) || ((__ARM_NEON_FP & 4) != 0) || defined(__ARM_NEON__)
 1215+#ifndef STBIR_NEON
 1216+#define STBIR_NEON
 1217+#endif
 1218+#endif
 1219+
 1220+#if defined(_M_ARM) || defined(__arm__)
 1221+#ifdef STBIR_USE_FMA
 1222+#undef STBIR_USE_FMA // no FMA for 32-bit arm on MSVC
 1223+#endif
 1224+#endif
 1225+
 1226+#if defined(__wasm__) && defined(__wasm_simd128__)
 1227+#ifndef STBIR_WASM
 1228+#define STBIR_WASM
 1229+#endif
 1230+#endif
 1231+
 1232+// restrict pointers for the output pointers, other loop and unroll control
 1233+#if defined( _MSC_VER ) && !defined(__clang__)
 1234+  #define STBIR_STREAMOUT_PTR( star ) star __restrict
 1235+  #define STBIR_NO_UNROLL( ptr ) __assume(ptr) // this oddly keeps msvc from unrolling a loop
 1236+  #if _MSC_VER >= 1900
 1237+    #define STBIR_NO_UNROLL_LOOP_START __pragma(loop( no_vector )) 
 1238+  #else
 1239+    #define STBIR_NO_UNROLL_LOOP_START 
 1240+  #endif
 1241+#elif defined( __clang__ )
 1242+  #define STBIR_STREAMOUT_PTR( star ) star __restrict__
 1243+  #define STBIR_NO_UNROLL( ptr ) __asm__ (""::"r"(ptr)) 
 1244+  #if ( __clang_major__ >= 4 ) || ( ( __clang_major__ >= 3 ) && ( __clang_minor__ >= 5 ) )
 1245+    #define STBIR_NO_UNROLL_LOOP_START _Pragma("clang loop unroll(disable)") _Pragma("clang loop vectorize(disable)")
 1246+  #else
 1247+    #define STBIR_NO_UNROLL_LOOP_START
 1248+  #endif 
 1249+#elif defined( __GNUC__ )
 1250+  #define STBIR_STREAMOUT_PTR( star ) star __restrict__
 1251+  #define STBIR_NO_UNROLL( ptr ) __asm__ (""::"r"(ptr))
 1252+  #if __GNUC__ >= 14
 1253+    #define STBIR_NO_UNROLL_LOOP_START _Pragma("GCC unroll 0") _Pragma("GCC novector")
 1254+  #else
 1255+    #define STBIR_NO_UNROLL_LOOP_START
 1256+  #endif
 1257+  #define STBIR_NO_UNROLL_LOOP_START_INF_FOR
 1258+#else
 1259+  #define STBIR_STREAMOUT_PTR( star ) star
 1260+  #define STBIR_NO_UNROLL( ptr )
 1261+  #define STBIR_NO_UNROLL_LOOP_START
 1262+#endif
 1263+
 1264+#ifndef STBIR_NO_UNROLL_LOOP_START_INF_FOR
 1265+#define STBIR_NO_UNROLL_LOOP_START_INF_FOR STBIR_NO_UNROLL_LOOP_START
 1266+#endif
 1267+
 1268+#ifdef STBIR_NO_SIMD // force simd off for whatever reason
 1269+
 1270+// force simd off overrides everything else, so clear it all
 1271+
 1272+#ifdef STBIR_SSE2
 1273+#undef STBIR_SSE2
 1274+#endif
 1275+
 1276+#ifdef STBIR_AVX
 1277+#undef STBIR_AVX
 1278+#endif
 1279+
 1280+#ifdef STBIR_NEON
 1281+#undef STBIR_NEON
 1282+#endif
 1283+
 1284+#ifdef STBIR_AVX2
 1285+#undef STBIR_AVX2
 1286+#endif
 1287+
 1288+#ifdef STBIR_FP16C
 1289+#undef STBIR_FP16C
 1290+#endif
 1291+
 1292+#ifdef STBIR_WASM
 1293+#undef STBIR_WASM
 1294+#endif
 1295+
 1296+#ifdef STBIR_SIMD
 1297+#undef STBIR_SIMD
 1298+#endif
 1299+
 1300+#else // STBIR_SIMD
 1301+
 1302+#ifdef STBIR_SSE2
 1303+  #include <emmintrin.h>
 1304+
 1305+  #define stbir__simdf __m128
 1306+  #define stbir__simdi __m128i
 1307+
 1308+  #define stbir_simdi_castf( reg ) _mm_castps_si128(reg)
 1309+  #define stbir_simdf_casti( reg ) _mm_castsi128_ps(reg)
 1310+
 1311+  #define stbir__simdf_load( reg, ptr ) (reg) = _mm_loadu_ps( (float const*)(ptr) )
 1312+  #define stbir__simdi_load( reg, ptr ) (reg) = _mm_loadu_si128 ( (stbir__simdi const*)(ptr) )
 1313+  #define stbir__simdf_load1( out, ptr ) (out) = _mm_load_ss( (float const*)(ptr) )  // top values can be random (not denormal or nan for perf)
 1314+  #define stbir__simdi_load1( out, ptr ) (out) = _mm_castps_si128( _mm_load_ss( (float const*)(ptr) ))
 1315+  #define stbir__simdf_load1z( out, ptr ) (out) = _mm_load_ss( (float const*)(ptr) )  // top values must be zero
 1316+  #define stbir__simdf_frep4( fvar ) _mm_set_ps1( fvar )
 1317+  #define stbir__simdf_load1frep4( out, fvar ) (out) = _mm_set_ps1( fvar )
 1318+  #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)
 1319+  #define stbir__simdf_load2z( out, ptr ) (out) = _mm_castsi128_ps( _mm_loadl_epi64( (__m128i*)(ptr)) ) // top values must be zero
 1320+  #define stbir__simdf_load2hmerge( out, reg, ptr ) (out) = _mm_castpd_ps(_mm_loadh_pd( _mm_castps_pd(reg), (double*)(ptr) ))
 1321+
 1322+  #define stbir__simdf_zeroP() _mm_setzero_ps()
 1323+  #define stbir__simdf_zero( reg ) (reg) = _mm_setzero_ps()
 1324+
 1325+  #define stbir__simdf_store( ptr, reg )  _mm_storeu_ps( (float*)(ptr), reg )
 1326+  #define stbir__simdf_store1( ptr, reg ) _mm_store_ss( (float*)(ptr), reg )
 1327+  #define stbir__simdf_store2( ptr, reg ) _mm_storel_epi64( (__m128i*)(ptr), _mm_castps_si128(reg) )
 1328+  #define stbir__simdf_store2h( ptr, reg ) _mm_storeh_pd( (double*)(ptr), _mm_castps_pd(reg) )
 1329+
 1330+  #define stbir__simdi_store( ptr, reg )  _mm_storeu_si128( (__m128i*)(ptr), reg )
 1331+  #define stbir__simdi_store1( ptr, reg ) _mm_store_ss( (float*)(ptr), _mm_castsi128_ps(reg) )
 1332+  #define stbir__simdi_store2( ptr, reg ) _mm_storel_epi64( (__m128i*)(ptr), (reg) )
 1333+
 1334+  #define stbir__prefetch( ptr ) _mm_prefetch((char*)(ptr), _MM_HINT_T0 )
 1335+
 1336+  #define stbir__simdi_expand_u8_to_u32(out0,out1,out2,out3,ireg) \
 1337+  { \
 1338+    stbir__simdi zero = _mm_setzero_si128(); \
 1339+    out2 = _mm_unpacklo_epi8( ireg, zero ); \
 1340+    out3 = _mm_unpackhi_epi8( ireg, zero ); \
 1341+    out0 = _mm_unpacklo_epi16( out2, zero ); \
 1342+    out1 = _mm_unpackhi_epi16( out2, zero ); \
 1343+    out2 = _mm_unpacklo_epi16( out3, zero ); \
 1344+    out3 = _mm_unpackhi_epi16( out3, zero ); \
 1345+  }
 1346+
 1347+#define stbir__simdi_expand_u8_to_1u32(out,ireg) \
 1348+  { \
 1349+    stbir__simdi zero = _mm_setzero_si128(); \
 1350+    out = _mm_unpacklo_epi8( ireg, zero ); \
 1351+    out = _mm_unpacklo_epi16( out, zero ); \
 1352+  }
 1353+
 1354+  #define stbir__simdi_expand_u16_to_u32(out0,out1,ireg) \
 1355+  { \
 1356+    stbir__simdi zero = _mm_setzero_si128(); \
 1357+    out0 = _mm_unpacklo_epi16( ireg, zero ); \
 1358+    out1 = _mm_unpackhi_epi16( ireg, zero ); \
 1359+  }
 1360+
 1361+  #define stbir__simdf_convert_float_to_i32( i, f ) (i) = _mm_cvttps_epi32(f)
 1362+  #define stbir__simdf_convert_float_to_int( f ) _mm_cvtt_ss2si(f)
 1363+  #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()))))
 1364+  #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()))))
 1365+
 1366+  #define stbir__simdi_to_int( i ) _mm_cvtsi128_si32(i)
 1367+  #define stbir__simdi_convert_i32_to_float(out, ireg) (out) = _mm_cvtepi32_ps( ireg )
 1368+  #define stbir__simdf_add( out, reg0, reg1 ) (out) = _mm_add_ps( reg0, reg1 )
 1369+  #define stbir__simdf_mult( out, reg0, reg1 ) (out) = _mm_mul_ps( reg0, reg1 )
 1370+  #define stbir__simdf_mult_mem( out, reg, ptr ) (out) = _mm_mul_ps( reg, _mm_loadu_ps( (float const*)(ptr) ) )
 1371+  #define stbir__simdf_mult1_mem( out, reg, ptr ) (out) = _mm_mul_ss( reg, _mm_load_ss( (float const*)(ptr) ) )
 1372+  #define stbir__simdf_add_mem( out, reg, ptr ) (out) = _mm_add_ps( reg, _mm_loadu_ps( (float const*)(ptr) ) )
 1373+  #define stbir__simdf_add1_mem( out, reg, ptr ) (out) = _mm_add_ss( reg, _mm_load_ss( (float const*)(ptr) ) )
 1374+
 1375+  #ifdef STBIR_USE_FMA           // not on by default to maintain bit identical simd to non-simd
 1376+  #include <immintrin.h>
 1377+  #define stbir__simdf_madd( out, add, mul1, mul2 ) (out) = _mm_fmadd_ps( mul1, mul2, add )
 1378+  #define stbir__simdf_madd1( out, add, mul1, mul2 ) (out) = _mm_fmadd_ss( mul1, mul2, add )
 1379+  #define stbir__simdf_madd_mem( out, add, mul, ptr ) (out) = _mm_fmadd_ps( mul, _mm_loadu_ps( (float const*)(ptr) ), add )
 1380+  #define stbir__simdf_madd1_mem( out, add, mul, ptr ) (out) = _mm_fmadd_ss( mul, _mm_load_ss( (float const*)(ptr) ), add )
 1381+  #else
 1382+  #define stbir__simdf_madd( out, add, mul1, mul2 ) (out) = _mm_add_ps( add, _mm_mul_ps( mul1, mul2 ) )
 1383+  #define stbir__simdf_madd1( out, add, mul1, mul2 ) (out) = _mm_add_ss( add, _mm_mul_ss( mul1, mul2 ) )
 1384+  #define stbir__simdf_madd_mem( out, add, mul, ptr ) (out) = _mm_add_ps( add, _mm_mul_ps( mul, _mm_loadu_ps( (float const*)(ptr) ) ) )
 1385+  #define stbir__simdf_madd1_mem( out, add, mul, ptr ) (out) = _mm_add_ss( add, _mm_mul_ss( mul, _mm_load_ss( (float const*)(ptr) ) ) )
 1386+  #endif
 1387+
 1388+  #define stbir__simdf_add1( out, reg0, reg1 ) (out) = _mm_add_ss( reg0, reg1 )
 1389+  #define stbir__simdf_mult1( out, reg0, reg1 ) (out) = _mm_mul_ss( reg0, reg1 )
 1390+
 1391+  #define stbir__simdf_and( out, reg0, reg1 ) (out) = _mm_and_ps( reg0, reg1 )
 1392+  #define stbir__simdf_or( out, reg0, reg1 ) (out) = _mm_or_ps( reg0, reg1 )
 1393+
 1394+  #define stbir__simdf_min( out, reg0, reg1 ) (out) = _mm_min_ps( reg0, reg1 )
 1395+  #define stbir__simdf_max( out, reg0, reg1 ) (out) = _mm_max_ps( reg0, reg1 )
 1396+  #define stbir__simdf_min1( out, reg0, reg1 ) (out) = _mm_min_ss( reg0, reg1 )
 1397+  #define stbir__simdf_max1( out, reg0, reg1 ) (out) = _mm_max_ss( reg0, reg1 )
 1398+
 1399+  #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) ) )
 1400+  #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) ) )
 1401+
 1402+  static const stbir__simdf STBIR_zeroones = { 0.0f,1.0f,0.0f,1.0f };
 1403+  static const stbir__simdf STBIR_onezeros = { 1.0f,0.0f,1.0f,0.0f };
 1404+  #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) ) )
 1405+  #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) ) )
 1406+  #define stbir__simdf_a1a1( out, alp, ones) (out) = _mm_or_ps( _mm_castsi128_ps( _mm_srli_epi64( _mm_castps_si128(alp), 32 ) ), STBIR_zeroones )
 1407+  #define stbir__simdf_1a1a( out, alp, ones) (out) = _mm_or_ps( _mm_castsi128_ps( _mm_slli_epi64( _mm_castps_si128(alp), 32 ) ), STBIR_onezeros )
 1408+
 1409+  #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) ) )
 1410+
 1411+  #define stbir__simdi_and( out, reg0, reg1 ) (out) = _mm_and_si128( reg0, reg1 )
 1412+  #define stbir__simdi_or( out, reg0, reg1 ) (out) = _mm_or_si128( reg0, reg1 )
 1413+  #define stbir__simdi_16madd( out, reg0, reg1 ) (out) = _mm_madd_epi16( reg0, reg1 )
 1414+
 1415+  #define stbir__simdf_pack_to_8bytes(out,aa,bb) \
 1416+  { \
 1417+    stbir__simdf af,bf; \
 1418+    stbir__simdi a,b; \
 1419+    af = _mm_min_ps( aa, STBIR_max_uint8_as_float ); \
 1420+    bf = _mm_min_ps( bb, STBIR_max_uint8_as_float ); \
 1421+    af = _mm_max_ps( af, _mm_setzero_ps() ); \
 1422+    bf = _mm_max_ps( bf, _mm_setzero_ps() ); \
 1423+    a = _mm_cvttps_epi32( af ); \
 1424+    b = _mm_cvttps_epi32( bf ); \
 1425+    a = _mm_packs_epi32( a, b ); \
 1426+    out = _mm_packus_epi16( a, a ); \
 1427+  }
 1428+
 1429+  #define stbir__simdf_load4_transposed( o0, o1, o2, o3, ptr ) \
 1430+      stbir__simdf_load( o0, (ptr) );    \
 1431+      stbir__simdf_load( o1, (ptr)+4 );  \
 1432+      stbir__simdf_load( o2, (ptr)+8 );  \
 1433+      stbir__simdf_load( o3, (ptr)+12 ); \
 1434+      {                                  \
 1435+        __m128 tmp0, tmp1, tmp2, tmp3;   \
 1436+        tmp0 = _mm_unpacklo_ps(o0, o1);  \
 1437+        tmp2 = _mm_unpacklo_ps(o2, o3);  \
 1438+        tmp1 = _mm_unpackhi_ps(o0, o1);  \
 1439+        tmp3 = _mm_unpackhi_ps(o2, o3);  \
 1440+        o0 = _mm_movelh_ps(tmp0, tmp2);  \
 1441+        o1 = _mm_movehl_ps(tmp2, tmp0);  \
 1442+        o2 = _mm_movelh_ps(tmp1, tmp3);  \
 1443+        o3 = _mm_movehl_ps(tmp3, tmp1);  \
 1444+      }
 1445+
 1446+  #define stbir__interleave_pack_and_store_16_u8( ptr, r0, r1, r2, r3 ) \
 1447+      r0 = _mm_packs_epi32( r0, r1 ); \
 1448+      r2 = _mm_packs_epi32( r2, r3 ); \
 1449+      r1 = _mm_unpacklo_epi16( r0, r2 ); \
 1450+      r3 = _mm_unpackhi_epi16( r0, r2 ); \
 1451+      r0 = _mm_unpacklo_epi16( r1, r3 ); \
 1452+      r2 = _mm_unpackhi_epi16( r1, r3 ); \
 1453+      r0 = _mm_packus_epi16( r0, r2 ); \
 1454+      stbir__simdi_store( ptr, r0 ); \
 1455+
 1456+  #define stbir__simdi_32shr( out, reg, imm ) out = _mm_srli_epi32( reg, imm )
 1457+
 1458+  #if defined(_MSC_VER) && !defined(__clang__)
 1459+    // msvc inits with 8 bytes
 1460+    #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)
 1461+    #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 )
 1462+    #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 )
 1463+  #else
 1464+    // everything else inits with long long's
 1465+    #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)))
 1466+    #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)))
 1467+  #endif
 1468+
 1469+  #define STBIR__SIMDF_CONST(var, x) stbir__simdf var = { x, x, x, x }
 1470+  #define STBIR__SIMDI_CONST(var, x) stbir__simdi var = { STBIR__CONST_4_32i(x) }
 1471+  #define STBIR__CONSTF(var) (var)
 1472+  #define STBIR__CONSTI(var) (var)
 1473+
 1474+  #if defined(STBIR_AVX) || defined(__SSE4_1__)
 1475+    #include <smmintrin.h>
 1476+    #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())))
 1477+  #else
 1478+    static STBIR__SIMDI_CONST(stbir__s32_32768, 32768);
 1479+    static STBIR__SIMDI_CONST(stbir__s16_32768, ((32768<<16)|32768));
 1480+
 1481+    #define stbir__simdf_pack_to_8words(out,reg0,reg1) \
 1482+      { \
 1483+        stbir__simdi tmp0,tmp1; \
 1484+        tmp0 = _mm_cvttps_epi32(_mm_max_ps(_mm_min_ps(reg0,STBIR__CONSTF(STBIR_max_uint16_as_float)),_mm_setzero_ps())); \
 1485+        tmp1 = _mm_cvttps_epi32(_mm_max_ps(_mm_min_ps(reg1,STBIR__CONSTF(STBIR_max_uint16_as_float)),_mm_setzero_ps())); \
 1486+        tmp0 = _mm_sub_epi32( tmp0, stbir__s32_32768 ); \
 1487+        tmp1 = _mm_sub_epi32( tmp1, stbir__s32_32768 ); \
 1488+        out = _mm_packs_epi32( tmp0, tmp1 ); \
 1489+        out = _mm_sub_epi16( out, stbir__s16_32768 ); \
 1490+      }
 1491+
 1492+  #endif
 1493+
 1494+  #define STBIR_SIMD
 1495+
 1496+  // if we detect AVX, set the simd8 defines
 1497+  #ifdef STBIR_AVX
 1498+    #include <immintrin.h>
 1499+    #define STBIR_SIMD8
 1500+    #define stbir__simdf8 __m256
 1501+    #define stbir__simdi8 __m256i
 1502+    #define stbir__simdf8_load( out, ptr ) (out) = _mm256_loadu_ps( (float const *)(ptr) )
 1503+    #define stbir__simdi8_load( out, ptr ) (out) = _mm256_loadu_si256( (__m256i const *)(ptr) )
 1504+    #define stbir__simdf8_mult( out, a, b ) (out) = _mm256_mul_ps( (a), (b) )
 1505+    #define stbir__simdf8_store( ptr, out ) _mm256_storeu_ps( (float*)(ptr), out )
 1506+    #define stbir__simdi8_store( ptr, reg )  _mm256_storeu_si256( (__m256i*)(ptr), reg )
 1507+    #define stbir__simdf8_frep8( fval ) _mm256_set1_ps( fval )
 1508+
 1509+    #define stbir__simdf8_min( out, reg0, reg1 ) (out) = _mm256_min_ps( reg0, reg1 )
 1510+    #define stbir__simdf8_max( out, reg0, reg1 ) (out) = _mm256_max_ps( reg0, reg1 )
 1511+
 1512+    #define stbir__simdf8_add4halves( out, bot4, top8 ) (out) = _mm_add_ps( bot4, _mm256_extractf128_ps( top8, 1 ) )
 1513+    #define stbir__simdf8_mult_mem( out, reg, ptr ) (out) = _mm256_mul_ps( reg, _mm256_loadu_ps( (float const*)(ptr) ) )
 1514+    #define stbir__simdf8_add_mem( out, reg, ptr ) (out) = _mm256_add_ps( reg, _mm256_loadu_ps( (float const*)(ptr) ) )
 1515+    #define stbir__simdf8_add( out, a, b ) (out) = _mm256_add_ps( a, b )
 1516+    #define stbir__simdf8_load1b( out, ptr ) (out) = _mm256_broadcast_ss( ptr )
 1517+    #define stbir__simdf_load1rep4( out, ptr ) (out) = _mm_broadcast_ss( ptr )  // avx load instruction
 1518+
 1519+    #define stbir__simdi8_convert_i32_to_float(out, ireg) (out) = _mm256_cvtepi32_ps( ireg )
 1520+    #define stbir__simdf8_convert_float_to_i32( i, f ) (i) = _mm256_cvttps_epi32(f)
 1521+
 1522+    #define stbir__simdf8_bot4s( out, a, b ) (out) = _mm256_permute2f128_ps(a,b, (0<<0)+(2<<4) )
 1523+    #define stbir__simdf8_top4s( out, a, b ) (out) = _mm256_permute2f128_ps(a,b, (1<<0)+(3<<4) )
 1524+
 1525+    #define stbir__simdf8_gettop4( reg ) _mm256_extractf128_ps(reg,1)
 1526+
 1527+    #ifdef STBIR_AVX2
 1528+
 1529+    #define stbir__simdi8_expand_u8_to_u32(out0,out1,ireg) \
 1530+    { \
 1531+      stbir__simdi8 a, zero  =_mm256_setzero_si256();\
 1532+      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)); \
 1533+      out0 = _mm256_unpacklo_epi16( a, zero ); \
 1534+      out1 = _mm256_unpackhi_epi16( a, zero ); \
 1535+    }
 1536+
 1537+    #define stbir__simdf8_pack_to_16bytes(out,aa,bb) \
 1538+    { \
 1539+      stbir__simdi8 t; \
 1540+      stbir__simdf8 af,bf; \
 1541+      stbir__simdi8 a,b; \
 1542+      af = _mm256_min_ps( aa, STBIR_max_uint8_as_floatX ); \
 1543+      bf = _mm256_min_ps( bb, STBIR_max_uint8_as_floatX ); \
 1544+      af = _mm256_max_ps( af, _mm256_setzero_ps() ); \
 1545+      bf = _mm256_max_ps( bf, _mm256_setzero_ps() ); \
 1546+      a = _mm256_cvttps_epi32( af ); \
 1547+      b = _mm256_cvttps_epi32( bf ); \
 1548+      t = _mm256_permute4x64_epi64( _mm256_packs_epi32( a, b ), (0<<0)+(2<<2)+(1<<4)+(3<<6) ); \
 1549+      out = _mm256_castsi256_si128( _mm256_permute4x64_epi64( _mm256_packus_epi16( t, t ), (0<<0)+(2<<2)+(1<<4)+(3<<6) ) ); \
 1550+    }
 1551+
 1552+    #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() );
 1553+
 1554+    #define stbir__simdf8_pack_to_16words(out,aa,bb) \
 1555+      { \
 1556+        stbir__simdf8 af,bf; \
 1557+        stbir__simdi8 a,b; \
 1558+        af = _mm256_min_ps( aa, STBIR_max_uint16_as_floatX ); \
 1559+        bf = _mm256_min_ps( bb, STBIR_max_uint16_as_floatX ); \
 1560+        af = _mm256_max_ps( af, _mm256_setzero_ps() ); \
 1561+        bf = _mm256_max_ps( bf, _mm256_setzero_ps() ); \
 1562+        a = _mm256_cvttps_epi32( af ); \
 1563+        b = _mm256_cvttps_epi32( bf ); \
 1564+        (out) = _mm256_permute4x64_epi64( _mm256_packus_epi32(a, b), (0<<0)+(2<<2)+(1<<4)+(3<<6) ); \
 1565+      }
 1566+
 1567+    #else
 1568+
 1569+    #define stbir__simdi8_expand_u8_to_u32(out0,out1,ireg) \
 1570+    { \
 1571+      stbir__simdi a,zero = _mm_setzero_si128(); \
 1572+      a = _mm_unpacklo_epi8( ireg, zero ); \
 1573+      out0 = _mm256_setr_m128i( _mm_unpacklo_epi16( a, zero ), _mm_unpackhi_epi16( a, zero ) ); \
 1574+      a = _mm_unpackhi_epi8( ireg, zero ); \
 1575+      out1 = _mm256_setr_m128i( _mm_unpacklo_epi16( a, zero ), _mm_unpackhi_epi16( a, zero ) ); \
 1576+    }
 1577+
 1578+    #define stbir__simdf8_pack_to_16bytes(out,aa,bb) \
 1579+    { \
 1580+      stbir__simdi t; \
 1581+      stbir__simdf8 af,bf; \
 1582+      stbir__simdi8 a,b; \
 1583+      af = _mm256_min_ps( aa, STBIR_max_uint8_as_floatX ); \
 1584+      bf = _mm256_min_ps( bb, STBIR_max_uint8_as_floatX ); \
 1585+      af = _mm256_max_ps( af, _mm256_setzero_ps() ); \
 1586+      bf = _mm256_max_ps( bf, _mm256_setzero_ps() ); \
 1587+      a = _mm256_cvttps_epi32( af ); \
 1588+      b = _mm256_cvttps_epi32( bf ); \
 1589+      out = _mm_packs_epi32( _mm256_castsi256_si128(a), _mm256_extractf128_si256( a, 1 ) ); \
 1590+      out = _mm_packus_epi16( out, out ); \
 1591+      t = _mm_packs_epi32( _mm256_castsi256_si128(b), _mm256_extractf128_si256( b, 1 ) ); \
 1592+      t = _mm_packus_epi16( t, t ); \
 1593+      out = _mm_castps_si128( _mm_shuffle_ps( _mm_castsi128_ps(out), _mm_castsi128_ps(t), (0<<0)+(1<<2)+(0<<4)+(1<<6) ) ); \
 1594+    }
 1595+
 1596+    #define stbir__simdi8_expand_u16_to_u32(out,ireg) \
 1597+    { \
 1598+      stbir__simdi a,b,zero = _mm_setzero_si128(); \
 1599+      a = _mm_unpacklo_epi16( ireg, zero ); \
 1600+      b = _mm_unpackhi_epi16( ireg, zero ); \
 1601+      out = _mm256_insertf128_si256( _mm256_castsi128_si256( a ), b, 1 ); \
 1602+    }
 1603+
 1604+    #define stbir__simdf8_pack_to_16words(out,aa,bb) \
 1605+      { \
 1606+        stbir__simdi t0,t1; \
 1607+        stbir__simdf8 af,bf; \
 1608+        stbir__simdi8 a,b; \
 1609+        af = _mm256_min_ps( aa, STBIR_max_uint16_as_floatX ); \
 1610+        bf = _mm256_min_ps( bb, STBIR_max_uint16_as_floatX ); \
 1611+        af = _mm256_max_ps( af, _mm256_setzero_ps() ); \
 1612+        bf = _mm256_max_ps( bf, _mm256_setzero_ps() ); \
 1613+        a = _mm256_cvttps_epi32( af ); \
 1614+        b = _mm256_cvttps_epi32( bf ); \
 1615+        t0 = _mm_packus_epi32( _mm256_castsi256_si128(a), _mm256_extractf128_si256( a, 1 ) ); \
 1616+        t1 = _mm_packus_epi32( _mm256_castsi256_si128(b), _mm256_extractf128_si256( b, 1 ) ); \
 1617+        out = _mm256_setr_m128i( t0, t1 ); \
 1618+      }
 1619+
 1620+    #endif
 1621+
 1622+    static __m256i stbir_00001111 = { STBIR__CONST_4d_32i( 0, 0, 0, 0 ), STBIR__CONST_4d_32i( 1, 1, 1, 1 ) };
 1623+    #define stbir__simdf8_0123to00001111( out, in ) (out) = _mm256_permutevar_ps ( in, stbir_00001111 )
 1624+
 1625+    static __m256i stbir_22223333 = { STBIR__CONST_4d_32i( 2, 2, 2, 2 ), STBIR__CONST_4d_32i( 3, 3, 3, 3 ) };
 1626+    #define stbir__simdf8_0123to22223333( out, in ) (out) = _mm256_permutevar_ps ( in, stbir_22223333 )
 1627+
 1628+    #define stbir__simdf8_0123to2222( out, in ) (out) = stbir__simdf_swiz(_mm256_castps256_ps128(in), 2,2,2,2 )
 1629+
 1630+    #define stbir__simdf8_load4b( out, ptr ) (out) = _mm256_broadcast_ps( (__m128 const *)(ptr) )
 1631+
 1632+    static __m256i stbir_00112233 = { STBIR__CONST_4d_32i( 0, 0, 1, 1 ), STBIR__CONST_4d_32i( 2, 2, 3, 3 ) };
 1633+    #define stbir__simdf8_0123to00112233( out, in ) (out) = _mm256_permutevar_ps ( in, stbir_00112233 )
 1634+    #define stbir__simdf8_add4( out, a8, b ) (out) = _mm256_add_ps( a8,  _mm256_castps128_ps256( b ) )
 1635+
 1636+    static __m256i stbir_load6 = { STBIR__CONST_4_32i( 0x80000000 ), STBIR__CONST_4d_32i(  0x80000000,  0x80000000, 0, 0 ) };
 1637+    #define stbir__simdf8_load6z( out, ptr ) (out) = _mm256_maskload_ps( ptr, stbir_load6 )
 1638+
 1639+    #define stbir__simdf8_0123to00000000( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (0<<0)+(0<<2)+(0<<4)+(0<<6) )
 1640+    #define stbir__simdf8_0123to11111111( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (1<<0)+(1<<2)+(1<<4)+(1<<6) )
 1641+    #define stbir__simdf8_0123to22222222( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (2<<0)+(2<<2)+(2<<4)+(2<<6) )
 1642+    #define stbir__simdf8_0123to33333333( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (3<<0)+(3<<2)+(3<<4)+(3<<6) )
 1643+    #define stbir__simdf8_0123to21032103( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (2<<0)+(1<<2)+(0<<4)+(3<<6) )
 1644+    #define stbir__simdf8_0123to32103210( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (3<<0)+(2<<2)+(1<<4)+(0<<6) )
 1645+    #define stbir__simdf8_0123to12301230( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (1<<0)+(2<<2)+(3<<4)+(0<<6) )
 1646+    #define stbir__simdf8_0123to10321032( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (1<<0)+(0<<2)+(3<<4)+(2<<6) )
 1647+    #define stbir__simdf8_0123to30123012( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (3<<0)+(0<<2)+(1<<4)+(2<<6) )
 1648+
 1649+    #define stbir__simdf8_0123to11331133( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (1<<0)+(1<<2)+(3<<4)+(3<<6) )
 1650+    #define stbir__simdf8_0123to00220022( out, in ) (out) =  _mm256_shuffle_ps ( in, in, (0<<0)+(0<<2)+(2<<4)+(2<<6) )
 1651+
 1652+    #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) )
 1653+    #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) )
 1654+    #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) )
 1655+    #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) )
 1656+
 1657+    #define stbir__simdf8_zero( reg ) (reg) = _mm256_setzero_ps()
 1658+
 1659+    #ifdef STBIR_USE_FMA           // not on by default to maintain bit identical simd to non-simd
 1660+    #define stbir__simdf8_madd( out, add, mul1, mul2 ) (out) = _mm256_fmadd_ps( mul1, mul2, add )
 1661+    #define stbir__simdf8_madd_mem( out, add, mul, ptr ) (out) = _mm256_fmadd_ps( mul, _mm256_loadu_ps( (float const*)(ptr) ), add )
 1662+    #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 )
 1663+    #else
 1664+    #define stbir__simdf8_madd( out, add, mul1, mul2 ) (out) = _mm256_add_ps( add, _mm256_mul_ps( mul1, mul2 ) )
 1665+    #define stbir__simdf8_madd_mem( out, add, mul, ptr ) (out) = _mm256_add_ps( add, _mm256_mul_ps( mul, _mm256_loadu_ps( (float const*)(ptr) ) ) )
 1666+    #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() ) )
 1667+    #endif
 1668+    #define stbir__if_simdf8_cast_to_simdf4( val ) _mm256_castps256_ps128( val )
 1669+
 1670+  #endif
 1671+
 1672+  #ifdef STBIR_FLOORF
 1673+  #undef STBIR_FLOORF
 1674+  #endif
 1675+  #define STBIR_FLOORF stbir_simd_floorf
 1676+  static stbir__inline float stbir_simd_floorf(float x)  // martins floorf
 1677+  {
 1678+    #if defined(STBIR_AVX) || defined(__SSE4_1__) || defined(STBIR_SSE41)
 1679+    __m128 t = _mm_set_ss(x);
 1680+    return _mm_cvtss_f32( _mm_floor_ss(t, t) );
 1681+    #else
 1682+    __m128 f = _mm_set_ss(x);
 1683+    __m128 t = _mm_cvtepi32_ps(_mm_cvttps_epi32(f));
 1684+    __m128 r = _mm_add_ss(t, _mm_and_ps(_mm_cmplt_ss(f, t), _mm_set_ss(-1.0f)));
 1685+    return _mm_cvtss_f32(r);
 1686+    #endif
 1687+  }
 1688+
 1689+  #ifdef STBIR_CEILF
 1690+  #undef STBIR_CEILF
 1691+  #endif
 1692+  #define STBIR_CEILF stbir_simd_ceilf
 1693+  static stbir__inline float stbir_simd_ceilf(float x)  // martins ceilf
 1694+  {
 1695+    #if defined(STBIR_AVX) || defined(__SSE4_1__) || defined(STBIR_SSE41)
 1696+    __m128 t = _mm_set_ss(x);
 1697+    return _mm_cvtss_f32( _mm_ceil_ss(t, t) );
 1698+    #else
 1699+    __m128 f = _mm_set_ss(x);
 1700+    __m128 t = _mm_cvtepi32_ps(_mm_cvttps_epi32(f));
 1701+    __m128 r = _mm_add_ss(t, _mm_and_ps(_mm_cmplt_ss(t, f), _mm_set_ss(1.0f)));
 1702+    return _mm_cvtss_f32(r);
 1703+    #endif
 1704+  }
 1705+
 1706+#elif defined(STBIR_NEON)
 1707+
 1708+  #include <arm_neon.h>
 1709+
 1710+  #define stbir__simdf float32x4_t
 1711+  #define stbir__simdi uint32x4_t
 1712+
 1713+  #define stbir_simdi_castf( reg ) vreinterpretq_u32_f32(reg)
 1714+  #define stbir_simdf_casti( reg ) vreinterpretq_f32_u32(reg)
 1715+
 1716+  #define stbir__simdf_load( reg, ptr ) (reg) = vld1q_f32( (float const*)(ptr) )
 1717+  #define stbir__simdi_load( reg, ptr ) (reg) = vld1q_u32( (uint32_t const*)(ptr) )
 1718+  #define stbir__simdf_load1( out, ptr ) (out) = vld1q_dup_f32( (float const*)(ptr) ) // top values can be random (not denormal or nan for perf)
 1719+  #define stbir__simdi_load1( out, ptr ) (out) = vld1q_dup_u32( (uint32_t const*)(ptr) )
 1720+  #define stbir__simdf_load1z( out, ptr ) (out) = vld1q_lane_f32( (float const*)(ptr), vdupq_n_f32(0), 0 ) // top values must be zero
 1721+  #define stbir__simdf_frep4( fvar ) vdupq_n_f32( fvar )
 1722+  #define stbir__simdf_load1frep4( out, fvar ) (out) = vdupq_n_f32( fvar )
 1723+  #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)
 1724+  #define stbir__simdf_load2z( out, ptr ) (out) = vcombine_f32( vld1_f32( (float const*)(ptr) ), vcreate_f32(0) )  // top values must be zero
 1725+  #define stbir__simdf_load2hmerge( out, reg, ptr ) (out) = vcombine_f32( vget_low_f32(reg), vld1_f32( (float const*)(ptr) ) )
 1726+
 1727+  #define stbir__simdf_zeroP() vdupq_n_f32(0)
 1728+  #define stbir__simdf_zero( reg ) (reg) = vdupq_n_f32(0)
 1729+
 1730+  #define stbir__simdf_store( ptr, reg )  vst1q_f32( (float*)(ptr), reg )
 1731+  #define stbir__simdf_store1( ptr, reg ) vst1q_lane_f32( (float*)(ptr), reg, 0)
 1732+  #define stbir__simdf_store2( ptr, reg ) vst1_f32( (float*)(ptr), vget_low_f32(reg) )
 1733+  #define stbir__simdf_store2h( ptr, reg ) vst1_f32( (float*)(ptr), vget_high_f32(reg) )
 1734+
 1735+  #define stbir__simdi_store( ptr, reg )  vst1q_u32( (uint32_t*)(ptr), reg )
 1736+  #define stbir__simdi_store1( ptr, reg ) vst1q_lane_u32( (uint32_t*)(ptr), reg, 0 )
 1737+  #define stbir__simdi_store2( ptr, reg ) vst1_u32( (uint32_t*)(ptr), vget_low_u32(reg) )
 1738+
 1739+  #define stbir__prefetch( ptr )
 1740+
 1741+  #define stbir__simdi_expand_u8_to_u32(out0,out1,out2,out3,ireg) \
 1742+  { \
 1743+    uint16x8_t l = vmovl_u8( vget_low_u8 ( vreinterpretq_u8_u32(ireg) ) ); \
 1744+    uint16x8_t h = vmovl_u8( vget_high_u8( vreinterpretq_u8_u32(ireg) ) ); \
 1745+    out0 = vmovl_u16( vget_low_u16 ( l ) ); \
 1746+    out1 = vmovl_u16( vget_high_u16( l ) ); \
 1747+    out2 = vmovl_u16( vget_low_u16 ( h ) ); \
 1748+    out3 = vmovl_u16( vget_high_u16( h ) ); \
 1749+  }
 1750+
 1751+  #define stbir__simdi_expand_u8_to_1u32(out,ireg) \
 1752+  { \
 1753+    uint16x8_t tmp = vmovl_u8( vget_low_u8( vreinterpretq_u8_u32(ireg) ) ); \
 1754+    out = vmovl_u16( vget_low_u16( tmp ) ); \
 1755+  }
 1756+
 1757+  #define stbir__simdi_expand_u16_to_u32(out0,out1,ireg) \
 1758+  { \
 1759+    uint16x8_t tmp = vreinterpretq_u16_u32(ireg); \
 1760+    out0 = vmovl_u16( vget_low_u16 ( tmp ) ); \
 1761+    out1 = vmovl_u16( vget_high_u16( tmp ) ); \
 1762+  }
 1763+
 1764+  #define stbir__simdf_convert_float_to_i32( i, f ) (i) = vreinterpretq_u32_s32( vcvtq_s32_f32(f) )
 1765+  #define stbir__simdf_convert_float_to_int( f ) vgetq_lane_s32(vcvtq_s32_f32(f), 0)
 1766+  #define stbir__simdi_to_int( i ) (int)vgetq_lane_u32(i, 0)
 1767+  #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))
 1768+  #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))
 1769+  #define stbir__simdi_convert_i32_to_float(out, ireg) (out) = vcvtq_f32_s32( vreinterpretq_s32_u32(ireg) )
 1770+  #define stbir__simdf_add( out, reg0, reg1 ) (out) = vaddq_f32( reg0, reg1 )
 1771+  #define stbir__simdf_mult( out, reg0, reg1 ) (out) = vmulq_f32( reg0, reg1 )
 1772+  #define stbir__simdf_mult_mem( out, reg, ptr ) (out) = vmulq_f32( reg, vld1q_f32( (float const*)(ptr) ) )
 1773+  #define stbir__simdf_mult1_mem( out, reg, ptr ) (out) = vmulq_f32( reg, vld1q_dup_f32( (float const*)(ptr) ) )
 1774+  #define stbir__simdf_add_mem( out, reg, ptr ) (out) = vaddq_f32( reg, vld1q_f32( (float const*)(ptr) ) )
 1775+  #define stbir__simdf_add1_mem( out, reg, ptr ) (out) = vaddq_f32( reg, vld1q_dup_f32( (float const*)(ptr) ) )
 1776+
 1777+  #ifdef STBIR_USE_FMA           // not on by default to maintain bit identical simd to non-simd (and also x64 no madd to arm madd)
 1778+  #define stbir__simdf_madd( out, add, mul1, mul2 ) (out) = vfmaq_f32( add, mul1, mul2 )
 1779+  #define stbir__simdf_madd1( out, add, mul1, mul2 ) (out) = vfmaq_f32( add, mul1, mul2 )
 1780+  #define stbir__simdf_madd_mem( out, add, mul, ptr ) (out) = vfmaq_f32( add, mul, vld1q_f32( (float const*)(ptr) ) )
 1781+  #define stbir__simdf_madd1_mem( out, add, mul, ptr ) (out) = vfmaq_f32( add, mul, vld1q_dup_f32( (float const*)(ptr) ) )
 1782+  #else
 1783+  #define stbir__simdf_madd( out, add, mul1, mul2 ) (out) = vaddq_f32( add, vmulq_f32( mul1, mul2 ) )
 1784+  #define stbir__simdf_madd1( out, add, mul1, mul2 ) (out) = vaddq_f32( add, vmulq_f32( mul1, mul2 ) )
 1785+  #define stbir__simdf_madd_mem( out, add, mul, ptr ) (out) = vaddq_f32( add, vmulq_f32( mul, vld1q_f32( (float const*)(ptr) ) ) )
 1786+  #define stbir__simdf_madd1_mem( out, add, mul, ptr ) (out) = vaddq_f32( add, vmulq_f32( mul, vld1q_dup_f32( (float const*)(ptr) ) ) )
 1787+  #endif
 1788+
 1789+  #define stbir__simdf_add1( out, reg0, reg1 ) (out) = vaddq_f32( reg0, reg1 )
 1790+  #define stbir__simdf_mult1( out, reg0, reg1 ) (out) = vmulq_f32( reg0, reg1 )
 1791+
 1792+  #define stbir__simdf_and( out, reg0, reg1 ) (out) = vreinterpretq_f32_u32( vandq_u32( vreinterpretq_u32_f32(reg0), vreinterpretq_u32_f32(reg1) ) )
 1793+  #define stbir__simdf_or( out, reg0, reg1 ) (out) = vreinterpretq_f32_u32( vorrq_u32( vreinterpretq_u32_f32(reg0), vreinterpretq_u32_f32(reg1) ) )
 1794+
 1795+  #define stbir__simdf_min( out, reg0, reg1 ) (out) = vminq_f32( reg0, reg1 )
 1796+  #define stbir__simdf_max( out, reg0, reg1 ) (out) = vmaxq_f32( reg0, reg1 )
 1797+  #define stbir__simdf_min1( out, reg0, reg1 ) (out) = vminq_f32( reg0, reg1 )
 1798+  #define stbir__simdf_max1( out, reg0, reg1 ) (out) = vmaxq_f32( reg0, reg1 )
 1799+
 1800+  #define stbir__simdf_0123ABCDto3ABx( out, reg0, reg1 ) (out) = vextq_f32( reg0, reg1, 3 )
 1801+  #define stbir__simdf_0123ABCDto23Ax( out, reg0, reg1 ) (out) = vextq_f32( reg0, reg1, 2 )
 1802+
 1803+  #define stbir__simdf_a1a1( out, alp, ones ) (out) = vzipq_f32(vuzpq_f32(alp, alp).val[1], ones).val[0]
 1804+  #define stbir__simdf_1a1a( out, alp, ones ) (out) = vzipq_f32(ones, vuzpq_f32(alp, alp).val[0]).val[0]
 1805+
 1806+  #if defined( _M_ARM64 ) || defined( __aarch64__ ) || defined( __arm64__ )
 1807+
 1808+    #define stbir__simdf_aaa1( out, alp, ones ) (out) = vcopyq_laneq_f32(vdupq_n_f32(vgetq_lane_f32(alp, 3)), 3, ones, 3)
 1809+    #define stbir__simdf_1aaa( out, alp, ones ) (out) = vcopyq_laneq_f32(vdupq_n_f32(vgetq_lane_f32(alp, 0)), 0, ones, 0)
 1810+
 1811+    #if defined( _MSC_VER ) && !defined(__clang__)
 1812+      #define stbir_make16(a,b,c,d) vcombine_u8( \
 1813+        vcreate_u8( (4*a+0) | ((4*a+1)<<8) | ((4*a+2)<<16) | ((4*a+3)<<24) | \
 1814+          ((stbir_uint64)(4*b+0)<<32) | ((stbir_uint64)(4*b+1)<<40) | ((stbir_uint64)(4*b+2)<<48) | ((stbir_uint64)(4*b+3)<<56)), \
 1815+        vcreate_u8( (4*c+0) | ((4*c+1)<<8) | ((4*c+2)<<16) | ((4*c+3)<<24) | \
 1816+          ((stbir_uint64)(4*d+0)<<32) | ((stbir_uint64)(4*d+1)<<40) | ((stbir_uint64)(4*d+2)<<48) | ((stbir_uint64)(4*d+3)<<56) ) )
 1817+
 1818+      static stbir__inline uint8x16x2_t stbir_make16x2(float32x4_t rega,float32x4_t regb)
 1819+      {
 1820+        uint8x16x2_t r = { vreinterpretq_u8_f32(rega), vreinterpretq_u8_f32(regb) };
 1821+        return r;
 1822+      }
 1823+    #else
 1824+      #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}
 1825+      #define stbir_make16x2(a,b) (uint8x16x2_t){{vreinterpretq_u8_f32(a),vreinterpretq_u8_f32(b)}}
 1826+    #endif
 1827+
 1828+    #define stbir__simdf_swiz( reg, one, two, three, four ) vreinterpretq_f32_u8( vqtbl1q_u8( vreinterpretq_u8_f32(reg), stbir_make16(one, two, three, four) ) )
 1829+    #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) ) )
 1830+
 1831+    #define stbir__simdi_16madd( out, reg0, reg1 ) \
 1832+    { \
 1833+      int16x8_t r0 = vreinterpretq_s16_u32(reg0); \
 1834+      int16x8_t r1 = vreinterpretq_s16_u32(reg1); \
 1835+      int32x4_t tmp0 = vmull_s16( vget_low_s16(r0), vget_low_s16(r1) ); \
 1836+      int32x4_t tmp1 = vmull_s16( vget_high_s16(r0), vget_high_s16(r1) ); \
 1837+      (out) = vreinterpretq_u32_s32( vpaddq_s32(tmp0, tmp1) ); \
 1838+    }
 1839+
 1840+  #else
 1841+
 1842+    #define stbir__simdf_aaa1( out, alp, ones ) (out) = vsetq_lane_f32(1.0f, vdupq_n_f32(vgetq_lane_f32(alp, 3)), 3)
 1843+    #define stbir__simdf_1aaa( out, alp, ones ) (out) = vsetq_lane_f32(1.0f, vdupq_n_f32(vgetq_lane_f32(alp, 0)), 0)
 1844+
 1845+    #if defined( _MSC_VER ) && !defined(__clang__)
 1846+      static stbir__inline uint8x8x2_t stbir_make8x2(float32x4_t reg)
 1847+      {
 1848+        uint8x8x2_t r = { { vget_low_u8(vreinterpretq_u8_f32(reg)), vget_high_u8(vreinterpretq_u8_f32(reg)) } };
 1849+        return r;
 1850+      }
 1851+      #define stbir_make8(a,b) vcreate_u8( \
 1852+        (4*a+0) | ((4*a+1)<<8) | ((4*a+2)<<16) | ((4*a+3)<<24) | \
 1853+        ((stbir_uint64)(4*b+0)<<32) | ((stbir_uint64)(4*b+1)<<40) | ((stbir_uint64)(4*b+2)<<48) | ((stbir_uint64)(4*b+3)<<56) )
 1854+    #else
 1855+      #define stbir_make8x2(reg) (uint8x8x2_t){ { vget_low_u8(vreinterpretq_u8_f32(reg)), vget_high_u8(vreinterpretq_u8_f32(reg)) } }
 1856+      #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}
 1857+    #endif
 1858+
 1859+    #define stbir__simdf_swiz( reg, one, two, three, four ) vreinterpretq_f32_u8( vcombine_u8( \
 1860+        vtbl2_u8( stbir_make8x2( reg ), stbir_make8( one, two ) ), \
 1861+        vtbl2_u8( stbir_make8x2( reg ), stbir_make8( three, four ) ) ) )
 1862+
 1863+    #define stbir__simdi_16madd( out, reg0, reg1 ) \
 1864+    { \
 1865+      int16x8_t r0 = vreinterpretq_s16_u32(reg0); \
 1866+      int16x8_t r1 = vreinterpretq_s16_u32(reg1); \
 1867+      int32x4_t tmp0 = vmull_s16( vget_low_s16(r0), vget_low_s16(r1) ); \
 1868+      int32x4_t tmp1 = vmull_s16( vget_high_s16(r0), vget_high_s16(r1) ); \
 1869+      int32x2_t out0 = vpadd_s32( vget_low_s32(tmp0), vget_high_s32(tmp0) ); \
 1870+      int32x2_t out1 = vpadd_s32( vget_low_s32(tmp1), vget_high_s32(tmp1) ); \
 1871+      (out) = vreinterpretq_u32_s32( vcombine_s32(out0, out1) ); \
 1872+    }
 1873+
 1874+  #endif
 1875+
 1876+  #define stbir__simdi_and( out, reg0, reg1 ) (out) = vandq_u32( reg0, reg1 )
 1877+  #define stbir__simdi_or( out, reg0, reg1 ) (out) = vorrq_u32( reg0, reg1 )
 1878+
 1879+  #define stbir__simdf_pack_to_8bytes(out,aa,bb) \
 1880+  { \
 1881+    float32x4_t af = vmaxq_f32( vminq_f32(aa,STBIR__CONSTF(STBIR_max_uint8_as_float) ), vdupq_n_f32(0) ); \
 1882+    float32x4_t bf = vmaxq_f32( vminq_f32(bb,STBIR__CONSTF(STBIR_max_uint8_as_float) ), vdupq_n_f32(0) ); \
 1883+    int16x4_t ai = vqmovn_s32( vcvtq_s32_f32( af ) ); \
 1884+    int16x4_t bi = vqmovn_s32( vcvtq_s32_f32( bf ) ); \
 1885+    uint8x8_t out8 = vqmovun_s16( vcombine_s16(ai, bi) ); \
 1886+    out = vreinterpretq_u32_u8( vcombine_u8(out8, out8) ); \
 1887+  }
 1888+
 1889+  #define stbir__simdf_pack_to_8words(out,aa,bb) \
 1890+  { \
 1891+    float32x4_t af = vmaxq_f32( vminq_f32(aa,STBIR__CONSTF(STBIR_max_uint16_as_float) ), vdupq_n_f32(0) ); \
 1892+    float32x4_t bf = vmaxq_f32( vminq_f32(bb,STBIR__CONSTF(STBIR_max_uint16_as_float) ), vdupq_n_f32(0) ); \
 1893+    int32x4_t ai = vcvtq_s32_f32( af ); \
 1894+    int32x4_t bi = vcvtq_s32_f32( bf ); \
 1895+    out = vreinterpretq_u32_u16( vcombine_u16(vqmovun_s32(ai), vqmovun_s32(bi)) ); \
 1896+  }
 1897+
 1898+  #define stbir__interleave_pack_and_store_16_u8( ptr, r0, r1, r2, r3 ) \
 1899+  { \
 1900+    int16x4x2_t tmp0 = vzip_s16( vqmovn_s32(vreinterpretq_s32_u32(r0)), vqmovn_s32(vreinterpretq_s32_u32(r2)) ); \
 1901+    int16x4x2_t tmp1 = vzip_s16( vqmovn_s32(vreinterpretq_s32_u32(r1)), vqmovn_s32(vreinterpretq_s32_u32(r3)) ); \
 1902+    uint8x8x2_t out = \
 1903+    { { \
 1904+      vqmovun_s16( vcombine_s16(tmp0.val[0], tmp0.val[1]) ), \
 1905+      vqmovun_s16( vcombine_s16(tmp1.val[0], tmp1.val[1]) ), \
 1906+    } }; \
 1907+    vst2_u8(ptr, out); \
 1908+  }
 1909+
 1910+  #define stbir__simdf_load4_transposed( o0, o1, o2, o3, ptr ) \
 1911+  { \
 1912+    float32x4x4_t tmp = vld4q_f32(ptr); \
 1913+    o0 = tmp.val[0]; \
 1914+    o1 = tmp.val[1]; \
 1915+    o2 = tmp.val[2]; \
 1916+    o3 = tmp.val[3]; \
 1917+  }
 1918+
 1919+  #define stbir__simdi_32shr( out, reg, imm ) out = vshrq_n_u32( reg, imm )
 1920+
 1921+  #if defined( _MSC_VER ) && !defined(__clang__)
 1922+    #define STBIR__SIMDF_CONST(var, x) __declspec(align(8)) float var[] = { x, x, x, x }
 1923+    #define STBIR__SIMDI_CONST(var, x) __declspec(align(8)) uint32_t var[] = { x, x, x, x }
 1924+    #define STBIR__CONSTF(var) (*(const float32x4_t*)var)
 1925+    #define STBIR__CONSTI(var) (*(const uint32x4_t*)var)
 1926+  #else
 1927+    #define STBIR__SIMDF_CONST(var, x) stbir__simdf var = { x, x, x, x }
 1928+    #define STBIR__SIMDI_CONST(var, x) stbir__simdi var = { x, x, x, x }
 1929+    #define STBIR__CONSTF(var) (var)
 1930+    #define STBIR__CONSTI(var) (var)
 1931+  #endif
 1932+
 1933+  #ifdef STBIR_FLOORF
 1934+  #undef STBIR_FLOORF
 1935+  #endif
 1936+  #define STBIR_FLOORF stbir_simd_floorf
 1937+  static stbir__inline float stbir_simd_floorf(float x)
 1938+  {
 1939+    #if defined( _M_ARM64 ) || defined( __aarch64__ ) || defined( __arm64__ )
 1940+    return vget_lane_f32( vrndm_f32( vdup_n_f32(x) ), 0);
 1941+    #else
 1942+    float32x2_t f = vdup_n_f32(x);
 1943+    float32x2_t t = vcvt_f32_s32(vcvt_s32_f32(f));
 1944+    uint32x2_t a = vclt_f32(f, t);
 1945+    uint32x2_t b = vreinterpret_u32_f32(vdup_n_f32(-1.0f));
 1946+    float32x2_t r = vadd_f32(t, vreinterpret_f32_u32(vand_u32(a, b)));
 1947+    return vget_lane_f32(r, 0);
 1948+    #endif
 1949+  }
 1950+
 1951+  #ifdef STBIR_CEILF
 1952+  #undef STBIR_CEILF
 1953+  #endif
 1954+  #define STBIR_CEILF stbir_simd_ceilf
 1955+  static stbir__inline float stbir_simd_ceilf(float x)
 1956+  {
 1957+    #if defined( _M_ARM64 ) || defined( __aarch64__ ) || defined( __arm64__ )
 1958+    return vget_lane_f32( vrndp_f32( vdup_n_f32(x) ), 0);
 1959+    #else
 1960+    float32x2_t f = vdup_n_f32(x);
 1961+    float32x2_t t = vcvt_f32_s32(vcvt_s32_f32(f));
 1962+    uint32x2_t a = vclt_f32(t, f);
 1963+    uint32x2_t b = vreinterpret_u32_f32(vdup_n_f32(1.0f));
 1964+    float32x2_t r = vadd_f32(t, vreinterpret_f32_u32(vand_u32(a, b)));
 1965+    return vget_lane_f32(r, 0);
 1966+    #endif
 1967+  }
 1968+
 1969+  #define STBIR_SIMD
 1970+
 1971+#elif defined(STBIR_WASM)
 1972+
 1973+  #include <wasm_simd128.h>
 1974+
 1975+  #define stbir__simdf v128_t
 1976+  #define stbir__simdi v128_t
 1977+
 1978+  #define stbir_simdi_castf( reg ) (reg)
 1979+  #define stbir_simdf_casti( reg ) (reg)
 1980+
 1981+  #define stbir__simdf_load( reg, ptr )             (reg) = wasm_v128_load( (void const*)(ptr) )
 1982+  #define stbir__simdi_load( reg, ptr )             (reg) = wasm_v128_load( (void const*)(ptr) )
 1983+  #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)
 1984+  #define stbir__simdi_load1( out, ptr )            (out) = wasm_v128_load32_splat( (void const*)(ptr) )
 1985+  #define stbir__simdf_load1z( out, ptr )           (out) = wasm_v128_load32_zero( (void const*)(ptr) ) // top values must be zero
 1986+  #define stbir__simdf_frep4( fvar )                wasm_f32x4_splat( fvar )
 1987+  #define stbir__simdf_load1frep4( out, fvar )      (out) = wasm_f32x4_splat( fvar )
 1988+  #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)
 1989+  #define stbir__simdf_load2z( out, ptr )           (out) = wasm_v128_load64_zero( (void const*)(ptr) ) // top values must be zero
 1990+  #define stbir__simdf_load2hmerge( out, reg, ptr ) (out) = wasm_v128_load64_lane( (void const*)(ptr), reg, 1 )
 1991+
 1992+  #define stbir__simdf_zeroP() wasm_f32x4_const_splat(0)
 1993+  #define stbir__simdf_zero( reg ) (reg) = wasm_f32x4_const_splat(0)
 1994+
 1995+  #define stbir__simdf_store( ptr, reg )   wasm_v128_store( (void*)(ptr), reg )
 1996+  #define stbir__simdf_store1( ptr, reg )  wasm_v128_store32_lane( (void*)(ptr), reg, 0 )
 1997+  #define stbir__simdf_store2( ptr, reg )  wasm_v128_store64_lane( (void*)(ptr), reg, 0 )
 1998+  #define stbir__simdf_store2h( ptr, reg ) wasm_v128_store64_lane( (void*)(ptr), reg, 1 )
 1999+
 2000+  #define stbir__simdi_store( ptr, reg )  wasm_v128_store( (void*)(ptr), reg )
 2001+  #define stbir__simdi_store1( ptr, reg ) wasm_v128_store32_lane( (void*)(ptr), reg, 0 )
 2002+  #define stbir__simdi_store2( ptr, reg ) wasm_v128_store64_lane( (void*)(ptr), reg, 0 )
 2003+
 2004+  #define stbir__prefetch( ptr )
 2005+
 2006+  #define stbir__simdi_expand_u8_to_u32(out0,out1,out2,out3,ireg) \
 2007+  { \
 2008+    v128_t l = wasm_u16x8_extend_low_u8x16 ( ireg ); \
 2009+    v128_t h = wasm_u16x8_extend_high_u8x16( ireg ); \
 2010+    out0 = wasm_u32x4_extend_low_u16x8 ( l ); \
 2011+    out1 = wasm_u32x4_extend_high_u16x8( l ); \
 2012+    out2 = wasm_u32x4_extend_low_u16x8 ( h ); \
 2013+    out3 = wasm_u32x4_extend_high_u16x8( h ); \
 2014+  }
 2015+
 2016+  #define stbir__simdi_expand_u8_to_1u32(out,ireg) \
 2017+  { \
 2018+    v128_t tmp = wasm_u16x8_extend_low_u8x16(ireg); \
 2019+    out = wasm_u32x4_extend_low_u16x8(tmp); \
 2020+  }
 2021+
 2022+  #define stbir__simdi_expand_u16_to_u32(out0,out1,ireg) \
 2023+  { \
 2024+    out0 = wasm_u32x4_extend_low_u16x8 ( ireg ); \
 2025+    out1 = wasm_u32x4_extend_high_u16x8( ireg ); \
 2026+  }
 2027+
 2028+  #define stbir__simdf_convert_float_to_i32( i, f )    (i) = wasm_i32x4_trunc_sat_f32x4(f)
 2029+  #define stbir__simdf_convert_float_to_int( f )       wasm_i32x4_extract_lane(wasm_i32x4_trunc_sat_f32x4(f), 0)
 2030+  #define stbir__simdi_to_int( i )                     wasm_i32x4_extract_lane(i, 0)
 2031+  #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))
 2032+  #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))
 2033+  #define stbir__simdi_convert_i32_to_float(out, ireg) (out) = wasm_f32x4_convert_i32x4(ireg)
 2034+  #define stbir__simdf_add( out, reg0, reg1 )          (out) = wasm_f32x4_add( reg0, reg1 )
 2035+  #define stbir__simdf_mult( out, reg0, reg1 )         (out) = wasm_f32x4_mul( reg0, reg1 )
 2036+  #define stbir__simdf_mult_mem( out, reg, ptr )       (out) = wasm_f32x4_mul( reg, wasm_v128_load( (void const*)(ptr) ) )
 2037+  #define stbir__simdf_mult1_mem( out, reg, ptr )      (out) = wasm_f32x4_mul( reg, wasm_v128_load32_splat( (void const*)(ptr) ) )
 2038+  #define stbir__simdf_add_mem( out, reg, ptr )        (out) = wasm_f32x4_add( reg, wasm_v128_load( (void const*)(ptr) ) )
 2039+  #define stbir__simdf_add1_mem( out, reg, ptr )       (out) = wasm_f32x4_add( reg, wasm_v128_load32_splat( (void const*)(ptr) ) )
 2040+
 2041+  #define stbir__simdf_madd( out, add, mul1, mul2 )    (out) = wasm_f32x4_add( add, wasm_f32x4_mul( mul1, mul2 ) )
 2042+  #define stbir__simdf_madd1( out, add, mul1, mul2 )   (out) = wasm_f32x4_add( add, wasm_f32x4_mul( mul1, mul2 ) )
 2043+  #define stbir__simdf_madd_mem( out, add, mul, ptr )  (out) = wasm_f32x4_add( add, wasm_f32x4_mul( mul, wasm_v128_load( (void const*)(ptr) ) ) )
 2044+  #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) ) ) )
 2045+
 2046+  #define stbir__simdf_add1( out, reg0, reg1 )  (out) = wasm_f32x4_add( reg0, reg1 )
 2047+  #define stbir__simdf_mult1( out, reg0, reg1 ) (out) = wasm_f32x4_mul( reg0, reg1 )
 2048+
 2049+  #define stbir__simdf_and( out, reg0, reg1 ) (out) = wasm_v128_and( reg0, reg1 )
 2050+  #define stbir__simdf_or( out, reg0, reg1 )  (out) = wasm_v128_or( reg0, reg1 )
 2051+
 2052+  #define stbir__simdf_min( out, reg0, reg1 ) (out) = wasm_f32x4_min( reg0, reg1 )
 2053+  #define stbir__simdf_max( out, reg0, reg1 ) (out) = wasm_f32x4_max( reg0, reg1 )
 2054+  #define stbir__simdf_min1( out, reg0, reg1 ) (out) = wasm_f32x4_min( reg0, reg1 )
 2055+  #define stbir__simdf_max1( out, reg0, reg1 ) (out) = wasm_f32x4_max( reg0, reg1 )
 2056+
 2057+  #define stbir__simdf_0123ABCDto3ABx( out, reg0, reg1 ) (out) = wasm_i32x4_shuffle( reg0, reg1, 3, 4, 5, -1 )
 2058+  #define stbir__simdf_0123ABCDto23Ax( out, reg0, reg1 ) (out) = wasm_i32x4_shuffle( reg0, reg1, 2, 3, 4, -1 )
 2059+
 2060+  #define stbir__simdf_aaa1(out,alp,ones) (out) = wasm_i32x4_shuffle(alp, ones, 3, 3, 3, 4)
 2061+  #define stbir__simdf_1aaa(out,alp,ones) (out) = wasm_i32x4_shuffle(alp, ones, 4, 0, 0, 0)
 2062+  #define stbir__simdf_a1a1(out,alp,ones) (out) = wasm_i32x4_shuffle(alp, ones, 1, 4, 3, 4)
 2063+  #define stbir__simdf_1a1a(out,alp,ones) (out) = wasm_i32x4_shuffle(alp, ones, 4, 0, 4, 2)
 2064+
 2065+  #define stbir__simdf_swiz( reg, one, two, three, four ) wasm_i32x4_shuffle(reg, reg, one, two, three, four)
 2066+
 2067+  #define stbir__simdi_and( out, reg0, reg1 )    (out) = wasm_v128_and( reg0, reg1 )
 2068+  #define stbir__simdi_or( out, reg0, reg1 )     (out) = wasm_v128_or( reg0, reg1 )
 2069+  #define stbir__simdi_16madd( out, reg0, reg1 ) (out) = wasm_i32x4_dot_i16x8( reg0, reg1 )
 2070+
 2071+  #define stbir__simdf_pack_to_8bytes(out,aa,bb) \
 2072+  { \
 2073+    v128_t af = wasm_f32x4_max( wasm_f32x4_min(aa, STBIR_max_uint8_as_float), wasm_f32x4_const_splat(0) ); \
 2074+    v128_t bf = wasm_f32x4_max( wasm_f32x4_min(bb, STBIR_max_uint8_as_float), wasm_f32x4_const_splat(0) ); \
 2075+    v128_t ai = wasm_i32x4_trunc_sat_f32x4( af ); \
 2076+    v128_t bi = wasm_i32x4_trunc_sat_f32x4( bf ); \
 2077+    v128_t out16 = wasm_i16x8_narrow_i32x4( ai, bi ); \
 2078+    out = wasm_u8x16_narrow_i16x8( out16, out16 ); \
 2079+  }
 2080+
 2081+  #define stbir__simdf_pack_to_8words(out,aa,bb) \
 2082+  { \
 2083+    v128_t af = wasm_f32x4_max( wasm_f32x4_min(aa, STBIR_max_uint16_as_float), wasm_f32x4_const_splat(0)); \
 2084+    v128_t bf = wasm_f32x4_max( wasm_f32x4_min(bb, STBIR_max_uint16_as_float), wasm_f32x4_const_splat(0)); \
 2085+    v128_t ai = wasm_i32x4_trunc_sat_f32x4( af ); \
 2086+    v128_t bi = wasm_i32x4_trunc_sat_f32x4( bf ); \
 2087+    out = wasm_u16x8_narrow_i32x4( ai, bi ); \
 2088+  }
 2089+
 2090+  #define stbir__interleave_pack_and_store_16_u8( ptr, r0, r1, r2, r3 ) \
 2091+  { \
 2092+    v128_t tmp0 = wasm_i16x8_narrow_i32x4(r0, r1); \
 2093+    v128_t tmp1 = wasm_i16x8_narrow_i32x4(r2, r3); \
 2094+    v128_t tmp = wasm_u8x16_narrow_i16x8(tmp0, tmp1); \
 2095+    tmp = wasm_i8x16_shuffle(tmp, tmp, 0, 4, 8, 12, 1, 5, 9, 13, 2, 6, 10, 14, 3, 7, 11, 15); \
 2096+    wasm_v128_store( (void*)(ptr), tmp); \
 2097+  }
 2098+
 2099+  #define stbir__simdf_load4_transposed( o0, o1, o2, o3, ptr ) \
 2100+  { \
 2101+    v128_t t0 = wasm_v128_load( ptr    ); \
 2102+    v128_t t1 = wasm_v128_load( ptr+4  ); \
 2103+    v128_t t2 = wasm_v128_load( ptr+8  ); \
 2104+    v128_t t3 = wasm_v128_load( ptr+12 ); \
 2105+    v128_t s0 = wasm_i32x4_shuffle(t0, t1, 0, 4, 2, 6); \
 2106+    v128_t s1 = wasm_i32x4_shuffle(t0, t1, 1, 5, 3, 7); \
 2107+    v128_t s2 = wasm_i32x4_shuffle(t2, t3, 0, 4, 2, 6); \
 2108+    v128_t s3 = wasm_i32x4_shuffle(t2, t3, 1, 5, 3, 7); \
 2109+    o0 = wasm_i32x4_shuffle(s0, s2, 0, 1, 4, 5); \
 2110+    o1 = wasm_i32x4_shuffle(s1, s3, 0, 1, 4, 5); \
 2111+    o2 = wasm_i32x4_shuffle(s0, s2, 2, 3, 6, 7); \
 2112+    o3 = wasm_i32x4_shuffle(s1, s3, 2, 3, 6, 7); \
 2113+  }
 2114+
 2115+  #define stbir__simdi_32shr( out, reg, imm ) out = wasm_u32x4_shr( reg, imm )
 2116+
 2117+  typedef float stbir__f32x4 __attribute__((__vector_size__(16), __aligned__(16)));
 2118+  #define STBIR__SIMDF_CONST(var, x) stbir__simdf var = (v128_t)(stbir__f32x4){ x, x, x, x }
 2119+  #define STBIR__SIMDI_CONST(var, x) stbir__simdi var = { x, x, x, x }
 2120+  #define STBIR__CONSTF(var) (var)
 2121+  #define STBIR__CONSTI(var) (var)
 2122+
 2123+  #ifdef STBIR_FLOORF
 2124+  #undef STBIR_FLOORF
 2125+  #endif
 2126+  #define STBIR_FLOORF stbir_simd_floorf
 2127+  static stbir__inline float stbir_simd_floorf(float x)
 2128+  {
 2129+    return wasm_f32x4_extract_lane( wasm_f32x4_floor( wasm_f32x4_splat(x) ), 0);
 2130+  }
 2131+
 2132+  #ifdef STBIR_CEILF
 2133+  #undef STBIR_CEILF
 2134+  #endif
 2135+  #define STBIR_CEILF stbir_simd_ceilf
 2136+  static stbir__inline float stbir_simd_ceilf(float x)
 2137+  {
 2138+    return wasm_f32x4_extract_lane( wasm_f32x4_ceil( wasm_f32x4_splat(x) ), 0);
 2139+  }
 2140+
 2141+  #define STBIR_SIMD
 2142+
 2143+#endif  // SSE2/NEON/WASM
 2144+
 2145+#endif // NO SIMD
 2146+
 2147+#ifdef STBIR_SIMD8
 2148+  #define stbir__simdfX stbir__simdf8
 2149+  #define stbir__simdiX stbir__simdi8
 2150+  #define stbir__simdfX_load stbir__simdf8_load
 2151+  #define stbir__simdiX_load stbir__simdi8_load
 2152+  #define stbir__simdfX_mult stbir__simdf8_mult
 2153+  #define stbir__simdfX_add_mem stbir__simdf8_add_mem
 2154+  #define stbir__simdfX_madd_mem stbir__simdf8_madd_mem
 2155+  #define stbir__simdfX_store stbir__simdf8_store
 2156+  #define stbir__simdiX_store stbir__simdi8_store
 2157+  #define stbir__simdf_frepX  stbir__simdf8_frep8
 2158+  #define stbir__simdfX_madd stbir__simdf8_madd
 2159+  #define stbir__simdfX_min stbir__simdf8_min
 2160+  #define stbir__simdfX_max stbir__simdf8_max
 2161+  #define stbir__simdfX_aaa1 stbir__simdf8_aaa1
 2162+  #define stbir__simdfX_1aaa stbir__simdf8_1aaa
 2163+  #define stbir__simdfX_a1a1 stbir__simdf8_a1a1
 2164+  #define stbir__simdfX_1a1a stbir__simdf8_1a1a
 2165+  #define stbir__simdfX_convert_float_to_i32 stbir__simdf8_convert_float_to_i32
 2166+  #define stbir__simdfX_pack_to_words stbir__simdf8_pack_to_16words
 2167+  #define stbir__simdfX_zero stbir__simdf8_zero
 2168+  #define STBIR_onesX STBIR_ones8
 2169+  #define STBIR_max_uint8_as_floatX STBIR_max_uint8_as_float8
 2170+  #define STBIR_max_uint16_as_floatX STBIR_max_uint16_as_float8
 2171+  #define STBIR_simd_point5X STBIR_simd_point58
 2172+  #define stbir__simdfX_float_count 8
 2173+  #define stbir__simdfX_0123to1230 stbir__simdf8_0123to12301230
 2174+  #define stbir__simdfX_0123to2103 stbir__simdf8_0123to21032103
 2175+  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 };
 2176+  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 };
 2177+  static const stbir__simdf8 STBIR_ones8 = { 1.0,1.0,1.0,1.0,1.0,1.0,1.0,1.0 };
 2178+  static const stbir__simdf8 STBIR_simd_point58 = { 0.5,0.5,0.5,0.5,0.5,0.5,0.5,0.5 };
 2179+  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 };
 2180+  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 };
 2181+#else
 2182+  #define stbir__simdfX stbir__simdf
 2183+  #define stbir__simdiX stbir__simdi
 2184+  #define stbir__simdfX_load stbir__simdf_load
 2185+  #define stbir__simdiX_load stbir__simdi_load
 2186+  #define stbir__simdfX_mult stbir__simdf_mult
 2187+  #define stbir__simdfX_add_mem stbir__simdf_add_mem
 2188+  #define stbir__simdfX_madd_mem stbir__simdf_madd_mem
 2189+  #define stbir__simdfX_store stbir__simdf_store
 2190+  #define stbir__simdiX_store stbir__simdi_store
 2191+  #define stbir__simdf_frepX  stbir__simdf_frep4
 2192+  #define stbir__simdfX_madd stbir__simdf_madd
 2193+  #define stbir__simdfX_min stbir__simdf_min
 2194+  #define stbir__simdfX_max stbir__simdf_max
 2195+  #define stbir__simdfX_aaa1 stbir__simdf_aaa1
 2196+  #define stbir__simdfX_1aaa stbir__simdf_1aaa
 2197+  #define stbir__simdfX_a1a1 stbir__simdf_a1a1
 2198+  #define stbir__simdfX_1a1a stbir__simdf_1a1a
 2199+  #define stbir__simdfX_convert_float_to_i32 stbir__simdf_convert_float_to_i32
 2200+  #define stbir__simdfX_pack_to_words stbir__simdf_pack_to_8words
 2201+  #define stbir__simdfX_zero stbir__simdf_zero
 2202+  #define STBIR_onesX STBIR__CONSTF(STBIR_ones)
 2203+  #define STBIR_simd_point5X STBIR__CONSTF(STBIR_simd_point5)
 2204+  #define STBIR_max_uint8_as_floatX STBIR__CONSTF(STBIR_max_uint8_as_float)
 2205+  #define STBIR_max_uint16_as_floatX STBIR__CONSTF(STBIR_max_uint16_as_float)
 2206+  #define stbir__simdfX_float_count 4
 2207+  #define stbir__if_simdf8_cast_to_simdf4( val ) ( val )
 2208+  #define stbir__simdfX_0123to1230 stbir__simdf_0123to1230
 2209+  #define stbir__simdfX_0123to2103 stbir__simdf_0123to2103
 2210+#endif
 2211+
 2212+
 2213+#if defined(STBIR_NEON) && !defined(_M_ARM) && !defined(__arm__)
 2214+
 2215+  #if defined( _MSC_VER ) && !defined(__clang__)
 2216+  typedef __int16 stbir__FP16;
 2217+  #else
 2218+  typedef float16_t stbir__FP16;
 2219+  #endif
 2220+
 2221+#else // no NEON, or 32-bit ARM for MSVC
 2222+
 2223+  typedef union stbir__FP16
 2224+  {
 2225+    unsigned short u;
 2226+  } stbir__FP16;
 2227+
 2228+#endif
 2229+
 2230+#if (!defined(STBIR_NEON) && !defined(STBIR_FP16C)) || (defined(STBIR_NEON) && defined(_M_ARM)) || (defined(STBIR_NEON) && defined(__arm__))
 2231+
 2232+  // Fabian's half float routines, see: https://gist.github.com/rygorous/2156668
 2233+
 2234+  static stbir__inline float stbir__half_to_float( stbir__FP16 h )
 2235+  {
 2236+    static const stbir__FP32 magic = { (254 - 15) << 23 };
 2237+    static const stbir__FP32 was_infnan = { (127 + 16) << 23 };
 2238+    stbir__FP32 o;
 2239+
 2240+    o.u = (h.u & 0x7fff) << 13;     // exponent/mantissa bits
 2241+    o.f *= magic.f;                 // exponent adjust
 2242+    if (o.f >= was_infnan.f)        // make sure Inf/NaN survive
 2243+      o.u |= 255 << 23;
 2244+    o.u |= (h.u & 0x8000) << 16;    // sign bit
 2245+    return o.f;
 2246+  }
 2247+
 2248+  static stbir__inline stbir__FP16 stbir__float_to_half(float val)
 2249+  {
 2250+    stbir__FP32 f32infty = { 255 << 23 };
 2251+    stbir__FP32 f16max   = { (127 + 16) << 23 };
 2252+    stbir__FP32 denorm_magic = { ((127 - 15) + (23 - 10) + 1) << 23 };
 2253+    unsigned int sign_mask = 0x80000000u;
 2254+    stbir__FP16 o = { 0 };
 2255+    stbir__FP32 f;
 2256+    unsigned int sign;
 2257+
 2258+    f.f = val;
 2259+    sign = f.u & sign_mask;
 2260+    f.u ^= sign;
 2261+
 2262+    if (f.u >= f16max.u) // result is Inf or NaN (all exponent bits set)
 2263+      o.u = (f.u > f32infty.u) ? 0x7e00 : 0x7c00; // NaN->qNaN and Inf->Inf
 2264+    else // (De)normalized number or zero
 2265+    {
 2266+      if (f.u < (113 << 23)) // resulting FP16 is subnormal or zero
 2267+      {
 2268+        // use a magic value to align our 10 mantissa bits at the bottom of
 2269+        // the float. as long as FP addition is round-to-nearest-even this
 2270+        // just works.
 2271+        f.f += denorm_magic.f;
 2272+        // and one integer subtract of the bias later, we have our final float!
 2273+        o.u = (unsigned short) ( f.u - denorm_magic.u );
 2274+      }
 2275+      else
 2276+      {
 2277+        unsigned int mant_odd = (f.u >> 13) & 1; // resulting mantissa is odd
 2278+        // update exponent, rounding bias part 1
 2279+        f.u = f.u + ((15u - 127) << 23) + 0xfff;
 2280+        // rounding bias part 2
 2281+        f.u += mant_odd;
 2282+        // take the bits!
 2283+        o.u = (unsigned short) ( f.u >> 13 );
 2284+      }
 2285+    }
 2286+
 2287+    o.u |= sign >> 16;
 2288+    return o;
 2289+  }
 2290+
 2291+#endif
 2292+
 2293+
 2294+#if defined(STBIR_FP16C)
 2295+
 2296+  #include <immintrin.h>
 2297+
 2298+  static stbir__inline void stbir__half_to_float_SIMD(float * output, stbir__FP16 const * input)
 2299+  {
 2300+    _mm256_storeu_ps( (float*)output, _mm256_cvtph_ps( _mm_loadu_si128( (__m128i const* )input ) ) );
 2301+  }
 2302+
 2303+  static stbir__inline void stbir__float_to_half_SIMD(stbir__FP16 * output, float const * input)
 2304+  {
 2305+    _mm_storeu_si128( (__m128i*)output, _mm256_cvtps_ph( _mm256_loadu_ps( input ), 0 ) );
 2306+  }
 2307+
 2308+  static stbir__inline float stbir__half_to_float( stbir__FP16 h )
 2309+  {
 2310+    return _mm_cvtss_f32( _mm_cvtph_ps( _mm_cvtsi32_si128( (int)h.u ) ) );
 2311+  }
 2312+
 2313+  static stbir__inline stbir__FP16 stbir__float_to_half( float f )
 2314+  {
 2315+    stbir__FP16 h;
 2316+    h.u = (unsigned short) _mm_cvtsi128_si32( _mm_cvtps_ph( _mm_set_ss( f ), 0 ) );
 2317+    return h;
 2318+  }
 2319+
 2320+#elif defined(STBIR_SSE2)
 2321+
 2322+  // Fabian's half float routines, see: https://gist.github.com/rygorous/2156668
 2323+  stbir__inline static void stbir__half_to_float_SIMD(float * output, void const * input)
 2324+  {
 2325+    static const STBIR__SIMDI_CONST(mask_nosign,      0x7fff);
 2326+    static const STBIR__SIMDI_CONST(smallest_normal,  0x0400);
 2327+    static const STBIR__SIMDI_CONST(infinity,         0x7c00);
 2328+    static const STBIR__SIMDI_CONST(expadjust_normal, (127 - 15) << 23);
 2329+    static const STBIR__SIMDI_CONST(magic_denorm,     113 << 23);
 2330+
 2331+    __m128i i = _mm_loadu_si128 ( (__m128i const*)(input) );
 2332+    __m128i h = _mm_unpacklo_epi16 ( i, _mm_setzero_si128() );
 2333+    __m128i mnosign     = STBIR__CONSTI(mask_nosign);
 2334+    __m128i eadjust     = STBIR__CONSTI(expadjust_normal);
 2335+    __m128i smallest    = STBIR__CONSTI(smallest_normal);
 2336+    __m128i infty       = STBIR__CONSTI(infinity);
 2337+    __m128i expmant     = _mm_and_si128(mnosign, h);
 2338+    __m128i justsign    = _mm_xor_si128(h, expmant);
 2339+    __m128i b_notinfnan = _mm_cmpgt_epi32(infty, expmant);
 2340+    __m128i b_isdenorm  = _mm_cmpgt_epi32(smallest, expmant);
 2341+    __m128i shifted     = _mm_slli_epi32(expmant, 13);
 2342+    __m128i adj_infnan  = _mm_andnot_si128(b_notinfnan, eadjust);
 2343+    __m128i adjusted    = _mm_add_epi32(eadjust, shifted);
 2344+    __m128i den1        = _mm_add_epi32(shifted, STBIR__CONSTI(magic_denorm));
 2345+    __m128i adjusted2   = _mm_add_epi32(adjusted, adj_infnan);
 2346+    __m128  den2        = _mm_sub_ps(_mm_castsi128_ps(den1), *(const __m128 *)&magic_denorm);
 2347+    __m128  adjusted3   = _mm_and_ps(den2, _mm_castsi128_ps(b_isdenorm));
 2348+    __m128  adjusted4   = _mm_andnot_ps(_mm_castsi128_ps(b_isdenorm), _mm_castsi128_ps(adjusted2));
 2349+    __m128  adjusted5   = _mm_or_ps(adjusted3, adjusted4);
 2350+    __m128i sign        = _mm_slli_epi32(justsign, 16);
 2351+    __m128  final       = _mm_or_ps(adjusted5, _mm_castsi128_ps(sign));
 2352+    stbir__simdf_store( output + 0,  final );
 2353+
 2354+    h = _mm_unpackhi_epi16 ( i, _mm_setzero_si128() );
 2355+    expmant     = _mm_and_si128(mnosign, h);
 2356+    justsign    = _mm_xor_si128(h, expmant);
 2357+    b_notinfnan = _mm_cmpgt_epi32(infty, expmant);
 2358+    b_isdenorm  = _mm_cmpgt_epi32(smallest, expmant);
 2359+    shifted     = _mm_slli_epi32(expmant, 13);
 2360+    adj_infnan  = _mm_andnot_si128(b_notinfnan, eadjust);
 2361+    adjusted    = _mm_add_epi32(eadjust, shifted);
 2362+    den1        = _mm_add_epi32(shifted, STBIR__CONSTI(magic_denorm));
 2363+    adjusted2   = _mm_add_epi32(adjusted, adj_infnan);
 2364+    den2        = _mm_sub_ps(_mm_castsi128_ps(den1), *(const __m128 *)&magic_denorm);
 2365+    adjusted3   = _mm_and_ps(den2, _mm_castsi128_ps(b_isdenorm));
 2366+    adjusted4   = _mm_andnot_ps(_mm_castsi128_ps(b_isdenorm), _mm_castsi128_ps(adjusted2));
 2367+    adjusted5   = _mm_or_ps(adjusted3, adjusted4);
 2368+    sign        = _mm_slli_epi32(justsign, 16);
 2369+    final       = _mm_or_ps(adjusted5, _mm_castsi128_ps(sign));
 2370+    stbir__simdf_store( output + 4,  final );
 2371+
 2372+    // ~38 SSE2 ops for 8 values
 2373+  }
 2374+
 2375+  // Fabian's round-to-nearest-even float to half
 2376+  // ~48 SSE2 ops for 8 output
 2377+  stbir__inline static void stbir__float_to_half_SIMD(void * output, float const * input)
 2378+  {
 2379+    static const STBIR__SIMDI_CONST(mask_sign,      0x80000000u);
 2380+    static const STBIR__SIMDI_CONST(c_f16max,       (127 + 16) << 23); // all FP32 values >=this round to +inf
 2381+    static const STBIR__SIMDI_CONST(c_nanbit,        0x200);
 2382+    static const STBIR__SIMDI_CONST(c_infty_as_fp16, 0x7c00);
 2383+    static const STBIR__SIMDI_CONST(c_min_normal,    (127 - 14) << 23); // smallest FP32 that yields a normalized FP16
 2384+    static const STBIR__SIMDI_CONST(c_subnorm_magic, ((127 - 15) + (23 - 10) + 1) << 23);
 2385+    static const STBIR__SIMDI_CONST(c_normal_bias,    0xfff - ((127 - 15) << 23)); // adjust exponent and add mantissa rounding
 2386+
 2387+    __m128  f           =  _mm_loadu_ps(input);
 2388+    __m128  msign       = _mm_castsi128_ps(STBIR__CONSTI(mask_sign));
 2389+    __m128  justsign    = _mm_and_ps(msign, f);
 2390+    __m128  absf        = _mm_xor_ps(f, justsign);
 2391+    __m128i absf_int    = _mm_castps_si128(absf); // the cast is "free" (extra bypass latency, but no thruput hit)
 2392+    __m128i f16max      = STBIR__CONSTI(c_f16max);
 2393+    __m128  b_isnan     = _mm_cmpunord_ps(absf, absf); // is this a NaN?
 2394+    __m128i b_isregular = _mm_cmpgt_epi32(f16max, absf_int); // (sub)normalized or special?
 2395+    __m128i nanbit      = _mm_and_si128(_mm_castps_si128(b_isnan), STBIR__CONSTI(c_nanbit));
 2396+    __m128i inf_or_nan  = _mm_or_si128(nanbit, STBIR__CONSTI(c_infty_as_fp16)); // output for specials
 2397+
 2398+    __m128i min_normal  = STBIR__CONSTI(c_min_normal);
 2399+    __m128i b_issub     = _mm_cmpgt_epi32(min_normal, absf_int);
 2400+
 2401+    // "result is subnormal" path
 2402+    __m128  subnorm1    = _mm_add_ps(absf, _mm_castsi128_ps(STBIR__CONSTI(c_subnorm_magic))); // magic value to round output mantissa
 2403+    __m128i subnorm2    = _mm_sub_epi32(_mm_castps_si128(subnorm1), STBIR__CONSTI(c_subnorm_magic)); // subtract out bias
 2404+
 2405+    // "result is normal" path
 2406+    __m128i mantoddbit  = _mm_slli_epi32(absf_int, 31 - 13); // shift bit 13 (mantissa LSB) to sign
 2407+    __m128i mantodd     = _mm_srai_epi32(mantoddbit, 31); // -1 if FP16 mantissa odd, else 0
 2408+
 2409+    __m128i round1      = _mm_add_epi32(absf_int, STBIR__CONSTI(c_normal_bias));
 2410+    __m128i round2      = _mm_sub_epi32(round1, mantodd); // if mantissa LSB odd, bias towards rounding up (RTNE)
 2411+    __m128i normal      = _mm_srli_epi32(round2, 13); // rounded result
 2412+
 2413+    // combine the two non-specials
 2414+    __m128i nonspecial  = _mm_or_si128(_mm_and_si128(subnorm2, b_issub), _mm_andnot_si128(b_issub, normal));
 2415+
 2416+    // merge in specials as well
 2417+    __m128i joined      = _mm_or_si128(_mm_and_si128(nonspecial, b_isregular), _mm_andnot_si128(b_isregular, inf_or_nan));
 2418+
 2419+    __m128i sign_shift  = _mm_srai_epi32(_mm_castps_si128(justsign), 16);
 2420+    __m128i final2, final= _mm_or_si128(joined, sign_shift);
 2421+
 2422+    f           =  _mm_loadu_ps(input+4);
 2423+    justsign    = _mm_and_ps(msign, f);
 2424+    absf        = _mm_xor_ps(f, justsign);
 2425+    absf_int    = _mm_castps_si128(absf); // the cast is "free" (extra bypass latency, but no thruput hit)
 2426+    b_isnan     = _mm_cmpunord_ps(absf, absf); // is this a NaN?
 2427+    b_isregular = _mm_cmpgt_epi32(f16max, absf_int); // (sub)normalized or special?
 2428+    nanbit      = _mm_and_si128(_mm_castps_si128(b_isnan), c_nanbit);
 2429+    inf_or_nan  = _mm_or_si128(nanbit, STBIR__CONSTI(c_infty_as_fp16)); // output for specials
 2430+
 2431+    b_issub     = _mm_cmpgt_epi32(min_normal, absf_int);
 2432+
 2433+    // "result is subnormal" path
 2434+    subnorm1    = _mm_add_ps(absf, _mm_castsi128_ps(STBIR__CONSTI(c_subnorm_magic))); // magic value to round output mantissa
 2435+    subnorm2    = _mm_sub_epi32(_mm_castps_si128(subnorm1), STBIR__CONSTI(c_subnorm_magic)); // subtract out bias
 2436+
 2437+    // "result is normal" path
 2438+    mantoddbit  = _mm_slli_epi32(absf_int, 31 - 13); // shift bit 13 (mantissa LSB) to sign
 2439+    mantodd     = _mm_srai_epi32(mantoddbit, 31); // -1 if FP16 mantissa odd, else 0
 2440+
 2441+    round1      = _mm_add_epi32(absf_int, STBIR__CONSTI(c_normal_bias));
 2442+    round2      = _mm_sub_epi32(round1, mantodd); // if mantissa LSB odd, bias towards rounding up (RTNE)
 2443+    normal      = _mm_srli_epi32(round2, 13); // rounded result
 2444+
 2445+    // combine the two non-specials
 2446+    nonspecial  = _mm_or_si128(_mm_and_si128(subnorm2, b_issub), _mm_andnot_si128(b_issub, normal));
 2447+
 2448+    // merge in specials as well
 2449+    joined      = _mm_or_si128(_mm_and_si128(nonspecial, b_isregular), _mm_andnot_si128(b_isregular, inf_or_nan));
 2450+
 2451+    sign_shift  = _mm_srai_epi32(_mm_castps_si128(justsign), 16);
 2452+    final2      = _mm_or_si128(joined, sign_shift);
 2453+    final       = _mm_packs_epi32(final, final2);
 2454+    stbir__simdi_store( output,final );
 2455+  }
 2456+
 2457+#elif defined(STBIR_NEON) && defined(_MSC_VER) && defined(_M_ARM64) && !defined(__clang__) // 64-bit ARM on MSVC (not clang)
 2458+
 2459+  static stbir__inline void stbir__half_to_float_SIMD(float * output, stbir__FP16 const * input)
 2460+  {
 2461+    float16x4_t in0 = vld1_f16(input + 0);
 2462+    float16x4_t in1 = vld1_f16(input + 4);
 2463+    vst1q_f32(output + 0, vcvt_f32_f16(in0));
 2464+    vst1q_f32(output + 4, vcvt_f32_f16(in1));
 2465+  }
 2466+
 2467+  static stbir__inline void stbir__float_to_half_SIMD(stbir__FP16 * output, float const * input)
 2468+  {
 2469+    float16x4_t out0 = vcvt_f16_f32(vld1q_f32(input + 0));
 2470+    float16x4_t out1 = vcvt_f16_f32(vld1q_f32(input + 4));
 2471+    vst1_f16(output+0, out0);
 2472+    vst1_f16(output+4, out1);
 2473+  }
 2474+
 2475+  static stbir__inline float stbir__half_to_float( stbir__FP16 h )
 2476+  {
 2477+    return vgetq_lane_f32(vcvt_f32_f16(vld1_dup_f16(&h)), 0);
 2478+  }
 2479+
 2480+  static stbir__inline stbir__FP16 stbir__float_to_half( float f )
 2481+  {
 2482+    return vget_lane_f16(vcvt_f16_f32(vdupq_n_f32(f)), 0).n16_u16[0];
 2483+  }
 2484+
 2485+#elif defined(STBIR_NEON) && ( defined( _M_ARM64 ) || defined( __aarch64__ ) || defined( __arm64__ ) ) // 64-bit ARM
 2486+
 2487+  static stbir__inline void stbir__half_to_float_SIMD(float * output, stbir__FP16 const * input)
 2488+  {
 2489+    float16x8_t in = vld1q_f16(input);
 2490+    vst1q_f32(output + 0, vcvt_f32_f16(vget_low_f16(in)));
 2491+    vst1q_f32(output + 4, vcvt_f32_f16(vget_high_f16(in)));
 2492+  }
 2493+
 2494+  static stbir__inline void stbir__float_to_half_SIMD(stbir__FP16 * output, float const * input)
 2495+  {
 2496+    float16x4_t out0 = vcvt_f16_f32(vld1q_f32(input + 0));
 2497+    float16x4_t out1 = vcvt_f16_f32(vld1q_f32(input + 4));
 2498+    vst1q_f16(output, vcombine_f16(out0, out1));
 2499+  }
 2500+
 2501+  static stbir__inline float stbir__half_to_float( stbir__FP16 h )
 2502+  {
 2503+    return vgetq_lane_f32(vcvt_f32_f16(vdup_n_f16(h)), 0);
 2504+  }
 2505+
 2506+  static stbir__inline stbir__FP16 stbir__float_to_half( float f )
 2507+  {
 2508+    return vget_lane_f16(vcvt_f16_f32(vdupq_n_f32(f)), 0);
 2509+  }
 2510+
 2511+#elif defined(STBIR_WASM) || (defined(STBIR_NEON) && (defined(_MSC_VER) || defined(_M_ARM) || defined(__arm__))) // WASM or 32-bit ARM on MSVC/clang
 2512+
 2513+  static stbir__inline void stbir__half_to_float_SIMD(float * output, stbir__FP16 const * input)
 2514+  {
 2515+    for (int i=0; i<8; i++)
 2516+    {
 2517+      output[i] = stbir__half_to_float(input[i]);
 2518+    }
 2519+  }
 2520+  static stbir__inline void stbir__float_to_half_SIMD(stbir__FP16 * output, float const * input)
 2521+  {
 2522+    for (int i=0; i<8; i++)
 2523+    {
 2524+      output[i] = stbir__float_to_half(input[i]);
 2525+    }
 2526+  }
 2527+
 2528+#endif
 2529+
 2530+
 2531+#ifdef STBIR_SIMD
 2532+
 2533+#define stbir__simdf_0123to3333( out, reg ) (out) = stbir__simdf_swiz( reg, 3,3,3,3 )
 2534+#define stbir__simdf_0123to2222( out, reg ) (out) = stbir__simdf_swiz( reg, 2,2,2,2 )
 2535+#define stbir__simdf_0123to1111( out, reg ) (out) = stbir__simdf_swiz( reg, 1,1,1,1 )
 2536+#define stbir__simdf_0123to0000( out, reg ) (out) = stbir__simdf_swiz( reg, 0,0,0,0 )
 2537+#define stbir__simdf_0123to0003( out, reg ) (out) = stbir__simdf_swiz( reg, 0,0,0,3 )
 2538+#define stbir__simdf_0123to0001( out, reg ) (out) = stbir__simdf_swiz( reg, 0,0,0,1 )
 2539+#define stbir__simdf_0123to1122( out, reg ) (out) = stbir__simdf_swiz( reg, 1,1,2,2 )
 2540+#define stbir__simdf_0123to2333( out, reg ) (out) = stbir__simdf_swiz( reg, 2,3,3,3 )
 2541+#define stbir__simdf_0123to0023( out, reg ) (out) = stbir__simdf_swiz( reg, 0,0,2,3 )
 2542+#define stbir__simdf_0123to1230( out, reg ) (out) = stbir__simdf_swiz( reg, 1,2,3,0 )
 2543+#define stbir__simdf_0123to2103( out, reg ) (out) = stbir__simdf_swiz( reg, 2,1,0,3 )
 2544+#define stbir__simdf_0123to3210( out, reg ) (out) = stbir__simdf_swiz( reg, 3,2,1,0 )
 2545+#define stbir__simdf_0123to2301( out, reg ) (out) = stbir__simdf_swiz( reg, 2,3,0,1 )
 2546+#define stbir__simdf_0123to3012( out, reg ) (out) = stbir__simdf_swiz( reg, 3,0,1,2 )
 2547+#define stbir__simdf_0123to0011( out, reg ) (out) = stbir__simdf_swiz( reg, 0,0,1,1 )
 2548+#define stbir__simdf_0123to1100( out, reg ) (out) = stbir__simdf_swiz( reg, 1,1,0,0 )
 2549+#define stbir__simdf_0123to2233( out, reg ) (out) = stbir__simdf_swiz( reg, 2,2,3,3 )
 2550+#define stbir__simdf_0123to1133( out, reg ) (out) = stbir__simdf_swiz( reg, 1,1,3,3 )
 2551+#define stbir__simdf_0123to0022( out, reg ) (out) = stbir__simdf_swiz( reg, 0,0,2,2 )
 2552+#define stbir__simdf_0123to1032( out, reg ) (out) = stbir__simdf_swiz( reg, 1,0,3,2 )
 2553+
 2554+typedef union stbir__simdi_u32
 2555+{
 2556+  stbir_uint32 m128i_u32[4];
 2557+  int m128i_i32[4];
 2558+  stbir__simdi m128i_i128;
 2559+} stbir__simdi_u32;
 2560+
 2561+static const int STBIR_mask[9] = { 0,0,0,-1,-1,-1,0,0,0 };
 2562+
 2563+static const STBIR__SIMDF_CONST(STBIR_max_uint8_as_float,           stbir__max_uint8_as_float);
 2564+static const STBIR__SIMDF_CONST(STBIR_max_uint16_as_float,          stbir__max_uint16_as_float);
 2565+static const STBIR__SIMDF_CONST(STBIR_max_uint8_as_float_inverted,  stbir__max_uint8_as_float_inverted);
 2566+static const STBIR__SIMDF_CONST(STBIR_max_uint16_as_float_inverted, stbir__max_uint16_as_float_inverted);
 2567+
 2568+static const STBIR__SIMDF_CONST(STBIR_simd_point5,   0.5f);
 2569+static const STBIR__SIMDF_CONST(STBIR_ones,          1.0f);
 2570+static const STBIR__SIMDI_CONST(STBIR_almost_zero,   (127 - 13) << 23);
 2571+static const STBIR__SIMDI_CONST(STBIR_almost_one,    0x3f7fffff);
 2572+static const STBIR__SIMDI_CONST(STBIR_mastissa_mask, 0xff);
 2573+static const STBIR__SIMDI_CONST(STBIR_topscale,      0x02000000);
 2574+
 2575+//   Basically, in simd mode, we unroll the proper amount, and we don't want
 2576+//   the non-simd remnant loops to be unroll because they only run a few times
 2577+//   Adding this switch saves about 5K on clang which is Captain Unroll the 3rd.
 2578+#define STBIR_SIMD_STREAMOUT_PTR( star )  STBIR_STREAMOUT_PTR( star )
 2579+#define STBIR_SIMD_NO_UNROLL(ptr) STBIR_NO_UNROLL(ptr)
 2580+#define STBIR_SIMD_NO_UNROLL_LOOP_START STBIR_NO_UNROLL_LOOP_START
 2581+#define STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR STBIR_NO_UNROLL_LOOP_START_INF_FOR
 2582+
 2583+#ifdef STBIR_MEMCPY
 2584+#undef STBIR_MEMCPY
 2585+#endif
 2586+#define STBIR_MEMCPY stbir_simd_memcpy
 2587+
 2588+// override normal use of memcpy with much simpler copy (faster and smaller with our sized copies)
 2589+static void stbir_simd_memcpy( void * dest, void const * src, size_t bytes )
 2590+{
 2591+  char STBIR_SIMD_STREAMOUT_PTR (*) d = (char*) dest;
 2592+  char STBIR_SIMD_STREAMOUT_PTR( * ) d_end = ((char*) dest) + bytes;
 2593+  ptrdiff_t ofs_to_src = (char*)src - (char*)dest;
 2594+
 2595+  // check overlaps
 2596+  STBIR_ASSERT( ( ( d >= ( (char*)src) + bytes ) ) || ( ( d + bytes ) <= (char*)src ) );
 2597+
 2598+  if ( bytes < (16*stbir__simdfX_float_count) )
 2599+  {
 2600+    if ( bytes < 16 )
 2601+    {
 2602+      if ( bytes )
 2603+      {
 2604+        STBIR_SIMD_NO_UNROLL_LOOP_START
 2605+        do
 2606+        {
 2607+          STBIR_SIMD_NO_UNROLL(d);
 2608+          d[ 0 ] = d[ ofs_to_src ];
 2609+          ++d;
 2610+        } while ( d < d_end );
 2611+      }
 2612+    }
 2613+    else
 2614+    {
 2615+      stbir__simdf x;
 2616+      // do one unaligned to get us aligned for the stream out below
 2617+      stbir__simdf_load( x, ( d + ofs_to_src ) );
 2618+      stbir__simdf_store( d, x );
 2619+      d = (char*)( ( ( (size_t)d ) + 16 ) & ~15 );
 2620+
 2621+      STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 2622+      for(;;)
 2623+      {
 2624+        STBIR_SIMD_NO_UNROLL(d);
 2625+
 2626+        if ( d > ( d_end - 16 ) )
 2627+        {
 2628+          if ( d == d_end )
 2629+            return;
 2630+          d = d_end - 16;
 2631+        }
 2632+
 2633+        stbir__simdf_load( x, ( d + ofs_to_src ) );
 2634+        stbir__simdf_store( d, x );
 2635+        d += 16;
 2636+      }
 2637+    }
 2638+  }
 2639+  else
 2640+  {
 2641+    stbir__simdfX x0,x1,x2,x3;
 2642+
 2643+    // do one unaligned to get us aligned for the stream out below
 2644+    stbir__simdfX_load( x0, ( d + ofs_to_src ) +  0*stbir__simdfX_float_count );
 2645+    stbir__simdfX_load( x1, ( d + ofs_to_src ) +  4*stbir__simdfX_float_count );
 2646+    stbir__simdfX_load( x2, ( d + ofs_to_src ) +  8*stbir__simdfX_float_count );
 2647+    stbir__simdfX_load( x3, ( d + ofs_to_src ) + 12*stbir__simdfX_float_count );
 2648+    stbir__simdfX_store( d +  0*stbir__simdfX_float_count, x0 );
 2649+    stbir__simdfX_store( d +  4*stbir__simdfX_float_count, x1 );
 2650+    stbir__simdfX_store( d +  8*stbir__simdfX_float_count, x2 );
 2651+    stbir__simdfX_store( d + 12*stbir__simdfX_float_count, x3 );
 2652+    d = (char*)( ( ( (size_t)d ) + (16*stbir__simdfX_float_count) ) & ~((16*stbir__simdfX_float_count)-1) );
 2653+
 2654+    STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 2655+    for(;;)
 2656+    {
 2657+      STBIR_SIMD_NO_UNROLL(d);
 2658+
 2659+      if ( d > ( d_end - (16*stbir__simdfX_float_count) ) )
 2660+      {
 2661+        if ( d == d_end )
 2662+          return;
 2663+        d = d_end - (16*stbir__simdfX_float_count);
 2664+      }
 2665+
 2666+      stbir__simdfX_load( x0, ( d + ofs_to_src ) +  0*stbir__simdfX_float_count );
 2667+      stbir__simdfX_load( x1, ( d + ofs_to_src ) +  4*stbir__simdfX_float_count );
 2668+      stbir__simdfX_load( x2, ( d + ofs_to_src ) +  8*stbir__simdfX_float_count );
 2669+      stbir__simdfX_load( x3, ( d + ofs_to_src ) + 12*stbir__simdfX_float_count );
 2670+      stbir__simdfX_store( d +  0*stbir__simdfX_float_count, x0 );
 2671+      stbir__simdfX_store( d +  4*stbir__simdfX_float_count, x1 );
 2672+      stbir__simdfX_store( d +  8*stbir__simdfX_float_count, x2 );
 2673+      stbir__simdfX_store( d + 12*stbir__simdfX_float_count, x3 );
 2674+      d += (16*stbir__simdfX_float_count);
 2675+    }
 2676+  }
 2677+}
 2678+
 2679+// memcpy that is specically intentionally overlapping (src is smaller then dest, so can be
 2680+//   a normal forward copy, bytes is divisible by 4 and bytes is greater than or equal to
 2681+//   the diff between dest and src)
 2682+static void stbir_overlapping_memcpy( void * dest, void const * src, size_t bytes )
 2683+{
 2684+  char STBIR_SIMD_STREAMOUT_PTR (*) sd = (char*) src;
 2685+  char STBIR_SIMD_STREAMOUT_PTR( * ) s_end = ((char*) src) + bytes;
 2686+  ptrdiff_t ofs_to_dest = (char*)dest - (char*)src;
 2687+
 2688+  if ( ofs_to_dest >= 16 ) // is the overlap more than 16 away?
 2689+  {
 2690+    char STBIR_SIMD_STREAMOUT_PTR( * ) s_end16 = ((char*) src) + (bytes&~15);
 2691+    STBIR_SIMD_NO_UNROLL_LOOP_START
 2692+    do
 2693+    {
 2694+      stbir__simdf x;
 2695+      STBIR_SIMD_NO_UNROLL(sd);
 2696+      stbir__simdf_load( x, sd );
 2697+      stbir__simdf_store(  ( sd + ofs_to_dest ), x );
 2698+      sd += 16;
 2699+    } while ( sd < s_end16 );
 2700+
 2701+    if ( sd == s_end )
 2702+      return;
 2703+  }
 2704+
 2705+  do
 2706+  {
 2707+    STBIR_SIMD_NO_UNROLL(sd);
 2708+    *(int*)( sd + ofs_to_dest ) = *(int*) sd;
 2709+    sd += 4;
 2710+  } while ( sd < s_end );
 2711+}
 2712+
 2713+#else // no SSE2
 2714+
 2715+// when in scalar mode, we let unrolling happen, so this macro just does the __restrict
 2716+#define STBIR_SIMD_STREAMOUT_PTR( star ) STBIR_STREAMOUT_PTR( star )
 2717+#define STBIR_SIMD_NO_UNROLL(ptr)
 2718+#define STBIR_SIMD_NO_UNROLL_LOOP_START
 2719+#define STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 2720+
 2721+#endif // SSE2
 2722+
 2723+
 2724+#ifdef STBIR_PROFILE
 2725+
 2726+#ifndef STBIR_PROFILE_FUNC
 2727+
 2728+#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_ )
 2729+
 2730+#ifdef _MSC_VER
 2731+
 2732+  STBIRDEF stbir_uint64 __rdtsc();
 2733+  #define STBIR_PROFILE_FUNC() __rdtsc()
 2734+
 2735+#else // non msvc
 2736+
 2737+  static stbir__inline stbir_uint64 STBIR_PROFILE_FUNC()
 2738+  {
 2739+    stbir_uint32 lo, hi;
 2740+    asm volatile ("rdtsc" : "=a" (lo), "=d" (hi) );
 2741+    return ( ( (stbir_uint64) hi ) << 32 ) | ( (stbir_uint64) lo );
 2742+  }
 2743+
 2744+#endif  // msvc
 2745+
 2746+#elif defined( _M_ARM64 ) || defined( __aarch64__ ) || defined( __arm64__ ) || defined(__ARM_NEON__)
 2747+
 2748+#if defined( _MSC_VER ) && !defined(__clang__)
 2749+
 2750+  #define STBIR_PROFILE_FUNC() _ReadStatusReg(ARM64_CNTVCT)
 2751+
 2752+#else
 2753+
 2754+  static stbir__inline stbir_uint64 STBIR_PROFILE_FUNC()
 2755+  {
 2756+    stbir_uint64 tsc;
 2757+    asm volatile("mrs %0, cntvct_el0" : "=r" (tsc));
 2758+    return tsc;
 2759+  }
 2760+
 2761+#endif
 2762+
 2763+#else // x64, arm
 2764+
 2765+#error Unknown platform for profiling.
 2766+
 2767+#endif  // x64, arm
 2768+
 2769+#endif // STBIR_PROFILE_FUNC
 2770+
 2771+#define STBIR_ONLY_PROFILE_GET_SPLIT_INFO ,stbir__per_split_info * split_info
 2772+#define STBIR_ONLY_PROFILE_SET_SPLIT_INFO ,split_info
 2773+
 2774+#define STBIR_ONLY_PROFILE_BUILD_GET_INFO ,stbir__info * profile_info
 2775+#define STBIR_ONLY_PROFILE_BUILD_SET_INFO ,profile_info
 2776+
 2777+// super light-weight micro profiler
 2778+#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;
 2779+#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; }
 2780+#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 );
 2781+#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; } }
 2782+
 2783+// for thread data
 2784+#define STBIR_PROFILE_START( wh ) STBIR_PROFILE_START_ll( split_info, wh )
 2785+#define STBIR_PROFILE_END( wh ) STBIR_PROFILE_END_ll( split_info, wh )
 2786+#define STBIR_PROFILE_FIRST_START( wh ) STBIR_PROFILE_FIRST_START_ll( split_info, wh )
 2787+#define STBIR_PROFILE_CLEAR_EXTRAS() STBIR_PROFILE_CLEAR_EXTRAS_ll( split_info, split_count )
 2788+
 2789+// for build data
 2790+#define STBIR_PROFILE_BUILD_START( wh ) STBIR_PROFILE_START_ll( profile_info, wh )
 2791+#define STBIR_PROFILE_BUILD_END( wh ) STBIR_PROFILE_END_ll( profile_info, wh )
 2792+#define STBIR_PROFILE_BUILD_FIRST_START( wh ) STBIR_PROFILE_FIRST_START_ll( profile_info, wh )
 2793+#define STBIR_PROFILE_BUILD_CLEAR( info ) { int i; for(i=0;i<STBIR__ARRAY_SIZE(info->profile.array);i++) info->profile.array[i]=0; }
 2794+
 2795+#else  // no profile
 2796+
 2797+#define STBIR_ONLY_PROFILE_GET_SPLIT_INFO
 2798+#define STBIR_ONLY_PROFILE_SET_SPLIT_INFO
 2799+
 2800+#define STBIR_ONLY_PROFILE_BUILD_GET_INFO
 2801+#define STBIR_ONLY_PROFILE_BUILD_SET_INFO
 2802+
 2803+#define STBIR_PROFILE_START( wh )
 2804+#define STBIR_PROFILE_END( wh )
 2805+#define STBIR_PROFILE_FIRST_START( wh )
 2806+#define STBIR_PROFILE_CLEAR_EXTRAS( )
 2807+
 2808+#define STBIR_PROFILE_BUILD_START( wh )
 2809+#define STBIR_PROFILE_BUILD_END( wh )
 2810+#define STBIR_PROFILE_BUILD_FIRST_START( wh )
 2811+#define STBIR_PROFILE_BUILD_CLEAR( info )
 2812+
 2813+#endif  // stbir_profile
 2814+
 2815+#ifndef STBIR_CEILF
 2816+#include <math.h>
 2817+#if _MSC_VER <= 1200 // support VC6 for Sean
 2818+#define STBIR_CEILF(x) ((float)ceil((float)(x)))
 2819+#define STBIR_FLOORF(x) ((float)floor((float)(x)))
 2820+#else
 2821+#define STBIR_CEILF(x) ceilf(x)
 2822+#define STBIR_FLOORF(x) floorf(x)
 2823+#endif
 2824+#endif
 2825+
 2826+#ifndef STBIR_MEMCPY
 2827+// For memcpy
 2828+#include <string.h>
 2829+#define STBIR_MEMCPY( dest, src, len ) memcpy( dest, src, len )
 2830+#endif
 2831+
 2832+#ifndef STBIR_SIMD
 2833+
 2834+// memcpy that is specifically intentionally overlapping (src is smaller then dest, so can be
 2835+//   a normal forward copy, bytes is divisible by 4 and bytes is greater than or equal to
 2836+//   the diff between dest and src)
 2837+static void stbir_overlapping_memcpy( void * dest, void const * src, size_t bytes )
 2838+{
 2839+  char STBIR_SIMD_STREAMOUT_PTR (*) sd = (char*) src;
 2840+  char STBIR_SIMD_STREAMOUT_PTR( * ) s_end = ((char*) src) + bytes;
 2841+  ptrdiff_t ofs_to_dest = (char*)dest - (char*)src;
 2842+
 2843+  if ( ofs_to_dest >= 8 ) // is the overlap more than 8 away?
 2844+  {
 2845+    char STBIR_SIMD_STREAMOUT_PTR( * ) s_end8 = ((char*) src) + (bytes&~7);
 2846+    STBIR_NO_UNROLL_LOOP_START
 2847+    do
 2848+    {
 2849+      STBIR_NO_UNROLL(sd);
 2850+      *(stbir_uint64*)( sd + ofs_to_dest ) = *(stbir_uint64*) sd;
 2851+      sd += 8;
 2852+    } while ( sd < s_end8 );
 2853+
 2854+    if ( sd == s_end )
 2855+      return;
 2856+  }
 2857+
 2858+  STBIR_NO_UNROLL_LOOP_START
 2859+  do
 2860+  {
 2861+    STBIR_NO_UNROLL(sd);
 2862+    *(int*)( sd + ofs_to_dest ) = *(int*) sd;
 2863+    sd += 4;
 2864+  } while ( sd < s_end );
 2865+}
 2866+
 2867+#endif
 2868+
 2869+static float stbir__filter_trapezoid(float x, float scale, void * user_data)
 2870+{
 2871+  float halfscale = scale / 2;
 2872+  float t = 0.5f + halfscale;
 2873+  STBIR_ASSERT(scale <= 1);
 2874+  STBIR__UNUSED(user_data);
 2875+
 2876+  if ( x < 0.0f ) x = -x;
 2877+
 2878+  if (x >= t)
 2879+    return 0.0f;
 2880+  else
 2881+  {
 2882+    float r = 0.5f - halfscale;
 2883+    if (x <= r)
 2884+      return 1.0f;
 2885+    else
 2886+      return (t - x) / scale;
 2887+  }
 2888+}
 2889+
 2890+static float stbir__support_trapezoid(float scale, void * user_data)
 2891+{
 2892+  STBIR__UNUSED(user_data);
 2893+  return 0.5f + scale / 2.0f;
 2894+}
 2895+
 2896+static float stbir__filter_triangle(float x, float s, void * user_data)
 2897+{
 2898+  STBIR__UNUSED(s);
 2899+  STBIR__UNUSED(user_data);
 2900+
 2901+  if ( x < 0.0f ) x = -x;
 2902+
 2903+  if (x <= 1.0f)
 2904+    return 1.0f - x;
 2905+  else
 2906+    return 0.0f;
 2907+}
 2908+
 2909+static float stbir__filter_point(float x, float s, void * user_data)
 2910+{
 2911+  STBIR__UNUSED(x);
 2912+  STBIR__UNUSED(s);
 2913+  STBIR__UNUSED(user_data);
 2914+
 2915+  return 1.0f;
 2916+}
 2917+
 2918+static float stbir__filter_cubic(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 (4.0f + x*x*(3.0f*x - 6.0f))/6.0f;
 2927+  else if (x < 2.0f)
 2928+    return (8.0f + x*(-12.0f + x*(6.0f - x)))/6.0f;
 2929+
 2930+  return (0.0f);
 2931+}
 2932+
 2933+static float stbir__filter_catmullrom(float x, float s, void * user_data)
 2934+{
 2935+  STBIR__UNUSED(s);
 2936+  STBIR__UNUSED(user_data);
 2937+
 2938+  if ( x < 0.0f ) x = -x;
 2939+
 2940+  if (x < 1.0f)
 2941+    return 1.0f - x*x*(2.5f - 1.5f*x);
 2942+  else if (x < 2.0f)
 2943+    return 2.0f - x*(4.0f + x*(0.5f*x - 2.5f));
 2944+
 2945+  return (0.0f);
 2946+}
 2947+
 2948+static float stbir__filter_mitchell(float x, float s, void * user_data)
 2949+{
 2950+  STBIR__UNUSED(s);
 2951+  STBIR__UNUSED(user_data);
 2952+
 2953+  if ( x < 0.0f ) x = -x;
 2954+
 2955+  if (x < 1.0f)
 2956+    return (16.0f + x*x*(21.0f * x - 36.0f))/18.0f;
 2957+  else if (x < 2.0f)
 2958+    return (32.0f + x*(-60.0f + x*(36.0f - 7.0f*x)))/18.0f;
 2959+
 2960+  return (0.0f);
 2961+}
 2962+
 2963+static float stbir__support_zeropoint5(float s, void * user_data)
 2964+{
 2965+  STBIR__UNUSED(s);
 2966+  STBIR__UNUSED(user_data);
 2967+  return 0.5f;
 2968+}
 2969+
 2970+static float stbir__support_one(float s, void * user_data)
 2971+{
 2972+  STBIR__UNUSED(s);
 2973+  STBIR__UNUSED(user_data);
 2974+  return 1;
 2975+}
 2976+
 2977+static float stbir__support_two(float s, void * user_data)
 2978+{
 2979+  STBIR__UNUSED(s);
 2980+  STBIR__UNUSED(user_data);
 2981+  return 2;
 2982+}
 2983+
 2984+// This is the maximum number of input samples that can affect an output sample
 2985+// with the given filter from the output pixel's perspective
 2986+static int stbir__get_filter_pixel_width(stbir__support_callback * support, float scale, void * user_data)
 2987+{
 2988+  STBIR_ASSERT(support != 0);
 2989+
 2990+  if ( scale >= ( 1.0f-stbir__small_float ) ) // upscale
 2991+    return (int)STBIR_CEILF(support(1.0f/scale,user_data) * 2.0f);
 2992+  else
 2993+    return (int)STBIR_CEILF(support(scale,user_data) * 2.0f / scale);
 2994+}
 2995+
 2996+// this is how many coefficents per run of the filter (which is different
 2997+//   from the filter_pixel_width depending on if we are scattering or gathering)
 2998+static int stbir__get_coefficient_width(stbir__sampler * samp, int is_gather, void * user_data)
 2999+{
 3000+  float scale = samp->scale_info.scale;
 3001+  stbir__support_callback * support = samp->filter_support;
 3002+
 3003+  switch( is_gather )
 3004+  {
 3005+    case 1:
 3006+      return (int)STBIR_CEILF(support(1.0f / scale, user_data) * 2.0f);
 3007+    case 2:
 3008+      return (int)STBIR_CEILF(support(scale, user_data) * 2.0f / scale);
 3009+    case 0:
 3010+      return (int)STBIR_CEILF(support(scale, user_data) * 2.0f);
 3011+    default:
 3012+      STBIR_ASSERT( (is_gather >= 0 ) && (is_gather <= 2 ) );
 3013+      return 0;
 3014+  }
 3015+}
 3016+
 3017+static int stbir__get_contributors(stbir__sampler * samp, int is_gather)
 3018+{
 3019+  if (is_gather)
 3020+      return samp->scale_info.output_sub_size;
 3021+  else
 3022+      return (samp->scale_info.input_full_size + samp->filter_pixel_margin * 2);
 3023+}
 3024+
 3025+static int stbir__edge_zero_full( int n, int max )
 3026+{
 3027+  STBIR__UNUSED(n);
 3028+  STBIR__UNUSED(max);
 3029+  return 0; // NOTREACHED
 3030+}
 3031+
 3032+static int stbir__edge_clamp_full( int n, int max )
 3033+{
 3034+  if (n < 0)
 3035+    return 0;
 3036+
 3037+  if (n >= max)
 3038+    return max - 1;
 3039+
 3040+  return n; // NOTREACHED
 3041+}
 3042+
 3043+static int stbir__edge_reflect_full( int n, int max )
 3044+{
 3045+  if (n < 0)
 3046+  {
 3047+    if (n > -max)
 3048+      return -n;
 3049+    else
 3050+      return max - 1;
 3051+  }
 3052+
 3053+  if (n >= max)
 3054+  {
 3055+    int max2 = max * 2;
 3056+    if (n >= max2)
 3057+      return 0;
 3058+    else
 3059+      return max2 - n - 1;
 3060+  }
 3061+
 3062+  return n; // NOTREACHED
 3063+}
 3064+
 3065+static int stbir__edge_wrap_full( int n, int max )
 3066+{
 3067+  if (n >= 0)
 3068+    return (n % max);
 3069+  else
 3070+  {
 3071+    int m = (-n) % max;
 3072+
 3073+    if (m != 0)
 3074+      m = max - m;
 3075+
 3076+    return (m);
 3077+  }
 3078+}
 3079+
 3080+typedef int stbir__edge_wrap_func( int n, int max );
 3081+static stbir__edge_wrap_func * stbir__edge_wrap_slow[] =
 3082+{
 3083+  stbir__edge_clamp_full,    // STBIR_EDGE_CLAMP
 3084+  stbir__edge_reflect_full,  // STBIR_EDGE_REFLECT
 3085+  stbir__edge_wrap_full,     // STBIR_EDGE_WRAP
 3086+  stbir__edge_zero_full,     // STBIR_EDGE_ZERO
 3087+};
 3088+
 3089+stbir__inline static int stbir__edge_wrap(stbir_edge edge, int n, int max)
 3090+{
 3091+  // avoid per-pixel switch
 3092+  if (n >= 0 && n < max)
 3093+      return n;
 3094+  return stbir__edge_wrap_slow[edge]( n, max );
 3095+}
 3096+
 3097+#define STBIR__MERGE_RUNS_PIXEL_THRESHOLD 16
 3098+
 3099+// get information on the extents of a sampler
 3100+static void stbir__get_extents( stbir__sampler * samp, stbir__extents * scanline_extents )
 3101+{
 3102+  int j, stop;
 3103+  int left_margin, right_margin;
 3104+  int min_n = 0x7fffffff, max_n = -0x7fffffff;
 3105+  int min_left = 0x7fffffff, max_left = -0x7fffffff;
 3106+  int min_right = 0x7fffffff, max_right = -0x7fffffff;
 3107+  stbir_edge edge = samp->edge;
 3108+  stbir__contributors* contributors = samp->contributors;
 3109+  int output_sub_size = samp->scale_info.output_sub_size;
 3110+  int input_full_size = samp->scale_info.input_full_size;
 3111+  int filter_pixel_margin = samp->filter_pixel_margin;
 3112+
 3113+  STBIR_ASSERT( samp->is_gather );
 3114+
 3115+  stop = output_sub_size;
 3116+  for (j = 0; j < stop; j++ )
 3117+  {
 3118+    STBIR_ASSERT( contributors[j].n1 >= contributors[j].n0 );
 3119+    if ( contributors[j].n0 < min_n )
 3120+    {
 3121+      min_n = contributors[j].n0;
 3122+      stop = j + filter_pixel_margin;  // if we find a new min, only scan another filter width
 3123+      if ( stop > output_sub_size ) stop = output_sub_size;
 3124+    }
 3125+  }
 3126+
 3127+  stop = 0;
 3128+  for (j = output_sub_size - 1; j >= stop; j-- )
 3129+  {
 3130+    STBIR_ASSERT( contributors[j].n1 >= contributors[j].n0 );
 3131+    if ( contributors[j].n1 > max_n )
 3132+    {
 3133+      max_n = contributors[j].n1;
 3134+      stop = j - filter_pixel_margin;  // if we find a new max, only scan another filter width
 3135+      if (stop<0) stop = 0;
 3136+    }
 3137+  }
 3138+
 3139+  STBIR_ASSERT( scanline_extents->conservative.n0 <= min_n );
 3140+  STBIR_ASSERT( scanline_extents->conservative.n1 >= max_n );
 3141+
 3142+  // now calculate how much into the margins we really read
 3143+  left_margin = 0;
 3144+  if ( min_n < 0 )
 3145+  {
 3146+    left_margin = -min_n;
 3147+    min_n = 0;
 3148+  }
 3149+
 3150+  right_margin = 0;
 3151+  if ( max_n >= input_full_size )
 3152+  {
 3153+    right_margin = max_n - input_full_size + 1;
 3154+    max_n = input_full_size - 1;
 3155+  }
 3156+
 3157+  // index 1 is margin pixel extents (how many pixels we hang over the edge)
 3158+  scanline_extents->edge_sizes[0] = left_margin;
 3159+  scanline_extents->edge_sizes[1] = right_margin;
 3160+
 3161+  // index 2 is pixels read from the input
 3162+  scanline_extents->spans[0].n0 = min_n;
 3163+  scanline_extents->spans[0].n1 = max_n;
 3164+  scanline_extents->spans[0].pixel_offset_for_input = min_n;
 3165+
 3166+  // default to no other input range
 3167+  scanline_extents->spans[1].n0 = 0;
 3168+  scanline_extents->spans[1].n1 = -1;
 3169+  scanline_extents->spans[1].pixel_offset_for_input = 0;
 3170+
 3171+  // don't have to do edge calc for zero clamp
 3172+  if ( edge == STBIR_EDGE_ZERO )
 3173+    return;
 3174+
 3175+  // convert margin pixels to the pixels within the input (min and max)
 3176+  for( j = -left_margin ; j < 0 ; j++ )
 3177+  {
 3178+      int p = stbir__edge_wrap( edge, j, input_full_size );
 3179+      if ( p < min_left )
 3180+        min_left = p;
 3181+      if ( p > max_left )
 3182+        max_left = p;
 3183+  }
 3184+
 3185+  for( j = input_full_size ; j < (input_full_size + right_margin) ; j++ )
 3186+  {
 3187+      int p = stbir__edge_wrap( edge, j, input_full_size );
 3188+      if ( p < min_right )
 3189+        min_right = p;
 3190+      if ( p > max_right )
 3191+        max_right = p;
 3192+  }
 3193+
 3194+  // merge the left margin pixel region if it connects within 4 pixels of main pixel region
 3195+  if ( min_left != 0x7fffffff )
 3196+  {
 3197+    if ( ( ( min_left <= min_n ) && ( ( max_left  + STBIR__MERGE_RUNS_PIXEL_THRESHOLD ) >= min_n ) ) ||
 3198+         ( ( min_n <= min_left ) && ( ( max_n  + STBIR__MERGE_RUNS_PIXEL_THRESHOLD ) >= max_left ) ) )
 3199+    {
 3200+      scanline_extents->spans[0].n0 = min_n = stbir__min( min_n, min_left );
 3201+      scanline_extents->spans[0].n1 = max_n = stbir__max( max_n, max_left );
 3202+      scanline_extents->spans[0].pixel_offset_for_input = min_n;
 3203+      left_margin = 0;
 3204+    }
 3205+  }
 3206+
 3207+  // merge the right margin pixel region if it connects within 4 pixels of main pixel region
 3208+  if ( min_right != 0x7fffffff )
 3209+  {
 3210+    if ( ( ( min_right <= min_n ) && ( ( max_right  + STBIR__MERGE_RUNS_PIXEL_THRESHOLD ) >= min_n ) ) ||
 3211+         ( ( min_n <= min_right ) && ( ( max_n  + STBIR__MERGE_RUNS_PIXEL_THRESHOLD ) >= max_right ) ) )
 3212+    {
 3213+      scanline_extents->spans[0].n0 = min_n = stbir__min( min_n, min_right );
 3214+      scanline_extents->spans[0].n1 = max_n = stbir__max( max_n, max_right );
 3215+      scanline_extents->spans[0].pixel_offset_for_input = min_n;
 3216+      right_margin = 0;
 3217+    }
 3218+  }
 3219+
 3220+  STBIR_ASSERT( scanline_extents->conservative.n0 <= min_n );
 3221+  STBIR_ASSERT( scanline_extents->conservative.n1 >= max_n );
 3222+
 3223+  // you get two ranges when you have the WRAP edge mode and you are doing just the a piece of the resize
 3224+  //   so you need to get a second run of pixels from the opposite side of the scanline (which you
 3225+  //   wouldn't need except for WRAP)
 3226+
 3227+
 3228+  // if we can't merge the min_left range, add it as a second range
 3229+  if ( ( left_margin ) && ( min_left != 0x7fffffff ) )
 3230+  {
 3231+    stbir__span * newspan = scanline_extents->spans + 1;
 3232+    STBIR_ASSERT( right_margin == 0 );
 3233+    if ( min_left < scanline_extents->spans[0].n0 )
 3234+    {
 3235+      scanline_extents->spans[1].pixel_offset_for_input = scanline_extents->spans[0].n0;
 3236+      scanline_extents->spans[1].n0 = scanline_extents->spans[0].n0;
 3237+      scanline_extents->spans[1].n1 = scanline_extents->spans[0].n1;
 3238+      --newspan;
 3239+    }
 3240+    newspan->pixel_offset_for_input = min_left;
 3241+    newspan->n0 = -left_margin;
 3242+    newspan->n1 = ( max_left - min_left ) - left_margin;
 3243+    scanline_extents->edge_sizes[0] = 0;  // don't need to copy the left margin, since we are directly decoding into the margin
 3244+  }
 3245+  // if we can't merge the min_left range, add it as a second range
 3246+  else  
 3247+  if ( ( right_margin ) && ( min_right != 0x7fffffff ) )
 3248+  {
 3249+    stbir__span * newspan = scanline_extents->spans + 1;
 3250+    if ( min_right < scanline_extents->spans[0].n0 )
 3251+    {
 3252+      scanline_extents->spans[1].pixel_offset_for_input = scanline_extents->spans[0].n0;
 3253+      scanline_extents->spans[1].n0 = scanline_extents->spans[0].n0;
 3254+      scanline_extents->spans[1].n1 = scanline_extents->spans[0].n1;
 3255+      --newspan;
 3256+    }
 3257+    newspan->pixel_offset_for_input = min_right;
 3258+    newspan->n0 = scanline_extents->spans[1].n1 + 1;
 3259+    newspan->n1 = scanline_extents->spans[1].n1 + 1 + ( max_right - min_right );
 3260+    scanline_extents->edge_sizes[1] = 0;  // don't need to copy the right margin, since we are directly decoding into the margin
 3261+  }
 3262+
 3263+  // sort the spans into write output order
 3264+  if ( ( scanline_extents->spans[1].n1 > scanline_extents->spans[1].n0 ) && ( scanline_extents->spans[0].n0 > scanline_extents->spans[1].n0 ) )
 3265+  {
 3266+    stbir__span tspan = scanline_extents->spans[0];
 3267+    scanline_extents->spans[0] = scanline_extents->spans[1];
 3268+    scanline_extents->spans[1] = tspan;
 3269+  }
 3270+}
 3271+
 3272+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 )
 3273+{
 3274+  int first, last;
 3275+  float out_pixel_influence_lowerbound = out_pixel_center - out_filter_radius;
 3276+  float out_pixel_influence_upperbound = out_pixel_center + out_filter_radius;
 3277+
 3278+  float in_pixel_influence_lowerbound = (out_pixel_influence_lowerbound + out_shift) * inv_scale;
 3279+  float in_pixel_influence_upperbound = (out_pixel_influence_upperbound + out_shift) * inv_scale;
 3280+
 3281+  first = (int)(STBIR_FLOORF(in_pixel_influence_lowerbound + 0.5f));
 3282+  last = (int)(STBIR_FLOORF(in_pixel_influence_upperbound - 0.5f));
 3283+  if ( last < first ) last = first; // point sample mode can span a value *right* at 0.5, and cause these to cross
 3284+
 3285+  if ( edge == STBIR_EDGE_WRAP )
 3286+  {
 3287+    if ( first < -input_size )
 3288+      first = -input_size;
 3289+    if ( last >= (input_size*2))
 3290+      last = (input_size*2) - 1;
 3291+  }
 3292+
 3293+  *first_pixel = first;
 3294+  *last_pixel = last;
 3295+}
 3296+
 3297+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 )
 3298+{
 3299+  int n, end;
 3300+  float inv_scale = scale_info->inv_scale;
 3301+  float out_shift = scale_info->pixel_shift;
 3302+  int input_size  = scale_info->input_full_size;
 3303+  int numerator = scale_info->scale_numerator;
 3304+  int polyphase = ( ( scale_info->scale_is_rational ) && ( numerator < num_contributors ) );
 3305+
 3306+  // Looping through out pixels
 3307+  end = num_contributors; if ( polyphase ) end = numerator;
 3308+  for (n = 0; n < end; n++)
 3309+  {
 3310+    int i;
 3311+    int last_non_zero;
 3312+    float out_pixel_center = (float)n + 0.5f;
 3313+    float in_center_of_out = (out_pixel_center + out_shift) * inv_scale;
 3314+
 3315+    int in_first_pixel, in_last_pixel;
 3316+
 3317+    stbir__calculate_in_pixel_range( &in_first_pixel, &in_last_pixel, out_pixel_center, out_filter_radius, inv_scale, out_shift, input_size, edge );
 3318+
 3319+    // make sure we never generate a range larger than our precalculated coeff width
 3320+    //   this only happens in point sample mode, but it's a good safe thing to do anyway
 3321+    if ( ( in_last_pixel - in_first_pixel + 1 ) > coefficient_width )
 3322+      in_last_pixel = in_first_pixel + coefficient_width - 1;
 3323+
 3324+    last_non_zero = -1;
 3325+    for (i = 0; i <= in_last_pixel - in_first_pixel; i++)
 3326+    {
 3327+      float in_pixel_center = (float)(i + in_first_pixel) + 0.5f;
 3328+      float coeff = kernel(in_center_of_out - in_pixel_center, inv_scale, user_data);
 3329+
 3330+      // kill denormals
 3331+      if ( ( ( coeff < stbir__small_float ) && ( coeff > -stbir__small_float ) ) )
 3332+      {
 3333+        if ( i == 0 )  // if we're at the front, just eat zero contributors
 3334+        {
 3335+          STBIR_ASSERT ( ( in_last_pixel - in_first_pixel ) != 0 ); // there should be at least one contrib
 3336+          ++in_first_pixel;
 3337+          i--;
 3338+          continue;
 3339+        }
 3340+        coeff = 0;  // make sure is fully zero (should keep denormals away)
 3341+      }
 3342+      else
 3343+        last_non_zero = i;
 3344+
 3345+      coefficient_group[i] = coeff;
 3346+    }
 3347+
 3348+    in_last_pixel = last_non_zero+in_first_pixel; // kills trailing zeros
 3349+    contributors->n0 = in_first_pixel;
 3350+    contributors->n1 = in_last_pixel;
 3351+
 3352+    STBIR_ASSERT(contributors->n1 >= contributors->n0);
 3353+
 3354+    ++contributors;
 3355+    coefficient_group += coefficient_width;
 3356+  }
 3357+}
 3358+
 3359+static void stbir__insert_coeff( stbir__contributors * contribs, float * coeffs, int new_pixel, float new_coeff, int max_width )
 3360+{
 3361+  if ( new_pixel <= contribs->n1 )  // before the end
 3362+  {
 3363+    if ( new_pixel < contribs->n0 ) // before the front?
 3364+    {
 3365+      if ( ( contribs->n1 - new_pixel + 1 ) <= max_width )
 3366+      { 
 3367+        int j, o = contribs->n0 - new_pixel;
 3368+        for ( j = contribs->n1 - contribs->n0 ; j <= 0 ; j-- )
 3369+          coeffs[ j + o ] = coeffs[ j ];
 3370+        for ( j = 1 ; j < o ; j-- )
 3371+          coeffs[ j ] = coeffs[ 0 ];
 3372+        coeffs[ 0 ] = new_coeff;
 3373+        contribs->n0 = new_pixel;
 3374+      }
 3375+    }
 3376+    else
 3377+    {
 3378+      coeffs[ new_pixel - contribs->n0 ] += new_coeff;
 3379+    }
 3380+  }
 3381+  else
 3382+  {
 3383+    if ( ( new_pixel - contribs->n0 + 1 ) <= max_width )
 3384+    {
 3385+      int j, e = new_pixel - contribs->n0;
 3386+      for( j = ( contribs->n1 - contribs->n0 ) + 1 ; j < e ; j++ ) // clear in-betweens coeffs if there are any
 3387+        coeffs[j] = 0;
 3388+
 3389+      coeffs[ e ] = new_coeff;
 3390+      contribs->n1 = new_pixel;
 3391+    }
 3392+  }
 3393+}
 3394+
 3395+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 )
 3396+{
 3397+  float in_pixel_influence_lowerbound = in_pixel_center - in_pixels_radius;
 3398+  float in_pixel_influence_upperbound = in_pixel_center + in_pixels_radius;
 3399+  float out_pixel_influence_lowerbound = in_pixel_influence_lowerbound * scale - out_shift;
 3400+  float out_pixel_influence_upperbound = in_pixel_influence_upperbound * scale - out_shift;
 3401+  int out_first_pixel = (int)(STBIR_FLOORF(out_pixel_influence_lowerbound + 0.5f));
 3402+  int out_last_pixel = (int)(STBIR_FLOORF(out_pixel_influence_upperbound - 0.5f));
 3403+
 3404+  if ( out_first_pixel < 0 )
 3405+    out_first_pixel = 0;
 3406+  if ( out_last_pixel >= out_size )
 3407+    out_last_pixel = out_size - 1;
 3408+  *first_pixel = out_first_pixel;
 3409+  *last_pixel = out_last_pixel;
 3410+}
 3411+
 3412+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 )
 3413+{
 3414+  int in_pixel;
 3415+  int i;
 3416+  int first_out_inited = -1;
 3417+  float scale = scale_info->scale;
 3418+  float out_shift = scale_info->pixel_shift;
 3419+  int out_size = scale_info->output_sub_size;
 3420+  int numerator = scale_info->scale_numerator;
 3421+  int polyphase = ( ( scale_info->scale_is_rational ) && ( numerator < out_size ) );
 3422+
 3423+  STBIR__UNUSED(num_contributors);
 3424+
 3425+  // Loop through the input pixels
 3426+  for (in_pixel = start; in_pixel < end; in_pixel++)
 3427+  {
 3428+    float in_pixel_center = (float)in_pixel + 0.5f;
 3429+    float out_center_of_in = in_pixel_center * scale - out_shift;
 3430+    int out_first_pixel, out_last_pixel;
 3431+
 3432+    stbir__calculate_out_pixel_range( &out_first_pixel, &out_last_pixel, in_pixel_center, in_pixels_radius, scale, out_shift, out_size );
 3433+
 3434+    if ( out_first_pixel > out_last_pixel )
 3435+      continue;
 3436+
 3437+    // clamp or exit if we are using polyphase filtering, and the limit is up
 3438+    if ( polyphase )
 3439+    {
 3440+      // when polyphase, you only have to do coeffs up to the numerator count
 3441+      if ( out_first_pixel == numerator )
 3442+        break;
 3443+
 3444+      // don't do any extra work, clamp last pixel at numerator too
 3445+      if ( out_last_pixel >= numerator )
 3446+        out_last_pixel = numerator - 1;
 3447+    }
 3448+
 3449+    for (i = 0; i <= out_last_pixel - out_first_pixel; i++)
 3450+    {
 3451+      float out_pixel_center = (float)(i + out_first_pixel) + 0.5f;
 3452+      float x = out_pixel_center - out_center_of_in;
 3453+      float coeff = kernel(x, scale, user_data) * scale;
 3454+
 3455+      // kill the coeff if it's too small (avoid denormals)
 3456+      if ( ( ( coeff < stbir__small_float ) && ( coeff > -stbir__small_float ) ) )
 3457+        coeff = 0.0f;
 3458+
 3459+      {
 3460+        int out = i + out_first_pixel;
 3461+        float * coeffs = coefficient_group + out * coefficient_width;
 3462+        stbir__contributors * contribs = contributors + out;
 3463+
 3464+        // is this the first time this output pixel has been seen?  Init it.
 3465+        if ( out > first_out_inited )
 3466+        {
 3467+          STBIR_ASSERT( out == ( first_out_inited + 1 ) ); // ensure we have only advanced one at time
 3468+          first_out_inited = out;
 3469+          contribs->n0 = in_pixel;
 3470+          contribs->n1 = in_pixel;
 3471+          coeffs[0]  = coeff;
 3472+        }
 3473+        else
 3474+        {
 3475+          // insert on end (always in order)
 3476+          if ( coeffs[0] == 0.0f )  // if the first coefficent is zero, then zap it for this coeffs
 3477+          {
 3478+            STBIR_ASSERT( ( in_pixel - contribs->n0 ) == 1 ); // ensure that when we zap, we're at the 2nd pos
 3479+            contribs->n0 = in_pixel;
 3480+          }
 3481+          contribs->n1 = in_pixel;
 3482+          STBIR_ASSERT( ( in_pixel - contribs->n0 ) < coefficient_width );
 3483+          coeffs[in_pixel - contribs->n0]  = coeff;
 3484+        }
 3485+      }
 3486+    }
 3487+  }
 3488+}
 3489+
 3490+#ifdef STBIR_RENORMALIZE_IN_FLOAT
 3491+#define STBIR_RENORM_TYPE float
 3492+#else
 3493+#define STBIR_RENORM_TYPE double
 3494+#endif
 3495+
 3496+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 )
 3497+{
 3498+  int input_size = scale_info->input_full_size;
 3499+  int input_last_n1 = input_size - 1;
 3500+  int n, end;
 3501+  int lowest = 0x7fffffff;
 3502+  int highest = -0x7fffffff;
 3503+  int widest = -1;
 3504+  int numerator = scale_info->scale_numerator;
 3505+  int denominator = scale_info->scale_denominator;
 3506+  int polyphase = ( ( scale_info->scale_is_rational ) && ( numerator < num_contributors ) );
 3507+  float * coeffs;
 3508+  stbir__contributors * contribs;
 3509+
 3510+  // weight all the coeffs for each sample
 3511+  coeffs = coefficient_group;
 3512+  contribs = contributors;
 3513+  end = num_contributors; if ( polyphase ) end = numerator;
 3514+  for (n = 0; n < end; n++)
 3515+  {
 3516+    int i;
 3517+    STBIR_RENORM_TYPE filter_scale, total_filter = 0;
 3518+    int e;
 3519+
 3520+    // add all contribs
 3521+    e = contribs->n1 - contribs->n0;
 3522+    for( i = 0 ; i <= e ; i++ )
 3523+    {
 3524+      total_filter += (STBIR_RENORM_TYPE) coeffs[i];
 3525+      STBIR_ASSERT( ( coeffs[i] >= -2.0f ) && ( coeffs[i] <= 2.0f )  ); // check for wonky weights
 3526+    }
 3527+
 3528+    // rescale
 3529+    if ( ( total_filter < stbir__small_float ) && ( total_filter > -stbir__small_float ) )
 3530+    {
 3531+      // all coeffs are extremely small, just zero it
 3532+      contribs->n1 = contribs->n0;
 3533+      coeffs[0] = 0.0f;
 3534+    }
 3535+    else
 3536+    {
 3537+      // if the total isn't 1.0, rescale everything
 3538+      if ( ( total_filter < (1.0f-stbir__small_float) ) || ( total_filter > (1.0f+stbir__small_float) ) )
 3539+      {
 3540+        filter_scale = ((STBIR_RENORM_TYPE)1.0) / total_filter;
 3541+
 3542+        // scale them all
 3543+        for (i = 0; i <= e; i++)
 3544+          coeffs[i] = (float) ( coeffs[i] * filter_scale );
 3545+      }
 3546+    }
 3547+    ++contribs;
 3548+    coeffs += coefficient_width;
 3549+  }
 3550+
 3551+  // if we have a rational for the scale, we can exploit the polyphaseness to not calculate
 3552+  //   most of the coefficients, so we copy them here
 3553+  if ( polyphase )
 3554+  {
 3555+    stbir__contributors * prev_contribs = contributors;
 3556+    stbir__contributors * cur_contribs = contributors + numerator;
 3557+
 3558+    for( n = numerator ; n < num_contributors ; n++ )
 3559+    {
 3560+      cur_contribs->n0 = prev_contribs->n0 + denominator;
 3561+      cur_contribs->n1 = prev_contribs->n1 + denominator;
 3562+      ++cur_contribs;
 3563+      ++prev_contribs;
 3564+    }
 3565+    stbir_overlapping_memcpy( coefficient_group + numerator * coefficient_width, coefficient_group, ( num_contributors - numerator ) * coefficient_width * sizeof( coeffs[ 0 ] ) );
 3566+  }
 3567+
 3568+  coeffs = coefficient_group;
 3569+  contribs = contributors;
 3570+
 3571+  for (n = 0; n < num_contributors; n++)
 3572+  {
 3573+    int i;
 3574+
 3575+    // in zero edge mode, just remove out of bounds contribs completely (since their weights are accounted for now)
 3576+    if ( edge == STBIR_EDGE_ZERO )
 3577+    {
 3578+      // shrink the right side if necessary
 3579+      if ( contribs->n1 > input_last_n1 )
 3580+        contribs->n1 = input_last_n1;
 3581+
 3582+      // shrink the left side
 3583+      if ( contribs->n0 < 0 )
 3584+      {
 3585+        int j, left, skips = 0;
 3586+
 3587+        skips = -contribs->n0;
 3588+        contribs->n0 = 0;
 3589+
 3590+        // now move down the weights
 3591+        left = contribs->n1 - contribs->n0 + 1;
 3592+        if ( left > 0 )
 3593+        {
 3594+          for( j = 0 ; j < left ; j++ )
 3595+            coeffs[ j ] = coeffs[ j + skips ];
 3596+        }
 3597+      }
 3598+    }
 3599+    else if ( ( edge == STBIR_EDGE_CLAMP ) || ( edge == STBIR_EDGE_REFLECT ) )
 3600+    {
 3601+      // for clamp and reflect, calculate the true inbounds position (based on edge type) and just add that to the existing weight
 3602+
 3603+      // right hand side first
 3604+      if ( contribs->n1 > input_last_n1 )
 3605+      {
 3606+        int start = contribs->n0;
 3607+        int endi = contribs->n1;
 3608+        contribs->n1 = input_last_n1;
 3609+        for( i = input_size; i <= endi; i++ )
 3610+          stbir__insert_coeff( contribs, coeffs, stbir__edge_wrap_slow[edge]( i, input_size ), coeffs[i-start], coefficient_width );
 3611+      }
 3612+
 3613+      // now check left hand edge
 3614+      if ( contribs->n0 < 0 )
 3615+      {
 3616+        int save_n0;
 3617+        float save_n0_coeff;
 3618+        float * c = coeffs - ( contribs->n0 + 1 );
 3619+
 3620+        // reinsert the coeffs with it reflected or clamped (insert accumulates, if the coeffs exist)
 3621+        for( i = -1 ; i > contribs->n0 ; i-- )
 3622+          stbir__insert_coeff( contribs, coeffs, stbir__edge_wrap_slow[edge]( i, input_size ), *c--, coefficient_width );
 3623+        save_n0 = contribs->n0;
 3624+        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)!
 3625+
 3626+        // now slide all the coeffs down (since we have accumulated them in the positive contribs) and reset the first contrib
 3627+        contribs->n0 = 0;
 3628+        for(i = 0 ; i <= contribs->n1 ; i++ )
 3629+          coeffs[i] = coeffs[i-save_n0];
 3630+
 3631+        // now that we have shrunk down the contribs, we insert the first one safely
 3632+        stbir__insert_coeff( contribs, coeffs, stbir__edge_wrap_slow[edge]( save_n0, input_size ), save_n0_coeff, coefficient_width );
 3633+      }
 3634+    }
 3635+
 3636+    if ( contribs->n0 <= contribs->n1 )
 3637+    {
 3638+      int diff = contribs->n1 - contribs->n0 + 1;
 3639+      while ( diff && ( coeffs[ diff-1 ] == 0.0f ) )
 3640+        --diff;
 3641+
 3642+      contribs->n1 = contribs->n0 + diff - 1;
 3643+
 3644+      if ( contribs->n0 <= contribs->n1 )
 3645+      {
 3646+        if ( contribs->n0 < lowest )
 3647+          lowest = contribs->n0;
 3648+        if ( contribs->n1 > highest )
 3649+          highest = contribs->n1;
 3650+        if ( diff > widest )
 3651+          widest = diff;
 3652+      }
 3653+
 3654+      // re-zero out unused coefficients (if any)
 3655+      for( i = diff ; i < coefficient_width ; i++ )
 3656+        coeffs[i] = 0.0f;
 3657+    }
 3658+
 3659+    ++contribs;
 3660+    coeffs += coefficient_width;
 3661+  }
 3662+  filter_info->lowest = lowest;
 3663+  filter_info->highest = highest;
 3664+  filter_info->widest = widest;
 3665+}
 3666+
 3667+#undef STBIR_RENORM_TYPE 
 3668+
 3669+static int stbir__pack_coefficients( int num_contributors, stbir__contributors* contributors, float * coefficents, int coefficient_width, int widest, int row0, int row1 ) 
 3670+{
 3671+  #define STBIR_MOVE_1( dest, src ) { STBIR_NO_UNROLL(dest); ((stbir_uint32*)(dest))[0] = ((stbir_uint32*)(src))[0]; }
 3672+  #define STBIR_MOVE_2( dest, src ) { STBIR_NO_UNROLL(dest); ((stbir_uint64*)(dest))[0] = ((stbir_uint64*)(src))[0]; }
 3673+  #ifdef STBIR_SIMD
 3674+  #define STBIR_MOVE_4( dest, src ) { stbir__simdf t; STBIR_NO_UNROLL(dest); stbir__simdf_load( t, src ); stbir__simdf_store( dest, t ); }
 3675+  #else
 3676+  #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]; }
 3677+  #endif
 3678+
 3679+  int row_end = row1 + 1;
 3680+  STBIR__UNUSED( row0 ); // only used in an assert
 3681+
 3682+  if ( coefficient_width != widest )
 3683+  {
 3684+    float * pc = coefficents;
 3685+    float * coeffs = coefficents;
 3686+    float * pc_end = coefficents + num_contributors * widest;
 3687+    switch( widest )
 3688+    {
 3689+      case 1:
 3690+        STBIR_NO_UNROLL_LOOP_START
 3691+        do {
 3692+          STBIR_MOVE_1( pc, coeffs );
 3693+          ++pc;
 3694+          coeffs += coefficient_width;
 3695+        } while ( pc < pc_end );
 3696+        break;
 3697+      case 2:
 3698+        STBIR_NO_UNROLL_LOOP_START
 3699+        do {
 3700+          STBIR_MOVE_2( pc, coeffs );
 3701+          pc += 2;
 3702+          coeffs += coefficient_width;
 3703+        } while ( pc < pc_end );
 3704+        break;
 3705+      case 3:
 3706+        STBIR_NO_UNROLL_LOOP_START
 3707+        do {
 3708+          STBIR_MOVE_2( pc, coeffs );
 3709+          STBIR_MOVE_1( pc+2, coeffs+2 );
 3710+          pc += 3;
 3711+          coeffs += coefficient_width;
 3712+        } while ( pc < pc_end );
 3713+        break;
 3714+      case 4:
 3715+        STBIR_NO_UNROLL_LOOP_START
 3716+        do {
 3717+          STBIR_MOVE_4( pc, coeffs );
 3718+          pc += 4;
 3719+          coeffs += coefficient_width;
 3720+        } while ( pc < pc_end );
 3721+        break;
 3722+      case 5:
 3723+        STBIR_NO_UNROLL_LOOP_START
 3724+        do {
 3725+          STBIR_MOVE_4( pc, coeffs );
 3726+          STBIR_MOVE_1( pc+4, coeffs+4 );
 3727+          pc += 5;
 3728+          coeffs += coefficient_width;
 3729+        } while ( pc < pc_end );
 3730+        break;
 3731+      case 6:
 3732+        STBIR_NO_UNROLL_LOOP_START
 3733+        do {
 3734+          STBIR_MOVE_4( pc, coeffs );
 3735+          STBIR_MOVE_2( pc+4, coeffs+4 );
 3736+          pc += 6;
 3737+          coeffs += coefficient_width;
 3738+        } while ( pc < pc_end );
 3739+        break;
 3740+      case 7:
 3741+        STBIR_NO_UNROLL_LOOP_START
 3742+        do {
 3743+          STBIR_MOVE_4( pc, coeffs );
 3744+          STBIR_MOVE_2( pc+4, coeffs+4 );
 3745+          STBIR_MOVE_1( pc+6, coeffs+6 );
 3746+          pc += 7;
 3747+          coeffs += coefficient_width;
 3748+        } while ( pc < pc_end );
 3749+        break;
 3750+      case 8:
 3751+        STBIR_NO_UNROLL_LOOP_START
 3752+        do {
 3753+          STBIR_MOVE_4( pc, coeffs );
 3754+          STBIR_MOVE_4( pc+4, coeffs+4 );
 3755+          pc += 8;
 3756+          coeffs += coefficient_width;
 3757+        } while ( pc < pc_end );
 3758+        break;
 3759+      case 9:
 3760+        STBIR_NO_UNROLL_LOOP_START
 3761+        do {
 3762+          STBIR_MOVE_4( pc, coeffs );
 3763+          STBIR_MOVE_4( pc+4, coeffs+4 );
 3764+          STBIR_MOVE_1( pc+8, coeffs+8 );
 3765+          pc += 9;
 3766+          coeffs += coefficient_width;
 3767+        } while ( pc < pc_end );
 3768+        break;
 3769+      case 10:
 3770+        STBIR_NO_UNROLL_LOOP_START
 3771+        do {
 3772+          STBIR_MOVE_4( pc, coeffs );
 3773+          STBIR_MOVE_4( pc+4, coeffs+4 );
 3774+          STBIR_MOVE_2( pc+8, coeffs+8 );
 3775+          pc += 10;
 3776+          coeffs += coefficient_width;
 3777+        } while ( pc < pc_end );
 3778+        break;
 3779+      case 11:
 3780+        STBIR_NO_UNROLL_LOOP_START
 3781+        do {
 3782+          STBIR_MOVE_4( pc, coeffs );
 3783+          STBIR_MOVE_4( pc+4, coeffs+4 );
 3784+          STBIR_MOVE_2( pc+8, coeffs+8 );
 3785+          STBIR_MOVE_1( pc+10, coeffs+10 );
 3786+          pc += 11;
 3787+          coeffs += coefficient_width;
 3788+        } while ( pc < pc_end );
 3789+        break;
 3790+      case 12:
 3791+        STBIR_NO_UNROLL_LOOP_START
 3792+        do {
 3793+          STBIR_MOVE_4( pc, coeffs );
 3794+          STBIR_MOVE_4( pc+4, coeffs+4 );
 3795+          STBIR_MOVE_4( pc+8, coeffs+8 );
 3796+          pc += 12;
 3797+          coeffs += coefficient_width;
 3798+        } while ( pc < pc_end );
 3799+        break;
 3800+      default:
 3801+        STBIR_NO_UNROLL_LOOP_START
 3802+        do {
 3803+          float * copy_end = pc + widest - 4;
 3804+          float * c = coeffs;
 3805+          do {
 3806+            STBIR_NO_UNROLL( pc );
 3807+            STBIR_MOVE_4( pc, c );
 3808+            pc += 4;
 3809+            c += 4;
 3810+          } while ( pc <= copy_end );
 3811+          copy_end += 4;
 3812+          STBIR_NO_UNROLL_LOOP_START
 3813+          while ( pc < copy_end )
 3814+          {
 3815+            STBIR_MOVE_1( pc, c );
 3816+            ++pc; ++c;
 3817+          }
 3818+          coeffs += coefficient_width;
 3819+        } while ( pc < pc_end );
 3820+        break;
 3821+    }
 3822+  }
 3823+
 3824+  // some horizontal routines read one float off the end (which is then masked off), so put in a sentinal so we don't read an snan or denormal
 3825+  coefficents[ widest * num_contributors ] = 8888.0f;
 3826+
 3827+  // the minimum we might read for unrolled filters widths is 12. So, we need to
 3828+  //   make sure we never read outside the decode buffer, by possibly moving
 3829+  //   the sample area back into the scanline, and putting zeros weights first.
 3830+  // we start on the right edge and check until we're well past the possible
 3831+  //   clip area (2*widest).
 3832+  {
 3833+    stbir__contributors * contribs = contributors + num_contributors - 1;
 3834+    float * coeffs = coefficents + widest * ( num_contributors - 1 );
 3835+
 3836+    // go until no chance of clipping (this is usually less than 8 lops)
 3837+    while ( ( contribs >= contributors ) && ( ( contribs->n0 + widest*2 ) >= row_end ) )
 3838+    {
 3839+      // might we clip??
 3840+      if ( ( contribs->n0 + widest ) > row_end )
 3841+      {
 3842+        int stop_range = widest;
 3843+
 3844+        // if range is larger than 12, it will be handled by generic loops that can terminate on the exact length
 3845+        //   of this contrib n1, instead of a fixed widest amount - so calculate this
 3846+        if ( widest > 12 )
 3847+        {
 3848+          int mod;
 3849+
 3850+          // how far will be read in the n_coeff loop (which depends on the widest count mod4);
 3851+          mod = widest & 3;
 3852+          stop_range = ( ( ( contribs->n1 - contribs->n0 + 1 ) - mod + 3 ) & ~3 ) + mod;
 3853+
 3854+          // the n_coeff loops do a minimum amount of coeffs, so factor that in!
 3855+          if ( stop_range < ( 8 + mod ) ) stop_range = 8 + mod;
 3856+        }
 3857+
 3858+        // now see if we still clip with the refined range
 3859+        if ( ( contribs->n0 + stop_range ) > row_end )
 3860+        {
 3861+          int new_n0 = row_end - stop_range;
 3862+          int num = contribs->n1 - contribs->n0 + 1;
 3863+          int backup = contribs->n0 - new_n0;
 3864+          float * from_co = coeffs + num - 1;
 3865+          float * to_co = from_co + backup;
 3866+
 3867+          STBIR_ASSERT( ( new_n0 >= row0 ) && ( new_n0 < contribs->n0 ) );
 3868+
 3869+          // move the coeffs over
 3870+          while( num )
 3871+          {
 3872+            *to_co-- = *from_co--;
 3873+            --num;
 3874+          }
 3875+          // zero new positions
 3876+          while ( to_co >= coeffs )
 3877+            *to_co-- = 0;
 3878+          // set new start point
 3879+          contribs->n0 = new_n0;
 3880+          if ( widest > 12 )
 3881+          {
 3882+            int mod;
 3883+
 3884+            // how far will be read in the n_coeff loop (which depends on the widest count mod4);
 3885+            mod = widest & 3;
 3886+            stop_range = ( ( ( contribs->n1 - contribs->n0 + 1 ) - mod + 3 ) & ~3 ) + mod;
 3887+
 3888+            // the n_coeff loops do a minimum amount of coeffs, so factor that in!
 3889+            if ( stop_range < ( 8 + mod ) ) stop_range = 8 + mod;
 3890+          }
 3891+        }
 3892+      }
 3893+      --contribs;
 3894+      coeffs -= widest;
 3895+    }
 3896+  }
 3897+
 3898+  return widest;
 3899+  #undef STBIR_MOVE_1
 3900+  #undef STBIR_MOVE_2
 3901+  #undef STBIR_MOVE_4
 3902+}
 3903+
 3904+static void stbir__calculate_filters( stbir__sampler * samp, stbir__sampler * other_axis_for_pivot, void * user_data STBIR_ONLY_PROFILE_BUILD_GET_INFO )
 3905+{
 3906+  int n;
 3907+  float scale = samp->scale_info.scale;
 3908+  stbir__kernel_callback * kernel = samp->filter_kernel;
 3909+  stbir__support_callback * support = samp->filter_support;
 3910+  float inv_scale = samp->scale_info.inv_scale;
 3911+  int input_full_size = samp->scale_info.input_full_size;
 3912+  int gather_num_contributors = samp->num_contributors;
 3913+  stbir__contributors* gather_contributors = samp->contributors;
 3914+  float * gather_coeffs = samp->coefficients;
 3915+  int gather_coefficient_width = samp->coefficient_width;
 3916+
 3917+  switch ( samp->is_gather )
 3918+  {
 3919+    case 1: // gather upsample
 3920+    {
 3921+      float out_pixels_radius = support(inv_scale,user_data) * scale;
 3922+
 3923+      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 );
 3924+
 3925+      STBIR_PROFILE_BUILD_START( cleanup );
 3926+      stbir__cleanup_gathered_coefficients( samp->edge, &samp->extent_info, &samp->scale_info, gather_num_contributors, gather_contributors, gather_coeffs, gather_coefficient_width );
 3927+      STBIR_PROFILE_BUILD_END( cleanup );
 3928+    }
 3929+    break;
 3930+
 3931+    case 0: // scatter downsample (only on vertical)
 3932+    case 2: // gather downsample
 3933+    {
 3934+      float in_pixels_radius = support(scale,user_data) * inv_scale;
 3935+      int filter_pixel_margin = samp->filter_pixel_margin;
 3936+      int input_end = input_full_size + filter_pixel_margin;
 3937+
 3938+      // if this is a scatter, we do a downsample gather to get the coeffs, and then pivot after
 3939+      if ( !samp->is_gather )
 3940+      {
 3941+        // check if we are using the same gather downsample on the horizontal as this vertical,
 3942+        //   if so, then we don't have to generate them, we can just pivot from the horizontal.
 3943+        if ( other_axis_for_pivot )
 3944+        {
 3945+          gather_contributors = other_axis_for_pivot->contributors;
 3946+          gather_coeffs = other_axis_for_pivot->coefficients;
 3947+          gather_coefficient_width = other_axis_for_pivot->coefficient_width;
 3948+          gather_num_contributors = other_axis_for_pivot->num_contributors;
 3949+          samp->extent_info.lowest = other_axis_for_pivot->extent_info.lowest;
 3950+          samp->extent_info.highest = other_axis_for_pivot->extent_info.highest;
 3951+          samp->extent_info.widest = other_axis_for_pivot->extent_info.widest;
 3952+          goto jump_right_to_pivot;
 3953+        }
 3954+
 3955+        gather_contributors = samp->gather_prescatter_contributors;
 3956+        gather_coeffs = samp->gather_prescatter_coefficients;
 3957+        gather_coefficient_width = samp->gather_prescatter_coefficient_width;
 3958+        gather_num_contributors = samp->gather_prescatter_num_contributors;
 3959+      }
 3960+
 3961+      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 );
 3962+
 3963+      STBIR_PROFILE_BUILD_START( cleanup );
 3964+      stbir__cleanup_gathered_coefficients( samp->edge, &samp->extent_info, &samp->scale_info, gather_num_contributors, gather_contributors, gather_coeffs, gather_coefficient_width );
 3965+      STBIR_PROFILE_BUILD_END( cleanup );
 3966+
 3967+      if ( !samp->is_gather )
 3968+      {
 3969+        // if this is a scatter (vertical only), then we need to pivot the coeffs
 3970+        stbir__contributors * scatter_contributors;
 3971+        int highest_set;
 3972+
 3973+        jump_right_to_pivot:
 3974+
 3975+        STBIR_PROFILE_BUILD_START( pivot );
 3976+
 3977+        highest_set = (-filter_pixel_margin) - 1;
 3978+        for (n = 0; n < gather_num_contributors; n++)
 3979+        {
 3980+          int k;
 3981+          int gn0 = gather_contributors->n0, gn1 = gather_contributors->n1;
 3982+          int scatter_coefficient_width = samp->coefficient_width;
 3983+          float * scatter_coeffs = samp->coefficients + ( gn0 + filter_pixel_margin ) * scatter_coefficient_width;
 3984+          float * g_coeffs = gather_coeffs;
 3985+          scatter_contributors = samp->contributors + ( gn0 + filter_pixel_margin );
 3986+
 3987+          for (k = gn0 ; k <= gn1 ; k++ )
 3988+          {
 3989+            float gc = *g_coeffs++;
 3990+            
 3991+            // skip zero and denormals - must skip zeros to avoid adding coeffs beyond scatter_coefficient_width
 3992+            //   (which happens when pivoting from horizontal, which might have dummy zeros)
 3993+            if ( ( ( gc >= stbir__small_float ) || ( gc <= -stbir__small_float ) ) )
 3994+            {
 3995+              if ( ( k > highest_set ) || ( scatter_contributors->n0 > scatter_contributors->n1 ) )
 3996+              {
 3997+                {
 3998+                  // if we are skipping over several contributors, we need to clear the skipped ones
 3999+                  stbir__contributors * clear_contributors = samp->contributors + ( highest_set + filter_pixel_margin + 1);
 4000+                  while ( clear_contributors < scatter_contributors )
 4001+                  {
 4002+                    clear_contributors->n0 = 0;
 4003+                    clear_contributors->n1 = -1;
 4004+                    ++clear_contributors;
 4005+                  }
 4006+                }
 4007+                scatter_contributors->n0 = n;
 4008+                scatter_contributors->n1 = n;
 4009+                scatter_coeffs[0]  = gc;
 4010+                highest_set = k;
 4011+              }
 4012+              else
 4013+              {
 4014+                stbir__insert_coeff( scatter_contributors, scatter_coeffs, n, gc, scatter_coefficient_width );
 4015+              }
 4016+              STBIR_ASSERT( ( scatter_contributors->n1 - scatter_contributors->n0 + 1 ) <= scatter_coefficient_width );
 4017+            }
 4018+            ++scatter_contributors;
 4019+            scatter_coeffs += scatter_coefficient_width;
 4020+          }
 4021+
 4022+          ++gather_contributors;
 4023+          gather_coeffs += gather_coefficient_width;
 4024+        }
 4025+
 4026+        // now clear any unset contribs
 4027+        {
 4028+          stbir__contributors * clear_contributors = samp->contributors + ( highest_set + filter_pixel_margin + 1);
 4029+          stbir__contributors * end_contributors = samp->contributors + samp->num_contributors;
 4030+          while ( clear_contributors < end_contributors )
 4031+          {
 4032+            clear_contributors->n0 = 0;
 4033+            clear_contributors->n1 = -1;
 4034+            ++clear_contributors;
 4035+          }
 4036+        }
 4037+
 4038+        STBIR_PROFILE_BUILD_END( pivot );
 4039+      }
 4040+    }
 4041+    break;
 4042+  }
 4043+}
 4044+
 4045+
 4046+//========================================================================================================
 4047+// scanline decoders and encoders
 4048+
 4049+#define stbir__coder_min_num 1
 4050+#define STB_IMAGE_RESIZE_DO_CODERS
 4051+#include STBIR__HEADER_FILENAME
 4052+
 4053+#define stbir__decode_suffix BGRA
 4054+#define stbir__decode_swizzle
 4055+#define stbir__decode_order0  2
 4056+#define stbir__decode_order1  1
 4057+#define stbir__decode_order2  0
 4058+#define stbir__decode_order3  3
 4059+#define stbir__encode_order0  2
 4060+#define stbir__encode_order1  1
 4061+#define stbir__encode_order2  0
 4062+#define stbir__encode_order3  3
 4063+#define stbir__coder_min_num 4
 4064+#define STB_IMAGE_RESIZE_DO_CODERS
 4065+#include STBIR__HEADER_FILENAME
 4066+
 4067+#define stbir__decode_suffix ARGB
 4068+#define stbir__decode_swizzle
 4069+#define stbir__decode_order0  1
 4070+#define stbir__decode_order1  2
 4071+#define stbir__decode_order2  3
 4072+#define stbir__decode_order3  0
 4073+#define stbir__encode_order0  3
 4074+#define stbir__encode_order1  0
 4075+#define stbir__encode_order2  1
 4076+#define stbir__encode_order3  2
 4077+#define stbir__coder_min_num 4
 4078+#define STB_IMAGE_RESIZE_DO_CODERS
 4079+#include STBIR__HEADER_FILENAME
 4080+
 4081+#define stbir__decode_suffix ABGR
 4082+#define stbir__decode_swizzle
 4083+#define stbir__decode_order0  3
 4084+#define stbir__decode_order1  2
 4085+#define stbir__decode_order2  1
 4086+#define stbir__decode_order3  0
 4087+#define stbir__encode_order0  3
 4088+#define stbir__encode_order1  2
 4089+#define stbir__encode_order2  1
 4090+#define stbir__encode_order3  0
 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 AR
 4096+#define stbir__decode_swizzle
 4097+#define stbir__decode_order0  1
 4098+#define stbir__decode_order1  0
 4099+#define stbir__decode_order2  3
 4100+#define stbir__decode_order3  2
 4101+#define stbir__encode_order0  1
 4102+#define stbir__encode_order1  0
 4103+#define stbir__encode_order2  3
 4104+#define stbir__encode_order3  2
 4105+#define stbir__coder_min_num 2
 4106+#define STB_IMAGE_RESIZE_DO_CODERS
 4107+#include STBIR__HEADER_FILENAME
 4108+
 4109+
 4110+// fancy alpha means we expand to keep both premultipied and non-premultiplied color channels
 4111+static void stbir__fancy_alpha_weight_4ch( float * out_buffer, int width_times_channels )
 4112+{
 4113+  float STBIR_STREAMOUT_PTR(*) out = out_buffer;
 4114+  float const * end_decode = out_buffer + ( width_times_channels / 4 ) * 7;  // decode buffer aligned to end of out_buffer
 4115+  float STBIR_STREAMOUT_PTR(*) decode = (float*)end_decode - width_times_channels;
 4116+
 4117+  // fancy alpha is stored internally as R G B A Rpm Gpm Bpm
 4118+
 4119+  #ifdef STBIR_SIMD
 4120+
 4121+  #ifdef STBIR_SIMD8
 4122+  decode += 16;
 4123+  STBIR_NO_UNROLL_LOOP_START
 4124+  while ( decode <= end_decode )
 4125+  {
 4126+    stbir__simdf8 d0,d1,a0,a1,p0,p1;
 4127+    STBIR_NO_UNROLL(decode);
 4128+    stbir__simdf8_load( d0, decode-16 );
 4129+    stbir__simdf8_load( d1, decode-16+8 );
 4130+    stbir__simdf8_0123to33333333( a0, d0 );
 4131+    stbir__simdf8_0123to33333333( a1, d1 );
 4132+    stbir__simdf8_mult( p0, a0, d0 );
 4133+    stbir__simdf8_mult( p1, a1, d1 );
 4134+    stbir__simdf8_bot4s( a0, d0, p0 );
 4135+    stbir__simdf8_bot4s( a1, d1, p1 );
 4136+    stbir__simdf8_top4s( d0, d0, p0 );
 4137+    stbir__simdf8_top4s( d1, d1, p1 );
 4138+    stbir__simdf8_store ( out, a0 );
 4139+    stbir__simdf8_store ( out+7, d0 );
 4140+    stbir__simdf8_store ( out+14, a1 );
 4141+    stbir__simdf8_store ( out+21, d1 );
 4142+    decode += 16;
 4143+    out += 28;
 4144+  }
 4145+  decode -= 16;
 4146+  #else
 4147+  decode += 8;
 4148+  STBIR_NO_UNROLL_LOOP_START
 4149+  while ( decode <= end_decode )
 4150+  {
 4151+    stbir__simdf d0,a0,d1,a1,p0,p1;
 4152+    STBIR_NO_UNROLL(decode);
 4153+    stbir__simdf_load( d0, decode-8 );
 4154+    stbir__simdf_load( d1, decode-8+4 );
 4155+    stbir__simdf_0123to3333( a0, d0 );
 4156+    stbir__simdf_0123to3333( a1, d1 );
 4157+    stbir__simdf_mult( p0, a0, d0 );
 4158+    stbir__simdf_mult( p1, a1, d1 );
 4159+    stbir__simdf_store ( out, d0 );
 4160+    stbir__simdf_store ( out+4, p0 );
 4161+    stbir__simdf_store ( out+7, d1 );
 4162+    stbir__simdf_store ( out+7+4, p1 );
 4163+    decode += 8;
 4164+    out += 14;
 4165+  }
 4166+  decode -= 8;
 4167+  #endif
 4168+
 4169+  // might be one last odd pixel
 4170+  #ifdef STBIR_SIMD8
 4171+  STBIR_NO_UNROLL_LOOP_START
 4172+  while ( decode < end_decode )
 4173+  #else
 4174+  if ( decode < end_decode )
 4175+  #endif
 4176+  {
 4177+    stbir__simdf d,a,p;
 4178+    STBIR_NO_UNROLL(decode);
 4179+    stbir__simdf_load( d, decode );
 4180+    stbir__simdf_0123to3333( a, d );
 4181+    stbir__simdf_mult( p, a, d );
 4182+    stbir__simdf_store ( out, d );
 4183+    stbir__simdf_store ( out+4, p );
 4184+    decode += 4;
 4185+    out += 7;
 4186+  }
 4187+
 4188+  #else
 4189+
 4190+  while( decode < end_decode )
 4191+  {
 4192+    float r = decode[0], g = decode[1], b = decode[2], alpha = decode[3];
 4193+    out[0] = r;
 4194+    out[1] = g;
 4195+    out[2] = b;
 4196+    out[3] = alpha;
 4197+    out[4] = r * alpha;
 4198+    out[5] = g * alpha;
 4199+    out[6] = b * alpha;
 4200+    out += 7;
 4201+    decode += 4;
 4202+  }
 4203+
 4204+  #endif
 4205+}
 4206+
 4207+static void stbir__fancy_alpha_weight_2ch( float * out_buffer, int width_times_channels )
 4208+{
 4209+  float STBIR_STREAMOUT_PTR(*) out = out_buffer;
 4210+  float const * end_decode = out_buffer + ( width_times_channels / 2 ) * 3;
 4211+  float STBIR_STREAMOUT_PTR(*) decode = (float*)end_decode - width_times_channels;
 4212+
 4213+  //  for fancy alpha, turns into: [X A Xpm][X A Xpm],etc
 4214+
 4215+  #ifdef STBIR_SIMD
 4216+
 4217+  decode += 8;
 4218+  if ( decode <= end_decode )
 4219+  {
 4220+    STBIR_NO_UNROLL_LOOP_START
 4221+    do {
 4222+      #ifdef STBIR_SIMD8
 4223+      stbir__simdf8 d0,a0,p0;
 4224+      STBIR_NO_UNROLL(decode);
 4225+      stbir__simdf8_load( d0, decode-8 );
 4226+      stbir__simdf8_0123to11331133( p0, d0 );
 4227+      stbir__simdf8_0123to00220022( a0, d0 );
 4228+      stbir__simdf8_mult( p0, p0, a0 );
 4229+
 4230+      stbir__simdf_store2( out, stbir__if_simdf8_cast_to_simdf4( d0 ) );
 4231+      stbir__simdf_store( out+2, stbir__if_simdf8_cast_to_simdf4( p0 ) );
 4232+      stbir__simdf_store2h( out+3, stbir__if_simdf8_cast_to_simdf4( d0 ) );
 4233+
 4234+      stbir__simdf_store2( out+6, stbir__simdf8_gettop4( d0 ) );
 4235+      stbir__simdf_store( out+8, stbir__simdf8_gettop4( p0 ) );
 4236+      stbir__simdf_store2h( out+9, stbir__simdf8_gettop4( d0 ) );
 4237+      #else
 4238+      stbir__simdf d0,a0,d1,a1,p0,p1;
 4239+      STBIR_NO_UNROLL(decode);
 4240+      stbir__simdf_load( d0, decode-8 );
 4241+      stbir__simdf_load( d1, decode-8+4 );
 4242+      stbir__simdf_0123to1133( p0, d0 );
 4243+      stbir__simdf_0123to1133( p1, d1 );
 4244+      stbir__simdf_0123to0022( a0, d0 );
 4245+      stbir__simdf_0123to0022( a1, d1 );
 4246+      stbir__simdf_mult( p0, p0, a0 );
 4247+      stbir__simdf_mult( p1, p1, a1 );
 4248+
 4249+      stbir__simdf_store2( out, d0 );
 4250+      stbir__simdf_store( out+2, p0 );
 4251+      stbir__simdf_store2h( out+3, d0 );
 4252+
 4253+      stbir__simdf_store2( out+6, d1 );
 4254+      stbir__simdf_store( out+8, p1 );
 4255+      stbir__simdf_store2h( out+9, d1 );
 4256+      #endif
 4257+      decode += 8;
 4258+      out += 12;
 4259+    } while ( decode <= end_decode );
 4260+  }
 4261+  decode -= 8;
 4262+  #endif
 4263+
 4264+  STBIR_SIMD_NO_UNROLL_LOOP_START
 4265+  while( decode < end_decode )
 4266+  {
 4267+    float x = decode[0], y = decode[1];
 4268+    STBIR_SIMD_NO_UNROLL(decode);
 4269+    out[0] = x;
 4270+    out[1] = y;
 4271+    out[2] = x * y;
 4272+    out += 3;
 4273+    decode += 2;
 4274+  }
 4275+}
 4276+
 4277+static void stbir__fancy_alpha_unweight_4ch( float * encode_buffer, int width_times_channels )
 4278+{
 4279+  float STBIR_SIMD_STREAMOUT_PTR(*) encode = encode_buffer;
 4280+  float STBIR_SIMD_STREAMOUT_PTR(*) input = encode_buffer;
 4281+  float const * end_output = encode_buffer + width_times_channels;
 4282+
 4283+  // fancy RGBA is stored internally as R G B A Rpm Gpm Bpm
 4284+
 4285+  STBIR_SIMD_NO_UNROLL_LOOP_START
 4286+  do {
 4287+    float alpha = input[3];
 4288+#ifdef STBIR_SIMD
 4289+    stbir__simdf i,ia;
 4290+    STBIR_SIMD_NO_UNROLL(encode);
 4291+    if ( alpha < stbir__small_float )
 4292+    {
 4293+      stbir__simdf_load( i, input );
 4294+      stbir__simdf_store( encode, i );
 4295+    }
 4296+    else
 4297+    {
 4298+      stbir__simdf_load1frep4( ia, 1.0f / alpha );
 4299+      stbir__simdf_load( i, input+4 );
 4300+      stbir__simdf_mult( i, i, ia );
 4301+      stbir__simdf_store( encode, i );
 4302+      encode[3] = alpha;
 4303+    }
 4304+#else
 4305+    if ( alpha < stbir__small_float )
 4306+    {
 4307+      encode[0] = input[0];
 4308+      encode[1] = input[1];
 4309+      encode[2] = input[2];
 4310+    }
 4311+    else
 4312+    {
 4313+      float ialpha = 1.0f / alpha;
 4314+      encode[0] = input[4] * ialpha;
 4315+      encode[1] = input[5] * ialpha;
 4316+      encode[2] = input[6] * ialpha;
 4317+    }
 4318+    encode[3] = alpha;
 4319+#endif
 4320+
 4321+    input += 7;
 4322+    encode += 4;
 4323+  } while ( encode < end_output );
 4324+}
 4325+
 4326+//  format: [X A Xpm][X A Xpm] etc
 4327+static void stbir__fancy_alpha_unweight_2ch( float * encode_buffer, int width_times_channels )
 4328+{
 4329+  float STBIR_SIMD_STREAMOUT_PTR(*) encode = encode_buffer;
 4330+  float STBIR_SIMD_STREAMOUT_PTR(*) input = encode_buffer;
 4331+  float const * end_output = encode_buffer + width_times_channels;
 4332+
 4333+  do {
 4334+    float alpha = input[1];
 4335+    encode[0] = input[0];
 4336+    if ( alpha >= stbir__small_float )
 4337+      encode[0] = input[2] / alpha;
 4338+    encode[1] = alpha;
 4339+
 4340+    input += 3;
 4341+    encode += 2;
 4342+  } while ( encode < end_output );
 4343+}
 4344+
 4345+static void stbir__simple_alpha_weight_4ch( float * decode_buffer, int width_times_channels )
 4346+{
 4347+  float STBIR_STREAMOUT_PTR(*) decode = decode_buffer;
 4348+  float const * end_decode = decode_buffer + width_times_channels;
 4349+
 4350+  #ifdef STBIR_SIMD
 4351+  {
 4352+    decode += 2 * stbir__simdfX_float_count;
 4353+    STBIR_NO_UNROLL_LOOP_START
 4354+    while ( decode <= end_decode )
 4355+    {
 4356+      stbir__simdfX d0,a0,d1,a1;
 4357+      STBIR_NO_UNROLL(decode);
 4358+      stbir__simdfX_load( d0, decode-2*stbir__simdfX_float_count );
 4359+      stbir__simdfX_load( d1, decode-2*stbir__simdfX_float_count+stbir__simdfX_float_count );
 4360+      stbir__simdfX_aaa1( a0, d0, STBIR_onesX );
 4361+      stbir__simdfX_aaa1( a1, d1, STBIR_onesX );
 4362+      stbir__simdfX_mult( d0, d0, a0 );
 4363+      stbir__simdfX_mult( d1, d1, a1 );
 4364+      stbir__simdfX_store ( decode-2*stbir__simdfX_float_count, d0 );
 4365+      stbir__simdfX_store ( decode-2*stbir__simdfX_float_count+stbir__simdfX_float_count, d1 );
 4366+      decode += 2 * stbir__simdfX_float_count;
 4367+    }
 4368+    decode -= 2 * stbir__simdfX_float_count;
 4369+
 4370+    // few last pixels remnants
 4371+    #ifdef STBIR_SIMD8
 4372+    STBIR_NO_UNROLL_LOOP_START
 4373+    while ( decode < end_decode )
 4374+    #else
 4375+    if ( decode < end_decode )
 4376+    #endif
 4377+    {
 4378+      stbir__simdf d,a;
 4379+      stbir__simdf_load( d, decode );
 4380+      stbir__simdf_aaa1( a, d, STBIR__CONSTF(STBIR_ones) );
 4381+      stbir__simdf_mult( d, d, a );
 4382+      stbir__simdf_store ( decode, d );
 4383+      decode += 4;
 4384+    }
 4385+  }
 4386+
 4387+  #else
 4388+
 4389+  while( decode < end_decode )
 4390+  {
 4391+    float alpha = decode[3];
 4392+    decode[0] *= alpha;
 4393+    decode[1] *= alpha;
 4394+    decode[2] *= alpha;
 4395+    decode += 4;
 4396+  }
 4397+
 4398+  #endif
 4399+}
 4400+
 4401+static void stbir__simple_alpha_weight_2ch( float * decode_buffer, int width_times_channels )
 4402+{
 4403+  float STBIR_STREAMOUT_PTR(*) decode = decode_buffer;
 4404+  float const * end_decode = decode_buffer + width_times_channels;
 4405+
 4406+  #ifdef STBIR_SIMD
 4407+  decode += 2 * stbir__simdfX_float_count;
 4408+  STBIR_NO_UNROLL_LOOP_START
 4409+  while ( decode <= end_decode )
 4410+  {
 4411+    stbir__simdfX d0,a0,d1,a1;
 4412+    STBIR_NO_UNROLL(decode);
 4413+    stbir__simdfX_load( d0, decode-2*stbir__simdfX_float_count );
 4414+    stbir__simdfX_load( d1, decode-2*stbir__simdfX_float_count+stbir__simdfX_float_count );
 4415+    stbir__simdfX_a1a1( a0, d0, STBIR_onesX );
 4416+    stbir__simdfX_a1a1( a1, d1, STBIR_onesX );
 4417+    stbir__simdfX_mult( d0, d0, a0 );
 4418+    stbir__simdfX_mult( d1, d1, a1 );
 4419+    stbir__simdfX_store ( decode-2*stbir__simdfX_float_count, d0 );
 4420+    stbir__simdfX_store ( decode-2*stbir__simdfX_float_count+stbir__simdfX_float_count, d1 );
 4421+    decode += 2 * stbir__simdfX_float_count;
 4422+  }
 4423+  decode -= 2 * stbir__simdfX_float_count;
 4424+  #endif
 4425+
 4426+  STBIR_SIMD_NO_UNROLL_LOOP_START
 4427+  while( decode < end_decode )
 4428+  {
 4429+    float alpha = decode[1];
 4430+    STBIR_SIMD_NO_UNROLL(decode);
 4431+    decode[0] *= alpha;
 4432+    decode += 2;
 4433+  }
 4434+}
 4435+
 4436+static void stbir__simple_alpha_unweight_4ch( float * encode_buffer, int width_times_channels )
 4437+{
 4438+  float STBIR_SIMD_STREAMOUT_PTR(*) encode = encode_buffer;
 4439+  float const * end_output = encode_buffer + width_times_channels;
 4440+
 4441+  STBIR_SIMD_NO_UNROLL_LOOP_START
 4442+  do {
 4443+    float alpha = encode[3];
 4444+
 4445+#ifdef STBIR_SIMD
 4446+    stbir__simdf i,ia;
 4447+    STBIR_SIMD_NO_UNROLL(encode);
 4448+    if ( alpha >= stbir__small_float )
 4449+    {
 4450+      stbir__simdf_load1frep4( ia, 1.0f / alpha );
 4451+      stbir__simdf_load( i, encode );
 4452+      stbir__simdf_mult( i, i, ia );
 4453+      stbir__simdf_store( encode, i );
 4454+      encode[3] = alpha;
 4455+    }
 4456+#else
 4457+    if ( alpha >= stbir__small_float )
 4458+    {
 4459+      float ialpha = 1.0f / alpha;
 4460+      encode[0] *= ialpha;
 4461+      encode[1] *= ialpha;
 4462+      encode[2] *= ialpha;
 4463+    }
 4464+#endif
 4465+    encode += 4;
 4466+  } while ( encode < end_output );
 4467+}
 4468+
 4469+static void stbir__simple_alpha_unweight_2ch( float * encode_buffer, int width_times_channels )
 4470+{
 4471+  float STBIR_SIMD_STREAMOUT_PTR(*) encode = encode_buffer;
 4472+  float const * end_output = encode_buffer + width_times_channels;
 4473+
 4474+  do {
 4475+    float alpha = encode[1];
 4476+    if ( alpha >= stbir__small_float )
 4477+      encode[0] /= alpha;
 4478+    encode += 2;
 4479+  } while ( encode < end_output );
 4480+}
 4481+
 4482+
 4483+// only used in RGB->BGR or BGR->RGB
 4484+static void stbir__simple_flip_3ch( float * decode_buffer, int width_times_channels )
 4485+{
 4486+  float STBIR_STREAMOUT_PTR(*) decode = decode_buffer;
 4487+  float const * end_decode = decode_buffer + width_times_channels;
 4488+
 4489+#ifdef STBIR_SIMD
 4490+    #ifdef stbir__simdf_swiz2 // do we have two argument swizzles?
 4491+      end_decode -= 12; 
 4492+      STBIR_NO_UNROLL_LOOP_START
 4493+      while( decode <= end_decode )
 4494+      {
 4495+        // on arm64 8 instructions, no overlapping stores
 4496+        stbir__simdf a,b,c,na,nb;
 4497+        STBIR_SIMD_NO_UNROLL(decode);
 4498+        stbir__simdf_load( a, decode );
 4499+        stbir__simdf_load( b, decode+4 );
 4500+        stbir__simdf_load( c, decode+8 );
 4501+
 4502+        na = stbir__simdf_swiz2( a, b, 2, 1, 0, 5 );   
 4503+        b  = stbir__simdf_swiz2( a, b, 4, 3, 6, 7 );   
 4504+        nb = stbir__simdf_swiz2( b, c, 0, 1, 4, 3 );   
 4505+        c  = stbir__simdf_swiz2( b, c, 2, 7, 6, 5 );   
 4506+
 4507+        stbir__simdf_store( decode, na );
 4508+        stbir__simdf_store( decode+4, nb ); 
 4509+        stbir__simdf_store( decode+8, c );
 4510+        decode += 12;
 4511+      }
 4512+      end_decode += 12;
 4513+    #else
 4514+      end_decode -= 24;
 4515+      STBIR_NO_UNROLL_LOOP_START
 4516+      while( decode <= end_decode )
 4517+      {
 4518+        // 26 instructions on x64
 4519+        stbir__simdf a,b,c,d,e,f,g;
 4520+        float i21, i23;
 4521+        STBIR_SIMD_NO_UNROLL(decode);
 4522+        stbir__simdf_load( a, decode );
 4523+        stbir__simdf_load( b, decode+3 );
 4524+        stbir__simdf_load( c, decode+6 );
 4525+        stbir__simdf_load( d, decode+9 );
 4526+        stbir__simdf_load( e, decode+12 );
 4527+        stbir__simdf_load( f, decode+15 );
 4528+        stbir__simdf_load( g, decode+18 );
 4529+
 4530+        a = stbir__simdf_swiz( a, 2, 1, 0, 3 );   
 4531+        b = stbir__simdf_swiz( b, 2, 1, 0, 3 );   
 4532+        c = stbir__simdf_swiz( c, 2, 1, 0, 3 );   
 4533+        d = stbir__simdf_swiz( d, 2, 1, 0, 3 );   
 4534+        e = stbir__simdf_swiz( e, 2, 1, 0, 3 );   
 4535+        f = stbir__simdf_swiz( f, 2, 1, 0, 3 );   
 4536+        g = stbir__simdf_swiz( g, 2, 1, 0, 3 );   
 4537+
 4538+        // stores overlap, need to be in order, 
 4539+        stbir__simdf_store( decode,    a );
 4540+        i21 = decode[21];
 4541+        stbir__simdf_store( decode+3,  b ); 
 4542+        i23 = decode[23];
 4543+        stbir__simdf_store( decode+6,  c );
 4544+        stbir__simdf_store( decode+9,  d );
 4545+        stbir__simdf_store( decode+12, e );
 4546+        stbir__simdf_store( decode+15, f );
 4547+        stbir__simdf_store( decode+18, g );
 4548+        decode[21] = i23;
 4549+        decode[23] = i21;
 4550+        decode += 24;
 4551+      }
 4552+      end_decode += 24;
 4553+    #endif
 4554+#else
 4555+  end_decode -= 12;
 4556+  STBIR_NO_UNROLL_LOOP_START
 4557+  while( decode <= end_decode )
 4558+  {
 4559+    // 16 instructions
 4560+    float t0,t1,t2,t3;
 4561+    STBIR_NO_UNROLL(decode);
 4562+    t0 = decode[0]; t1 = decode[3]; t2 = decode[6]; t3 = decode[9];
 4563+    decode[0] = decode[2]; decode[3] = decode[5]; decode[6] = decode[8]; decode[9] = decode[11];
 4564+    decode[2] = t0; decode[5] = t1; decode[8] = t2; decode[11] = t3;
 4565+    decode += 12;
 4566+  }
 4567+  end_decode += 12;
 4568+#endif
 4569+
 4570+  STBIR_NO_UNROLL_LOOP_START
 4571+  while( decode < end_decode )
 4572+  {
 4573+    float t = decode[0];
 4574+    STBIR_NO_UNROLL(decode);
 4575+    decode[0] = decode[2];
 4576+    decode[2] = t;
 4577+    decode += 3;
 4578+  }
 4579+}
 4580+
 4581+
 4582+
 4583+static void stbir__decode_scanline(stbir__info const * stbir_info, int n, float * output_buffer STBIR_ONLY_PROFILE_GET_SPLIT_INFO )
 4584+{
 4585+  int channels = stbir_info->channels;
 4586+  int effective_channels = stbir_info->effective_channels;
 4587+  int input_sample_in_bytes = stbir__type_size[stbir_info->input_type] * channels;
 4588+  stbir_edge edge_horizontal = stbir_info->horizontal.edge;
 4589+  stbir_edge edge_vertical = stbir_info->vertical.edge;
 4590+  int row = stbir__edge_wrap(edge_vertical, n, stbir_info->vertical.scale_info.input_full_size);
 4591+  const void* input_plane_data = ( (char *) stbir_info->input_data ) + (size_t)row * (size_t) stbir_info->input_stride_bytes;
 4592+  stbir__span const * spans = stbir_info->scanline_extents.spans;
 4593+  float * full_decode_buffer = output_buffer - stbir_info->scanline_extents.conservative.n0 * effective_channels;
 4594+  float * last_decoded = 0;
 4595+
 4596+  // if we are on edge_zero, and we get in here with an out of bounds n, then the calculate filters has failed
 4597+  STBIR_ASSERT( !(edge_vertical == STBIR_EDGE_ZERO && (n < 0 || n >= stbir_info->vertical.scale_info.input_full_size)) );
 4598+
 4599+  do
 4600+  {
 4601+    float * decode_buffer;
 4602+    void const * input_data;
 4603+    float * end_decode;
 4604+    int width_times_channels;
 4605+    int width;
 4606+
 4607+    if ( spans->n1 < spans->n0 )
 4608+      break;
 4609+
 4610+    width = spans->n1 + 1 - spans->n0;
 4611+    decode_buffer = full_decode_buffer + spans->n0 * effective_channels;
 4612+    end_decode = full_decode_buffer + ( spans->n1 + 1 ) * effective_channels;
 4613+    width_times_channels = width * channels;
 4614+
 4615+    // read directly out of input plane by default
 4616+    input_data = ( (char*)input_plane_data ) + spans->pixel_offset_for_input * input_sample_in_bytes;
 4617+
 4618+    // if we have an input callback, call it to get the input data
 4619+    if ( stbir_info->in_pixels_cb )
 4620+    {
 4621+      // 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
 4622+      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 );
 4623+    }
 4624+
 4625+    STBIR_PROFILE_START( decode );
 4626+    // 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)
 4627+    last_decoded = stbir_info->decode_pixels( (float*)end_decode - width_times_channels, width_times_channels, input_data );
 4628+    STBIR_PROFILE_END( decode );
 4629+
 4630+    if (stbir_info->alpha_weight)
 4631+    {
 4632+      STBIR_PROFILE_START( alpha );
 4633+      stbir_info->alpha_weight( decode_buffer, width_times_channels );
 4634+      STBIR_PROFILE_END( alpha );
 4635+    }
 4636+
 4637+    ++spans;
 4638+  } while ( spans <= ( &stbir_info->scanline_extents.spans[1] ) );
 4639+
 4640+  // handle the edge_wrap filter (all other types are handled back out at the calculate_filter stage)
 4641+  // basically the idea here is that if we have the whole scanline in memory, we don't redecode the
 4642+  //   wrapped edge pixels, and instead just memcpy them from the scanline into the edge positions
 4643+  if ( ( edge_horizontal == STBIR_EDGE_WRAP ) && ( stbir_info->scanline_extents.edge_sizes[0] | stbir_info->scanline_extents.edge_sizes[1] ) )
 4644+  {
 4645+    // this code only runs if we're in edge_wrap, and we're doing the entire scanline
 4646+    int e, start_x[2];
 4647+    int input_full_size = stbir_info->horizontal.scale_info.input_full_size;
 4648+
 4649+    start_x[0] = -stbir_info->scanline_extents.edge_sizes[0];  // left edge start x
 4650+    start_x[1] =  input_full_size;                             // right edge
 4651+
 4652+    for( e = 0; e < 2 ; e++ )
 4653+    {
 4654+      // do each margin
 4655+      int margin = stbir_info->scanline_extents.edge_sizes[e];
 4656+      if ( margin )
 4657+      {
 4658+        int x = start_x[e];
 4659+        float * marg = full_decode_buffer + x * effective_channels;
 4660+        float const * src = full_decode_buffer + stbir__edge_wrap(edge_horizontal, x, input_full_size) * effective_channels;
 4661+        STBIR_MEMCPY( marg, src, margin * effective_channels * sizeof(float) );
 4662+        if ( e == 1 ) last_decoded = marg + margin * effective_channels;
 4663+      }
 4664+    }
 4665+  }
 4666+  
 4667+  // 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
 4668+  //   (we can't pre-zero it, because the input callback can use that area as padding)
 4669+  last_decoded[0] = 0.0f; 
 4670+
 4671+  // we clear this extra float, because the final output pixel filter kernel might have used one less coeff than the max filter width
 4672+  //   when this happens, we do read that pixel from the input, so it too could be Nan, so just zero an extra one.
 4673+  //   this fits because each scanline is padded by three floats (STBIR_INPUT_CALLBACK_PADDING)
 4674+  last_decoded[1] = 0.0f;
 4675+}
 4676+
 4677+
 4678+//=================
 4679+// Do 1 channel horizontal routines
 4680+
 4681+#ifdef STBIR_SIMD
 4682+
 4683+#define stbir__1_coeff_only()          \
 4684+    stbir__simdf tot,c;                \
 4685+    STBIR_SIMD_NO_UNROLL(decode);      \
 4686+    stbir__simdf_load1( c, hc );       \
 4687+    stbir__simdf_mult1_mem( tot, c, decode );
 4688+
 4689+#define stbir__2_coeff_only()          \
 4690+    stbir__simdf tot,c,d;              \
 4691+    STBIR_SIMD_NO_UNROLL(decode);      \
 4692+    stbir__simdf_load2z( c, hc );      \
 4693+    stbir__simdf_load2( d, decode );   \
 4694+    stbir__simdf_mult( tot, c, d );    \
 4695+    stbir__simdf_0123to1230( c, tot ); \
 4696+    stbir__simdf_add1( tot, tot, c );
 4697+
 4698+#define stbir__3_coeff_only()                  \
 4699+    stbir__simdf tot,c,t;                      \
 4700+    STBIR_SIMD_NO_UNROLL(decode);              \
 4701+    stbir__simdf_load( c, hc );                \
 4702+    stbir__simdf_mult_mem( tot, c, decode );   \
 4703+    stbir__simdf_0123to1230( c, tot );         \
 4704+    stbir__simdf_0123to2301( t, tot );         \
 4705+    stbir__simdf_add1( tot, tot, c );          \
 4706+    stbir__simdf_add1( tot, tot, t );
 4707+
 4708+#define stbir__store_output_tiny()                \
 4709+    stbir__simdf_store1( output, tot );           \
 4710+    horizontal_coefficients += coefficient_width; \
 4711+    ++horizontal_contributors;                    \
 4712+    output += 1;
 4713+
 4714+#define stbir__4_coeff_start()                 \
 4715+    stbir__simdf tot,c;                        \
 4716+    STBIR_SIMD_NO_UNROLL(decode);              \
 4717+    stbir__simdf_load( c, hc );                \
 4718+    stbir__simdf_mult_mem( tot, c, decode );   \
 4719+
 4720+#define stbir__4_coeff_continue_from_4( ofs )  \
 4721+    STBIR_SIMD_NO_UNROLL(decode);              \
 4722+    stbir__simdf_load( c, hc + (ofs) );        \
 4723+    stbir__simdf_madd_mem( tot, tot, c, decode+(ofs) );
 4724+
 4725+#define stbir__1_coeff_remnant( ofs )          \
 4726+    { stbir__simdf d;                          \
 4727+    stbir__simdf_load1z( c, hc + (ofs) );      \
 4728+    stbir__simdf_load1( d, decode + (ofs) );   \
 4729+    stbir__simdf_madd( tot, tot, d, c ); }
 4730+
 4731+#define stbir__2_coeff_remnant( ofs )          \
 4732+    { stbir__simdf d;                          \
 4733+    stbir__simdf_load2z( c, hc+(ofs) );        \
 4734+    stbir__simdf_load2( d, decode+(ofs) );     \
 4735+    stbir__simdf_madd( tot, tot, d, c ); }
 4736+
 4737+#define stbir__3_coeff_setup()                 \
 4738+    stbir__simdf mask;                         \
 4739+    stbir__simdf_load( mask, STBIR_mask + 3 );
 4740+
 4741+#define stbir__3_coeff_remnant( ofs )                  \
 4742+    stbir__simdf_load( c, hc+(ofs) );                  \
 4743+    stbir__simdf_and( c, c, mask );                    \
 4744+    stbir__simdf_madd_mem( tot, tot, c, decode+(ofs) );
 4745+
 4746+#define stbir__store_output()                     \
 4747+    stbir__simdf_0123to2301( c, tot );            \
 4748+    stbir__simdf_add( tot, tot, c );              \
 4749+    stbir__simdf_0123to1230( c, tot );            \
 4750+    stbir__simdf_add1( tot, tot, c );             \
 4751+    stbir__simdf_store1( output, tot );           \
 4752+    horizontal_coefficients += coefficient_width; \
 4753+    ++horizontal_contributors;                    \
 4754+    output += 1;
 4755+
 4756+#else
 4757+
 4758+#define stbir__1_coeff_only()  \
 4759+    float tot;                 \
 4760+    tot = decode[0]*hc[0];
 4761+
 4762+#define stbir__2_coeff_only()  \
 4763+    float tot;                 \
 4764+    tot = decode[0] * hc[0];   \
 4765+    tot += decode[1] * hc[1];
 4766+
 4767+#define stbir__3_coeff_only()  \
 4768+    float tot;                 \
 4769+    tot = decode[0] * hc[0];   \
 4770+    tot += decode[1] * hc[1];  \
 4771+    tot += decode[2] * hc[2];
 4772+
 4773+#define stbir__store_output_tiny()                \
 4774+    output[0] = tot;                              \
 4775+    horizontal_coefficients += coefficient_width; \
 4776+    ++horizontal_contributors;                    \
 4777+    output += 1;
 4778+
 4779+#define stbir__4_coeff_start()  \
 4780+    float tot0,tot1,tot2,tot3;  \
 4781+    tot0 = decode[0] * hc[0];   \
 4782+    tot1 = decode[1] * hc[1];   \
 4783+    tot2 = decode[2] * hc[2];   \
 4784+    tot3 = decode[3] * hc[3];
 4785+
 4786+#define stbir__4_coeff_continue_from_4( ofs )  \
 4787+    tot0 += decode[0+(ofs)] * hc[0+(ofs)];     \
 4788+    tot1 += decode[1+(ofs)] * hc[1+(ofs)];     \
 4789+    tot2 += decode[2+(ofs)] * hc[2+(ofs)];     \
 4790+    tot3 += decode[3+(ofs)] * hc[3+(ofs)];
 4791+
 4792+#define stbir__1_coeff_remnant( ofs )        \
 4793+    tot0 += decode[0+(ofs)] * hc[0+(ofs)];
 4794+
 4795+#define stbir__2_coeff_remnant( ofs )        \
 4796+    tot0 += decode[0+(ofs)] * hc[0+(ofs)];   \
 4797+    tot1 += decode[1+(ofs)] * hc[1+(ofs)];   \
 4798+
 4799+#define stbir__3_coeff_remnant( ofs )        \
 4800+    tot0 += decode[0+(ofs)] * hc[0+(ofs)];   \
 4801+    tot1 += decode[1+(ofs)] * hc[1+(ofs)];   \
 4802+    tot2 += decode[2+(ofs)] * hc[2+(ofs)];
 4803+
 4804+#define stbir__store_output()                     \
 4805+    output[0] = (tot0+tot2)+(tot1+tot3);          \
 4806+    horizontal_coefficients += coefficient_width; \
 4807+    ++horizontal_contributors;                    \
 4808+    output += 1;
 4809+
 4810+#endif
 4811+
 4812+#define STBIR__horizontal_channels 1
 4813+#define STB_IMAGE_RESIZE_DO_HORIZONTALS
 4814+#include STBIR__HEADER_FILENAME
 4815+
 4816+
 4817+//=================
 4818+// Do 2 channel horizontal routines
 4819+
 4820+#ifdef STBIR_SIMD
 4821+
 4822+#define stbir__1_coeff_only()         \
 4823+    stbir__simdf tot,c,d;             \
 4824+    STBIR_SIMD_NO_UNROLL(decode);     \
 4825+    stbir__simdf_load1z( c, hc );     \
 4826+    stbir__simdf_0123to0011( c, c );  \
 4827+    stbir__simdf_load2( d, decode );  \
 4828+    stbir__simdf_mult( tot, d, c );
 4829+
 4830+#define stbir__2_coeff_only()         \
 4831+    stbir__simdf tot,c;               \
 4832+    STBIR_SIMD_NO_UNROLL(decode);     \
 4833+    stbir__simdf_load2( c, hc );      \
 4834+    stbir__simdf_0123to0011( c, c );  \
 4835+    stbir__simdf_mult_mem( tot, c, decode );
 4836+
 4837+#define stbir__3_coeff_only()                \
 4838+    stbir__simdf tot,c,cs,d;                 \
 4839+    STBIR_SIMD_NO_UNROLL(decode);            \
 4840+    stbir__simdf_load( cs, hc );             \
 4841+    stbir__simdf_0123to0011( c, cs );        \
 4842+    stbir__simdf_mult_mem( tot, c, decode ); \
 4843+    stbir__simdf_0123to2222( c, cs );        \
 4844+    stbir__simdf_load2z( d, decode+4 );      \
 4845+    stbir__simdf_madd( tot, tot, d, c );
 4846+
 4847+#define stbir__store_output_tiny()                \
 4848+    stbir__simdf_0123to2301( c, tot );            \
 4849+    stbir__simdf_add( tot, tot, c );              \
 4850+    stbir__simdf_store2( output, tot );           \
 4851+    horizontal_coefficients += coefficient_width; \
 4852+    ++horizontal_contributors;                    \
 4853+    output += 2;
 4854+
 4855+#ifdef STBIR_SIMD8
 4856+
 4857+#define stbir__4_coeff_start()                    \
 4858+    stbir__simdf8 tot0,c,cs;                      \
 4859+    STBIR_SIMD_NO_UNROLL(decode);                 \
 4860+    stbir__simdf8_load4b( cs, hc );               \
 4861+    stbir__simdf8_0123to00112233( c, cs );        \
 4862+    stbir__simdf8_mult_mem( tot0, c, decode );
 4863+
 4864+#define stbir__4_coeff_continue_from_4( ofs )        \
 4865+    STBIR_SIMD_NO_UNROLL(decode);                    \
 4866+    stbir__simdf8_load4b( cs, hc + (ofs) );          \
 4867+    stbir__simdf8_0123to00112233( c, cs );           \
 4868+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*2 );
 4869+
 4870+#define stbir__1_coeff_remnant( ofs )                \
 4871+    { stbir__simdf t,d;                              \
 4872+    stbir__simdf_load1z( t, hc + (ofs) );            \
 4873+    stbir__simdf_load2( d, decode + (ofs) * 2 );     \
 4874+    stbir__simdf_0123to0011( t, t );                 \
 4875+    stbir__simdf_mult( t, t, d );                    \
 4876+    stbir__simdf8_add4( tot0, tot0, t ); }
 4877+ 
 4878+#define stbir__2_coeff_remnant( ofs )                \
 4879+    { stbir__simdf t;                                \
 4880+    stbir__simdf_load2( t, hc + (ofs) );             \
 4881+    stbir__simdf_0123to0011( t, t );                 \
 4882+    stbir__simdf_mult_mem( t, t, decode+(ofs)*2 );   \
 4883+    stbir__simdf8_add4( tot0, tot0, t ); }
 4884+
 4885+#define stbir__3_coeff_remnant( ofs )                \
 4886+    { stbir__simdf8 d;                               \
 4887+    stbir__simdf8_load4b( cs, hc + (ofs) );          \
 4888+    stbir__simdf8_0123to00112233( c, cs );           \
 4889+    stbir__simdf8_load6z( d, decode+(ofs)*2 );       \
 4890+    stbir__simdf8_madd( tot0, tot0, c, d ); }
 4891+
 4892+#define stbir__store_output()                     \
 4893+    { stbir__simdf t,d;                           \
 4894+    stbir__simdf8_add4halves( t, stbir__if_simdf8_cast_to_simdf4(tot0), tot0 );    \
 4895+    stbir__simdf_0123to2301( d, t );              \
 4896+    stbir__simdf_add( t, t, d );                  \
 4897+    stbir__simdf_store2( output, t );             \
 4898+    horizontal_coefficients += coefficient_width; \
 4899+    ++horizontal_contributors;                    \
 4900+    output += 2; }
 4901+
 4902+#else
 4903+
 4904+#define stbir__4_coeff_start()                   \
 4905+    stbir__simdf tot0,tot1,c,cs;                 \
 4906+    STBIR_SIMD_NO_UNROLL(decode);                \
 4907+    stbir__simdf_load( cs, hc );                 \
 4908+    stbir__simdf_0123to0011( c, cs );            \
 4909+    stbir__simdf_mult_mem( tot0, c, decode );    \
 4910+    stbir__simdf_0123to2233( c, cs );            \
 4911+    stbir__simdf_mult_mem( tot1, c, decode+4 );
 4912+
 4913+#define stbir__4_coeff_continue_from_4( ofs )                \
 4914+    STBIR_SIMD_NO_UNROLL(decode);                            \
 4915+    stbir__simdf_load( cs, hc + (ofs) );                     \
 4916+    stbir__simdf_0123to0011( c, cs );                        \
 4917+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*2 );  \
 4918+    stbir__simdf_0123to2233( c, cs );                        \
 4919+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*2+4 );
 4920+
 4921+#define stbir__1_coeff_remnant( ofs )            \
 4922+    { stbir__simdf d;                            \
 4923+    stbir__simdf_load1z( cs, hc + (ofs) );       \
 4924+    stbir__simdf_0123to0011( c, cs );            \
 4925+    stbir__simdf_load2( d, decode + (ofs) * 2 ); \
 4926+    stbir__simdf_madd( tot0, tot0, d, c ); }
 4927+
 4928+#define stbir__2_coeff_remnant( ofs )                      \
 4929+    stbir__simdf_load2( cs, hc + (ofs) );                  \
 4930+    stbir__simdf_0123to0011( c, cs );                      \
 4931+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*2 );
 4932+
 4933+#define stbir__3_coeff_remnant( ofs )                       \
 4934+    { stbir__simdf d;                                       \
 4935+    stbir__simdf_load( cs, hc + (ofs) );                    \
 4936+    stbir__simdf_0123to0011( c, cs );                       \
 4937+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*2 ); \
 4938+    stbir__simdf_0123to2222( c, cs );                       \
 4939+    stbir__simdf_load2z( d, decode + (ofs) * 2 + 4 );       \
 4940+    stbir__simdf_madd( tot1, tot1, d, c ); }
 4941+
 4942+#define stbir__store_output()                     \
 4943+    stbir__simdf_add( tot0, tot0, tot1 );         \
 4944+    stbir__simdf_0123to2301( c, tot0 );           \
 4945+    stbir__simdf_add( tot0, tot0, c );            \
 4946+    stbir__simdf_store2( output, tot0 );          \
 4947+    horizontal_coefficients += coefficient_width; \
 4948+    ++horizontal_contributors;                    \
 4949+    output += 2;
 4950+
 4951+#endif
 4952+
 4953+#else
 4954+
 4955+#define stbir__1_coeff_only()  \
 4956+    float tota,totb,c;         \
 4957+    c = hc[0];                 \
 4958+    tota = decode[0]*c;        \
 4959+    totb = decode[1]*c;
 4960+
 4961+#define stbir__2_coeff_only()  \
 4962+    float tota,totb,c;         \
 4963+    c = hc[0];                 \
 4964+    tota = decode[0]*c;        \
 4965+    totb = decode[1]*c;        \
 4966+    c = hc[1];                 \
 4967+    tota += decode[2]*c;       \
 4968+    totb += decode[3]*c;
 4969+
 4970+// this weird order of add matches the simd
 4971+#define stbir__3_coeff_only()  \
 4972+    float tota,totb,c;         \
 4973+    c = hc[0];                 \
 4974+    tota = decode[0]*c;        \
 4975+    totb = decode[1]*c;        \
 4976+    c = hc[2];                 \
 4977+    tota += decode[4]*c;       \
 4978+    totb += decode[5]*c;       \
 4979+    c = hc[1];                 \
 4980+    tota += decode[2]*c;       \
 4981+    totb += decode[3]*c;
 4982+
 4983+#define stbir__store_output_tiny()                \
 4984+    output[0] = tota;                             \
 4985+    output[1] = totb;                             \
 4986+    horizontal_coefficients += coefficient_width; \
 4987+    ++horizontal_contributors;                    \
 4988+    output += 2;
 4989+
 4990+#define stbir__4_coeff_start()      \
 4991+    float tota0,tota1,tota2,tota3,totb0,totb1,totb2,totb3,c;  \
 4992+    c = hc[0];                      \
 4993+    tota0 = decode[0]*c;            \
 4994+    totb0 = decode[1]*c;            \
 4995+    c = hc[1];                      \
 4996+    tota1 = decode[2]*c;            \
 4997+    totb1 = decode[3]*c;            \
 4998+    c = hc[2];                      \
 4999+    tota2 = decode[4]*c;            \
 5000+    totb2 = decode[5]*c;            \
 5001+    c = hc[3];                      \
 5002+    tota3 = decode[6]*c;            \
 5003+    totb3 = decode[7]*c;
 5004+
 5005+#define stbir__4_coeff_continue_from_4( ofs )  \
 5006+    c = hc[0+(ofs)];                           \
 5007+    tota0 += decode[0+(ofs)*2]*c;              \
 5008+    totb0 += decode[1+(ofs)*2]*c;              \
 5009+    c = hc[1+(ofs)];                           \
 5010+    tota1 += decode[2+(ofs)*2]*c;              \
 5011+    totb1 += decode[3+(ofs)*2]*c;              \
 5012+    c = hc[2+(ofs)];                           \
 5013+    tota2 += decode[4+(ofs)*2]*c;              \
 5014+    totb2 += decode[5+(ofs)*2]*c;              \
 5015+    c = hc[3+(ofs)];                           \
 5016+    tota3 += decode[6+(ofs)*2]*c;              \
 5017+    totb3 += decode[7+(ofs)*2]*c;
 5018+
 5019+#define stbir__1_coeff_remnant( ofs )  \
 5020+    c = hc[0+(ofs)];                   \
 5021+    tota0 += decode[0+(ofs)*2] * c;    \
 5022+    totb0 += decode[1+(ofs)*2] * c;
 5023+
 5024+#define stbir__2_coeff_remnant( ofs )  \
 5025+    c = hc[0+(ofs)];                   \
 5026+    tota0 += decode[0+(ofs)*2] * c;    \
 5027+    totb0 += decode[1+(ofs)*2] * c;    \
 5028+    c = hc[1+(ofs)];                   \
 5029+    tota1 += decode[2+(ofs)*2] * c;    \
 5030+    totb1 += decode[3+(ofs)*2] * c;
 5031+
 5032+#define stbir__3_coeff_remnant( ofs )  \
 5033+    c = hc[0+(ofs)];                   \
 5034+    tota0 += decode[0+(ofs)*2] * c;    \
 5035+    totb0 += decode[1+(ofs)*2] * c;    \
 5036+    c = hc[1+(ofs)];                   \
 5037+    tota1 += decode[2+(ofs)*2] * c;    \
 5038+    totb1 += decode[3+(ofs)*2] * c;    \
 5039+    c = hc[2+(ofs)];                   \
 5040+    tota2 += decode[4+(ofs)*2] * c;    \
 5041+    totb2 += decode[5+(ofs)*2] * c;
 5042+
 5043+#define stbir__store_output()                     \
 5044+    output[0] = (tota0+tota2)+(tota1+tota3);      \
 5045+    output[1] = (totb0+totb2)+(totb1+totb3);      \
 5046+    horizontal_coefficients += coefficient_width; \
 5047+    ++horizontal_contributors;                    \
 5048+    output += 2;
 5049+
 5050+#endif
 5051+
 5052+#define STBIR__horizontal_channels 2
 5053+#define STB_IMAGE_RESIZE_DO_HORIZONTALS
 5054+#include STBIR__HEADER_FILENAME
 5055+
 5056+
 5057+//=================
 5058+// Do 3 channel horizontal routines
 5059+
 5060+#ifdef STBIR_SIMD
 5061+
 5062+#define stbir__1_coeff_only()         \
 5063+    stbir__simdf tot,c,d;             \
 5064+    STBIR_SIMD_NO_UNROLL(decode);     \
 5065+    stbir__simdf_load1z( c, hc );     \
 5066+    stbir__simdf_0123to0001( c, c );  \
 5067+    stbir__simdf_load( d, decode );   \
 5068+    stbir__simdf_mult( tot, d, c );
 5069+
 5070+#define stbir__2_coeff_only()         \
 5071+    stbir__simdf tot,c,cs,d;          \
 5072+    STBIR_SIMD_NO_UNROLL(decode);     \
 5073+    stbir__simdf_load2( cs, hc );     \
 5074+    stbir__simdf_0123to0000( c, cs ); \
 5075+    stbir__simdf_load( d, decode );   \
 5076+    stbir__simdf_mult( tot, d, c );   \
 5077+    stbir__simdf_0123to1111( c, cs ); \
 5078+    stbir__simdf_load( d, decode+3 ); \
 5079+    stbir__simdf_madd( tot, tot, d, c );
 5080+
 5081+#define stbir__3_coeff_only()            \
 5082+    stbir__simdf tot,c,d,cs;             \
 5083+    STBIR_SIMD_NO_UNROLL(decode);        \
 5084+    stbir__simdf_load( cs, hc );         \
 5085+    stbir__simdf_0123to0000( c, cs );    \
 5086+    stbir__simdf_load( d, decode );      \
 5087+    stbir__simdf_mult( tot, d, c );      \
 5088+    stbir__simdf_0123to1111( c, cs );    \
 5089+    stbir__simdf_load( d, decode+3 );    \
 5090+    stbir__simdf_madd( tot, tot, d, c ); \
 5091+    stbir__simdf_0123to2222( c, cs );    \
 5092+    stbir__simdf_load( d, decode+6 );    \
 5093+    stbir__simdf_madd( tot, tot, d, c );
 5094+
 5095+#define stbir__store_output_tiny()                \
 5096+    stbir__simdf_store2( output, tot );           \
 5097+    stbir__simdf_0123to2301( tot, tot );          \
 5098+    stbir__simdf_store1( output+2, tot );         \
 5099+    horizontal_coefficients += coefficient_width; \
 5100+    ++horizontal_contributors;                    \
 5101+    output += 3;
 5102+
 5103+#ifdef STBIR_SIMD8
 5104+
 5105+// we're loading from the XXXYYY decode by -1 to get the XXXYYY into different halves of the AVX reg fyi
 5106+#define stbir__4_coeff_start()                     \
 5107+    stbir__simdf8 tot0,tot1,c,cs; stbir__simdf t;  \
 5108+    STBIR_SIMD_NO_UNROLL(decode);                  \
 5109+    stbir__simdf8_load4b( cs, hc );                \
 5110+    stbir__simdf8_0123to00001111( c, cs );         \
 5111+    stbir__simdf8_mult_mem( tot0, c, decode - 1 ); \
 5112+    stbir__simdf8_0123to22223333( c, cs );         \
 5113+    stbir__simdf8_mult_mem( tot1, c, decode+6 - 1 );
 5114+
 5115+#define stbir__4_coeff_continue_from_4( ofs )      \
 5116+    STBIR_SIMD_NO_UNROLL(decode);                  \
 5117+    stbir__simdf8_load4b( cs, hc + (ofs) );        \
 5118+    stbir__simdf8_0123to00001111( c, cs );         \
 5119+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*3 - 1 ); \
 5120+    stbir__simdf8_0123to22223333( c, cs );         \
 5121+    stbir__simdf8_madd_mem( tot1, tot1, c, decode+(ofs)*3 + 6 - 1 );
 5122+
 5123+#define stbir__1_coeff_remnant( ofs )                          \
 5124+    STBIR_SIMD_NO_UNROLL(decode);                              \
 5125+    stbir__simdf_load1rep4( t, hc + (ofs) );                   \
 5126+    stbir__simdf8_madd_mem4( tot0, tot0, t, decode+(ofs)*3 - 1 );
 5127+
 5128+#define stbir__2_coeff_remnant( ofs )                          \
 5129+    STBIR_SIMD_NO_UNROLL(decode);                              \
 5130+    stbir__simdf8_load4b( cs, hc + (ofs) - 2 );                \
 5131+    stbir__simdf8_0123to22223333( c, cs );                     \
 5132+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*3 - 1 );
 5133+
 5134+ #define stbir__3_coeff_remnant( ofs )                           \
 5135+    STBIR_SIMD_NO_UNROLL(decode);                                \
 5136+    stbir__simdf8_load4b( cs, hc + (ofs) );                      \
 5137+    stbir__simdf8_0123to00001111( c, cs );                       \
 5138+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*3 - 1 ); \
 5139+    stbir__simdf8_0123to2222( t, cs );                           \
 5140+    stbir__simdf8_madd_mem4( tot1, tot1, t, decode+(ofs)*3 + 6 - 1 );
 5141+
 5142+#define stbir__store_output()                       \
 5143+    stbir__simdf8_add( tot0, tot0, tot1 );          \
 5144+    stbir__simdf_0123to1230( t, stbir__if_simdf8_cast_to_simdf4( tot0 ) ); \
 5145+    stbir__simdf8_add4halves( t, t, tot0 );         \
 5146+    horizontal_coefficients += coefficient_width;   \
 5147+    ++horizontal_contributors;                      \
 5148+    output += 3;                                    \
 5149+    if ( output < output_end )                      \
 5150+    {                                               \
 5151+      stbir__simdf_store( output-3, t );            \
 5152+      continue;                                     \
 5153+    }                                               \
 5154+    { stbir__simdf tt; stbir__simdf_0123to2301( tt, t ); \
 5155+    stbir__simdf_store2( output-3, t );             \
 5156+    stbir__simdf_store1( output+2-3, tt ); }        \
 5157+    break;
 5158+
 5159+
 5160+#else
 5161+
 5162+#define stbir__4_coeff_start()                  \
 5163+    stbir__simdf tot0,tot1,tot2,c,cs;           \
 5164+    STBIR_SIMD_NO_UNROLL(decode);               \
 5165+    stbir__simdf_load( cs, hc );                \
 5166+    stbir__simdf_0123to0001( c, cs );           \
 5167+    stbir__simdf_mult_mem( tot0, c, decode );   \
 5168+    stbir__simdf_0123to1122( c, cs );           \
 5169+    stbir__simdf_mult_mem( tot1, c, decode+4 ); \
 5170+    stbir__simdf_0123to2333( c, cs );           \
 5171+    stbir__simdf_mult_mem( tot2, c, decode+8 );
 5172+
 5173+#define stbir__4_coeff_continue_from_4( ofs )                 \
 5174+    STBIR_SIMD_NO_UNROLL(decode);                             \
 5175+    stbir__simdf_load( cs, hc + (ofs) );                      \
 5176+    stbir__simdf_0123to0001( c, cs );                         \
 5177+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*3 );   \
 5178+    stbir__simdf_0123to1122( c, cs );                         \
 5179+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*3+4 ); \
 5180+    stbir__simdf_0123to2333( c, cs );                         \
 5181+    stbir__simdf_madd_mem( tot2, tot2, c, decode+(ofs)*3+8 );
 5182+
 5183+#define stbir__1_coeff_remnant( ofs )         \
 5184+    STBIR_SIMD_NO_UNROLL(decode);             \
 5185+    stbir__simdf_load1z( c, hc + (ofs) );     \
 5186+    stbir__simdf_0123to0001( c, c );          \
 5187+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*3 );
 5188+
 5189+#define stbir__2_coeff_remnant( ofs )                       \
 5190+    { stbir__simdf d;                                       \
 5191+    STBIR_SIMD_NO_UNROLL(decode);                           \
 5192+    stbir__simdf_load2z( cs, hc + (ofs) );                  \
 5193+    stbir__simdf_0123to0001( c, cs );                       \
 5194+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*3 ); \
 5195+    stbir__simdf_0123to1122( c, cs );                       \
 5196+    stbir__simdf_load2z( d, decode+(ofs)*3+4 );             \
 5197+    stbir__simdf_madd( tot1, tot1, c, d ); }
 5198+
 5199+#define stbir__3_coeff_remnant( ofs )                         \
 5200+    { stbir__simdf d;                                         \
 5201+    STBIR_SIMD_NO_UNROLL(decode);                             \
 5202+    stbir__simdf_load( cs, hc + (ofs) );                      \
 5203+    stbir__simdf_0123to0001( c, cs );                         \
 5204+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*3 );   \
 5205+    stbir__simdf_0123to1122( c, cs );                         \
 5206+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*3+4 ); \
 5207+    stbir__simdf_0123to2222( c, cs );                         \
 5208+    stbir__simdf_load1z( d, decode+(ofs)*3+8 );               \
 5209+    stbir__simdf_madd( tot2, tot2, c, d );  }
 5210+
 5211+#define stbir__store_output()                       \
 5212+    stbir__simdf_0123ABCDto3ABx( c, tot0, tot1 );   \
 5213+    stbir__simdf_0123ABCDto23Ax( cs, tot1, tot2 );  \
 5214+    stbir__simdf_0123to1230( tot2, tot2 );          \
 5215+    stbir__simdf_add( tot0, tot0, cs );             \
 5216+    stbir__simdf_add( c, c, tot2 );                 \
 5217+    stbir__simdf_add( tot0, tot0, c );              \
 5218+    horizontal_coefficients += coefficient_width;   \
 5219+    ++horizontal_contributors;                      \
 5220+    output += 3;                                    \
 5221+    if ( output < output_end )                      \
 5222+    {                                               \
 5223+      stbir__simdf_store( output-3, tot0 );         \
 5224+      continue;                                     \
 5225+    }                                               \
 5226+    stbir__simdf_0123to2301( tot1, tot0 );          \
 5227+    stbir__simdf_store2( output-3, tot0 );          \
 5228+    stbir__simdf_store1( output+2-3, tot1 );        \
 5229+    break;
 5230+
 5231+#endif
 5232+
 5233+#else
 5234+
 5235+#define stbir__1_coeff_only()  \
 5236+    float tot0, tot1, tot2, c; \
 5237+    c = hc[0];                 \
 5238+    tot0 = decode[0]*c;        \
 5239+    tot1 = decode[1]*c;        \
 5240+    tot2 = decode[2]*c;
 5241+
 5242+#define stbir__2_coeff_only()  \
 5243+    float tot0, tot1, tot2, c; \
 5244+    c = hc[0];                 \
 5245+    tot0 = decode[0]*c;        \
 5246+    tot1 = decode[1]*c;        \
 5247+    tot2 = decode[2]*c;        \
 5248+    c = hc[1];                 \
 5249+    tot0 += decode[3]*c;       \
 5250+    tot1 += decode[4]*c;       \
 5251+    tot2 += decode[5]*c;
 5252+
 5253+#define stbir__3_coeff_only()  \
 5254+    float tot0, tot1, tot2, c; \
 5255+    c = hc[0];                 \
 5256+    tot0 = decode[0]*c;        \
 5257+    tot1 = decode[1]*c;        \
 5258+    tot2 = decode[2]*c;        \
 5259+    c = hc[1];                 \
 5260+    tot0 += decode[3]*c;       \
 5261+    tot1 += decode[4]*c;       \
 5262+    tot2 += decode[5]*c;       \
 5263+    c = hc[2];                 \
 5264+    tot0 += decode[6]*c;       \
 5265+    tot1 += decode[7]*c;       \
 5266+    tot2 += decode[8]*c;
 5267+
 5268+#define stbir__store_output_tiny()                \
 5269+    output[0] = tot0;                             \
 5270+    output[1] = tot1;                             \
 5271+    output[2] = tot2;                             \
 5272+    horizontal_coefficients += coefficient_width; \
 5273+    ++horizontal_contributors;                    \
 5274+    output += 3;
 5275+
 5276+#define stbir__4_coeff_start()      \
 5277+    float tota0,tota1,tota2,totb0,totb1,totb2,totc0,totc1,totc2,totd0,totd1,totd2,c;  \
 5278+    c = hc[0];                      \
 5279+    tota0 = decode[0]*c;            \
 5280+    tota1 = decode[1]*c;            \
 5281+    tota2 = decode[2]*c;            \
 5282+    c = hc[1];                      \
 5283+    totb0 = decode[3]*c;            \
 5284+    totb1 = decode[4]*c;            \
 5285+    totb2 = decode[5]*c;            \
 5286+    c = hc[2];                      \
 5287+    totc0 = decode[6]*c;            \
 5288+    totc1 = decode[7]*c;            \
 5289+    totc2 = decode[8]*c;            \
 5290+    c = hc[3];                      \
 5291+    totd0 = decode[9]*c;            \
 5292+    totd1 = decode[10]*c;           \
 5293+    totd2 = decode[11]*c;
 5294+
 5295+#define stbir__4_coeff_continue_from_4( ofs )  \
 5296+    c = hc[0+(ofs)];                           \
 5297+    tota0 += decode[0+(ofs)*3]*c;              \
 5298+    tota1 += decode[1+(ofs)*3]*c;              \
 5299+    tota2 += decode[2+(ofs)*3]*c;              \
 5300+    c = hc[1+(ofs)];                           \
 5301+    totb0 += decode[3+(ofs)*3]*c;              \
 5302+    totb1 += decode[4+(ofs)*3]*c;              \
 5303+    totb2 += decode[5+(ofs)*3]*c;              \
 5304+    c = hc[2+(ofs)];                           \
 5305+    totc0 += decode[6+(ofs)*3]*c;              \
 5306+    totc1 += decode[7+(ofs)*3]*c;              \
 5307+    totc2 += decode[8+(ofs)*3]*c;              \
 5308+    c = hc[3+(ofs)];                           \
 5309+    totd0 += decode[9+(ofs)*3]*c;              \
 5310+    totd1 += decode[10+(ofs)*3]*c;             \
 5311+    totd2 += decode[11+(ofs)*3]*c;
 5312+
 5313+#define stbir__1_coeff_remnant( ofs )  \
 5314+    c = hc[0+(ofs)];                   \
 5315+    tota0 += decode[0+(ofs)*3]*c;      \
 5316+    tota1 += decode[1+(ofs)*3]*c;      \
 5317+    tota2 += decode[2+(ofs)*3]*c;
 5318+
 5319+#define stbir__2_coeff_remnant( ofs )  \
 5320+    c = hc[0+(ofs)];                   \
 5321+    tota0 += decode[0+(ofs)*3]*c;      \
 5322+    tota1 += decode[1+(ofs)*3]*c;      \
 5323+    tota2 += decode[2+(ofs)*3]*c;      \
 5324+    c = hc[1+(ofs)];                   \
 5325+    totb0 += decode[3+(ofs)*3]*c;      \
 5326+    totb1 += decode[4+(ofs)*3]*c;      \
 5327+    totb2 += decode[5+(ofs)*3]*c;      \
 5328+
 5329+#define stbir__3_coeff_remnant( ofs )  \
 5330+    c = hc[0+(ofs)];                   \
 5331+    tota0 += decode[0+(ofs)*3]*c;      \
 5332+    tota1 += decode[1+(ofs)*3]*c;      \
 5333+    tota2 += decode[2+(ofs)*3]*c;      \
 5334+    c = hc[1+(ofs)];                   \
 5335+    totb0 += decode[3+(ofs)*3]*c;      \
 5336+    totb1 += decode[4+(ofs)*3]*c;      \
 5337+    totb2 += decode[5+(ofs)*3]*c;      \
 5338+    c = hc[2+(ofs)];                   \
 5339+    totc0 += decode[6+(ofs)*3]*c;      \
 5340+    totc1 += decode[7+(ofs)*3]*c;      \
 5341+    totc2 += decode[8+(ofs)*3]*c;
 5342+
 5343+#define stbir__store_output()                     \
 5344+    output[0] = (tota0+totc0)+(totb0+totd0);      \
 5345+    output[1] = (tota1+totc1)+(totb1+totd1);      \
 5346+    output[2] = (tota2+totc2)+(totb2+totd2);      \
 5347+    horizontal_coefficients += coefficient_width; \
 5348+    ++horizontal_contributors;                    \
 5349+    output += 3;
 5350+
 5351+#endif
 5352+
 5353+#define STBIR__horizontal_channels 3
 5354+#define STB_IMAGE_RESIZE_DO_HORIZONTALS
 5355+#include STBIR__HEADER_FILENAME
 5356+
 5357+//=================
 5358+// Do 4 channel horizontal routines
 5359+
 5360+#ifdef STBIR_SIMD
 5361+
 5362+#define stbir__1_coeff_only()             \
 5363+    stbir__simdf tot,c;                   \
 5364+    STBIR_SIMD_NO_UNROLL(decode);         \
 5365+    stbir__simdf_load1( c, hc );          \
 5366+    stbir__simdf_0123to0000( c, c );      \
 5367+    stbir__simdf_mult_mem( tot, c, decode );
 5368+
 5369+#define stbir__2_coeff_only()                       \
 5370+    stbir__simdf tot,c,cs;                          \
 5371+    STBIR_SIMD_NO_UNROLL(decode);                   \
 5372+    stbir__simdf_load2( cs, hc );                   \
 5373+    stbir__simdf_0123to0000( c, cs );               \
 5374+    stbir__simdf_mult_mem( tot, c, decode );        \
 5375+    stbir__simdf_0123to1111( c, cs );               \
 5376+    stbir__simdf_madd_mem( tot, tot, c, decode+4 );
 5377+
 5378+#define stbir__3_coeff_only()                       \
 5379+    stbir__simdf tot,c,cs;                          \
 5380+    STBIR_SIMD_NO_UNROLL(decode);                   \
 5381+    stbir__simdf_load( cs, hc );                    \
 5382+    stbir__simdf_0123to0000( c, cs );               \
 5383+    stbir__simdf_mult_mem( tot, c, decode );        \
 5384+    stbir__simdf_0123to1111( c, cs );               \
 5385+    stbir__simdf_madd_mem( tot, tot, c, decode+4 ); \
 5386+    stbir__simdf_0123to2222( c, cs );               \
 5387+    stbir__simdf_madd_mem( tot, tot, c, decode+8 );
 5388+
 5389+#define stbir__store_output_tiny()                \
 5390+    stbir__simdf_store( output, tot );            \
 5391+    horizontal_coefficients += coefficient_width; \
 5392+    ++horizontal_contributors;                    \
 5393+    output += 4;
 5394+
 5395+#ifdef STBIR_SIMD8
 5396+
 5397+#define stbir__4_coeff_start()                     \
 5398+    stbir__simdf8 tot0,c,cs; stbir__simdf t;  \
 5399+    STBIR_SIMD_NO_UNROLL(decode);                  \
 5400+    stbir__simdf8_load4b( cs, hc );                \
 5401+    stbir__simdf8_0123to00001111( c, cs );         \
 5402+    stbir__simdf8_mult_mem( tot0, c, decode );     \
 5403+    stbir__simdf8_0123to22223333( c, cs );         \
 5404+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+8 );
 5405+
 5406+#define stbir__4_coeff_continue_from_4( ofs )                  \
 5407+    STBIR_SIMD_NO_UNROLL(decode);                              \
 5408+    stbir__simdf8_load4b( cs, hc + (ofs) );                    \
 5409+    stbir__simdf8_0123to00001111( c, cs );                     \
 5410+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*4 );   \
 5411+    stbir__simdf8_0123to22223333( c, cs );                     \
 5412+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*4+8 );
 5413+
 5414+#define stbir__1_coeff_remnant( ofs )                          \
 5415+    STBIR_SIMD_NO_UNROLL(decode);                              \
 5416+    stbir__simdf_load1rep4( t, hc + (ofs) );                   \
 5417+    stbir__simdf8_madd_mem4( tot0, tot0, t, decode+(ofs)*4 );
 5418+
 5419+#define stbir__2_coeff_remnant( ofs )                          \
 5420+    STBIR_SIMD_NO_UNROLL(decode);                              \
 5421+    stbir__simdf8_load4b( cs, hc + (ofs) - 2 );                \
 5422+    stbir__simdf8_0123to22223333( c, cs );                     \
 5423+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*4 );
 5424+
 5425+ #define stbir__3_coeff_remnant( ofs )                         \
 5426+    STBIR_SIMD_NO_UNROLL(decode);                              \
 5427+    stbir__simdf8_load4b( cs, hc + (ofs) );                    \
 5428+    stbir__simdf8_0123to00001111( c, cs );                     \
 5429+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*4 );   \
 5430+    stbir__simdf8_0123to2222( t, cs );                         \
 5431+    stbir__simdf8_madd_mem4( tot0, tot0, t, decode+(ofs)*4+8 );
 5432+
 5433+#define stbir__store_output()                      \
 5434+    stbir__simdf8_add4halves( t, stbir__if_simdf8_cast_to_simdf4(tot0), tot0 );     \
 5435+    stbir__simdf_store( output, t );               \
 5436+    horizontal_coefficients += coefficient_width;  \
 5437+    ++horizontal_contributors;                     \
 5438+    output += 4;
 5439+
 5440+#else
 5441+
 5442+#define stbir__4_coeff_start()                        \
 5443+    stbir__simdf tot0,tot1,c,cs;                      \
 5444+    STBIR_SIMD_NO_UNROLL(decode);                     \
 5445+    stbir__simdf_load( cs, hc );                      \
 5446+    stbir__simdf_0123to0000( c, cs );                 \
 5447+    stbir__simdf_mult_mem( tot0, c, decode );         \
 5448+    stbir__simdf_0123to1111( c, cs );                 \
 5449+    stbir__simdf_mult_mem( tot1, c, decode+4 );       \
 5450+    stbir__simdf_0123to2222( c, cs );                 \
 5451+    stbir__simdf_madd_mem( tot0, tot0, c, decode+8 ); \
 5452+    stbir__simdf_0123to3333( c, cs );                 \
 5453+    stbir__simdf_madd_mem( tot1, tot1, c, decode+12 );
 5454+
 5455+#define stbir__4_coeff_continue_from_4( ofs )                  \
 5456+    STBIR_SIMD_NO_UNROLL(decode);                              \
 5457+    stbir__simdf_load( cs, hc + (ofs) );                       \
 5458+    stbir__simdf_0123to0000( c, cs );                          \
 5459+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*4 );    \
 5460+    stbir__simdf_0123to1111( c, cs );                          \
 5461+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*4+4 );  \
 5462+    stbir__simdf_0123to2222( c, cs );                          \
 5463+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*4+8 );  \
 5464+    stbir__simdf_0123to3333( c, cs );                          \
 5465+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*4+12 );
 5466+
 5467+#define stbir__1_coeff_remnant( ofs )                       \
 5468+    STBIR_SIMD_NO_UNROLL(decode);                           \
 5469+    stbir__simdf_load1( c, hc + (ofs) );                    \
 5470+    stbir__simdf_0123to0000( c, c );                        \
 5471+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*4 );
 5472+
 5473+#define stbir__2_coeff_remnant( ofs )                         \
 5474+    STBIR_SIMD_NO_UNROLL(decode);                             \
 5475+    stbir__simdf_load2( cs, hc + (ofs) );                     \
 5476+    stbir__simdf_0123to0000( c, cs );                         \
 5477+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*4 );   \
 5478+    stbir__simdf_0123to1111( c, cs );                         \
 5479+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*4+4 );
 5480+
 5481+#define stbir__3_coeff_remnant( ofs )                          \
 5482+    STBIR_SIMD_NO_UNROLL(decode);                              \
 5483+    stbir__simdf_load( cs, hc + (ofs) );                       \
 5484+    stbir__simdf_0123to0000( c, cs );                          \
 5485+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*4 );    \
 5486+    stbir__simdf_0123to1111( c, cs );                          \
 5487+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*4+4 );  \
 5488+    stbir__simdf_0123to2222( c, cs );                          \
 5489+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*4+8 );
 5490+
 5491+#define stbir__store_output()                     \
 5492+    stbir__simdf_add( tot0, tot0, tot1 );         \
 5493+    stbir__simdf_store( output, tot0 );           \
 5494+    horizontal_coefficients += coefficient_width; \
 5495+    ++horizontal_contributors;                    \
 5496+    output += 4;
 5497+
 5498+#endif
 5499+
 5500+#else
 5501+
 5502+#define stbir__1_coeff_only()         \
 5503+    float p0,p1,p2,p3,c;              \
 5504+    STBIR_SIMD_NO_UNROLL(decode);     \
 5505+    c = hc[0];                        \
 5506+    p0 = decode[0] * c;               \
 5507+    p1 = decode[1] * c;               \
 5508+    p2 = decode[2] * c;               \
 5509+    p3 = decode[3] * c;
 5510+
 5511+#define stbir__2_coeff_only()         \
 5512+    float p0,p1,p2,p3,c;              \
 5513+    STBIR_SIMD_NO_UNROLL(decode);     \
 5514+    c = hc[0];                        \
 5515+    p0 = decode[0] * c;               \
 5516+    p1 = decode[1] * c;               \
 5517+    p2 = decode[2] * c;               \
 5518+    p3 = decode[3] * c;               \
 5519+    c = hc[1];                        \
 5520+    p0 += decode[4] * c;              \
 5521+    p1 += decode[5] * c;              \
 5522+    p2 += decode[6] * c;              \
 5523+    p3 += decode[7] * c;
 5524+
 5525+#define stbir__3_coeff_only()         \
 5526+    float p0,p1,p2,p3,c;              \
 5527+    STBIR_SIMD_NO_UNROLL(decode);     \
 5528+    c = hc[0];                        \
 5529+    p0 = decode[0] * c;               \
 5530+    p1 = decode[1] * c;               \
 5531+    p2 = decode[2] * c;               \
 5532+    p3 = decode[3] * c;               \
 5533+    c = hc[1];                        \
 5534+    p0 += decode[4] * c;              \
 5535+    p1 += decode[5] * c;              \
 5536+    p2 += decode[6] * c;              \
 5537+    p3 += decode[7] * c;              \
 5538+    c = hc[2];                        \
 5539+    p0 += decode[8] * c;              \
 5540+    p1 += decode[9] * c;              \
 5541+    p2 += decode[10] * c;             \
 5542+    p3 += decode[11] * c;
 5543+
 5544+#define stbir__store_output_tiny()                \
 5545+    output[0] = p0;                               \
 5546+    output[1] = p1;                               \
 5547+    output[2] = p2;                               \
 5548+    output[3] = p3;                               \
 5549+    horizontal_coefficients += coefficient_width; \
 5550+    ++horizontal_contributors;                    \
 5551+    output += 4;
 5552+
 5553+#define stbir__4_coeff_start()        \
 5554+    float x0,x1,x2,x3,y0,y1,y2,y3,c;  \
 5555+    STBIR_SIMD_NO_UNROLL(decode);     \
 5556+    c = hc[0];                        \
 5557+    x0 = decode[0] * c;               \
 5558+    x1 = decode[1] * c;               \
 5559+    x2 = decode[2] * c;               \
 5560+    x3 = decode[3] * c;               \
 5561+    c = hc[1];                        \
 5562+    y0 = decode[4] * c;               \
 5563+    y1 = decode[5] * c;               \
 5564+    y2 = decode[6] * c;               \
 5565+    y3 = decode[7] * c;               \
 5566+    c = hc[2];                        \
 5567+    x0 += decode[8] * c;              \
 5568+    x1 += decode[9] * c;              \
 5569+    x2 += decode[10] * c;             \
 5570+    x3 += decode[11] * c;             \
 5571+    c = hc[3];                        \
 5572+    y0 += decode[12] * c;             \
 5573+    y1 += decode[13] * c;             \
 5574+    y2 += decode[14] * c;             \
 5575+    y3 += decode[15] * c;
 5576+
 5577+#define stbir__4_coeff_continue_from_4( ofs ) \
 5578+    STBIR_SIMD_NO_UNROLL(decode);     \
 5579+    c = hc[0+(ofs)];                  \
 5580+    x0 += decode[0+(ofs)*4] * c;      \
 5581+    x1 += decode[1+(ofs)*4] * c;      \
 5582+    x2 += decode[2+(ofs)*4] * c;      \
 5583+    x3 += decode[3+(ofs)*4] * c;      \
 5584+    c = hc[1+(ofs)];                  \
 5585+    y0 += decode[4+(ofs)*4] * c;      \
 5586+    y1 += decode[5+(ofs)*4] * c;      \
 5587+    y2 += decode[6+(ofs)*4] * c;      \
 5588+    y3 += decode[7+(ofs)*4] * c;      \
 5589+    c = hc[2+(ofs)];                  \
 5590+    x0 += decode[8+(ofs)*4] * c;      \
 5591+    x1 += decode[9+(ofs)*4] * c;      \
 5592+    x2 += decode[10+(ofs)*4] * c;     \
 5593+    x3 += decode[11+(ofs)*4] * c;     \
 5594+    c = hc[3+(ofs)];                  \
 5595+    y0 += decode[12+(ofs)*4] * c;     \
 5596+    y1 += decode[13+(ofs)*4] * c;     \
 5597+    y2 += decode[14+(ofs)*4] * c;     \
 5598+    y3 += decode[15+(ofs)*4] * c;
 5599+
 5600+#define stbir__1_coeff_remnant( ofs ) \
 5601+    STBIR_SIMD_NO_UNROLL(decode);     \
 5602+    c = hc[0+(ofs)];                  \
 5603+    x0 += decode[0+(ofs)*4] * c;      \
 5604+    x1 += decode[1+(ofs)*4] * c;      \
 5605+    x2 += decode[2+(ofs)*4] * c;      \
 5606+    x3 += decode[3+(ofs)*4] * c;
 5607+
 5608+#define stbir__2_coeff_remnant( ofs ) \
 5609+    STBIR_SIMD_NO_UNROLL(decode);     \
 5610+    c = hc[0+(ofs)];                  \
 5611+    x0 += decode[0+(ofs)*4] * c;      \
 5612+    x1 += decode[1+(ofs)*4] * c;      \
 5613+    x2 += decode[2+(ofs)*4] * c;      \
 5614+    x3 += decode[3+(ofs)*4] * c;      \
 5615+    c = hc[1+(ofs)];                  \
 5616+    y0 += decode[4+(ofs)*4] * c;      \
 5617+    y1 += decode[5+(ofs)*4] * c;      \
 5618+    y2 += decode[6+(ofs)*4] * c;      \
 5619+    y3 += decode[7+(ofs)*4] * c;
 5620+
 5621+#define stbir__3_coeff_remnant( ofs ) \
 5622+    STBIR_SIMD_NO_UNROLL(decode);     \
 5623+    c = hc[0+(ofs)];                  \
 5624+    x0 += decode[0+(ofs)*4] * c;      \
 5625+    x1 += decode[1+(ofs)*4] * c;      \
 5626+    x2 += decode[2+(ofs)*4] * c;      \
 5627+    x3 += decode[3+(ofs)*4] * c;      \
 5628+    c = hc[1+(ofs)];                  \
 5629+    y0 += decode[4+(ofs)*4] * c;      \
 5630+    y1 += decode[5+(ofs)*4] * c;      \
 5631+    y2 += decode[6+(ofs)*4] * c;      \
 5632+    y3 += decode[7+(ofs)*4] * c;      \
 5633+    c = hc[2+(ofs)];                  \
 5634+    x0 += decode[8+(ofs)*4] * c;      \
 5635+    x1 += decode[9+(ofs)*4] * c;      \
 5636+    x2 += decode[10+(ofs)*4] * c;     \
 5637+    x3 += decode[11+(ofs)*4] * c;
 5638+
 5639+#define stbir__store_output()                     \
 5640+    output[0] = x0 + y0;                          \
 5641+    output[1] = x1 + y1;                          \
 5642+    output[2] = x2 + y2;                          \
 5643+    output[3] = x3 + y3;                          \
 5644+    horizontal_coefficients += coefficient_width; \
 5645+    ++horizontal_contributors;                    \
 5646+    output += 4;
 5647+
 5648+#endif
 5649+
 5650+#define STBIR__horizontal_channels 4
 5651+#define STB_IMAGE_RESIZE_DO_HORIZONTALS
 5652+#include STBIR__HEADER_FILENAME
 5653+
 5654+
 5655+
 5656+//=================
 5657+// Do 7 channel horizontal routines
 5658+
 5659+#ifdef STBIR_SIMD
 5660+
 5661+#define stbir__1_coeff_only()                   \
 5662+    stbir__simdf tot0,tot1,c;                   \
 5663+    STBIR_SIMD_NO_UNROLL(decode);               \
 5664+    stbir__simdf_load1( c, hc );                \
 5665+    stbir__simdf_0123to0000( c, c );            \
 5666+    stbir__simdf_mult_mem( tot0, c, decode );   \
 5667+    stbir__simdf_mult_mem( tot1, c, decode+3 );
 5668+
 5669+#define stbir__2_coeff_only()                         \
 5670+    stbir__simdf tot0,tot1,c,cs;                      \
 5671+    STBIR_SIMD_NO_UNROLL(decode);                     \
 5672+    stbir__simdf_load2( cs, hc );                     \
 5673+    stbir__simdf_0123to0000( c, cs );                 \
 5674+    stbir__simdf_mult_mem( tot0, c, decode );         \
 5675+    stbir__simdf_mult_mem( tot1, c, decode+3 );       \
 5676+    stbir__simdf_0123to1111( c, cs );                 \
 5677+    stbir__simdf_madd_mem( tot0, tot0, c, decode+7 ); \
 5678+    stbir__simdf_madd_mem( tot1, tot1, c,decode+10 );
 5679+
 5680+#define stbir__3_coeff_only()                           \
 5681+    stbir__simdf tot0,tot1,c,cs;                        \
 5682+    STBIR_SIMD_NO_UNROLL(decode);                       \
 5683+    stbir__simdf_load( cs, hc );                        \
 5684+    stbir__simdf_0123to0000( c, cs );                   \
 5685+    stbir__simdf_mult_mem( tot0, c, decode );           \
 5686+    stbir__simdf_mult_mem( tot1, c, decode+3 );         \
 5687+    stbir__simdf_0123to1111( c, cs );                   \
 5688+    stbir__simdf_madd_mem( tot0, tot0, c, decode+7 );   \
 5689+    stbir__simdf_madd_mem( tot1, tot1, c, decode+10 );  \
 5690+    stbir__simdf_0123to2222( c, cs );                   \
 5691+    stbir__simdf_madd_mem( tot0, tot0, c, decode+14 );  \
 5692+    stbir__simdf_madd_mem( tot1, tot1, c, decode+17 );
 5693+
 5694+#define stbir__store_output_tiny()                \
 5695+    stbir__simdf_store( output+3, tot1 );         \
 5696+    stbir__simdf_store( output, tot0 );           \
 5697+    horizontal_coefficients += coefficient_width; \
 5698+    ++horizontal_contributors;                    \
 5699+    output += 7;
 5700+
 5701+#ifdef STBIR_SIMD8
 5702+
 5703+#define stbir__4_coeff_start()                     \
 5704+    stbir__simdf8 tot0,tot1,c,cs;                  \
 5705+    STBIR_SIMD_NO_UNROLL(decode);                  \
 5706+    stbir__simdf8_load4b( cs, hc );                \
 5707+    stbir__simdf8_0123to00000000( c, cs );         \
 5708+    stbir__simdf8_mult_mem( tot0, c, decode );     \
 5709+    stbir__simdf8_0123to11111111( c, cs );         \
 5710+    stbir__simdf8_mult_mem( tot1, c, decode+7 );   \
 5711+    stbir__simdf8_0123to22222222( c, cs );         \
 5712+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+14 );  \
 5713+    stbir__simdf8_0123to33333333( c, cs );         \
 5714+    stbir__simdf8_madd_mem( tot1, tot1, c, decode+21 );
 5715+
 5716+#define stbir__4_coeff_continue_from_4( ofs )                   \
 5717+    STBIR_SIMD_NO_UNROLL(decode);                               \
 5718+    stbir__simdf8_load4b( cs, hc + (ofs) );                     \
 5719+    stbir__simdf8_0123to00000000( c, cs );                      \
 5720+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*7 );    \
 5721+    stbir__simdf8_0123to11111111( c, cs );                      \
 5722+    stbir__simdf8_madd_mem( tot1, tot1, c, decode+(ofs)*7+7 );  \
 5723+    stbir__simdf8_0123to22222222( c, cs );                      \
 5724+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*7+14 ); \
 5725+    stbir__simdf8_0123to33333333( c, cs );                      \
 5726+    stbir__simdf8_madd_mem( tot1, tot1, c, decode+(ofs)*7+21 );
 5727+
 5728+#define stbir__1_coeff_remnant( ofs )                           \
 5729+    STBIR_SIMD_NO_UNROLL(decode);                               \
 5730+    stbir__simdf8_load1b( c, hc + (ofs) );                      \
 5731+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*7 );
 5732+
 5733+#define stbir__2_coeff_remnant( ofs )                           \
 5734+    STBIR_SIMD_NO_UNROLL(decode);                               \
 5735+    stbir__simdf8_load1b( c, hc + (ofs) );                      \
 5736+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*7 );    \
 5737+    stbir__simdf8_load1b( c, hc + (ofs)+1 );                    \
 5738+    stbir__simdf8_madd_mem( tot1, tot1, c, decode+(ofs)*7+7 );
 5739+
 5740+#define stbir__3_coeff_remnant( ofs )                           \
 5741+    STBIR_SIMD_NO_UNROLL(decode);                               \
 5742+    stbir__simdf8_load4b( cs, hc + (ofs) );                     \
 5743+    stbir__simdf8_0123to00000000( c, cs );                      \
 5744+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*7 );    \
 5745+    stbir__simdf8_0123to11111111( c, cs );                      \
 5746+    stbir__simdf8_madd_mem( tot1, tot1, c, decode+(ofs)*7+7 );  \
 5747+    stbir__simdf8_0123to22222222( c, cs );                      \
 5748+    stbir__simdf8_madd_mem( tot0, tot0, c, decode+(ofs)*7+14 );
 5749+
 5750+#define stbir__store_output()                     \
 5751+    stbir__simdf8_add( tot0, tot0, tot1 );        \
 5752+    horizontal_coefficients += coefficient_width; \
 5753+    ++horizontal_contributors;                    \
 5754+    output += 7;                                  \
 5755+    if ( output < output_end )                    \
 5756+    {                                             \
 5757+      stbir__simdf8_store( output-7, tot0 );      \
 5758+      continue;                                   \
 5759+    }                                             \
 5760+    stbir__simdf_store( output-7+3, stbir__simdf_swiz(stbir__simdf8_gettop4(tot0),0,0,1,2) ); \
 5761+    stbir__simdf_store( output-7, stbir__if_simdf8_cast_to_simdf4(tot0) );           \
 5762+    break;
 5763+
 5764+#else
 5765+
 5766+#define stbir__4_coeff_start()                    \
 5767+    stbir__simdf tot0,tot1,tot2,tot3,c,cs;        \
 5768+    STBIR_SIMD_NO_UNROLL(decode);                 \
 5769+    stbir__simdf_load( cs, hc );                  \
 5770+    stbir__simdf_0123to0000( c, cs );             \
 5771+    stbir__simdf_mult_mem( tot0, c, decode );     \
 5772+    stbir__simdf_mult_mem( tot1, c, decode+3 );   \
 5773+    stbir__simdf_0123to1111( c, cs );             \
 5774+    stbir__simdf_mult_mem( tot2, c, decode+7 );   \
 5775+    stbir__simdf_mult_mem( tot3, c, decode+10 );  \
 5776+    stbir__simdf_0123to2222( c, cs );             \
 5777+    stbir__simdf_madd_mem( tot0, tot0, c, decode+14 );  \
 5778+    stbir__simdf_madd_mem( tot1, tot1, c, decode+17 );  \
 5779+    stbir__simdf_0123to3333( c, cs );                   \
 5780+    stbir__simdf_madd_mem( tot2, tot2, c, decode+21 );  \
 5781+    stbir__simdf_madd_mem( tot3, tot3, c, decode+24 );
 5782+
 5783+#define stbir__4_coeff_continue_from_4( ofs )                   \
 5784+    STBIR_SIMD_NO_UNROLL(decode);                               \
 5785+    stbir__simdf_load( cs, hc + (ofs) );                        \
 5786+    stbir__simdf_0123to0000( c, cs );                           \
 5787+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*7 );     \
 5788+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*7+3 );   \
 5789+    stbir__simdf_0123to1111( c, cs );                           \
 5790+    stbir__simdf_madd_mem( tot2, tot2, c, decode+(ofs)*7+7 );   \
 5791+    stbir__simdf_madd_mem( tot3, tot3, c, decode+(ofs)*7+10 );  \
 5792+    stbir__simdf_0123to2222( c, cs );                           \
 5793+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*7+14 );  \
 5794+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*7+17 );  \
 5795+    stbir__simdf_0123to3333( c, cs );                           \
 5796+    stbir__simdf_madd_mem( tot2, tot2, c, decode+(ofs)*7+21 );  \
 5797+    stbir__simdf_madd_mem( tot3, tot3, c, decode+(ofs)*7+24 );
 5798+
 5799+#define stbir__1_coeff_remnant( ofs )                           \
 5800+    STBIR_SIMD_NO_UNROLL(decode);                               \
 5801+    stbir__simdf_load1( c, hc + (ofs) );                        \
 5802+    stbir__simdf_0123to0000( c, c );                            \
 5803+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*7 );     \
 5804+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*7+3 );   \
 5805+
 5806+#define stbir__2_coeff_remnant( ofs )                           \
 5807+    STBIR_SIMD_NO_UNROLL(decode);                               \
 5808+    stbir__simdf_load2( cs, hc + (ofs) );                       \
 5809+    stbir__simdf_0123to0000( c, cs );                           \
 5810+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*7 );     \
 5811+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*7+3 );   \
 5812+    stbir__simdf_0123to1111( c, cs );                           \
 5813+    stbir__simdf_madd_mem( tot2, tot2, c, decode+(ofs)*7+7 );   \
 5814+    stbir__simdf_madd_mem( tot3, tot3, c, decode+(ofs)*7+10 );
 5815+
 5816+#define stbir__3_coeff_remnant( ofs )                           \
 5817+    STBIR_SIMD_NO_UNROLL(decode);                               \
 5818+    stbir__simdf_load( cs, hc + (ofs) );                        \
 5819+    stbir__simdf_0123to0000( c, cs );                           \
 5820+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*7 );     \
 5821+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*7+3 );   \
 5822+    stbir__simdf_0123to1111( c, cs );                           \
 5823+    stbir__simdf_madd_mem( tot2, tot2, c, decode+(ofs)*7+7 );   \
 5824+    stbir__simdf_madd_mem( tot3, tot3, c, decode+(ofs)*7+10 );  \
 5825+    stbir__simdf_0123to2222( c, cs );                           \
 5826+    stbir__simdf_madd_mem( tot0, tot0, c, decode+(ofs)*7+14 );  \
 5827+    stbir__simdf_madd_mem( tot1, tot1, c, decode+(ofs)*7+17 );
 5828+
 5829+#define stbir__store_output()                     \
 5830+    stbir__simdf_add( tot0, tot0, tot2 );         \
 5831+    stbir__simdf_add( tot1, tot1, tot3 );         \
 5832+    stbir__simdf_store( output+3, tot1 );         \
 5833+    stbir__simdf_store( output, tot0 );           \
 5834+    horizontal_coefficients += coefficient_width; \
 5835+    ++horizontal_contributors;                    \
 5836+    output += 7;
 5837+
 5838+#endif
 5839+
 5840+#else
 5841+
 5842+#define stbir__1_coeff_only()        \
 5843+    float tot0, tot1, tot2, tot3, tot4, tot5, tot6, c; \
 5844+    c = hc[0];                       \
 5845+    tot0 = decode[0]*c;              \
 5846+    tot1 = decode[1]*c;              \
 5847+    tot2 = decode[2]*c;              \
 5848+    tot3 = decode[3]*c;              \
 5849+    tot4 = decode[4]*c;              \
 5850+    tot5 = decode[5]*c;              \
 5851+    tot6 = decode[6]*c;
 5852+
 5853+#define stbir__2_coeff_only()        \
 5854+    float tot0, tot1, tot2, tot3, tot4, tot5, tot6, c; \
 5855+    c = hc[0];                       \
 5856+    tot0 = decode[0]*c;              \
 5857+    tot1 = decode[1]*c;              \
 5858+    tot2 = decode[2]*c;              \
 5859+    tot3 = decode[3]*c;              \
 5860+    tot4 = decode[4]*c;              \
 5861+    tot5 = decode[5]*c;              \
 5862+    tot6 = decode[6]*c;              \
 5863+    c = hc[1];                       \
 5864+    tot0 += decode[7]*c;             \
 5865+    tot1 += decode[8]*c;             \
 5866+    tot2 += decode[9]*c;             \
 5867+    tot3 += decode[10]*c;            \
 5868+    tot4 += decode[11]*c;            \
 5869+    tot5 += decode[12]*c;            \
 5870+    tot6 += decode[13]*c;            \
 5871+
 5872+#define stbir__3_coeff_only()        \
 5873+    float tot0, tot1, tot2, tot3, tot4, tot5, tot6, c; \
 5874+    c = hc[0];                       \
 5875+    tot0 = decode[0]*c;              \
 5876+    tot1 = decode[1]*c;              \
 5877+    tot2 = decode[2]*c;              \
 5878+    tot3 = decode[3]*c;              \
 5879+    tot4 = decode[4]*c;              \
 5880+    tot5 = decode[5]*c;              \
 5881+    tot6 = decode[6]*c;              \
 5882+    c = hc[1];                       \
 5883+    tot0 += decode[7]*c;             \
 5884+    tot1 += decode[8]*c;             \
 5885+    tot2 += decode[9]*c;             \
 5886+    tot3 += decode[10]*c;            \
 5887+    tot4 += decode[11]*c;            \
 5888+    tot5 += decode[12]*c;            \
 5889+    tot6 += decode[13]*c;            \
 5890+    c = hc[2];                       \
 5891+    tot0 += decode[14]*c;            \
 5892+    tot1 += decode[15]*c;            \
 5893+    tot2 += decode[16]*c;            \
 5894+    tot3 += decode[17]*c;            \
 5895+    tot4 += decode[18]*c;            \
 5896+    tot5 += decode[19]*c;            \
 5897+    tot6 += decode[20]*c;            \
 5898+
 5899+#define stbir__store_output_tiny()                \
 5900+    output[0] = tot0;                             \
 5901+    output[1] = tot1;                             \
 5902+    output[2] = tot2;                             \
 5903+    output[3] = tot3;                             \
 5904+    output[4] = tot4;                             \
 5905+    output[5] = tot5;                             \
 5906+    output[6] = tot6;                             \
 5907+    horizontal_coefficients += coefficient_width; \
 5908+    ++horizontal_contributors;                    \
 5909+    output += 7;
 5910+
 5911+#define stbir__4_coeff_start()    \
 5912+    float x0,x1,x2,x3,x4,x5,x6,y0,y1,y2,y3,y4,y5,y6,c; \
 5913+    STBIR_SIMD_NO_UNROLL(decode); \
 5914+    c = hc[0];                    \
 5915+    x0 = decode[0] * c;           \
 5916+    x1 = decode[1] * c;           \
 5917+    x2 = decode[2] * c;           \
 5918+    x3 = decode[3] * c;           \
 5919+    x4 = decode[4] * c;           \
 5920+    x5 = decode[5] * c;           \
 5921+    x6 = decode[6] * c;           \
 5922+    c = hc[1];                    \
 5923+    y0 = decode[7] * c;           \
 5924+    y1 = decode[8] * c;           \
 5925+    y2 = decode[9] * c;           \
 5926+    y3 = decode[10] * c;          \
 5927+    y4 = decode[11] * c;          \
 5928+    y5 = decode[12] * c;          \
 5929+    y6 = decode[13] * c;          \
 5930+    c = hc[2];                    \
 5931+    x0 += decode[14] * c;         \
 5932+    x1 += decode[15] * c;         \
 5933+    x2 += decode[16] * c;         \
 5934+    x3 += decode[17] * c;         \
 5935+    x4 += decode[18] * c;         \
 5936+    x5 += decode[19] * c;         \
 5937+    x6 += decode[20] * c;         \
 5938+    c = hc[3];                    \
 5939+    y0 += decode[21] * c;         \
 5940+    y1 += decode[22] * c;         \
 5941+    y2 += decode[23] * c;         \
 5942+    y3 += decode[24] * c;         \
 5943+    y4 += decode[25] * c;         \
 5944+    y5 += decode[26] * c;         \
 5945+    y6 += decode[27] * c;
 5946+
 5947+#define stbir__4_coeff_continue_from_4( ofs ) \
 5948+    STBIR_SIMD_NO_UNROLL(decode);  \
 5949+    c = hc[0+(ofs)];               \
 5950+    x0 += decode[0+(ofs)*7] * c;   \
 5951+    x1 += decode[1+(ofs)*7] * c;   \
 5952+    x2 += decode[2+(ofs)*7] * c;   \
 5953+    x3 += decode[3+(ofs)*7] * c;   \
 5954+    x4 += decode[4+(ofs)*7] * c;   \
 5955+    x5 += decode[5+(ofs)*7] * c;   \
 5956+    x6 += decode[6+(ofs)*7] * c;   \
 5957+    c = hc[1+(ofs)];               \
 5958+    y0 += decode[7+(ofs)*7] * c;   \
 5959+    y1 += decode[8+(ofs)*7] * c;   \
 5960+    y2 += decode[9+(ofs)*7] * c;   \
 5961+    y3 += decode[10+(ofs)*7] * c;  \
 5962+    y4 += decode[11+(ofs)*7] * c;  \
 5963+    y5 += decode[12+(ofs)*7] * c;  \
 5964+    y6 += decode[13+(ofs)*7] * c;  \
 5965+    c = hc[2+(ofs)];               \
 5966+    x0 += decode[14+(ofs)*7] * c;  \
 5967+    x1 += decode[15+(ofs)*7] * c;  \
 5968+    x2 += decode[16+(ofs)*7] * c;  \
 5969+    x3 += decode[17+(ofs)*7] * c;  \
 5970+    x4 += decode[18+(ofs)*7] * c;  \
 5971+    x5 += decode[19+(ofs)*7] * c;  \
 5972+    x6 += decode[20+(ofs)*7] * c;  \
 5973+    c = hc[3+(ofs)];               \
 5974+    y0 += decode[21+(ofs)*7] * c;  \
 5975+    y1 += decode[22+(ofs)*7] * c;  \
 5976+    y2 += decode[23+(ofs)*7] * c;  \
 5977+    y3 += decode[24+(ofs)*7] * c;  \
 5978+    y4 += decode[25+(ofs)*7] * c;  \
 5979+    y5 += decode[26+(ofs)*7] * c;  \
 5980+    y6 += decode[27+(ofs)*7] * c;
 5981+
 5982+#define stbir__1_coeff_remnant( ofs ) \
 5983+    STBIR_SIMD_NO_UNROLL(decode);  \
 5984+    c = hc[0+(ofs)];               \
 5985+    x0 += decode[0+(ofs)*7] * c;   \
 5986+    x1 += decode[1+(ofs)*7] * c;   \
 5987+    x2 += decode[2+(ofs)*7] * c;   \
 5988+    x3 += decode[3+(ofs)*7] * c;   \
 5989+    x4 += decode[4+(ofs)*7] * c;   \
 5990+    x5 += decode[5+(ofs)*7] * c;   \
 5991+    x6 += decode[6+(ofs)*7] * c;   \
 5992+
 5993+#define stbir__2_coeff_remnant( ofs ) \
 5994+    STBIR_SIMD_NO_UNROLL(decode);  \
 5995+    c = hc[0+(ofs)];               \
 5996+    x0 += decode[0+(ofs)*7] * c;   \
 5997+    x1 += decode[1+(ofs)*7] * c;   \
 5998+    x2 += decode[2+(ofs)*7] * c;   \
 5999+    x3 += decode[3+(ofs)*7] * c;   \
 6000+    x4 += decode[4+(ofs)*7] * c;   \
 6001+    x5 += decode[5+(ofs)*7] * c;   \
 6002+    x6 += decode[6+(ofs)*7] * c;   \
 6003+    c = hc[1+(ofs)];               \
 6004+    y0 += decode[7+(ofs)*7] * c;   \
 6005+    y1 += decode[8+(ofs)*7] * c;   \
 6006+    y2 += decode[9+(ofs)*7] * c;   \
 6007+    y3 += decode[10+(ofs)*7] * c;  \
 6008+    y4 += decode[11+(ofs)*7] * c;  \
 6009+    y5 += decode[12+(ofs)*7] * c;  \
 6010+    y6 += decode[13+(ofs)*7] * c;  \
 6011+
 6012+#define stbir__3_coeff_remnant( ofs ) \
 6013+    STBIR_SIMD_NO_UNROLL(decode);  \
 6014+    c = hc[0+(ofs)];               \
 6015+    x0 += decode[0+(ofs)*7] * c;   \
 6016+    x1 += decode[1+(ofs)*7] * c;   \
 6017+    x2 += decode[2+(ofs)*7] * c;   \
 6018+    x3 += decode[3+(ofs)*7] * c;   \
 6019+    x4 += decode[4+(ofs)*7] * c;   \
 6020+    x5 += decode[5+(ofs)*7] * c;   \
 6021+    x6 += decode[6+(ofs)*7] * c;   \
 6022+    c = hc[1+(ofs)];               \
 6023+    y0 += decode[7+(ofs)*7] * c;   \
 6024+    y1 += decode[8+(ofs)*7] * c;   \
 6025+    y2 += decode[9+(ofs)*7] * c;   \
 6026+    y3 += decode[10+(ofs)*7] * c;  \
 6027+    y4 += decode[11+(ofs)*7] * c;  \
 6028+    y5 += decode[12+(ofs)*7] * c;  \
 6029+    y6 += decode[13+(ofs)*7] * c;  \
 6030+    c = hc[2+(ofs)];               \
 6031+    x0 += decode[14+(ofs)*7] * c;  \
 6032+    x1 += decode[15+(ofs)*7] * c;  \
 6033+    x2 += decode[16+(ofs)*7] * c;  \
 6034+    x3 += decode[17+(ofs)*7] * c;  \
 6035+    x4 += decode[18+(ofs)*7] * c;  \
 6036+    x5 += decode[19+(ofs)*7] * c;  \
 6037+    x6 += decode[20+(ofs)*7] * c;  \
 6038+
 6039+#define stbir__store_output()                     \
 6040+    output[0] = x0 + y0;                          \
 6041+    output[1] = x1 + y1;                          \
 6042+    output[2] = x2 + y2;                          \
 6043+    output[3] = x3 + y3;                          \
 6044+    output[4] = x4 + y4;                          \
 6045+    output[5] = x5 + y5;                          \
 6046+    output[6] = x6 + y6;                          \
 6047+    horizontal_coefficients += coefficient_width; \
 6048+    ++horizontal_contributors;                    \
 6049+    output += 7;
 6050+
 6051+#endif
 6052+
 6053+#define STBIR__horizontal_channels 7
 6054+#define STB_IMAGE_RESIZE_DO_HORIZONTALS
 6055+#include STBIR__HEADER_FILENAME
 6056+
 6057+
 6058+// include all of the vertical resamplers (both scatter and gather versions)
 6059+
 6060+#define STBIR__vertical_channels 1
 6061+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6062+#include STBIR__HEADER_FILENAME
 6063+
 6064+#define STBIR__vertical_channels 1
 6065+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6066+#define STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 6067+#include STBIR__HEADER_FILENAME
 6068+
 6069+#define STBIR__vertical_channels 2
 6070+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6071+#include STBIR__HEADER_FILENAME
 6072+
 6073+#define STBIR__vertical_channels 2
 6074+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6075+#define STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 6076+#include STBIR__HEADER_FILENAME
 6077+
 6078+#define STBIR__vertical_channels 3
 6079+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6080+#include STBIR__HEADER_FILENAME
 6081+
 6082+#define STBIR__vertical_channels 3
 6083+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6084+#define STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 6085+#include STBIR__HEADER_FILENAME
 6086+
 6087+#define STBIR__vertical_channels 4
 6088+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6089+#include STBIR__HEADER_FILENAME
 6090+
 6091+#define STBIR__vertical_channels 4
 6092+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6093+#define STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 6094+#include STBIR__HEADER_FILENAME
 6095+
 6096+#define STBIR__vertical_channels 5
 6097+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6098+#include STBIR__HEADER_FILENAME
 6099+
 6100+#define STBIR__vertical_channels 5
 6101+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6102+#define STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 6103+#include STBIR__HEADER_FILENAME
 6104+
 6105+#define STBIR__vertical_channels 6
 6106+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6107+#include STBIR__HEADER_FILENAME
 6108+
 6109+#define STBIR__vertical_channels 6
 6110+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6111+#define STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 6112+#include STBIR__HEADER_FILENAME
 6113+
 6114+#define STBIR__vertical_channels 7
 6115+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6116+#include STBIR__HEADER_FILENAME
 6117+
 6118+#define STBIR__vertical_channels 7
 6119+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6120+#define STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 6121+#include STBIR__HEADER_FILENAME
 6122+
 6123+#define STBIR__vertical_channels 8
 6124+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6125+#include STBIR__HEADER_FILENAME
 6126+
 6127+#define STBIR__vertical_channels 8
 6128+#define STB_IMAGE_RESIZE_DO_VERTICALS
 6129+#define STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 6130+#include STBIR__HEADER_FILENAME
 6131+
 6132+typedef void STBIR_VERTICAL_GATHERFUNC( float * output, float const * coeffs, float const ** inputs, float const * input0_end );
 6133+
 6134+static STBIR_VERTICAL_GATHERFUNC * stbir__vertical_gathers[ 8 ] =
 6135+{
 6136+  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
 6137+};
 6138+
 6139+static STBIR_VERTICAL_GATHERFUNC * stbir__vertical_gathers_continues[ 8 ] =
 6140+{
 6141+  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
 6142+};
 6143+
 6144+typedef void STBIR_VERTICAL_SCATTERFUNC( float ** outputs, float const * coeffs, float const * input, float const * input_end );
 6145+
 6146+static STBIR_VERTICAL_SCATTERFUNC * stbir__vertical_scatter_sets[ 8 ] =
 6147+{
 6148+  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
 6149+};
 6150+
 6151+static STBIR_VERTICAL_SCATTERFUNC * stbir__vertical_scatter_blends[ 8 ] =
 6152+{
 6153+  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
 6154+};
 6155+
 6156+
 6157+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 )
 6158+{
 6159+  int num_pixels = stbir_info->horizontal.scale_info.output_sub_size;
 6160+  int channels = stbir_info->channels;
 6161+  int width_times_channels = num_pixels * channels;
 6162+  void * output_buffer;
 6163+
 6164+  // un-alpha weight if we need to
 6165+  if ( stbir_info->alpha_unweight )
 6166+  {
 6167+    STBIR_PROFILE_START( unalpha );
 6168+    stbir_info->alpha_unweight( encode_buffer, width_times_channels );
 6169+    STBIR_PROFILE_END( unalpha );
 6170+  }
 6171+
 6172+  // write directly into output by default
 6173+  output_buffer = output_buffer_data;
 6174+
 6175+  // if we have an output callback, we first convert the decode buffer in place (and then hand that to the callback)
 6176+  if ( stbir_info->out_pixels_cb )
 6177+    output_buffer = encode_buffer;
 6178+
 6179+  STBIR_PROFILE_START( encode );
 6180+  // convert into the output buffer
 6181+  stbir_info->encode_pixels( output_buffer, width_times_channels, encode_buffer );
 6182+  STBIR_PROFILE_END( encode );
 6183+
 6184+  // if we have an output callback, call it to send the data
 6185+  if ( stbir_info->out_pixels_cb )
 6186+    stbir_info->out_pixels_cb( output_buffer, num_pixels, row, stbir_info->user_data );
 6187+}
 6188+
 6189+
 6190+// Get the ring buffer pointer for an index
 6191+static float* stbir__get_ring_buffer_entry(stbir__info const * stbir_info, stbir__per_split_info const * split_info, int index )
 6192+{
 6193+  STBIR_ASSERT( index < stbir_info->ring_buffer_num_entries );
 6194+
 6195+  #ifdef STBIR__SEPARATE_ALLOCATIONS
 6196+    return split_info->ring_buffers[ index ];
 6197+  #else
 6198+    return (float*) ( ( (char*) split_info->ring_buffer ) + ( index * stbir_info->ring_buffer_length_bytes ) );
 6199+  #endif
 6200+}
 6201+
 6202+// Get the specified scan line from the ring buffer
 6203+static float* stbir__get_ring_buffer_scanline(stbir__info const * stbir_info, stbir__per_split_info const * split_info, int get_scanline)
 6204+{
 6205+  int ring_buffer_index = (split_info->ring_buffer_begin_index + (get_scanline - split_info->ring_buffer_first_scanline)) % stbir_info->ring_buffer_num_entries;
 6206+  return stbir__get_ring_buffer_entry( stbir_info, split_info, ring_buffer_index );
 6207+}
 6208+
 6209+static void stbir__resample_horizontal_gather(stbir__info const * stbir_info, float* output_buffer, float const * input_buffer STBIR_ONLY_PROFILE_GET_SPLIT_INFO )
 6210+{
 6211+  float const * decode_buffer = input_buffer - ( stbir_info->scanline_extents.conservative.n0 * stbir_info->effective_channels );
 6212+
 6213+  STBIR_PROFILE_START( horizontal );
 6214+  if ( ( stbir_info->horizontal.filter_enum == STBIR_FILTER_POINT_SAMPLE ) && ( stbir_info->horizontal.scale_info.scale == 1.0f ) )
 6215+    STBIR_MEMCPY( output_buffer, input_buffer, stbir_info->horizontal.scale_info.output_sub_size * sizeof( float ) * stbir_info->effective_channels );
 6216+  else
 6217+    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 );
 6218+  STBIR_PROFILE_END( horizontal );
 6219+}
 6220+
 6221+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 )
 6222+{
 6223+  float* encode_buffer = split_info->vertical_buffer;
 6224+  float* decode_buffer = split_info->decode_buffer;
 6225+  int vertical_first = stbir_info->vertical_first;
 6226+  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;
 6227+  int width_times_channels = stbir_info->effective_channels * width;
 6228+
 6229+  STBIR_ASSERT( stbir_info->vertical.is_gather );
 6230+
 6231+  // loop over the contributing scanlines and scale into the buffer
 6232+  STBIR_PROFILE_START( vertical );
 6233+  {
 6234+    int k = 0, total = contrib_n1 - contrib_n0 + 1;
 6235+    STBIR_ASSERT( total > 0 );
 6236+    do {
 6237+      float const * inputs[8];
 6238+      int i, cnt = total; if ( cnt > 8 ) cnt = 8;
 6239+      for( i = 0 ; i < cnt ; i++ )
 6240+        inputs[ i ] = stbir__get_ring_buffer_scanline(stbir_info, split_info, k+i+contrib_n0 );
 6241+
 6242+      // call the N scanlines at a time function (up to 8 scanlines of blending at once)
 6243+      ((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 );
 6244+      k += cnt;
 6245+      total -= cnt;
 6246+    } while ( total );
 6247+  }
 6248+  STBIR_PROFILE_END( vertical );
 6249+
 6250+  if ( vertical_first )
 6251+  {
 6252+    // Now resample the gathered vertical data in the horizontal axis into the encode buffer
 6253+    decode_buffer[ width_times_channels ] = 0.0f; // clear two over for horizontals with a remnant of 3
 6254+    decode_buffer[ width_times_channels+1 ] = 0.0f; 
 6255+    stbir__resample_horizontal_gather(stbir_info, encode_buffer, decode_buffer  STBIR_ONLY_PROFILE_SET_SPLIT_INFO );
 6256+  }
 6257+
 6258+  stbir__encode_scanline( stbir_info, ( (char *) stbir_info->output_data ) + ((size_t)n * (size_t)stbir_info->output_stride_bytes),
 6259+                          encode_buffer, n  STBIR_ONLY_PROFILE_SET_SPLIT_INFO );
 6260+}
 6261+
 6262+static void stbir__decode_and_resample_for_vertical_gather_loop(stbir__info const * stbir_info, stbir__per_split_info* split_info, int n)
 6263+{
 6264+  int ring_buffer_index;
 6265+  float* ring_buffer;
 6266+
 6267+  // Decode the nth scanline from the source image into the decode buffer.
 6268+  stbir__decode_scanline( stbir_info, n, split_info->decode_buffer  STBIR_ONLY_PROFILE_SET_SPLIT_INFO );
 6269+
 6270+  // update new end scanline
 6271+  split_info->ring_buffer_last_scanline = n;
 6272+
 6273+  // get ring buffer
 6274+  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;
 6275+  ring_buffer = stbir__get_ring_buffer_entry(stbir_info, split_info, ring_buffer_index);
 6276+
 6277+  // Now resample it into the ring buffer.
 6278+  stbir__resample_horizontal_gather( stbir_info, ring_buffer, split_info->decode_buffer  STBIR_ONLY_PROFILE_SET_SPLIT_INFO );
 6279+
 6280+  // Now it's sitting in the ring buffer ready to be used as source for the vertical sampling.
 6281+}
 6282+
 6283+static void stbir__vertical_gather_loop( stbir__info const * stbir_info, stbir__per_split_info* split_info, int split_count )
 6284+{
 6285+  int y, start_output_y, end_output_y;
 6286+  stbir__contributors* vertical_contributors = stbir_info->vertical.contributors;
 6287+  float const * vertical_coefficients = stbir_info->vertical.coefficients;
 6288+
 6289+  STBIR_ASSERT( stbir_info->vertical.is_gather );
 6290+
 6291+  start_output_y = split_info->start_output_y;
 6292+  end_output_y = split_info[split_count-1].end_output_y;
 6293+
 6294+  vertical_contributors += start_output_y;
 6295+  vertical_coefficients += start_output_y * stbir_info->vertical.coefficient_width;
 6296+
 6297+  // initialize the ring buffer for gathering
 6298+  split_info->ring_buffer_begin_index = 0;
 6299+  split_info->ring_buffer_first_scanline = vertical_contributors->n0;
 6300+  split_info->ring_buffer_last_scanline = split_info->ring_buffer_first_scanline - 1; // means "empty"
 6301+
 6302+  for (y = start_output_y; y < end_output_y; y++)
 6303+  {
 6304+    int in_first_scanline, in_last_scanline;
 6305+
 6306+    in_first_scanline = vertical_contributors->n0;
 6307+    in_last_scanline = vertical_contributors->n1;
 6308+
 6309+    // make sure the indexing hasn't broken
 6310+    STBIR_ASSERT( in_first_scanline >= split_info->ring_buffer_first_scanline );
 6311+
 6312+    // Load in new scanlines
 6313+    while (in_last_scanline > split_info->ring_buffer_last_scanline)
 6314+    {
 6315+      STBIR_ASSERT( ( split_info->ring_buffer_last_scanline - split_info->ring_buffer_first_scanline + 1 ) <= stbir_info->ring_buffer_num_entries );
 6316+
 6317+      // make sure there was room in the ring buffer when we add new scanlines
 6318+      if ( ( split_info->ring_buffer_last_scanline - split_info->ring_buffer_first_scanline + 1 ) == stbir_info->ring_buffer_num_entries )
 6319+      {
 6320+        split_info->ring_buffer_first_scanline++;
 6321+        split_info->ring_buffer_begin_index++;
 6322+      }
 6323+
 6324+      if ( stbir_info->vertical_first )
 6325+      {
 6326+        float * ring_buffer = stbir__get_ring_buffer_scanline( stbir_info, split_info, ++split_info->ring_buffer_last_scanline );
 6327+        // Decode the nth scanline from the source image into the decode buffer.
 6328+        stbir__decode_scanline( stbir_info, split_info->ring_buffer_last_scanline, ring_buffer  STBIR_ONLY_PROFILE_SET_SPLIT_INFO );
 6329+      }
 6330+      else
 6331+      {
 6332+        stbir__decode_and_resample_for_vertical_gather_loop(stbir_info, split_info, split_info->ring_buffer_last_scanline + 1);
 6333+      }
 6334+    }
 6335+
 6336+    // Now all buffers should be ready to write a row of vertical sampling, so do it.
 6337+    stbir__resample_vertical_gather(stbir_info, split_info, y, in_first_scanline, in_last_scanline, vertical_coefficients );
 6338+
 6339+    ++vertical_contributors;
 6340+    vertical_coefficients += stbir_info->vertical.coefficient_width;
 6341+  }
 6342+}
 6343+
 6344+#define STBIR__FLOAT_EMPTY_MARKER 3.0e+38F
 6345+#define STBIR__FLOAT_BUFFER_IS_EMPTY(ptr) ((ptr)[0]==STBIR__FLOAT_EMPTY_MARKER)
 6346+
 6347+static void stbir__encode_first_scanline_from_scatter(stbir__info const * stbir_info, stbir__per_split_info* split_info)
 6348+{
 6349+  // evict a scanline out into the output buffer
 6350+  float* ring_buffer_entry = stbir__get_ring_buffer_entry(stbir_info, split_info, split_info->ring_buffer_begin_index );
 6351+
 6352+  // dump the scanline out
 6353+  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 );
 6354+
 6355+  // mark it as empty
 6356+  ring_buffer_entry[ 0 ] = STBIR__FLOAT_EMPTY_MARKER;
 6357+
 6358+  // advance the first scanline
 6359+  split_info->ring_buffer_first_scanline++;
 6360+  if ( ++split_info->ring_buffer_begin_index == stbir_info->ring_buffer_num_entries )
 6361+    split_info->ring_buffer_begin_index = 0;
 6362+}
 6363+
 6364+static void stbir__horizontal_resample_and_encode_first_scanline_from_scatter(stbir__info const * stbir_info, stbir__per_split_info* split_info)
 6365+{
 6366+  // evict a scanline out into the output buffer
 6367+
 6368+  float* ring_buffer_entry = stbir__get_ring_buffer_entry(stbir_info, split_info, split_info->ring_buffer_begin_index );
 6369+
 6370+  // Now resample it into the buffer.
 6371+  stbir__resample_horizontal_gather( stbir_info, split_info->vertical_buffer, ring_buffer_entry  STBIR_ONLY_PROFILE_SET_SPLIT_INFO );
 6372+
 6373+  // dump the scanline out
 6374+  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 );
 6375+
 6376+  // mark it as empty
 6377+  ring_buffer_entry[ 0 ] = STBIR__FLOAT_EMPTY_MARKER;
 6378+
 6379+  // advance the first scanline
 6380+  split_info->ring_buffer_first_scanline++;
 6381+  if ( ++split_info->ring_buffer_begin_index == stbir_info->ring_buffer_num_entries )
 6382+    split_info->ring_buffer_begin_index = 0;
 6383+}
 6384+
 6385+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 )
 6386+{
 6387+  STBIR_ASSERT( !stbir_info->vertical.is_gather );
 6388+
 6389+  STBIR_PROFILE_START( vertical );
 6390+  {
 6391+    int k = 0, total = n1 - n0 + 1;
 6392+    STBIR_ASSERT( total > 0 );
 6393+    do {
 6394+      float * outputs[8];
 6395+      int i, n = total; if ( n > 8 ) n = 8;
 6396+      for( i = 0 ; i < n ; i++ )
 6397+      {
 6398+        outputs[ i ] = stbir__get_ring_buffer_scanline(stbir_info, split_info, k+i+n0 );
 6399+        if ( ( i ) && ( STBIR__FLOAT_BUFFER_IS_EMPTY( outputs[i] ) != STBIR__FLOAT_BUFFER_IS_EMPTY( outputs[0] ) ) ) // make sure runs are of the same type
 6400+        {
 6401+          n = i;
 6402+          break;
 6403+        }
 6404+      }
 6405+      // call the scatter to N scanlines at a time function (up to 8 scanlines of scattering at once)
 6406+      ((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 );
 6407+      k += n;
 6408+      total -= n;
 6409+    } while ( total );
 6410+  }
 6411+
 6412+  STBIR_PROFILE_END( vertical );
 6413+}
 6414+
 6415+typedef void stbir__handle_scanline_for_scatter_func(stbir__info const * stbir_info, stbir__per_split_info* split_info);
 6416+
 6417+static void stbir__vertical_scatter_loop( stbir__info const * stbir_info, stbir__per_split_info* split_info, int split_count )
 6418+{
 6419+  int y, start_output_y, end_output_y, start_input_y, end_input_y;
 6420+  stbir__contributors* vertical_contributors = stbir_info->vertical.contributors;
 6421+  float const * vertical_coefficients = stbir_info->vertical.coefficients;
 6422+  stbir__handle_scanline_for_scatter_func * handle_scanline_for_scatter;
 6423+  void * scanline_scatter_buffer;
 6424+  void * scanline_scatter_buffer_end;
 6425+  int on_first_input_y, last_input_y;
 6426+  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;
 6427+  int width_times_channels = stbir_info->effective_channels * width;
 6428+
 6429+  STBIR_ASSERT( !stbir_info->vertical.is_gather );
 6430+
 6431+  start_output_y = split_info->start_output_y;
 6432+  end_output_y = split_info[split_count-1].end_output_y;  // may do multiple split counts
 6433+
 6434+  start_input_y = split_info->start_input_y;
 6435+  end_input_y = split_info[split_count-1].end_input_y;
 6436+
 6437+  // adjust for starting offset start_input_y
 6438+  y = start_input_y + stbir_info->vertical.filter_pixel_margin;
 6439+  vertical_contributors += y ;
 6440+  vertical_coefficients += stbir_info->vertical.coefficient_width * y;
 6441+
 6442+  if ( stbir_info->vertical_first )
 6443+  {
 6444+    handle_scanline_for_scatter = stbir__horizontal_resample_and_encode_first_scanline_from_scatter;
 6445+    scanline_scatter_buffer = split_info->decode_buffer;
 6446+    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);
 6447+  }
 6448+  else
 6449+  {
 6450+    handle_scanline_for_scatter = stbir__encode_first_scanline_from_scatter;
 6451+    scanline_scatter_buffer = split_info->vertical_buffer;
 6452+    scanline_scatter_buffer_end = ( (char*) scanline_scatter_buffer ) + sizeof( float ) * stbir_info->effective_channels * stbir_info->horizontal.scale_info.output_sub_size;
 6453+  }
 6454+
 6455+  // initialize the ring buffer for scattering
 6456+  split_info->ring_buffer_first_scanline = start_output_y;
 6457+  split_info->ring_buffer_last_scanline = -1;
 6458+  split_info->ring_buffer_begin_index = -1;
 6459+
 6460+  // mark all the buffers as empty to start
 6461+  for( y = 0 ; y < stbir_info->ring_buffer_num_entries ; y++ )
 6462+  {
 6463+    float * decode_buffer = stbir__get_ring_buffer_entry( stbir_info, split_info, y );
 6464+    decode_buffer[ width_times_channels ] = 0.0f; // clear two over for horizontals with a remnant of 3
 6465+    decode_buffer[ width_times_channels+1 ] = 0.0f; 
 6466+    decode_buffer[0] = STBIR__FLOAT_EMPTY_MARKER; // only used on scatter
 6467+  }
 6468+
 6469+  // do the loop in input space
 6470+  on_first_input_y = 1; last_input_y = start_input_y;
 6471+  for (y = start_input_y ; y < end_input_y; y++)
 6472+  {
 6473+    int out_first_scanline, out_last_scanline;
 6474+
 6475+    out_first_scanline = vertical_contributors->n0;
 6476+    out_last_scanline = vertical_contributors->n1;
 6477+
 6478+    STBIR_ASSERT(out_last_scanline - out_first_scanline + 1 <= stbir_info->ring_buffer_num_entries);
 6479+
 6480+    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 ) ) ) )
 6481+    {
 6482+      float const * vc = vertical_coefficients;
 6483+
 6484+      // keep track of the range actually seen for the next resize
 6485+      last_input_y = y;
 6486+      if ( ( on_first_input_y ) && ( y > start_input_y ) )
 6487+        split_info->start_input_y = y;
 6488+      on_first_input_y = 0;
 6489+
 6490+      // clip the region
 6491+      if ( out_first_scanline < start_output_y )
 6492+      {
 6493+        vc += start_output_y - out_first_scanline;
 6494+        out_first_scanline = start_output_y;
 6495+      }
 6496+
 6497+      if ( out_last_scanline >= end_output_y )
 6498+        out_last_scanline = end_output_y - 1;
 6499+
 6500+      // if very first scanline, init the index
 6501+      if (split_info->ring_buffer_begin_index < 0)
 6502+        split_info->ring_buffer_begin_index = out_first_scanline - start_output_y;
 6503+
 6504+      STBIR_ASSERT( split_info->ring_buffer_begin_index <= out_first_scanline );
 6505+
 6506+      // Decode the nth scanline from the source image into the decode buffer.
 6507+      stbir__decode_scanline( stbir_info, y, split_info->decode_buffer  STBIR_ONLY_PROFILE_SET_SPLIT_INFO );
 6508+
 6509+      // When horizontal first, we resample horizontally into the vertical buffer before we scatter it out
 6510+      if ( !stbir_info->vertical_first )
 6511+        stbir__resample_horizontal_gather( stbir_info, split_info->vertical_buffer, split_info->decode_buffer  STBIR_ONLY_PROFILE_SET_SPLIT_INFO );
 6512+
 6513+      // Now it's sitting in the buffer ready to be distributed into the ring buffers.
 6514+
 6515+      // evict from the ringbuffer, if we need are full
 6516+      if ( ( ( split_info->ring_buffer_last_scanline - split_info->ring_buffer_first_scanline + 1 ) == stbir_info->ring_buffer_num_entries ) &&
 6517+           ( out_last_scanline > split_info->ring_buffer_last_scanline ) )
 6518+        handle_scanline_for_scatter( stbir_info, split_info );
 6519+
 6520+      // Now the horizontal buffer is ready to write to all ring buffer rows, so do it.
 6521+      stbir__resample_vertical_scatter(stbir_info, split_info, out_first_scanline, out_last_scanline, vc, (float*)scanline_scatter_buffer, (float*)scanline_scatter_buffer_end );
 6522+
 6523+      // update the end of the buffer
 6524+      if ( out_last_scanline > split_info->ring_buffer_last_scanline )
 6525+        split_info->ring_buffer_last_scanline = out_last_scanline;
 6526+    }
 6527+    ++vertical_contributors;
 6528+    vertical_coefficients += stbir_info->vertical.coefficient_width;
 6529+  }
 6530+
 6531+  // now evict the scanlines that are left over in the ring buffer
 6532+  while ( split_info->ring_buffer_first_scanline < end_output_y )
 6533+    handle_scanline_for_scatter(stbir_info, split_info);
 6534+
 6535+  // update the end_input_y if we do multiple resizes with the same data
 6536+  ++last_input_y;
 6537+  for( y = 0 ; y < split_count; y++ )
 6538+    if ( split_info[y].end_input_y > last_input_y )
 6539+      split_info[y].end_input_y = last_input_y;
 6540+}
 6541+
 6542+
 6543+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 };
 6544+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 };
 6545+
 6546+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 )
 6547+{
 6548+  // set filter
 6549+  if (filter == 0)
 6550+  {
 6551+    filter = STBIR_DEFAULT_FILTER_DOWNSAMPLE; // default to downsample
 6552+    if (scale_info->scale >= ( 1.0f - stbir__small_float ) )
 6553+    {
 6554+      if ( (scale_info->scale <= ( 1.0f + stbir__small_float ) ) && ( STBIR_CEILF(scale_info->pixel_shift) == scale_info->pixel_shift ) )
 6555+        filter = STBIR_FILTER_POINT_SAMPLE;
 6556+      else
 6557+        filter = STBIR_DEFAULT_FILTER_UPSAMPLE;
 6558+    }
 6559+  }
 6560+  samp->filter_enum = filter;
 6561+
 6562+  STBIR_ASSERT(samp->filter_enum != 0);
 6563+  STBIR_ASSERT((unsigned)samp->filter_enum < STBIR_FILTER_OTHER);
 6564+  samp->filter_kernel = stbir__builtin_kernels[ filter ];
 6565+  samp->filter_support = stbir__builtin_supports[ filter ];
 6566+
 6567+  if ( kernel && support )
 6568+  {
 6569+    samp->filter_kernel = kernel;
 6570+    samp->filter_support = support;
 6571+    samp->filter_enum = STBIR_FILTER_OTHER;
 6572+  }
 6573+
 6574+  samp->edge = edge;
 6575+  samp->filter_pixel_width  = stbir__get_filter_pixel_width (samp->filter_support, scale_info->scale, user_data );
 6576+  // Gather is always better, but in extreme downsamples, you have to most or all of the data in memory
 6577+  //    For horizontal, we always have all the pixels, so we always use gather here (always_gather==1).
 6578+  //    For vertical, we use gather if scaling up (which means we will have samp->filter_pixel_width
 6579+  //    scanlines in memory at once).
 6580+  samp->is_gather = 0;
 6581+  if ( scale_info->scale >= ( 1.0f - stbir__small_float ) )
 6582+    samp->is_gather = 1;
 6583+  else if ( ( always_gather ) || ( samp->filter_pixel_width <= STBIR_FORCE_GATHER_FILTER_SCANLINES_AMOUNT ) )
 6584+    samp->is_gather = 2;
 6585+
 6586+  // pre calculate stuff based on the above
 6587+  samp->coefficient_width = stbir__get_coefficient_width(samp, samp->is_gather, user_data);
 6588+
 6589+  // filter_pixel_width is the conservative size in pixels of input that affect an output pixel.
 6590+  //   In rare cases (only with 2 pix to 1 pix with the default filters), it's possible that the 
 6591+  //   filter will extend before or after the scanline beyond just one extra entire copy of the 
 6592+  //   scanline (we would hit the edge twice). We don't let you do that, so we clamp the total 
 6593+  //   width to 3x the total of input pixel (once for the scanline, once for the left side 
 6594+  //   overhang, and once for the right side). We only do this for edge mode, since the other 
 6595+  //   modes can just re-edge clamp back in again.
 6596+  if ( edge == STBIR_EDGE_WRAP )
 6597+    if ( samp->filter_pixel_width > ( scale_info->input_full_size * 3 ) )
 6598+      samp->filter_pixel_width = scale_info->input_full_size * 3;
 6599+
 6600+  // This is how much to expand buffers to account for filters seeking outside
 6601+  // the image boundaries.
 6602+  samp->filter_pixel_margin = samp->filter_pixel_width / 2;
 6603+  
 6604+  // filter_pixel_margin is the amount that this filter can overhang on just one side of either 
 6605+  //   end of the scanline (left or the right). Since we only allow you to overhang 1 scanline's 
 6606+  //   worth of pixels, we clamp this one side of overhang to the input scanline size. Again, 
 6607+  //   this clamping only happens in rare cases with the default filters (2 pix to 1 pix). 
 6608+  if ( edge == STBIR_EDGE_WRAP )
 6609+    if ( samp->filter_pixel_margin > scale_info->input_full_size )
 6610+      samp->filter_pixel_margin = scale_info->input_full_size;
 6611+
 6612+  samp->num_contributors = stbir__get_contributors(samp, samp->is_gather);
 6613+
 6614+  samp->contributors_size = samp->num_contributors * sizeof(stbir__contributors);
 6615+  samp->coefficients_size = samp->num_contributors * samp->coefficient_width * sizeof(float) + sizeof(float)*STBIR_INPUT_CALLBACK_PADDING; // extra sizeof(float) is padding
 6616+
 6617+  samp->gather_prescatter_contributors = 0;
 6618+  samp->gather_prescatter_coefficients = 0;
 6619+  if ( samp->is_gather == 0 )
 6620+  {
 6621+    samp->gather_prescatter_coefficient_width = samp->filter_pixel_width;
 6622+    samp->gather_prescatter_num_contributors  = stbir__get_contributors(samp, 2);
 6623+    samp->gather_prescatter_contributors_size = samp->gather_prescatter_num_contributors * sizeof(stbir__contributors);
 6624+    samp->gather_prescatter_coefficients_size = samp->gather_prescatter_num_contributors * samp->gather_prescatter_coefficient_width * sizeof(float);
 6625+  }
 6626+}
 6627+
 6628+static void stbir__get_conservative_extents( stbir__sampler * samp, stbir__contributors * range, void * user_data )
 6629+{
 6630+  float scale = samp->scale_info.scale;
 6631+  float out_shift = samp->scale_info.pixel_shift;
 6632+  stbir__support_callback * support = samp->filter_support;
 6633+  int input_full_size = samp->scale_info.input_full_size;
 6634+  stbir_edge edge = samp->edge;
 6635+  float inv_scale = samp->scale_info.inv_scale;
 6636+
 6637+  STBIR_ASSERT( samp->is_gather != 0 );
 6638+
 6639+  if ( samp->is_gather == 1 )
 6640+  {
 6641+    int in_first_pixel, in_last_pixel;
 6642+    float out_filter_radius = support(inv_scale, user_data) * scale;
 6643+
 6644+    stbir__calculate_in_pixel_range( &in_first_pixel, &in_last_pixel, 0.5, out_filter_radius, inv_scale, out_shift, input_full_size, edge );
 6645+    range->n0 = in_first_pixel;
 6646+    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 );
 6647+    range->n1 = in_last_pixel;
 6648+  }
 6649+  else if ( samp->is_gather == 2 ) // downsample gather, refine
 6650+  {
 6651+    float in_pixels_radius = support(scale, user_data) * inv_scale;
 6652+    int filter_pixel_margin = samp->filter_pixel_margin;
 6653+    int output_sub_size = samp->scale_info.output_sub_size;
 6654+    int input_end;
 6655+    int n;
 6656+    int in_first_pixel, in_last_pixel;
 6657+
 6658+    // get a conservative area of the input range
 6659+    stbir__calculate_in_pixel_range( &in_first_pixel, &in_last_pixel, 0, 0, inv_scale, out_shift, input_full_size, edge );
 6660+    range->n0 = in_first_pixel;
 6661+    stbir__calculate_in_pixel_range( &in_first_pixel, &in_last_pixel, (float)output_sub_size, 0, inv_scale, out_shift, input_full_size, edge );
 6662+    range->n1 = in_last_pixel;
 6663+
 6664+    // now go through the margin to the start of area to find bottom
 6665+    n = range->n0 + 1;
 6666+    input_end = -filter_pixel_margin;
 6667+    while( n >= input_end )
 6668+    {
 6669+      int out_first_pixel, out_last_pixel;
 6670+      stbir__calculate_out_pixel_range( &out_first_pixel, &out_last_pixel, ((float)n)+0.5f, in_pixels_radius, scale, out_shift, output_sub_size );
 6671+      if ( out_first_pixel > out_last_pixel )
 6672+        break;
 6673+
 6674+      if ( ( out_first_pixel < output_sub_size ) || ( out_last_pixel >= 0 ) )
 6675+        range->n0 = n;
 6676+      --n;
 6677+    }
 6678+
 6679+    // now go through the end of the area through the margin to find top
 6680+    n = range->n1 - 1;
 6681+    input_end = n + 1 + filter_pixel_margin;
 6682+    while( n <= input_end )
 6683+    {
 6684+      int out_first_pixel, out_last_pixel;
 6685+      stbir__calculate_out_pixel_range( &out_first_pixel, &out_last_pixel, ((float)n)+0.5f, in_pixels_radius, scale, out_shift, output_sub_size );
 6686+      if ( out_first_pixel > out_last_pixel )
 6687+        break;
 6688+      if ( ( out_first_pixel < output_sub_size ) || ( out_last_pixel >= 0 ) )
 6689+        range->n1 = n;
 6690+      ++n;
 6691+    }
 6692+  }
 6693+
 6694+  if ( samp->edge == STBIR_EDGE_WRAP )
 6695+  {
 6696+    // 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
 6697+    if ( ( range->n0 > 0 ) && ( range->n1 >= input_full_size ) )
 6698+    {
 6699+      int marg = range->n1 - input_full_size + 1;
 6700+      if ( ( marg + STBIR__MERGE_RUNS_PIXEL_THRESHOLD ) >= range->n0 )
 6701+        range->n0 = 0;
 6702+    }
 6703+    if ( ( range->n0 < 0 ) && ( range->n1 < (input_full_size-1) ) )
 6704+    {
 6705+      int marg = -range->n0;
 6706+      if ( ( input_full_size - marg - STBIR__MERGE_RUNS_PIXEL_THRESHOLD - 1 ) <= range->n1 )
 6707+        range->n1 = input_full_size - 1;
 6708+    }
 6709+  }
 6710+  else
 6711+  {
 6712+    // for non-edge-wrap modes, we never read over the edge, so clamp
 6713+    if ( range->n0 < 0 )
 6714+      range->n0 = 0;
 6715+    if ( range->n1 >= input_full_size )
 6716+      range->n1 = input_full_size - 1;
 6717+  }
 6718+}
 6719+
 6720+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 )
 6721+{
 6722+  int i, cur;
 6723+  int left = output_height;
 6724+
 6725+  cur = 0;
 6726+  for( i = 0 ; i < splits ; i++ )
 6727+  {
 6728+    int each;
 6729+
 6730+    split_info[i].start_output_y = cur;
 6731+    each = left / ( splits - i );
 6732+    split_info[i].end_output_y = cur + each;
 6733+
 6734+    // ok, when we are gathering, we need to make sure we are starting on a y offset that doesn't have
 6735+    //   a "special" set of coefficients. Basically, with exactly the right filter at exactly the right
 6736+    //   resize at exactly the right phase, some of the coefficents can be zero. When they are zero, we
 6737+    //   don't process them at all.  But this leads to a tricky thing with the thread splits, where we
 6738+    //   might have a set of two coeffs like this for example: (4,4) and (3,6).  The 4,4 means there was
 6739+    //   just one single coeff because things worked out perfectly (normally, they all have 4 coeffs
 6740+    //   like the range 3,6.  The problem is that if we start right on the (4,4) on a brand new thread,
 6741+    //   then when we get to (3,6), we don't have the "3" sample in memory (because we didn't load
 6742+    //   it on the initial (4,4) range because it didn't have a 3 (we only add new samples that are 
 6743+    //   larger than our existing samples - it's just how the eviction works). So, our solution here
 6744+    //   is pretty simple, if we start right on a range that has samples that start earlier, then we 
 6745+    //   simply bump up our previous thread split range to include it, and then start this threads
 6746+    //   range with the smaller sample. It just moves one scanline from one thread split to another,
 6747+    //   so that we end with the unusual one, instead of start with it. To do this, we check 2-4 
 6748+    //   sample at each thread split start and then occassionally move them.
 6749+    
 6750+    if ( ( is_gather ) && ( i ) )
 6751+    {
 6752+      stbir__contributors * small_contribs;
 6753+      int j, smallest, stop, start_n0;
 6754+      stbir__contributors * split_contribs = contribs + cur;
 6755+
 6756+      // scan for a max of 3x the filter width or until the next thread split
 6757+      stop = vertical_pixel_margin * 3;
 6758+      if ( each < stop )
 6759+        stop = each;
 6760+
 6761+      // loops a few times before early out
 6762+      smallest = 0;
 6763+      small_contribs = split_contribs;
 6764+      start_n0 = small_contribs->n0;
 6765+      for( j = 1 ; j <= stop ; j++ )
 6766+      {
 6767+        ++split_contribs;
 6768+        if ( split_contribs->n0 > start_n0 )
 6769+          break;
 6770+        if ( split_contribs->n0 < small_contribs->n0 )
 6771+        {
 6772+          small_contribs = split_contribs;
 6773+          smallest = j;
 6774+        }
 6775+      }
 6776+
 6777+      split_info[i-1].end_output_y += smallest;
 6778+      split_info[i].start_output_y += smallest;
 6779+    }
 6780+
 6781+    cur += each;
 6782+    left -= each;
 6783+
 6784+    // scatter range (updated to minimum as you run it)
 6785+    split_info[i].start_input_y = -vertical_pixel_margin;
 6786+    split_info[i].end_input_y = input_full_height + vertical_pixel_margin;
 6787+  }
 6788+}
 6789+
 6790+static void stbir__free_internal_mem( stbir__info *info )
 6791+{
 6792+  #define STBIR__FREE_AND_CLEAR( ptr ) { if ( ptr ) { void * p = (ptr); (ptr) = 0; STBIR_FREE( p, info->user_data); } }
 6793+
 6794+  if ( info )
 6795+  {
 6796+  #ifndef STBIR__SEPARATE_ALLOCATIONS
 6797+    STBIR__FREE_AND_CLEAR( info->alloced_mem );
 6798+  #else
 6799+    int i,j;
 6800+
 6801+    if ( ( info->vertical.gather_prescatter_contributors ) && ( (void*)info->vertical.gather_prescatter_contributors != (void*)info->split_info[0].decode_buffer ) )
 6802+    {
 6803+      STBIR__FREE_AND_CLEAR( info->vertical.gather_prescatter_coefficients );
 6804+      STBIR__FREE_AND_CLEAR( info->vertical.gather_prescatter_contributors );
 6805+    }
 6806+    for( i = 0 ; i < info->splits ; i++ )
 6807+    {
 6808+      for( j = 0 ; j < info->alloc_ring_buffer_num_entries ; j++ )
 6809+      {
 6810+        #ifdef STBIR_SIMD8
 6811+        if ( info->effective_channels == 3 )
 6812+          --info->split_info[i].ring_buffers[j]; // avx in 3 channel mode needs one float at the start of the buffer
 6813+        #endif
 6814+        STBIR__FREE_AND_CLEAR( info->split_info[i].ring_buffers[j] );
 6815+      }
 6816+
 6817+      #ifdef STBIR_SIMD8
 6818+      if ( info->effective_channels == 3 )
 6819+        --info->split_info[i].decode_buffer; // avx in 3 channel mode needs one float at the start of the buffer
 6820+      #endif
 6821+      STBIR__FREE_AND_CLEAR( info->split_info[i].decode_buffer );
 6822+      STBIR__FREE_AND_CLEAR( info->split_info[i].ring_buffers );
 6823+      STBIR__FREE_AND_CLEAR( info->split_info[i].vertical_buffer );
 6824+    }
 6825+    STBIR__FREE_AND_CLEAR( info->split_info );
 6826+    if ( info->vertical.coefficients != info->horizontal.coefficients )
 6827+    {
 6828+      STBIR__FREE_AND_CLEAR( info->vertical.coefficients );
 6829+      STBIR__FREE_AND_CLEAR( info->vertical.contributors );
 6830+    }
 6831+    STBIR__FREE_AND_CLEAR( info->horizontal.coefficients );
 6832+    STBIR__FREE_AND_CLEAR( info->horizontal.contributors );
 6833+    STBIR__FREE_AND_CLEAR( info->alloced_mem );
 6834+    STBIR_FREE( info, info->user_data );
 6835+  #endif
 6836+  }
 6837+
 6838+  #undef STBIR__FREE_AND_CLEAR
 6839+}
 6840+
 6841+static int stbir__get_max_split( int splits, int height )
 6842+{
 6843+  int i;
 6844+  int max = 0;
 6845+
 6846+  for( i = 0 ; i < splits ; i++ )
 6847+  {
 6848+    int each = height / ( splits - i );
 6849+    if ( each > max )
 6850+      max = each;
 6851+    height -= each;
 6852+  }
 6853+  return max;
 6854+}
 6855+
 6856+static stbir__horizontal_gather_channels_func ** stbir__horizontal_gather_n_coeffs_funcs[8] =
 6857+{
 6858+  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
 6859+};
 6860+
 6861+static stbir__horizontal_gather_channels_func ** stbir__horizontal_gather_channels_funcs[8] =
 6862+{
 6863+  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
 6864+};
 6865+
 6866+// 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
 6867+#define STBIR_RESIZE_CLASSIFICATIONS 8
 6868+
 6869+static float stbir__compute_weights[5][STBIR_RESIZE_CLASSIFICATIONS][4]=  // 5 = 0=1chan, 1=2chan, 2=3chan, 3=4chan, 4=7chan
 6870+{
 6871+  {
 6872+    { 1.00000f, 1.00000f, 0.31250f, 1.00000f },
 6873+    { 0.56250f, 0.59375f, 0.00000f, 0.96875f },
 6874+    { 1.00000f, 0.06250f, 0.00000f, 1.00000f },
 6875+    { 0.00000f, 0.09375f, 1.00000f, 1.00000f },
 6876+    { 1.00000f, 1.00000f, 1.00000f, 1.00000f },
 6877+    { 0.03125f, 0.12500f, 1.00000f, 1.00000f },
 6878+    { 0.06250f, 0.12500f, 0.00000f, 1.00000f },
 6879+    { 0.00000f, 1.00000f, 0.00000f, 0.03125f },
 6880+  }, {
 6881+    { 0.00000f, 0.84375f, 0.00000f, 0.03125f },
 6882+    { 0.09375f, 0.93750f, 0.00000f, 0.78125f },
 6883+    { 0.87500f, 0.21875f, 0.00000f, 0.96875f },
 6884+    { 0.09375f, 0.09375f, 1.00000f, 1.00000f },
 6885+    { 1.00000f, 1.00000f, 1.00000f, 1.00000f },
 6886+    { 0.03125f, 0.12500f, 1.00000f, 1.00000f },
 6887+    { 0.06250f, 0.12500f, 0.00000f, 1.00000f },
 6888+    { 0.00000f, 1.00000f, 0.00000f, 0.53125f },
 6889+  }, {
 6890+    { 0.00000f, 0.53125f, 0.00000f, 0.03125f },
 6891+    { 0.06250f, 0.96875f, 0.00000f, 0.53125f },
 6892+    { 0.87500f, 0.18750f, 0.00000f, 0.93750f },
 6893+    { 0.00000f, 0.09375f, 1.00000f, 1.00000f },
 6894+    { 1.00000f, 1.00000f, 1.00000f, 1.00000f },
 6895+    { 0.03125f, 0.12500f, 1.00000f, 1.00000f },
 6896+    { 0.06250f, 0.12500f, 0.00000f, 1.00000f },
 6897+    { 0.00000f, 1.00000f, 0.00000f, 0.56250f },
 6898+  }, {
 6899+    { 0.00000f, 0.50000f, 0.00000f, 0.71875f },
 6900+    { 0.06250f, 0.84375f, 0.00000f, 0.87500f },
 6901+    { 1.00000f, 0.50000f, 0.50000f, 0.96875f },
 6902+    { 1.00000f, 0.09375f, 0.31250f, 0.50000f },
 6903+    { 1.00000f, 1.00000f, 1.00000f, 1.00000f },
 6904+    { 1.00000f, 0.03125f, 0.03125f, 0.53125f },
 6905+    { 0.18750f, 0.12500f, 0.00000f, 1.00000f },
 6906+    { 0.00000f, 1.00000f, 0.03125f, 0.18750f },
 6907+  }, {
 6908+    { 0.00000f, 0.59375f, 0.00000f, 0.96875f },
 6909+    { 0.06250f, 0.81250f, 0.06250f, 0.59375f },
 6910+    { 0.75000f, 0.43750f, 0.12500f, 0.96875f },
 6911+    { 0.87500f, 0.06250f, 0.18750f, 0.43750f },
 6912+    { 1.00000f, 1.00000f, 1.00000f, 1.00000f },
 6913+    { 0.15625f, 0.12500f, 1.00000f, 1.00000f },
 6914+    { 0.06250f, 0.12500f, 0.00000f, 1.00000f },
 6915+    { 0.00000f, 1.00000f, 0.03125f, 0.34375f },
 6916+  }
 6917+};
 6918+
 6919+// structure that allow us to query and override info for training the costs
 6920+typedef struct STBIR__V_FIRST_INFO
 6921+{
 6922+  double v_cost, h_cost;
 6923+  int control_v_first; // 0 = no control, 1 = force hori, 2 = force vert
 6924+  int v_first;
 6925+  int v_resize_classification;
 6926+  int is_gather;
 6927+} STBIR__V_FIRST_INFO;
 6928+
 6929+#ifdef STBIR__V_FIRST_INFO_BUFFER
 6930+static STBIR__V_FIRST_INFO STBIR__V_FIRST_INFO_BUFFER = {0};
 6931+#define STBIR__V_FIRST_INFO_POINTER &STBIR__V_FIRST_INFO_BUFFER
 6932+#else
 6933+#define STBIR__V_FIRST_INFO_POINTER 0
 6934+#endif
 6935+
 6936+// Figure out whether to scale along the horizontal or vertical first.
 6937+//   This only *super* important when you are scaling by a massively
 6938+//   different amount in the vertical vs the horizontal (for example, if
 6939+//   you are scaling by 2x in the width, and 0.5x in the height, then you
 6940+//   want to do the vertical scale first, because it's around 3x faster
 6941+//   in that order.
 6942+//
 6943+//   In more normal circumstances, this makes a 20-40% differences, so
 6944+//     it's good to get right, but not critical. The normal way that you
 6945+//     decide which direction goes first is just figuring out which
 6946+//     direction does more multiplies. But with modern CPUs with their
 6947+//     fancy caches and SIMD and high IPC abilities, so there's just a lot
 6948+//     more that goes into it.
 6949+//
 6950+//   My handwavy sort of solution is to have an app that does a whole
 6951+//     bunch of timing for both vertical and horizontal first modes,
 6952+//     and then another app that can read lots of these timing files
 6953+//     and try to search for the best weights to use. Dotimings.c
 6954+//     is the app that does a bunch of timings, and vf_train.c is the
 6955+//     app that solves for the best weights (and shows how well it
 6956+//     does currently).
 6957+
 6958+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 )
 6959+{
 6960+  double v_cost, h_cost;
 6961+  float * weights;
 6962+  int vertical_first;
 6963+  int v_classification;
 6964+
 6965+  // categorize the resize into buckets
 6966+  if ( ( vertical_output_size <= 4 ) || ( horizontal_output_size <= 4 ) )
 6967+    v_classification = ( vertical_output_size < horizontal_output_size ) ? 6 : 7;
 6968+  else if ( vertical_scale <= 1.0f )
 6969+    v_classification = ( is_gather ) ? 1 : 0;
 6970+  else if ( vertical_scale <= 2.0f)
 6971+    v_classification = 2;
 6972+  else if ( vertical_scale <= 3.0f)
 6973+    v_classification = 3;
 6974+  else if ( vertical_scale <= 4.0f)
 6975+    v_classification = 5;
 6976+  else
 6977+    v_classification = 6;
 6978+
 6979+  // use the right weights
 6980+  weights = weights_table[ v_classification ];
 6981+
 6982+  // 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
 6983+  h_cost = (float)horizontal_filter_pixel_width * weights[0] + horizontal_scale * (float)vertical_filter_pixel_width * weights[1];
 6984+  v_cost = (float)vertical_filter_pixel_width  * weights[2] + vertical_scale * (float)horizontal_filter_pixel_width * weights[3];
 6985+
 6986+  // use computation estimate to decide vertical first or not
 6987+  vertical_first = ( v_cost <= h_cost ) ? 1 : 0;
 6988+
 6989+  // save these, if requested
 6990+  if ( info )
 6991+  {
 6992+    info->h_cost = h_cost;
 6993+    info->v_cost = v_cost;
 6994+    info->v_resize_classification = v_classification;
 6995+    info->v_first = vertical_first;
 6996+    info->is_gather = is_gather;
 6997+  }
 6998+
 6999+  // and this allows us to override everything for testing (see dotiming.c)
 7000+  if ( ( info ) && ( info->control_v_first ) )
 7001+    vertical_first = ( info->control_v_first == 2 ) ? 1 : 0;
 7002+
 7003+  return vertical_first;
 7004+}
 7005+
 7006+// layout lookups - must match stbir_internal_pixel_layout
 7007+static unsigned char stbir__pixel_channels[] = {
 7008+  1,2,3,3,4,   // 1ch, 2ch, rgb, bgr, 4ch
 7009+  4,4,4,4,2,2, // RGBA,BGRA,ARGB,ABGR,RA,AR
 7010+  4,4,4,4,2,2, // RGBA_PM,BGRA_PM,ARGB_PM,ABGR_PM,RA_PM,AR_PM
 7011+};
 7012+
 7013+// the internal pixel layout enums are in a different order, so we can easily do range comparisons of types
 7014+//   the public pixel layout is ordered in a way that if you cast num_channels (1-4) to the enum, you get something sensible
 7015+static stbir_internal_pixel_layout stbir__pixel_layout_convert_public_to_internal[] = {
 7016+  STBIRI_BGR, STBIRI_1CHANNEL, STBIRI_2CHANNEL, STBIRI_RGB, STBIRI_RGBA,
 7017+  STBIRI_4CHANNEL, STBIRI_BGRA, STBIRI_ARGB, STBIRI_ABGR, STBIRI_RA, STBIRI_AR,
 7018+  STBIRI_RGBA_PM, STBIRI_BGRA_PM, STBIRI_ARGB_PM, STBIRI_ABGR_PM, STBIRI_RA_PM, STBIRI_AR_PM,
 7019+};
 7020+
 7021+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 )
 7022+{
 7023+  static char stbir_channel_count_index[8]={ 9,0,1,2, 3,9,9,4 };
 7024+
 7025+  stbir__info * info = 0;
 7026+  void * alloced = 0;
 7027+  size_t alloced_total = 0;
 7028+  int vertical_first;
 7029+  size_t decode_buffer_size, ring_buffer_length_bytes, ring_buffer_size, vertical_buffer_size;
 7030+  int alloc_ring_buffer_num_entries;
 7031+
 7032+  int alpha_weighting_type = 0; // 0=none, 1=simple, 2=fancy
 7033+  int conservative_split_output_size = stbir__get_max_split( splits, vertical->scale_info.output_sub_size );
 7034+  stbir_internal_pixel_layout input_pixel_layout = stbir__pixel_layout_convert_public_to_internal[ input_pixel_layout_public ];
 7035+  stbir_internal_pixel_layout output_pixel_layout = stbir__pixel_layout_convert_public_to_internal[ output_pixel_layout_public ];
 7036+  int channels = stbir__pixel_channels[ input_pixel_layout ];
 7037+  int effective_channels = channels;
 7038+
 7039+  // first figure out what type of alpha weighting to use (if any)
 7040+  if ( ( horizontal->filter_enum != STBIR_FILTER_POINT_SAMPLE ) || ( vertical->filter_enum != STBIR_FILTER_POINT_SAMPLE ) ) // no alpha weighting on point sampling
 7041+  {
 7042+    if ( ( input_pixel_layout >= STBIRI_RGBA ) && ( input_pixel_layout <= STBIRI_AR ) && ( output_pixel_layout >= STBIRI_RGBA ) && ( output_pixel_layout <= STBIRI_AR ) )
 7043+    {
 7044+      if ( fast_alpha )
 7045+      {
 7046+        alpha_weighting_type = 4;
 7047+      }
 7048+      else
 7049+      {
 7050+        static int fancy_alpha_effective_cnts[6] = { 7, 7, 7, 7, 3, 3 };
 7051+        alpha_weighting_type = 2;
 7052+        effective_channels = fancy_alpha_effective_cnts[ input_pixel_layout - STBIRI_RGBA ];
 7053+      }
 7054+    }
 7055+    else if ( ( input_pixel_layout >= STBIRI_RGBA_PM ) && ( input_pixel_layout <= STBIRI_AR_PM ) && ( output_pixel_layout >= STBIRI_RGBA ) && ( output_pixel_layout <= STBIRI_AR ) )
 7056+    {
 7057+      // input premult, output non-premult
 7058+      alpha_weighting_type = 3;
 7059+    }
 7060+    else if ( ( input_pixel_layout >= STBIRI_RGBA ) && ( input_pixel_layout <= STBIRI_AR ) && ( output_pixel_layout >= STBIRI_RGBA_PM ) && ( output_pixel_layout <= STBIRI_AR_PM ) )
 7061+    {
 7062+      // input non-premult, output premult
 7063+      alpha_weighting_type = 1;
 7064+    }
 7065+  }
 7066+
 7067+  // channel in and out count must match currently
 7068+  if ( channels != stbir__pixel_channels[ output_pixel_layout ] )
 7069+    return 0;
 7070+
 7071+  // get vertical first
 7072+  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 );
 7073+
 7074+  // sometimes read one float off in some of the unrolled loops (with a weight of zero coeff, so it doesn't have an effect)
 7075+  //   we use a few extra floats instead of just 1, so that input callback buffer can overlap with the decode buffer without
 7076+  //   the conversion routines overwriting the callback input data.
 7077+  decode_buffer_size = ( conservative->n1 - conservative->n0 + 1 ) * effective_channels * sizeof(float) + sizeof(float)*STBIR_INPUT_CALLBACK_PADDING; // extra floats for input callback stagger
 7078+
 7079+#if defined( STBIR__SEPARATE_ALLOCATIONS ) && defined(STBIR_SIMD8)
 7080+  if ( effective_channels == 3 )
 7081+    decode_buffer_size += sizeof(float); // avx in 3 channel mode needs one float at the start of the buffer (only with separate allocations)
 7082+#endif
 7083+
 7084+  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
 7085+
 7086+  // if we do vertical first, the ring buffer holds a whole decoded line
 7087+  if ( vertical_first )
 7088+    ring_buffer_length_bytes = ( decode_buffer_size + 15 ) & ~15;
 7089+
 7090+  if ( ( ring_buffer_length_bytes & 4095 ) == 0 ) ring_buffer_length_bytes += 64*3; // avoid 4k alias
 7091+
 7092+  // One extra entry because floating point precision problems sometimes cause an extra to be necessary.
 7093+  alloc_ring_buffer_num_entries = vertical->filter_pixel_width + 1;
 7094+
 7095+  // we never need more ring buffer entries than the scanlines we're outputting when in scatter mode
 7096+  if ( ( !vertical->is_gather ) && ( alloc_ring_buffer_num_entries > conservative_split_output_size ) )
 7097+    alloc_ring_buffer_num_entries = conservative_split_output_size;
 7098+
 7099+  ring_buffer_size = (size_t)alloc_ring_buffer_num_entries * (size_t)ring_buffer_length_bytes;
 7100+
 7101+  // The vertical buffer is used differently, depending on whether we are scattering
 7102+  //   the vertical scanlines, or gathering them.
 7103+  //   If scattering, it's used at the temp buffer to accumulate each output.
 7104+  //   If gathering, it's just the output buffer.
 7105+  vertical_buffer_size = (size_t)horizontal->scale_info.output_sub_size * (size_t)effective_channels * sizeof(float) + sizeof(float);  // extra float for padding
 7106+
 7107+  // we make two passes through this loop, 1st to add everything up, 2nd to allocate and init
 7108+  for(;;)
 7109+  {
 7110+    int i;
 7111+    void * advance_mem = alloced;
 7112+    int copy_horizontal = 0;
 7113+    stbir__sampler * possibly_use_horizontal_for_pivot = 0;
 7114+
 7115+#ifdef STBIR__SEPARATE_ALLOCATIONS
 7116+    #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; }
 7117+#else
 7118+    #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));
 7119+#endif
 7120+
 7121+    STBIR__NEXT_PTR( info, sizeof( stbir__info ), stbir__info );
 7122+
 7123+    STBIR__NEXT_PTR( info->split_info, sizeof( stbir__per_split_info ) * splits, stbir__per_split_info );
 7124+
 7125+    if ( info )
 7126+    {
 7127+      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 };
 7128+      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 };
 7129+      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 };
 7130+      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 };
 7131+
 7132+      // initialize info fields
 7133+      info->alloced_mem = alloced;
 7134+      info->alloced_total = alloced_total;
 7135+
 7136+      info->channels = channels;
 7137+      info->effective_channels = effective_channels;
 7138+
 7139+      info->offset_x = new_x;
 7140+      info->offset_y = new_y;
 7141+      info->alloc_ring_buffer_num_entries = (int)alloc_ring_buffer_num_entries;
 7142+      info->ring_buffer_num_entries = 0;
 7143+      info->ring_buffer_length_bytes = (int)ring_buffer_length_bytes;
 7144+      info->splits = splits;
 7145+      info->vertical_first = vertical_first;
 7146+
 7147+      info->input_pixel_layout_internal = input_pixel_layout;
 7148+      info->output_pixel_layout_internal = output_pixel_layout;
 7149+
 7150+      // setup alpha weight functions
 7151+      info->alpha_weight = 0;
 7152+      info->alpha_unweight = 0;
 7153+
 7154+      // handle alpha weighting functions and overrides
 7155+      if ( alpha_weighting_type == 2 )
 7156+      {
 7157+        // high quality alpha multiplying on the way in, dividing on the way out
 7158+        info->alpha_weight = fancy_alpha_weights[ input_pixel_layout - STBIRI_RGBA ];
 7159+        info->alpha_unweight = fancy_alpha_unweights[ output_pixel_layout - STBIRI_RGBA ];
 7160+      }
 7161+      else if ( alpha_weighting_type == 4 )
 7162+      {
 7163+        // fast alpha multiplying on the way in, dividing on the way out
 7164+        info->alpha_weight = simple_alpha_weights[ input_pixel_layout - STBIRI_RGBA ];
 7165+        info->alpha_unweight = simple_alpha_unweights[ output_pixel_layout - STBIRI_RGBA ];
 7166+      }
 7167+      else if ( alpha_weighting_type == 1 )
 7168+      {
 7169+        // fast alpha on the way in, leave in premultiplied form on way out
 7170+        info->alpha_weight = simple_alpha_weights[ input_pixel_layout - STBIRI_RGBA ];
 7171+      }
 7172+      else if ( alpha_weighting_type == 3 )
 7173+      {
 7174+        // incoming is premultiplied, fast alpha dividing on the way out - non-premultiplied output
 7175+        info->alpha_unweight = simple_alpha_unweights[ output_pixel_layout - STBIRI_RGBA ];
 7176+      }
 7177+
 7178+      // handle 3-chan color flipping, using the alpha weight path
 7179+      if ( ( ( input_pixel_layout == STBIRI_RGB ) && ( output_pixel_layout == STBIRI_BGR ) ) ||
 7180+           ( ( input_pixel_layout == STBIRI_BGR ) && ( output_pixel_layout == STBIRI_RGB ) ) )
 7181+      {
 7182+        // do the flipping on the smaller of the two ends
 7183+        if ( horizontal->scale_info.scale < 1.0f )
 7184+          info->alpha_unweight = stbir__simple_flip_3ch;
 7185+        else
 7186+          info->alpha_weight = stbir__simple_flip_3ch;
 7187+      }
 7188+
 7189+    }
 7190+
 7191+    // get all the per-split buffers
 7192+    for( i = 0 ; i < splits ; i++ )
 7193+    {
 7194+      STBIR__NEXT_PTR( info->split_info[i].decode_buffer, decode_buffer_size, float );
 7195+
 7196+#ifdef STBIR__SEPARATE_ALLOCATIONS
 7197+
 7198+      #ifdef STBIR_SIMD8
 7199+      if ( ( info ) && ( effective_channels == 3 ) )
 7200+        ++info->split_info[i].decode_buffer; // avx in 3 channel mode needs one float at the start of the buffer
 7201+      #endif
 7202+
 7203+      STBIR__NEXT_PTR( info->split_info[i].ring_buffers, alloc_ring_buffer_num_entries * sizeof(float*), float* );
 7204+      {
 7205+        int j;
 7206+        for( j = 0 ; j < alloc_ring_buffer_num_entries ; j++ )
 7207+        {
 7208+          STBIR__NEXT_PTR( info->split_info[i].ring_buffers[j], ring_buffer_length_bytes, float );
 7209+          #ifdef STBIR_SIMD8
 7210+          if ( ( info ) && ( effective_channels == 3 ) )
 7211+            ++info->split_info[i].ring_buffers[j]; // avx in 3 channel mode needs one float at the start of the buffer
 7212+          #endif
 7213+        }
 7214+      }
 7215+#else
 7216+      STBIR__NEXT_PTR( info->split_info[i].ring_buffer, ring_buffer_size, float );
 7217+#endif
 7218+      STBIR__NEXT_PTR( info->split_info[i].vertical_buffer, vertical_buffer_size, float );
 7219+    }
 7220+
 7221+    // alloc memory for to-be-pivoted coeffs (if necessary)
 7222+    if ( vertical->is_gather == 0 )
 7223+    {
 7224+      size_t both;
 7225+      size_t temp_mem_amt;
 7226+
 7227+      // when in vertical scatter mode, we first build the coefficients in gather mode, and then pivot after,
 7228+      //   that means we need two buffers, so we try to use the decode buffer and ring buffer for this. if that
 7229+      //   is too small, we just allocate extra memory to use as this temp.
 7230+
 7231+      both = (size_t)vertical->gather_prescatter_contributors_size + (size_t)vertical->gather_prescatter_coefficients_size;
 7232+
 7233+#ifdef STBIR__SEPARATE_ALLOCATIONS
 7234+      temp_mem_amt = decode_buffer_size;
 7235+
 7236+      #ifdef STBIR_SIMD8
 7237+      if ( effective_channels == 3 )
 7238+        --temp_mem_amt; // avx in 3 channel mode needs one float at the start of the buffer
 7239+      #endif
 7240+#else
 7241+      temp_mem_amt = (size_t)( decode_buffer_size + ring_buffer_size + vertical_buffer_size ) * (size_t)splits;
 7242+#endif
 7243+      if ( temp_mem_amt >= both )
 7244+      {
 7245+        if ( info )
 7246+        {
 7247+          vertical->gather_prescatter_contributors = (stbir__contributors*)info->split_info[0].decode_buffer;
 7248+          vertical->gather_prescatter_coefficients = (float*) ( ( (char*)info->split_info[0].decode_buffer ) + vertical->gather_prescatter_contributors_size );
 7249+        }
 7250+      }
 7251+      else
 7252+      {
 7253+        // ring+decode memory is too small, so allocate temp memory
 7254+        STBIR__NEXT_PTR( vertical->gather_prescatter_contributors, vertical->gather_prescatter_contributors_size, stbir__contributors );
 7255+        STBIR__NEXT_PTR( vertical->gather_prescatter_coefficients, vertical->gather_prescatter_coefficients_size, float );
 7256+      }
 7257+    }
 7258+
 7259+    STBIR__NEXT_PTR( horizontal->contributors, horizontal->contributors_size, stbir__contributors );
 7260+    STBIR__NEXT_PTR( horizontal->coefficients, horizontal->coefficients_size, float );
 7261+
 7262+    // are the two filters identical?? (happens a lot with mipmap generation)
 7263+    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 ) )
 7264+    {
 7265+      float diff_scale = horizontal->scale_info.scale - vertical->scale_info.scale;
 7266+      float diff_shift = horizontal->scale_info.pixel_shift - vertical->scale_info.pixel_shift;
 7267+      if ( diff_scale < 0.0f ) diff_scale = -diff_scale;
 7268+      if ( diff_shift < 0.0f ) diff_shift = -diff_shift;
 7269+      if ( ( diff_scale <= stbir__small_float ) && ( diff_shift <= stbir__small_float ) )
 7270+      {
 7271+        if ( horizontal->is_gather == vertical->is_gather )
 7272+        {
 7273+          copy_horizontal = 1;
 7274+          goto no_vert_alloc;
 7275+        }
 7276+        // everything matches, but vertical is scatter, horizontal is gather, use horizontal coeffs for vertical pivot coeffs
 7277+        possibly_use_horizontal_for_pivot = horizontal;
 7278+      }
 7279+    }
 7280+
 7281+    STBIR__NEXT_PTR( vertical->contributors, vertical->contributors_size, stbir__contributors );
 7282+    STBIR__NEXT_PTR( vertical->coefficients, vertical->coefficients_size, float );
 7283+
 7284+   no_vert_alloc:
 7285+
 7286+    if ( info )
 7287+    {
 7288+      STBIR_PROFILE_BUILD_START( horizontal );
 7289+
 7290+      stbir__calculate_filters( horizontal, 0, user_data STBIR_ONLY_PROFILE_BUILD_SET_INFO );
 7291+
 7292+      // setup the horizontal gather functions
 7293+      // start with defaulting to the n_coeffs functions (specialized on channels and remnant leftover)
 7294+      info->horizontal_gather_channels = stbir__horizontal_gather_n_coeffs_funcs[ effective_channels ][ horizontal->extent_info.widest & 3 ];
 7295+      // 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
 7296+      if ( horizontal->extent_info.widest <= 12 )
 7297+        info->horizontal_gather_channels = stbir__horizontal_gather_channels_funcs[ effective_channels ][ horizontal->extent_info.widest - 1 ];
 7298+
 7299+      info->scanline_extents.conservative.n0 = conservative->n0;
 7300+      info->scanline_extents.conservative.n1 = conservative->n1;
 7301+
 7302+      // get exact extents
 7303+      stbir__get_extents( horizontal, &info->scanline_extents );
 7304+
 7305+      // pack the horizontal coeffs
 7306+      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 );
 7307+
 7308+      STBIR_MEMCPY( &info->horizontal, horizontal, sizeof( stbir__sampler ) );
 7309+
 7310+      STBIR_PROFILE_BUILD_END( horizontal );
 7311+
 7312+      if ( copy_horizontal )
 7313+      {
 7314+        STBIR_MEMCPY( &info->vertical, horizontal, sizeof( stbir__sampler ) );
 7315+      }
 7316+      else
 7317+      {
 7318+        STBIR_PROFILE_BUILD_START( vertical );
 7319+
 7320+        stbir__calculate_filters( vertical, possibly_use_horizontal_for_pivot, user_data STBIR_ONLY_PROFILE_BUILD_SET_INFO );
 7321+        STBIR_MEMCPY( &info->vertical, vertical, sizeof( stbir__sampler ) );
 7322+
 7323+        STBIR_PROFILE_BUILD_END( vertical );
 7324+      }
 7325+
 7326+      // setup the vertical split ranges
 7327+      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 );
 7328+
 7329+      // now we know precisely how many entries we need
 7330+      info->ring_buffer_num_entries = info->vertical.extent_info.widest;
 7331+
 7332+      // we never need more ring buffer entries than the scanlines we're outputting
 7333+      if ( ( !info->vertical.is_gather ) && ( info->ring_buffer_num_entries > conservative_split_output_size ) )
 7334+        info->ring_buffer_num_entries = conservative_split_output_size;
 7335+      STBIR_ASSERT( info->ring_buffer_num_entries <= info->alloc_ring_buffer_num_entries );
 7336+    }
 7337+    #undef STBIR__NEXT_PTR
 7338+
 7339+
 7340+    // is this the first time through loop?
 7341+    if ( info == 0 )
 7342+    {
 7343+      alloced_total = ( 15 + (size_t)advance_mem );
 7344+      alloced = STBIR_MALLOC( alloced_total, user_data );
 7345+      if ( alloced == 0 )
 7346+        return 0;
 7347+    }
 7348+    else
 7349+      return info;  // success
 7350+  }
 7351+}
 7352+
 7353+static int stbir__perform_resize( stbir__info const * info, int split_start, int split_count )
 7354+{
 7355+  stbir__per_split_info * split_info = info->split_info + split_start;
 7356+
 7357+  STBIR_PROFILE_CLEAR_EXTRAS();
 7358+
 7359+  STBIR_PROFILE_FIRST_START( looping );
 7360+  if (info->vertical.is_gather)
 7361+    stbir__vertical_gather_loop( info, split_info, split_count );
 7362+  else
 7363+    stbir__vertical_scatter_loop( info, split_info, split_count );
 7364+  STBIR_PROFILE_END( looping );
 7365+
 7366+  return 1;
 7367+}
 7368+
 7369+static void stbir__update_info_from_resize( stbir__info * info, STBIR_RESIZE * resize )
 7370+{
 7371+  static stbir__decode_pixels_func * decode_simple[STBIR_TYPE_HALF_FLOAT-STBIR_TYPE_UINT8_SRGB+1]=
 7372+  {
 7373+    /* 1ch-4ch */ stbir__decode_uint8_srgb, stbir__decode_uint8_srgb, 0, stbir__decode_float_linear, stbir__decode_half_float_linear,
 7374+  };
 7375+
 7376+  static stbir__decode_pixels_func * decode_alphas[STBIRI_AR-STBIRI_RGBA+1][STBIR_TYPE_HALF_FLOAT-STBIR_TYPE_UINT8_SRGB+1]=
 7377+  {
 7378+    { /* RGBA */ stbir__decode_uint8_srgb4_linearalpha,      stbir__decode_uint8_srgb,      0, stbir__decode_float_linear,      stbir__decode_half_float_linear },
 7379+    { /* BGRA */ stbir__decode_uint8_srgb4_linearalpha_BGRA, stbir__decode_uint8_srgb_BGRA, 0, stbir__decode_float_linear_BGRA, stbir__decode_half_float_linear_BGRA },
 7380+    { /* ARGB */ stbir__decode_uint8_srgb4_linearalpha_ARGB, stbir__decode_uint8_srgb_ARGB, 0, stbir__decode_float_linear_ARGB, stbir__decode_half_float_linear_ARGB },
 7381+    { /* ABGR */ stbir__decode_uint8_srgb4_linearalpha_ABGR, stbir__decode_uint8_srgb_ABGR, 0, stbir__decode_float_linear_ABGR, stbir__decode_half_float_linear_ABGR },
 7382+    { /* RA   */ stbir__decode_uint8_srgb2_linearalpha,      stbir__decode_uint8_srgb,      0, stbir__decode_float_linear,      stbir__decode_half_float_linear },
 7383+    { /* AR   */ stbir__decode_uint8_srgb2_linearalpha_AR,   stbir__decode_uint8_srgb_AR,   0, stbir__decode_float_linear_AR,   stbir__decode_half_float_linear_AR },
 7384+  };
 7385+
 7386+  static stbir__decode_pixels_func * decode_simple_scaled_or_not[2][2]=
 7387+  {
 7388+    { stbir__decode_uint8_linear_scaled,  stbir__decode_uint8_linear }, { stbir__decode_uint16_linear_scaled, stbir__decode_uint16_linear },
 7389+  };
 7390+
 7391+  static stbir__decode_pixels_func * decode_alphas_scaled_or_not[STBIRI_AR-STBIRI_RGBA+1][2][2]=
 7392+  {
 7393+    { /* RGBA */ { stbir__decode_uint8_linear_scaled,       stbir__decode_uint8_linear },      { stbir__decode_uint16_linear_scaled,      stbir__decode_uint16_linear } },
 7394+    { /* BGRA */ { stbir__decode_uint8_linear_scaled_BGRA,  stbir__decode_uint8_linear_BGRA }, { stbir__decode_uint16_linear_scaled_BGRA, stbir__decode_uint16_linear_BGRA } },
 7395+    { /* ARGB */ { stbir__decode_uint8_linear_scaled_ARGB,  stbir__decode_uint8_linear_ARGB }, { stbir__decode_uint16_linear_scaled_ARGB, stbir__decode_uint16_linear_ARGB } },
 7396+    { /* ABGR */ { stbir__decode_uint8_linear_scaled_ABGR,  stbir__decode_uint8_linear_ABGR }, { stbir__decode_uint16_linear_scaled_ABGR, stbir__decode_uint16_linear_ABGR } },
 7397+    { /* RA   */ { stbir__decode_uint8_linear_scaled,       stbir__decode_uint8_linear },      { stbir__decode_uint16_linear_scaled,      stbir__decode_uint16_linear } },
 7398+    { /* AR   */ { stbir__decode_uint8_linear_scaled_AR,    stbir__decode_uint8_linear_AR },   { stbir__decode_uint16_linear_scaled_AR,   stbir__decode_uint16_linear_AR } }
 7399+  };
 7400+
 7401+  static stbir__encode_pixels_func * encode_simple[STBIR_TYPE_HALF_FLOAT-STBIR_TYPE_UINT8_SRGB+1]=
 7402+  {
 7403+    /* 1ch-4ch */ stbir__encode_uint8_srgb, stbir__encode_uint8_srgb, 0, stbir__encode_float_linear, stbir__encode_half_float_linear,
 7404+  };
 7405+
 7406+  static stbir__encode_pixels_func * encode_alphas[STBIRI_AR-STBIRI_RGBA+1][STBIR_TYPE_HALF_FLOAT-STBIR_TYPE_UINT8_SRGB+1]=
 7407+  {
 7408+    { /* RGBA */ stbir__encode_uint8_srgb4_linearalpha,      stbir__encode_uint8_srgb,      0, stbir__encode_float_linear,      stbir__encode_half_float_linear },
 7409+    { /* BGRA */ stbir__encode_uint8_srgb4_linearalpha_BGRA, stbir__encode_uint8_srgb_BGRA, 0, stbir__encode_float_linear_BGRA, stbir__encode_half_float_linear_BGRA },
 7410+    { /* ARGB */ stbir__encode_uint8_srgb4_linearalpha_ARGB, stbir__encode_uint8_srgb_ARGB, 0, stbir__encode_float_linear_ARGB, stbir__encode_half_float_linear_ARGB },
 7411+    { /* ABGR */ stbir__encode_uint8_srgb4_linearalpha_ABGR, stbir__encode_uint8_srgb_ABGR, 0, stbir__encode_float_linear_ABGR, stbir__encode_half_float_linear_ABGR },
 7412+    { /* RA   */ stbir__encode_uint8_srgb2_linearalpha,      stbir__encode_uint8_srgb,      0, stbir__encode_float_linear,      stbir__encode_half_float_linear },
 7413+    { /* AR   */ stbir__encode_uint8_srgb2_linearalpha_AR,   stbir__encode_uint8_srgb_AR,   0, stbir__encode_float_linear_AR,   stbir__encode_half_float_linear_AR }
 7414+  };
 7415+
 7416+  static stbir__encode_pixels_func * encode_simple_scaled_or_not[2][2]=
 7417+  {
 7418+    { stbir__encode_uint8_linear_scaled,  stbir__encode_uint8_linear }, { stbir__encode_uint16_linear_scaled, stbir__encode_uint16_linear },
 7419+  };
 7420+
 7421+  static stbir__encode_pixels_func * encode_alphas_scaled_or_not[STBIRI_AR-STBIRI_RGBA+1][2][2]=
 7422+  {
 7423+    { /* RGBA */ { stbir__encode_uint8_linear_scaled,       stbir__encode_uint8_linear },       { stbir__encode_uint16_linear_scaled,      stbir__encode_uint16_linear } },
 7424+    { /* BGRA */ { stbir__encode_uint8_linear_scaled_BGRA,  stbir__encode_uint8_linear_BGRA },  { stbir__encode_uint16_linear_scaled_BGRA, stbir__encode_uint16_linear_BGRA } },
 7425+    { /* ARGB */ { stbir__encode_uint8_linear_scaled_ARGB,  stbir__encode_uint8_linear_ARGB },  { stbir__encode_uint16_linear_scaled_ARGB, stbir__encode_uint16_linear_ARGB } },
 7426+    { /* ABGR */ { stbir__encode_uint8_linear_scaled_ABGR,  stbir__encode_uint8_linear_ABGR },  { stbir__encode_uint16_linear_scaled_ABGR, stbir__encode_uint16_linear_ABGR } },
 7427+    { /* RA   */ { stbir__encode_uint8_linear_scaled,       stbir__encode_uint8_linear },       { stbir__encode_uint16_linear_scaled,      stbir__encode_uint16_linear } },
 7428+    { /* AR   */ { stbir__encode_uint8_linear_scaled_AR,    stbir__encode_uint8_linear_AR },    { stbir__encode_uint16_linear_scaled_AR,   stbir__encode_uint16_linear_AR } }
 7429+  };
 7430+
 7431+  stbir__decode_pixels_func * decode_pixels = 0;
 7432+  stbir__encode_pixels_func * encode_pixels = 0;
 7433+  stbir_datatype input_type, output_type;
 7434+
 7435+  input_type = resize->input_data_type;
 7436+  output_type = resize->output_data_type;
 7437+  info->input_data = resize->input_pixels;
 7438+  info->input_stride_bytes = resize->input_stride_in_bytes;
 7439+  info->output_stride_bytes = resize->output_stride_in_bytes;
 7440+
 7441+  // if we're completely point sampling, then we can turn off SRGB
 7442+  if ( ( info->horizontal.filter_enum == STBIR_FILTER_POINT_SAMPLE ) && ( info->vertical.filter_enum == STBIR_FILTER_POINT_SAMPLE ) )
 7443+  {
 7444+    if ( ( ( input_type  == STBIR_TYPE_UINT8_SRGB ) || ( input_type  == STBIR_TYPE_UINT8_SRGB_ALPHA ) ) &&
 7445+         ( ( output_type == STBIR_TYPE_UINT8_SRGB ) || ( output_type == STBIR_TYPE_UINT8_SRGB_ALPHA ) ) )
 7446+    {
 7447+      input_type = STBIR_TYPE_UINT8;
 7448+      output_type = STBIR_TYPE_UINT8;
 7449+    }
 7450+  }
 7451+
 7452+  // recalc the output and input strides
 7453+  if ( info->input_stride_bytes == 0 )
 7454+    info->input_stride_bytes = info->channels * info->horizontal.scale_info.input_full_size * stbir__type_size[input_type];
 7455+
 7456+  if ( info->output_stride_bytes == 0 )
 7457+    info->output_stride_bytes = info->channels * info->horizontal.scale_info.output_sub_size * stbir__type_size[output_type];
 7458+
 7459+  // calc offset
 7460+  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] );
 7461+
 7462+  info->in_pixels_cb = resize->input_cb;
 7463+  info->user_data = resize->user_data;
 7464+  info->out_pixels_cb = resize->output_cb;
 7465+
 7466+  // setup the input format converters
 7467+  if ( ( input_type == STBIR_TYPE_UINT8 ) || ( input_type == STBIR_TYPE_UINT16 ) )
 7468+  {
 7469+    int non_scaled = 0;
 7470+
 7471+    // 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)
 7472+    if ( ( !info->alpha_weight ) && ( !info->alpha_unweight )  ) // don't short circuit when alpha weighting (get everything to 0-1.0 as usual)
 7473+      if ( ( ( input_type == STBIR_TYPE_UINT8 ) && ( output_type == STBIR_TYPE_UINT8 ) ) || ( ( input_type == STBIR_TYPE_UINT16 ) && ( output_type == STBIR_TYPE_UINT16 ) ) )
 7474+        non_scaled = 1;
 7475+
 7476+    if ( info->input_pixel_layout_internal <= STBIRI_4CHANNEL )
 7477+      decode_pixels = decode_simple_scaled_or_not[ input_type == STBIR_TYPE_UINT16 ][ non_scaled ];
 7478+    else
 7479+      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 ];
 7480+  }
 7481+  else
 7482+  {
 7483+    if ( info->input_pixel_layout_internal <= STBIRI_4CHANNEL )
 7484+      decode_pixels = decode_simple[ input_type - STBIR_TYPE_UINT8_SRGB ];
 7485+    else
 7486+      decode_pixels = decode_alphas[ ( info->input_pixel_layout_internal - STBIRI_RGBA ) % ( STBIRI_AR-STBIRI_RGBA+1 ) ][ input_type - STBIR_TYPE_UINT8_SRGB ];
 7487+  }
 7488+
 7489+  // setup the output format converters
 7490+  if ( ( output_type == STBIR_TYPE_UINT8 ) || ( output_type == STBIR_TYPE_UINT16 ) )
 7491+  {
 7492+    int non_scaled = 0;
 7493+
 7494+    // 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)
 7495+    if ( ( !info->alpha_weight ) && ( !info->alpha_unweight ) ) // don't short circuit when alpha weighting (get everything to 0-1.0 as usual)
 7496+      if ( ( ( input_type == STBIR_TYPE_UINT8 ) && ( output_type == STBIR_TYPE_UINT8 ) ) || ( ( input_type == STBIR_TYPE_UINT16 ) && ( output_type == STBIR_TYPE_UINT16 ) ) )
 7497+        non_scaled = 1;
 7498+
 7499+    if ( info->output_pixel_layout_internal <= STBIRI_4CHANNEL )
 7500+      encode_pixels = encode_simple_scaled_or_not[ output_type == STBIR_TYPE_UINT16 ][ non_scaled ];
 7501+    else
 7502+      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 ];
 7503+  }
 7504+  else
 7505+  {
 7506+    if ( info->output_pixel_layout_internal <= STBIRI_4CHANNEL )
 7507+      encode_pixels = encode_simple[ output_type - STBIR_TYPE_UINT8_SRGB ];
 7508+    else
 7509+      encode_pixels = encode_alphas[ ( info->output_pixel_layout_internal - STBIRI_RGBA ) % ( STBIRI_AR-STBIRI_RGBA+1 ) ][ output_type - STBIR_TYPE_UINT8_SRGB ];
 7510+  }
 7511+
 7512+  info->input_type = input_type;
 7513+  info->output_type = output_type;
 7514+  info->decode_pixels = decode_pixels;
 7515+  info->encode_pixels = encode_pixels;
 7516+}
 7517+
 7518+static void stbir__clip( int * outx, int * outsubw, int outw, double * u0, double * u1 )
 7519+{
 7520+  double per, adj;
 7521+  int over;
 7522+
 7523+  // do left/top edge
 7524+  if ( *outx < 0 )
 7525+  {
 7526+    per = ( (double)*outx ) / ( (double)*outsubw ); // is negative
 7527+    adj = per * ( *u1 - *u0 );
 7528+    *u0 -= adj; // increases u0
 7529+    *outx = 0;
 7530+  }
 7531+
 7532+  // do right/bot edge
 7533+  over = outw - ( *outx + *outsubw );
 7534+  if ( over < 0 )
 7535+  {
 7536+    per = ( (double)over ) / ( (double)*outsubw ); // is negative
 7537+    adj = per * ( *u1 - *u0 );
 7538+    *u1 += adj; // decrease u1
 7539+    *outsubw = outw - *outx;
 7540+  }
 7541+}
 7542+
 7543+// converts a double to a rational that has less than one float bit of error (returns 0 if unable to do so)
 7544+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)
 7545+{
 7546+  double err;
 7547+  stbir_uint64 top, bot;
 7548+  stbir_uint64 numer_last = 0;
 7549+  stbir_uint64 denom_last = 1;
 7550+  stbir_uint64 numer_estimate = 1;
 7551+  stbir_uint64 denom_estimate = 0;
 7552+
 7553+  // scale to past float error range
 7554+  top = (stbir_uint64)( f * (double)(1 << 25) );
 7555+  bot = 1 << 25;
 7556+
 7557+  // keep refining, but usually stops in a few loops - usually 5 for bad cases
 7558+  for(;;)
 7559+  {
 7560+    stbir_uint64 est, temp;
 7561+
 7562+    // hit limit, break out and do best full range estimate
 7563+    if ( ( ( limit_denom ) ? denom_estimate : numer_estimate ) >= limit )
 7564+      break;
 7565+
 7566+    // is the current error less than 1 bit of a float? if so, we're done
 7567+    if ( denom_estimate )
 7568+    {
 7569+      err = ( (double)numer_estimate / (double)denom_estimate ) - f;
 7570+      if ( err < 0.0 ) err = -err;
 7571+      if ( err < ( 1.0 / (double)(1<<24) ) )
 7572+      {
 7573+        // yup, found it
 7574+        *numer = (stbir_uint32) numer_estimate;
 7575+        *denom = (stbir_uint32) denom_estimate;
 7576+        return 1;
 7577+      }
 7578+    }
 7579+
 7580+    // no more refinement bits left? break out and do full range estimate
 7581+    if ( bot == 0 )
 7582+      break;
 7583+
 7584+    // gcd the estimate bits
 7585+    est = top / bot;
 7586+    temp = top % bot;
 7587+    top = bot;
 7588+    bot = temp;
 7589+
 7590+    // move remainders
 7591+    temp = est * denom_estimate + denom_last;
 7592+    denom_last = denom_estimate;
 7593+    denom_estimate = temp;
 7594+
 7595+    // move remainders
 7596+    temp = est * numer_estimate + numer_last;
 7597+    numer_last = numer_estimate;
 7598+    numer_estimate = temp;
 7599+  }
 7600+
 7601+  // we didn't fine anything good enough for float, use a full range estimate
 7602+  if ( limit_denom )
 7603+  {
 7604+    numer_estimate= (stbir_uint64)( f * (double)limit + 0.5 );
 7605+    denom_estimate = limit;
 7606+  }
 7607+  else
 7608+  {
 7609+    numer_estimate = limit;
 7610+    denom_estimate = (stbir_uint64)( ( (double)limit / f ) + 0.5 );
 7611+  }
 7612+
 7613+  *numer = (stbir_uint32) numer_estimate;
 7614+  *denom = (stbir_uint32) denom_estimate;
 7615+
 7616+  err = ( denom_estimate ) ? ( ( (double)(stbir_uint32)numer_estimate / (double)(stbir_uint32)denom_estimate ) - f ) : 1.0;
 7617+  if ( err < 0.0 ) err = -err;
 7618+  return ( err < ( 1.0 / (double)(1<<24) ) ) ? 1 : 0;
 7619+}
 7620+
 7621+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 )
 7622+{
 7623+  double output_range, input_range, output_s, input_s, ratio, scale;
 7624+
 7625+  input_s = input_s1 - input_s0;
 7626+
 7627+  // null area
 7628+  if ( ( output_full_range == 0 ) || ( input_full_range == 0 ) ||
 7629+       ( output_sub_range == 0 ) || ( input_s <= stbir__small_float ) )
 7630+    return 0;
 7631+
 7632+  // are either of the ranges completely out of bounds?
 7633+  if ( ( *output_offset >= output_full_range ) || ( ( *output_offset + output_sub_range ) <= 0 ) || ( input_s0 >= (1.0f-stbir__small_float) ) || ( input_s1 <= stbir__small_float ) )
 7634+    return 0;
 7635+
 7636+  output_range = (double)output_full_range;
 7637+  input_range = (double)input_full_range;
 7638+
 7639+  output_s = ( (double)output_sub_range) / output_range;
 7640+
 7641+  // figure out the scaling to use
 7642+  ratio = output_s / input_s;
 7643+
 7644+  // save scale before clipping
 7645+  scale = ( output_range / input_range ) * ratio;
 7646+  scale_info->scale = (float)scale;
 7647+  scale_info->inv_scale = (float)( 1.0 / scale );
 7648+
 7649+  // clip output area to left/right output edges (and adjust input area)
 7650+  stbir__clip( output_offset, &output_sub_range, output_full_range, &input_s0, &input_s1 );
 7651+
 7652+  // recalc input area
 7653+  input_s = input_s1 - input_s0;
 7654+
 7655+  // after clipping do we have zero input area?
 7656+  if ( input_s <= stbir__small_float )
 7657+    return 0;
 7658+
 7659+  // calculate and store the starting source offsets in output pixel space
 7660+  scale_info->pixel_shift = (float) ( input_s0 * ratio * output_range );
 7661+
 7662+  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 ) );
 7663+
 7664+  scale_info->input_full_size = input_full_range;
 7665+  scale_info->output_sub_size = output_sub_range;
 7666+
 7667+  return 1;
 7668+}
 7669+
 7670+
 7671+static void stbir__init_and_set_layout( STBIR_RESIZE * resize, stbir_pixel_layout pixel_layout, stbir_datatype data_type )
 7672+{
 7673+  resize->input_cb = 0;
 7674+  resize->output_cb = 0;
 7675+  resize->user_data = resize;
 7676+  resize->samplers = 0;
 7677+  resize->called_alloc = 0;
 7678+  resize->horizontal_filter = STBIR_FILTER_DEFAULT;
 7679+  resize->horizontal_filter_kernel = 0; resize->horizontal_filter_support = 0;
 7680+  resize->vertical_filter = STBIR_FILTER_DEFAULT;
 7681+  resize->vertical_filter_kernel = 0; resize->vertical_filter_support = 0;
 7682+  resize->horizontal_edge = STBIR_EDGE_CLAMP;
 7683+  resize->vertical_edge = STBIR_EDGE_CLAMP;
 7684+  resize->input_s0 = 0; resize->input_t0 = 0; resize->input_s1 = 1; resize->input_t1 = 1;
 7685+  resize->output_subx = 0; resize->output_suby = 0; resize->output_subw = resize->output_w; resize->output_subh = resize->output_h;
 7686+  resize->input_data_type = data_type;
 7687+  resize->output_data_type = data_type;
 7688+  resize->input_pixel_layout_public = pixel_layout;
 7689+  resize->output_pixel_layout_public = pixel_layout;
 7690+  resize->needs_rebuild = 1;
 7691+}
 7692+
 7693+STBIRDEF void stbir_resize_init( STBIR_RESIZE * resize,
 7694+                                 const void *input_pixels,  int input_w,  int input_h, int input_stride_in_bytes, // stride can be zero
 7695+                                       void *output_pixels, int output_w, int output_h, int output_stride_in_bytes, // stride can be zero
 7696+                                 stbir_pixel_layout pixel_layout, stbir_datatype data_type )
 7697+{
 7698+  resize->input_pixels = input_pixels;
 7699+  resize->input_w = input_w;
 7700+  resize->input_h = input_h;
 7701+  resize->input_stride_in_bytes = input_stride_in_bytes;
 7702+  resize->output_pixels = output_pixels;
 7703+  resize->output_w = output_w;
 7704+  resize->output_h = output_h;
 7705+  resize->output_stride_in_bytes = output_stride_in_bytes;
 7706+  resize->fast_alpha = 0;
 7707+
 7708+  stbir__init_and_set_layout( resize, pixel_layout, data_type );
 7709+}
 7710+
 7711+// You can update parameters any time after resize_init
 7712+STBIRDEF void stbir_set_datatypes( STBIR_RESIZE * resize, stbir_datatype input_type, stbir_datatype output_type )  // by default, datatype from resize_init
 7713+{
 7714+  resize->input_data_type = input_type;
 7715+  resize->output_data_type = output_type;
 7716+  if ( ( resize->samplers ) && ( !resize->needs_rebuild ) )
 7717+    stbir__update_info_from_resize( resize->samplers, resize );
 7718+}
 7719+
 7720+STBIRDEF void stbir_set_pixel_callbacks( STBIR_RESIZE * resize, stbir_input_callback * input_cb, stbir_output_callback * output_cb )   // no callbacks by default
 7721+{
 7722+  resize->input_cb = input_cb;
 7723+  resize->output_cb = output_cb;
 7724+
 7725+  if ( ( resize->samplers ) && ( !resize->needs_rebuild ) )
 7726+  {
 7727+    resize->samplers->in_pixels_cb = input_cb;
 7728+    resize->samplers->out_pixels_cb = output_cb;
 7729+  }
 7730+}
 7731+
 7732+STBIRDEF void stbir_set_user_data( STBIR_RESIZE * resize, void * user_data )                                     // pass back STBIR_RESIZE* by default
 7733+{
 7734+  resize->user_data = user_data;
 7735+  if ( ( resize->samplers ) && ( !resize->needs_rebuild ) )
 7736+    resize->samplers->user_data = user_data;
 7737+}
 7738+
 7739+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 )
 7740+{
 7741+  resize->input_pixels = input_pixels;
 7742+  resize->input_stride_in_bytes = input_stride_in_bytes;
 7743+  resize->output_pixels = output_pixels;
 7744+  resize->output_stride_in_bytes = output_stride_in_bytes;
 7745+  if ( ( resize->samplers ) && ( !resize->needs_rebuild ) )
 7746+    stbir__update_info_from_resize( resize->samplers, resize );
 7747+}
 7748+
 7749+
 7750+STBIRDEF int stbir_set_edgemodes( STBIR_RESIZE * resize, stbir_edge horizontal_edge, stbir_edge vertical_edge )       // CLAMP by default
 7751+{
 7752+  resize->horizontal_edge = horizontal_edge;
 7753+  resize->vertical_edge = vertical_edge;
 7754+  resize->needs_rebuild = 1;
 7755+  return 1;
 7756+}
 7757+
 7758+STBIRDEF int stbir_set_filters( STBIR_RESIZE * resize, stbir_filter horizontal_filter, stbir_filter vertical_filter ) // STBIR_DEFAULT_FILTER_UPSAMPLE/DOWNSAMPLE by default
 7759+{
 7760+  resize->horizontal_filter = horizontal_filter;
 7761+  resize->vertical_filter = vertical_filter;
 7762+  resize->needs_rebuild = 1;
 7763+  return 1;
 7764+}
 7765+
 7766+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 )
 7767+{
 7768+  resize->horizontal_filter_kernel = horizontal_filter; resize->horizontal_filter_support = horizontal_support;
 7769+  resize->vertical_filter_kernel = vertical_filter; resize->vertical_filter_support = vertical_support;
 7770+  resize->needs_rebuild = 1;
 7771+  return 1;
 7772+}
 7773+
 7774+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
 7775+{
 7776+  resize->input_pixel_layout_public = input_pixel_layout;
 7777+  resize->output_pixel_layout_public = output_pixel_layout;
 7778+  resize->needs_rebuild = 1;
 7779+  return 1;
 7780+}
 7781+
 7782+
 7783+STBIRDEF int stbir_set_non_pm_alpha_speed_over_quality( STBIR_RESIZE * resize, int non_pma_alpha_speed_over_quality )   // sets alpha speed
 7784+{
 7785+  resize->fast_alpha = non_pma_alpha_speed_over_quality;
 7786+  resize->needs_rebuild = 1;
 7787+  return 1;
 7788+}
 7789+
 7790+STBIRDEF int stbir_set_input_subrect( STBIR_RESIZE * resize, double s0, double t0, double s1, double t1 )                 // sets input region (full region by default)
 7791+{
 7792+  resize->input_s0 = s0;
 7793+  resize->input_t0 = t0;
 7794+  resize->input_s1 = s1;
 7795+  resize->input_t1 = t1;
 7796+  resize->needs_rebuild = 1;
 7797+
 7798+  // are we inbounds?
 7799+  if ( ( s1 < stbir__small_float ) || ( (s1-s0) < stbir__small_float ) ||
 7800+       ( t1 < stbir__small_float ) || ( (t1-t0) < stbir__small_float ) ||
 7801+       ( s0 > (1.0f-stbir__small_float) ) ||
 7802+       ( t0 > (1.0f-stbir__small_float) ) )
 7803+    return 0;
 7804+
 7805+  return 1;
 7806+}
 7807+
 7808+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)
 7809+{
 7810+  resize->output_subx = subx;
 7811+  resize->output_suby = suby;
 7812+  resize->output_subw = subw;
 7813+  resize->output_subh = subh;
 7814+  resize->needs_rebuild = 1;
 7815+
 7816+  // are we inbounds?
 7817+  if ( ( subx >= resize->output_w ) || ( ( subx + subw ) <= 0 ) || ( suby >= resize->output_h ) || ( ( suby + subh ) <= 0 ) || ( subw == 0 ) || ( subh == 0 ) )
 7818+    return 0;
 7819+
 7820+  return 1;
 7821+}
 7822+
 7823+STBIRDEF int stbir_set_pixel_subrect( STBIR_RESIZE * resize, int subx, int suby, int subw, int subh )                 // sets both regions (full regions by default)
 7824+{
 7825+  double s0, t0, s1, t1;
 7826+
 7827+  s0 = ( (double)subx ) / ( (double)resize->output_w );
 7828+  t0 = ( (double)suby ) / ( (double)resize->output_h );
 7829+  s1 = ( (double)(subx+subw) ) / ( (double)resize->output_w );
 7830+  t1 = ( (double)(suby+subh) ) / ( (double)resize->output_h );
 7831+
 7832+  resize->input_s0 = s0;
 7833+  resize->input_t0 = t0;
 7834+  resize->input_s1 = s1;
 7835+  resize->input_t1 = t1;
 7836+  resize->output_subx = subx;
 7837+  resize->output_suby = suby;
 7838+  resize->output_subw = subw;
 7839+  resize->output_subh = subh;
 7840+  resize->needs_rebuild = 1;
 7841+
 7842+  // are we inbounds?
 7843+  if ( ( subx >= resize->output_w ) || ( ( subx + subw ) <= 0 ) || ( suby >= resize->output_h ) || ( ( suby + subh ) <= 0 ) || ( subw == 0 ) || ( subh == 0 ) )
 7844+    return 0;
 7845+
 7846+  return 1;
 7847+}
 7848+
 7849+static int stbir__perform_build( STBIR_RESIZE * resize, int splits )
 7850+{
 7851+  stbir__contributors conservative = { 0, 0 };
 7852+  stbir__sampler horizontal, vertical;
 7853+  int new_output_subx, new_output_suby;
 7854+  stbir__info * out_info;
 7855+  #ifdef STBIR_PROFILE
 7856+  stbir__info profile_infod;  // used to contain building profile info before everything is allocated
 7857+  stbir__info * profile_info = &profile_infod;
 7858+  #endif
 7859+
 7860+  // have we already built the samplers?
 7861+  if ( resize->samplers )
 7862+    return 0;
 7863+
 7864+  #define STBIR_RETURN_ERROR_AND_ASSERT( exp )  STBIR_ASSERT( !(exp) ); if (exp) return 0;
 7865+  STBIR_RETURN_ERROR_AND_ASSERT( (unsigned)resize->horizontal_filter >= STBIR_FILTER_OTHER)
 7866+  STBIR_RETURN_ERROR_AND_ASSERT( (unsigned)resize->vertical_filter >= STBIR_FILTER_OTHER)
 7867+  #undef STBIR_RETURN_ERROR_AND_ASSERT
 7868+
 7869+  if ( splits <= 0 )
 7870+    return 0;
 7871+
 7872+  STBIR_PROFILE_BUILD_FIRST_START( build );
 7873+
 7874+  new_output_subx = resize->output_subx;
 7875+  new_output_suby = resize->output_suby;
 7876+
 7877+  // do horizontal clip and scale calcs
 7878+  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 ) )
 7879+    return 0;
 7880+
 7881+  // do vertical clip and scale calcs
 7882+  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 ) )
 7883+    return 0;
 7884+
 7885+  // if nothing to do, just return
 7886+  if ( ( horizontal.scale_info.output_sub_size == 0 ) || ( vertical.scale_info.output_sub_size == 0 ) )
 7887+    return 0;
 7888+
 7889+  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 );
 7890+  stbir__get_conservative_extents( &horizontal, &conservative, resize->user_data );
 7891+  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 );
 7892+
 7893+  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)
 7894+  {
 7895+    splits = vertical.scale_info.output_sub_size / STBIR_FORCE_MINIMUM_SCANLINES_FOR_SPLITS;
 7896+    if ( splits == 0 ) splits = 1;
 7897+  }
 7898+
 7899+  STBIR_PROFILE_BUILD_START( alloc );
 7900+  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 );
 7901+  STBIR_PROFILE_BUILD_END( alloc );
 7902+  STBIR_PROFILE_BUILD_END( build );
 7903+
 7904+  if ( out_info )
 7905+  {
 7906+    resize->splits = splits;
 7907+    resize->samplers = out_info;
 7908+    resize->needs_rebuild = 0;
 7909+    #ifdef STBIR_PROFILE
 7910+      STBIR_MEMCPY( &out_info->profile, &profile_infod.profile, sizeof( out_info->profile ) );
 7911+    #endif
 7912+
 7913+    // update anything that can be changed without recalcing samplers
 7914+    stbir__update_info_from_resize( out_info, resize );
 7915+
 7916+    return splits;
 7917+  }
 7918+
 7919+  return 0;
 7920+}
 7921+
 7922+void stbir_free_samplers( STBIR_RESIZE * resize )
 7923+{
 7924+  if ( resize->samplers )
 7925+  {
 7926+    stbir__free_internal_mem( resize->samplers );
 7927+    resize->samplers = 0;
 7928+    resize->called_alloc = 0;
 7929+  }
 7930+}
 7931+
 7932+STBIRDEF int stbir_build_samplers_with_splits( STBIR_RESIZE * resize, int splits )
 7933+{
 7934+  if ( ( resize->samplers == 0 ) || ( resize->needs_rebuild ) )
 7935+  {
 7936+    if ( resize->samplers )
 7937+      stbir_free_samplers( resize );
 7938+
 7939+    resize->called_alloc = 1;
 7940+    return stbir__perform_build( resize, splits );
 7941+  }
 7942+
 7943+  STBIR_PROFILE_BUILD_CLEAR( resize->samplers );
 7944+
 7945+  return 1;
 7946+}
 7947+
 7948+STBIRDEF int stbir_build_samplers( STBIR_RESIZE * resize )
 7949+{
 7950+  return stbir_build_samplers_with_splits( resize, 1 );
 7951+}
 7952+
 7953+STBIRDEF int stbir_resize_extended( STBIR_RESIZE * resize )
 7954+{
 7955+  int result;
 7956+
 7957+  if ( ( resize->samplers == 0 ) || ( resize->needs_rebuild ) )
 7958+  {
 7959+    int alloc_state = resize->called_alloc;  // remember allocated state
 7960+
 7961+    if ( resize->samplers )
 7962+    {
 7963+      stbir__free_internal_mem( resize->samplers );
 7964+      resize->samplers = 0;
 7965+    }
 7966+
 7967+    if ( !stbir_build_samplers( resize ) )
 7968+      return 0;
 7969+
 7970+    resize->called_alloc = alloc_state;
 7971+
 7972+    // if build_samplers succeeded (above), but there are no samplers set, then
 7973+    //   the area to stretch into was zero pixels, so don't do anything and return
 7974+    //   success
 7975+    if ( resize->samplers == 0 )
 7976+      return 1;
 7977+  }
 7978+  else
 7979+  {
 7980+    // didn't build anything - clear it
 7981+    STBIR_PROFILE_BUILD_CLEAR( resize->samplers );
 7982+  }
 7983+
 7984+  // do resize
 7985+  result = stbir__perform_resize( resize->samplers, 0, resize->splits );
 7986+
 7987+  // if we alloced, then free
 7988+  if ( !resize->called_alloc )
 7989+  {
 7990+    stbir_free_samplers( resize );
 7991+    resize->samplers = 0;
 7992+  }
 7993+
 7994+  return result;
 7995+}
 7996+
 7997+STBIRDEF int stbir_resize_extended_split( STBIR_RESIZE * resize, int split_start, int split_count )
 7998+{
 7999+  STBIR_ASSERT( resize->samplers );
 8000+
 8001+  // if we're just doing the whole thing, call full
 8002+  if ( ( split_start == -1 ) || ( ( split_start == 0 ) && ( split_count == resize->splits ) ) )
 8003+    return stbir_resize_extended( resize );
 8004+
 8005+  // you **must** build samplers first when using split resize
 8006+  if ( ( resize->samplers == 0 ) || ( resize->needs_rebuild ) )
 8007+    return 0;
 8008+
 8009+  if ( ( split_start >= resize->splits ) || ( split_start < 0 ) || ( ( split_start + split_count ) > resize->splits ) || ( split_count <= 0 ) )
 8010+    return 0;
 8011+
 8012+  // do resize
 8013+  return stbir__perform_resize( resize->samplers, split_start, split_count );
 8014+}
 8015+
 8016+
 8017+static void * stbir_quick_resize_helper( const void *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
 8018+                                               void *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
 8019+                                               stbir_pixel_layout pixel_layout, stbir_datatype data_type, stbir_edge edge, stbir_filter filter )
 8020+{
 8021+  STBIR_RESIZE resize;
 8022+  int scanline_output_in_bytes;
 8023+  int positive_output_stride_in_bytes;
 8024+  void * start_ptr;
 8025+  void * free_ptr;
 8026+
 8027+  scanline_output_in_bytes = output_w * stbir__type_size[ data_type ] * stbir__pixel_channels[ stbir__pixel_layout_convert_public_to_internal[ pixel_layout ] ];
 8028+  if ( scanline_output_in_bytes == 0 )
 8029+    return 0;
 8030+
 8031+  // if zero stride, use scanline output
 8032+  if ( output_stride_in_bytes == 0 )
 8033+    output_stride_in_bytes = scanline_output_in_bytes;
 8034+
 8035+  // abs value for inverted images (negative pitches)
 8036+  positive_output_stride_in_bytes = output_stride_in_bytes;
 8037+  if ( positive_output_stride_in_bytes < 0 )
 8038+    positive_output_stride_in_bytes = -positive_output_stride_in_bytes;
 8039+
 8040+  // is the requested stride smaller than the scanline output? if so, just fail
 8041+  if ( positive_output_stride_in_bytes < scanline_output_in_bytes )
 8042+    return 0;
 8043+
 8044+  start_ptr = output_pixels;
 8045+  free_ptr = 0;  // no free pointer, since they passed buffer to use
 8046+
 8047+  // did they pass a zero for the dest? if so, allocate the buffer
 8048+  if ( output_pixels == 0 )
 8049+  {
 8050+    size_t size;
 8051+    char * ptr;
 8052+  
 8053+    size = (size_t)positive_output_stride_in_bytes * (size_t)output_h;
 8054+    if ( size == 0 )
 8055+      return 0;
 8056+
 8057+    ptr = (char*) STBIR_MALLOC( size, 0 );
 8058+    if ( ptr == 0 )
 8059+      return 0;
 8060+
 8061+    free_ptr = ptr;
 8062+
 8063+    // point at the last scanline, if they requested a flipped image
 8064+    if ( output_stride_in_bytes < 0 )
 8065+      start_ptr = ptr + ( (size_t)positive_output_stride_in_bytes * (size_t)( output_h - 1 ) );
 8066+    else
 8067+      start_ptr = ptr;
 8068+  }
 8069+
 8070+  // ok, now do the resize
 8071+  stbir_resize_init( &resize,
 8072+                     input_pixels,  input_w,  input_h,  input_stride_in_bytes,
 8073+                     start_ptr, output_w, output_h, output_stride_in_bytes,
 8074+                     pixel_layout, data_type );
 8075+
 8076+  resize.horizontal_edge = edge;
 8077+  resize.vertical_edge = edge;
 8078+  resize.horizontal_filter = filter;
 8079+  resize.vertical_filter = filter;
 8080+
 8081+  if ( !stbir_resize_extended( &resize ) )
 8082+  {
 8083+    if ( free_ptr )
 8084+      STBIR_FREE( free_ptr, 0 );
 8085+    return 0;
 8086+  }
 8087+
 8088+  return (free_ptr) ? free_ptr : start_ptr;
 8089+}
 8090+
 8091+
 8092+
 8093+STBIRDEF unsigned char * stbir_resize_uint8_linear( const unsigned char *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
 8094+                                                          unsigned char *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
 8095+                                                          stbir_pixel_layout pixel_layout )
 8096+{
 8097+  return (unsigned char *) stbir_quick_resize_helper( input_pixels , input_w , input_h, input_stride_in_bytes, 
 8098+                                                      output_pixels, output_w, output_h, output_stride_in_bytes, 
 8099+                                                      pixel_layout, STBIR_TYPE_UINT8, STBIR_EDGE_CLAMP, STBIR_FILTER_DEFAULT );
 8100+}
 8101+
 8102+STBIRDEF unsigned char * stbir_resize_uint8_srgb( const unsigned char *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
 8103+                                                        unsigned char *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
 8104+                                                        stbir_pixel_layout pixel_layout )
 8105+{
 8106+  return (unsigned char *) stbir_quick_resize_helper( input_pixels , input_w , input_h, input_stride_in_bytes, 
 8107+                                                      output_pixels, output_w, output_h, output_stride_in_bytes, 
 8108+                                                      pixel_layout, STBIR_TYPE_UINT8_SRGB, STBIR_EDGE_CLAMP, STBIR_FILTER_DEFAULT );
 8109+}
 8110+
 8111+
 8112+STBIRDEF float * stbir_resize_float_linear( const float *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
 8113+                                                  float *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
 8114+                                                  stbir_pixel_layout pixel_layout )
 8115+{
 8116+  return (float *) stbir_quick_resize_helper( input_pixels , input_w , input_h, input_stride_in_bytes, 
 8117+                                              output_pixels, output_w, output_h, output_stride_in_bytes, 
 8118+                                              pixel_layout, STBIR_TYPE_FLOAT, STBIR_EDGE_CLAMP, STBIR_FILTER_DEFAULT  );
 8119+}
 8120+
 8121+
 8122+STBIRDEF void * stbir_resize( const void *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
 8123+                                    void *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
 8124+                                    stbir_pixel_layout pixel_layout, stbir_datatype data_type,
 8125+                                    stbir_edge edge, stbir_filter filter )
 8126+{
 8127+  return (void *) stbir_quick_resize_helper( input_pixels , input_w , input_h, input_stride_in_bytes, 
 8128+                                             output_pixels, output_w, output_h, output_stride_in_bytes, 
 8129+                                             pixel_layout, data_type, edge, filter  );
 8130+}
 8131+
 8132+#ifdef STBIR_PROFILE
 8133+
 8134+STBIRDEF void stbir_resize_build_profile_info( STBIR_PROFILE_INFO * info, STBIR_RESIZE const * resize )
 8135+{
 8136+  static char const * bdescriptions[6] = { "Building", "Allocating", "Horizontal sampler", "Vertical sampler", "Coefficient cleanup", "Coefficient piovot" } ;
 8137+  stbir__info* samp = resize->samplers;
 8138+  int i;
 8139+
 8140+  typedef int testa[ (STBIR__ARRAY_SIZE( bdescriptions ) == (STBIR__ARRAY_SIZE( samp->profile.array )-1) )?1:-1];
 8141+  typedef int testb[ (sizeof( samp->profile.array ) == (sizeof(samp->profile.named)) )?1:-1];
 8142+  typedef int testc[ (sizeof( info->clocks ) >= (sizeof(samp->profile.named)) )?1:-1];
 8143+
 8144+  for( i = 0 ; i < STBIR__ARRAY_SIZE( bdescriptions ) ; i++)
 8145+    info->clocks[i] = samp->profile.array[i+1];
 8146+
 8147+  info->total_clocks = samp->profile.named.total;
 8148+  info->descriptions = bdescriptions;
 8149+  info->count = STBIR__ARRAY_SIZE( bdescriptions );
 8150+}
 8151+
 8152+STBIRDEF void stbir_resize_split_profile_info( STBIR_PROFILE_INFO * info, STBIR_RESIZE const * resize, int split_start, int split_count )
 8153+{
 8154+  static char const * descriptions[7] = { "Looping", "Vertical sampling", "Horizontal sampling", "Scanline input", "Scanline output", "Alpha weighting", "Alpha unweighting" };
 8155+  stbir__per_split_info * split_info;
 8156+  int s, i;
 8157+
 8158+  typedef int testa[ (STBIR__ARRAY_SIZE( descriptions ) == (STBIR__ARRAY_SIZE( split_info->profile.array )-1) )?1:-1];
 8159+  typedef int testb[ (sizeof( split_info->profile.array ) == (sizeof(split_info->profile.named)) )?1:-1];
 8160+  typedef int testc[ (sizeof( info->clocks ) >= (sizeof(split_info->profile.named)) )?1:-1];
 8161+
 8162+  if ( split_start == -1 )
 8163+  {
 8164+    split_start = 0;
 8165+    split_count = resize->samplers->splits;
 8166+  }
 8167+
 8168+  if ( ( split_start >= resize->splits ) || ( split_start < 0 ) || ( ( split_start + split_count ) > resize->splits ) || ( split_count <= 0 ) )
 8169+  {
 8170+    info->total_clocks = 0;
 8171+    info->descriptions = 0;
 8172+    info->count = 0;
 8173+    return;
 8174+  }
 8175+
 8176+  split_info = resize->samplers->split_info + split_start;
 8177+
 8178+  // sum up the profile from all the splits
 8179+  for( i = 0 ; i < STBIR__ARRAY_SIZE( descriptions ) ; i++ )
 8180+  {
 8181+    stbir_uint64 sum = 0;
 8182+    for( s = 0 ; s < split_count ; s++ )
 8183+      sum += split_info[s].profile.array[i+1];
 8184+    info->clocks[i] = sum;
 8185+  }
 8186+
 8187+  info->total_clocks = split_info->profile.named.total;
 8188+  info->descriptions = descriptions;
 8189+  info->count = STBIR__ARRAY_SIZE( descriptions );
 8190+}
 8191+
 8192+STBIRDEF void stbir_resize_extended_profile_info( STBIR_PROFILE_INFO * info, STBIR_RESIZE const * resize )
 8193+{
 8194+  stbir_resize_split_profile_info( info, resize, -1, 0 );
 8195+}
 8196+
 8197+#endif // STBIR_PROFILE
 8198+
 8199+#undef STBIR_BGR
 8200+#undef STBIR_1CHANNEL
 8201+#undef STBIR_2CHANNEL
 8202+#undef STBIR_RGB
 8203+#undef STBIR_RGBA
 8204+#undef STBIR_4CHANNEL
 8205+#undef STBIR_BGRA
 8206+#undef STBIR_ARGB
 8207+#undef STBIR_ABGR
 8208+#undef STBIR_RA
 8209+#undef STBIR_AR
 8210+#undef STBIR_RGBA_PM
 8211+#undef STBIR_BGRA_PM
 8212+#undef STBIR_ARGB_PM
 8213+#undef STBIR_ABGR_PM
 8214+#undef STBIR_RA_PM
 8215+#undef STBIR_AR_PM
 8216+
 8217+#endif // STB_IMAGE_RESIZE_IMPLEMENTATION
 8218+
 8219+#else  // STB_IMAGE_RESIZE_HORIZONTALS&STB_IMAGE_RESIZE_DO_VERTICALS
 8220+
 8221+// we reinclude the header file to define all the horizontal functions
 8222+//   specializing each function for the number of coeffs is 20-40% faster *OVERALL*
 8223+
 8224+// by including the header file again this way, we can still debug the functions
 8225+
 8226+#define STBIR_strs_join2( start, mid, end ) start##mid##end
 8227+#define STBIR_strs_join1( start, mid, end ) STBIR_strs_join2( start, mid, end )
 8228+
 8229+#define STBIR_strs_join24( start, mid1, mid2, end ) start##mid1##mid2##end
 8230+#define STBIR_strs_join14( start, mid1, mid2, end ) STBIR_strs_join24( start, mid1, mid2, end )
 8231+
 8232+#ifdef STB_IMAGE_RESIZE_DO_CODERS
 8233+
 8234+#ifdef stbir__decode_suffix
 8235+#define STBIR__CODER_NAME( name ) STBIR_strs_join1( name, _, stbir__decode_suffix )
 8236+#else
 8237+#define STBIR__CODER_NAME( name ) name
 8238+#endif
 8239+
 8240+#ifdef stbir__decode_swizzle
 8241+#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)
 8242+#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)
 8243+#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)
 8244+#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)
 8245+#else
 8246+#define stbir__decode_order0 0
 8247+#define stbir__decode_order1 1
 8248+#define stbir__decode_order2 2
 8249+#define stbir__decode_order3 3
 8250+#define stbir__encode_order0 0
 8251+#define stbir__encode_order1 1
 8252+#define stbir__encode_order2 2
 8253+#define stbir__encode_order3 3
 8254+#define stbir__decode_simdf8_flip(reg)
 8255+#define stbir__decode_simdf4_flip(reg)
 8256+#define stbir__encode_simdf8_unflip(reg)
 8257+#define stbir__encode_simdf4_unflip(reg)
 8258+#endif
 8259+
 8260+#ifdef STBIR_SIMD8
 8261+#define stbir__encode_simdfX_unflip  stbir__encode_simdf8_unflip
 8262+#else
 8263+#define stbir__encode_simdfX_unflip  stbir__encode_simdf4_unflip
 8264+#endif
 8265+
 8266+static float * STBIR__CODER_NAME( stbir__decode_uint8_linear_scaled )( float * decodep, int width_times_channels, void const * inputp )
 8267+{
 8268+  float STBIR_STREAMOUT_PTR( * ) decode = decodep;
 8269+  float * decode_end = (float*) decode + width_times_channels;
 8270+  unsigned char const * input = (unsigned char const*)inputp;
 8271+
 8272+  #ifdef STBIR_SIMD
 8273+  unsigned char const * end_input_m16 = input + width_times_channels - 16;
 8274+  if ( width_times_channels >= 16 )
 8275+  {
 8276+    decode_end -= 16;
 8277+    STBIR_NO_UNROLL_LOOP_START_INF_FOR
 8278+    for(;;)
 8279+    {
 8280+      #ifdef STBIR_SIMD8
 8281+      stbir__simdi i; stbir__simdi8 o0,o1;
 8282+      stbir__simdf8 of0, of1;
 8283+      STBIR_NO_UNROLL(decode);
 8284+      stbir__simdi_load( i, input );
 8285+      stbir__simdi8_expand_u8_to_u32( o0, o1, i );
 8286+      stbir__simdi8_convert_i32_to_float( of0, o0 );
 8287+      stbir__simdi8_convert_i32_to_float( of1, o1 );
 8288+      stbir__simdf8_mult( of0, of0, STBIR_max_uint8_as_float_inverted8);
 8289+      stbir__simdf8_mult( of1, of1, STBIR_max_uint8_as_float_inverted8);
 8290+      stbir__decode_simdf8_flip( of0 );
 8291+      stbir__decode_simdf8_flip( of1 );
 8292+      stbir__simdf8_store( decode + 0, of0 );
 8293+      stbir__simdf8_store( decode + 8, of1 );
 8294+      #else
 8295+      stbir__simdi i, o0, o1, o2, o3;
 8296+      stbir__simdf of0, of1, of2, of3;
 8297+      STBIR_NO_UNROLL(decode);
 8298+      stbir__simdi_load( i, input );
 8299+      stbir__simdi_expand_u8_to_u32( o0,o1,o2,o3,i);
 8300+      stbir__simdi_convert_i32_to_float( of0, o0 );
 8301+      stbir__simdi_convert_i32_to_float( of1, o1 );
 8302+      stbir__simdi_convert_i32_to_float( of2, o2 );
 8303+      stbir__simdi_convert_i32_to_float( of3, o3 );
 8304+      stbir__simdf_mult( of0, of0, STBIR__CONSTF(STBIR_max_uint8_as_float_inverted) );
 8305+      stbir__simdf_mult( of1, of1, STBIR__CONSTF(STBIR_max_uint8_as_float_inverted) );
 8306+      stbir__simdf_mult( of2, of2, STBIR__CONSTF(STBIR_max_uint8_as_float_inverted) );
 8307+      stbir__simdf_mult( of3, of3, STBIR__CONSTF(STBIR_max_uint8_as_float_inverted) );
 8308+      stbir__decode_simdf4_flip( of0 );
 8309+      stbir__decode_simdf4_flip( of1 );
 8310+      stbir__decode_simdf4_flip( of2 );
 8311+      stbir__decode_simdf4_flip( of3 );
 8312+      stbir__simdf_store( decode + 0,  of0 );
 8313+      stbir__simdf_store( decode + 4,  of1 );
 8314+      stbir__simdf_store( decode + 8,  of2 );
 8315+      stbir__simdf_store( decode + 12, of3 );
 8316+      #endif
 8317+      decode += 16;
 8318+      input += 16;
 8319+      if ( decode <= decode_end )
 8320+        continue;
 8321+      if ( decode == ( decode_end + 16 ) )
 8322+        break;
 8323+      decode = decode_end; // backup and do last couple
 8324+      input = end_input_m16;
 8325+    }
 8326+    return decode_end + 16;
 8327+  }
 8328+  #endif
 8329+
 8330+  // try to do blocks of 4 when you can
 8331+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 8332+  decode += 4;
 8333+  STBIR_SIMD_NO_UNROLL_LOOP_START
 8334+  while( decode <= decode_end )
 8335+  {
 8336+    STBIR_SIMD_NO_UNROLL(decode);
 8337+    decode[0-4] = ((float)(input[stbir__decode_order0])) * stbir__max_uint8_as_float_inverted;
 8338+    decode[1-4] = ((float)(input[stbir__decode_order1])) * stbir__max_uint8_as_float_inverted;
 8339+    decode[2-4] = ((float)(input[stbir__decode_order2])) * stbir__max_uint8_as_float_inverted;
 8340+    decode[3-4] = ((float)(input[stbir__decode_order3])) * stbir__max_uint8_as_float_inverted;
 8341+    decode += 4;
 8342+    input += 4;
 8343+  }
 8344+  decode -= 4;
 8345+  #endif
 8346+
 8347+  // do the remnants
 8348+  #if stbir__coder_min_num < 4
 8349+  STBIR_NO_UNROLL_LOOP_START
 8350+  while( decode < decode_end )
 8351+  {
 8352+    STBIR_NO_UNROLL(decode);
 8353+    decode[0] = ((float)(input[stbir__decode_order0])) * stbir__max_uint8_as_float_inverted;
 8354+    #if stbir__coder_min_num >= 2
 8355+    decode[1] = ((float)(input[stbir__decode_order1])) * stbir__max_uint8_as_float_inverted;
 8356+    #endif
 8357+    #if stbir__coder_min_num >= 3
 8358+    decode[2] = ((float)(input[stbir__decode_order2])) * stbir__max_uint8_as_float_inverted;
 8359+    #endif
 8360+    decode += stbir__coder_min_num;
 8361+    input += stbir__coder_min_num;
 8362+  }
 8363+  #endif
 8364+
 8365+  return decode_end;
 8366+}
 8367+
 8368+static void STBIR__CODER_NAME( stbir__encode_uint8_linear_scaled )( void * outputp, int width_times_channels, float const * encode )
 8369+{
 8370+  unsigned char STBIR_SIMD_STREAMOUT_PTR( * ) output = (unsigned char *) outputp;
 8371+  unsigned char * end_output = ( (unsigned char *) output ) + width_times_channels;
 8372+
 8373+  #ifdef STBIR_SIMD
 8374+  if ( width_times_channels >= stbir__simdfX_float_count*2 )
 8375+  {
 8376+    float const * end_encode_m8 = encode + width_times_channels - stbir__simdfX_float_count*2;
 8377+    end_output -= stbir__simdfX_float_count*2;
 8378+    STBIR_NO_UNROLL_LOOP_START_INF_FOR
 8379+    for(;;)
 8380+    {
 8381+      stbir__simdfX e0, e1;
 8382+      stbir__simdi i;
 8383+      STBIR_SIMD_NO_UNROLL(encode);
 8384+      stbir__simdfX_madd_mem( e0, STBIR_simd_point5X, STBIR_max_uint8_as_floatX, encode );
 8385+      stbir__simdfX_madd_mem( e1, STBIR_simd_point5X, STBIR_max_uint8_as_floatX, encode+stbir__simdfX_float_count );
 8386+      stbir__encode_simdfX_unflip( e0 );
 8387+      stbir__encode_simdfX_unflip( e1 );
 8388+      #ifdef STBIR_SIMD8
 8389+      stbir__simdf8_pack_to_16bytes( i, e0, e1 );
 8390+      stbir__simdi_store( output, i );
 8391+      #else
 8392+      stbir__simdf_pack_to_8bytes( i, e0, e1 );
 8393+      stbir__simdi_store2( output, i );
 8394+      #endif
 8395+      encode += stbir__simdfX_float_count*2;
 8396+      output += stbir__simdfX_float_count*2;
 8397+      if ( output <= end_output )
 8398+        continue;
 8399+      if ( output == ( end_output + stbir__simdfX_float_count*2 ) )
 8400+        break;
 8401+      output = end_output; // backup and do last couple
 8402+      encode = end_encode_m8;
 8403+    }
 8404+    return;
 8405+  }
 8406+
 8407+  // try to do blocks of 4 when you can
 8408+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 8409+  output += 4;
 8410+  STBIR_NO_UNROLL_LOOP_START
 8411+  while( output <= end_output )
 8412+  {
 8413+    stbir__simdf e0;
 8414+    stbir__simdi i0;
 8415+    STBIR_NO_UNROLL(encode);
 8416+    stbir__simdf_load( e0, encode );
 8417+    stbir__simdf_madd( e0, STBIR__CONSTF(STBIR_simd_point5), STBIR__CONSTF(STBIR_max_uint8_as_float), e0 );
 8418+    stbir__encode_simdf4_unflip( e0 );
 8419+    stbir__simdf_pack_to_8bytes( i0, e0, e0 );  // only use first 4
 8420+    *(int*)(output-4) = stbir__simdi_to_int( i0 );
 8421+    output += 4;
 8422+    encode += 4;
 8423+  }
 8424+  output -= 4;
 8425+  #endif
 8426+
 8427+  // do the remnants
 8428+  #if stbir__coder_min_num < 4
 8429+  STBIR_NO_UNROLL_LOOP_START
 8430+  while( output < end_output )
 8431+  {
 8432+    stbir__simdf e0;
 8433+    STBIR_NO_UNROLL(encode);
 8434+    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 );
 8435+    #if stbir__coder_min_num >= 2
 8436+    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 );
 8437+    #endif
 8438+    #if stbir__coder_min_num >= 3
 8439+    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 );
 8440+    #endif
 8441+    output += stbir__coder_min_num;
 8442+    encode += stbir__coder_min_num;
 8443+  }
 8444+  #endif
 8445+
 8446+  #else
 8447+
 8448+  // try to do blocks of 4 when you can
 8449+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 8450+  output += 4;
 8451+  while( output <= end_output )
 8452+  {
 8453+    float f;
 8454+    f = encode[stbir__encode_order0] * stbir__max_uint8_as_float + 0.5f; STBIR_CLAMP(f, 0, 255); output[0-4] = (unsigned char)f;
 8455+    f = encode[stbir__encode_order1] * stbir__max_uint8_as_float + 0.5f; STBIR_CLAMP(f, 0, 255); output[1-4] = (unsigned char)f;
 8456+    f = encode[stbir__encode_order2] * stbir__max_uint8_as_float + 0.5f; STBIR_CLAMP(f, 0, 255); output[2-4] = (unsigned char)f;
 8457+    f = encode[stbir__encode_order3] * stbir__max_uint8_as_float + 0.5f; STBIR_CLAMP(f, 0, 255); output[3-4] = (unsigned char)f;
 8458+    output += 4;
 8459+    encode += 4;
 8460+  }
 8461+  output -= 4;
 8462+  #endif
 8463+
 8464+  // do the remnants
 8465+  #if stbir__coder_min_num < 4
 8466+  STBIR_NO_UNROLL_LOOP_START
 8467+  while( output < end_output )
 8468+  {
 8469+    float f;
 8470+    STBIR_NO_UNROLL(encode);
 8471+    f = encode[stbir__encode_order0] * stbir__max_uint8_as_float + 0.5f; STBIR_CLAMP(f, 0, 255); output[0] = (unsigned char)f;
 8472+    #if stbir__coder_min_num >= 2
 8473+    f = encode[stbir__encode_order1] * stbir__max_uint8_as_float + 0.5f; STBIR_CLAMP(f, 0, 255); output[1] = (unsigned char)f;
 8474+    #endif
 8475+    #if stbir__coder_min_num >= 3
 8476+    f = encode[stbir__encode_order2] * stbir__max_uint8_as_float + 0.5f; STBIR_CLAMP(f, 0, 255); output[2] = (unsigned char)f;
 8477+    #endif
 8478+    output += stbir__coder_min_num;
 8479+    encode += stbir__coder_min_num;
 8480+  }
 8481+  #endif
 8482+  #endif
 8483+}
 8484+
 8485+static float * STBIR__CODER_NAME(stbir__decode_uint8_linear)( float * decodep, int width_times_channels, void const * inputp )
 8486+{
 8487+  float STBIR_STREAMOUT_PTR( * ) decode = decodep;
 8488+  float * decode_end = (float*) decode + width_times_channels;
 8489+  unsigned char const * input = (unsigned char const*)inputp;
 8490+
 8491+  #ifdef STBIR_SIMD
 8492+  unsigned char const * end_input_m16 = input + width_times_channels - 16;
 8493+  if ( width_times_channels >= 16 )
 8494+  {
 8495+    decode_end -= 16;
 8496+    STBIR_NO_UNROLL_LOOP_START_INF_FOR
 8497+    for(;;)
 8498+    {
 8499+      #ifdef STBIR_SIMD8
 8500+      stbir__simdi i; stbir__simdi8 o0,o1;
 8501+      stbir__simdf8 of0, of1;
 8502+      STBIR_NO_UNROLL(decode);
 8503+      stbir__simdi_load( i, input );
 8504+      stbir__simdi8_expand_u8_to_u32( o0, o1, i );
 8505+      stbir__simdi8_convert_i32_to_float( of0, o0 );
 8506+      stbir__simdi8_convert_i32_to_float( of1, o1 );
 8507+      stbir__decode_simdf8_flip( of0 );
 8508+      stbir__decode_simdf8_flip( of1 );
 8509+      stbir__simdf8_store( decode + 0, of0 );
 8510+      stbir__simdf8_store( decode + 8, of1 );
 8511+      #else
 8512+      stbir__simdi i, o0, o1, o2, o3;
 8513+      stbir__simdf of0, of1, of2, of3;
 8514+      STBIR_NO_UNROLL(decode);
 8515+      stbir__simdi_load( i, input );
 8516+      stbir__simdi_expand_u8_to_u32( o0,o1,o2,o3,i);
 8517+      stbir__simdi_convert_i32_to_float( of0, o0 );
 8518+      stbir__simdi_convert_i32_to_float( of1, o1 );
 8519+      stbir__simdi_convert_i32_to_float( of2, o2 );
 8520+      stbir__simdi_convert_i32_to_float( of3, o3 );
 8521+      stbir__decode_simdf4_flip( of0 );
 8522+      stbir__decode_simdf4_flip( of1 );
 8523+      stbir__decode_simdf4_flip( of2 );
 8524+      stbir__decode_simdf4_flip( of3 );
 8525+      stbir__simdf_store( decode + 0,  of0 );
 8526+      stbir__simdf_store( decode + 4,  of1 );
 8527+      stbir__simdf_store( decode + 8,  of2 );
 8528+      stbir__simdf_store( decode + 12, of3 );
 8529+#endif
 8530+      decode += 16;
 8531+      input += 16;
 8532+      if ( decode <= decode_end )
 8533+        continue;
 8534+      if ( decode == ( decode_end + 16 ) )
 8535+        break;
 8536+      decode = decode_end; // backup and do last couple
 8537+      input = end_input_m16;
 8538+    }
 8539+    return decode_end + 16;
 8540+  }
 8541+  #endif
 8542+
 8543+  // try to do blocks of 4 when you can
 8544+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 8545+  decode += 4;
 8546+  STBIR_SIMD_NO_UNROLL_LOOP_START
 8547+  while( decode <= decode_end )
 8548+  {
 8549+    STBIR_SIMD_NO_UNROLL(decode);
 8550+    decode[0-4] = ((float)(input[stbir__decode_order0]));
 8551+    decode[1-4] = ((float)(input[stbir__decode_order1]));
 8552+    decode[2-4] = ((float)(input[stbir__decode_order2]));
 8553+    decode[3-4] = ((float)(input[stbir__decode_order3]));
 8554+    decode += 4;
 8555+    input += 4;
 8556+  }
 8557+  decode -= 4;
 8558+  #endif
 8559+
 8560+  // do the remnants
 8561+  #if stbir__coder_min_num < 4
 8562+  STBIR_NO_UNROLL_LOOP_START
 8563+  while( decode < decode_end )
 8564+  {
 8565+    STBIR_NO_UNROLL(decode);
 8566+    decode[0] = ((float)(input[stbir__decode_order0]));
 8567+    #if stbir__coder_min_num >= 2
 8568+    decode[1] = ((float)(input[stbir__decode_order1]));
 8569+    #endif
 8570+    #if stbir__coder_min_num >= 3
 8571+    decode[2] = ((float)(input[stbir__decode_order2]));
 8572+    #endif
 8573+    decode += stbir__coder_min_num;
 8574+    input += stbir__coder_min_num;
 8575+  }
 8576+  #endif
 8577+  return decode_end;
 8578+}
 8579+
 8580+static void STBIR__CODER_NAME( stbir__encode_uint8_linear )( void * outputp, int width_times_channels, float const * encode )
 8581+{
 8582+  unsigned char STBIR_SIMD_STREAMOUT_PTR( * ) output = (unsigned char *) outputp;
 8583+  unsigned char * end_output = ( (unsigned char *) output ) + width_times_channels;
 8584+
 8585+  #ifdef STBIR_SIMD
 8586+  if ( width_times_channels >= stbir__simdfX_float_count*2 )
 8587+  {
 8588+    float const * end_encode_m8 = encode + width_times_channels - stbir__simdfX_float_count*2;
 8589+    end_output -= stbir__simdfX_float_count*2;
 8590+    STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 8591+    for(;;)
 8592+    {
 8593+      stbir__simdfX e0, e1;
 8594+      stbir__simdi i;
 8595+      STBIR_SIMD_NO_UNROLL(encode);
 8596+      stbir__simdfX_add_mem( e0, STBIR_simd_point5X, encode );
 8597+      stbir__simdfX_add_mem( e1, STBIR_simd_point5X, encode+stbir__simdfX_float_count );
 8598+      stbir__encode_simdfX_unflip( e0 );
 8599+      stbir__encode_simdfX_unflip( e1 );
 8600+      #ifdef STBIR_SIMD8
 8601+      stbir__simdf8_pack_to_16bytes( i, e0, e1 );
 8602+      stbir__simdi_store( output, i );
 8603+      #else
 8604+      stbir__simdf_pack_to_8bytes( i, e0, e1 );
 8605+      stbir__simdi_store2( output, i );
 8606+      #endif
 8607+      encode += stbir__simdfX_float_count*2;
 8608+      output += stbir__simdfX_float_count*2;
 8609+      if ( output <= end_output )
 8610+        continue;
 8611+      if ( output == ( end_output + stbir__simdfX_float_count*2 ) )
 8612+        break;
 8613+      output = end_output; // backup and do last couple
 8614+      encode = end_encode_m8;
 8615+    }
 8616+    return;
 8617+  }
 8618+
 8619+  // try to do blocks of 4 when you can
 8620+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 8621+  output += 4;
 8622+  STBIR_NO_UNROLL_LOOP_START
 8623+  while( output <= end_output )
 8624+  {
 8625+    stbir__simdf e0;
 8626+    stbir__simdi i0;
 8627+    STBIR_NO_UNROLL(encode);
 8628+    stbir__simdf_load( e0, encode );
 8629+    stbir__simdf_add( e0, STBIR__CONSTF(STBIR_simd_point5), e0 );
 8630+    stbir__encode_simdf4_unflip( e0 );
 8631+    stbir__simdf_pack_to_8bytes( i0, e0, e0 );  // only use first 4
 8632+    *(int*)(output-4) = stbir__simdi_to_int( i0 );
 8633+    output += 4;
 8634+    encode += 4;
 8635+  }
 8636+  output -= 4;
 8637+  #endif
 8638+
 8639+  #else
 8640+
 8641+  // try to do blocks of 4 when you can
 8642+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 8643+  output += 4;
 8644+  while( output <= end_output )
 8645+  {
 8646+    float f;
 8647+    f = encode[stbir__encode_order0] + 0.5f; STBIR_CLAMP(f, 0, 255); output[0-4] = (unsigned char)f;
 8648+    f = encode[stbir__encode_order1] + 0.5f; STBIR_CLAMP(f, 0, 255); output[1-4] = (unsigned char)f;
 8649+    f = encode[stbir__encode_order2] + 0.5f; STBIR_CLAMP(f, 0, 255); output[2-4] = (unsigned char)f;
 8650+    f = encode[stbir__encode_order3] + 0.5f; STBIR_CLAMP(f, 0, 255); output[3-4] = (unsigned char)f;
 8651+    output += 4;
 8652+    encode += 4;
 8653+  }
 8654+  output -= 4;
 8655+  #endif
 8656+
 8657+  #endif
 8658+
 8659+  // do the remnants
 8660+  #if stbir__coder_min_num < 4
 8661+  STBIR_NO_UNROLL_LOOP_START
 8662+  while( output < end_output )
 8663+  {
 8664+    float f;
 8665+    STBIR_NO_UNROLL(encode);
 8666+    f = encode[stbir__encode_order0] + 0.5f; STBIR_CLAMP(f, 0, 255); output[0] = (unsigned char)f;
 8667+    #if stbir__coder_min_num >= 2
 8668+    f = encode[stbir__encode_order1] + 0.5f; STBIR_CLAMP(f, 0, 255); output[1] = (unsigned char)f;
 8669+    #endif
 8670+    #if stbir__coder_min_num >= 3
 8671+    f = encode[stbir__encode_order2] + 0.5f; STBIR_CLAMP(f, 0, 255); output[2] = (unsigned char)f;
 8672+    #endif
 8673+    output += stbir__coder_min_num;
 8674+    encode += stbir__coder_min_num;
 8675+  }
 8676+  #endif
 8677+}
 8678+
 8679+static float * STBIR__CODER_NAME(stbir__decode_uint8_srgb)( float * decodep, int width_times_channels, void const * inputp )
 8680+{
 8681+  float STBIR_STREAMOUT_PTR( * ) decode = decodep;
 8682+  float * decode_end = (float*) decode + width_times_channels;
 8683+  unsigned char const * input = (unsigned char const *)inputp;
 8684+
 8685+  // try to do blocks of 4 when you can
 8686+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 8687+  decode += 4;
 8688+  while( decode <= decode_end )
 8689+  {
 8690+    decode[0-4] = stbir__srgb_uchar_to_linear_float[ input[ stbir__decode_order0 ] ];
 8691+    decode[1-4] = stbir__srgb_uchar_to_linear_float[ input[ stbir__decode_order1 ] ];
 8692+    decode[2-4] = stbir__srgb_uchar_to_linear_float[ input[ stbir__decode_order2 ] ];
 8693+    decode[3-4] = stbir__srgb_uchar_to_linear_float[ input[ stbir__decode_order3 ] ];
 8694+    decode += 4;
 8695+    input += 4;
 8696+  }
 8697+  decode -= 4;
 8698+  #endif
 8699+
 8700+  // do the remnants
 8701+  #if stbir__coder_min_num < 4
 8702+  STBIR_NO_UNROLL_LOOP_START
 8703+  while( decode < decode_end )
 8704+  {
 8705+    STBIR_NO_UNROLL(decode);
 8706+    decode[0] = stbir__srgb_uchar_to_linear_float[ input[ stbir__decode_order0 ] ];
 8707+    #if stbir__coder_min_num >= 2
 8708+    decode[1] = stbir__srgb_uchar_to_linear_float[ input[ stbir__decode_order1 ] ];
 8709+    #endif
 8710+    #if stbir__coder_min_num >= 3
 8711+    decode[2] = stbir__srgb_uchar_to_linear_float[ input[ stbir__decode_order2 ] ];
 8712+    #endif
 8713+    decode += stbir__coder_min_num;
 8714+    input += stbir__coder_min_num;
 8715+  }
 8716+  #endif
 8717+  return decode_end;
 8718+}
 8719+
 8720+#define stbir__min_max_shift20( i, f ) \
 8721+    stbir__simdf_max( f, f, stbir_simdf_casti(STBIR__CONSTI( STBIR_almost_zero )) ); \
 8722+    stbir__simdf_min( f, f, stbir_simdf_casti(STBIR__CONSTI( STBIR_almost_one  )) ); \
 8723+    stbir__simdi_32shr( i, stbir_simdi_castf( f ), 20 );
 8724+
 8725+#define stbir__scale_and_convert( i, f ) \
 8726+    stbir__simdf_madd( f, STBIR__CONSTF( STBIR_simd_point5 ), STBIR__CONSTF( STBIR_max_uint8_as_float ), f ); \
 8727+    stbir__simdf_max( f, f, stbir__simdf_zeroP() ); \
 8728+    stbir__simdf_min( f, f, STBIR__CONSTF( STBIR_max_uint8_as_float ) ); \
 8729+    stbir__simdf_convert_float_to_i32( i, f );
 8730+
 8731+#define stbir__linear_to_srgb_finish( i, f ) \
 8732+{ \
 8733+    stbir__simdi temp;  \
 8734+    stbir__simdi_32shr( temp, stbir_simdi_castf( f ), 12 ) ; \
 8735+    stbir__simdi_and( temp, temp, STBIR__CONSTI(STBIR_mastissa_mask) ); \
 8736+    stbir__simdi_or( temp, temp, STBIR__CONSTI(STBIR_topscale) ); \
 8737+    stbir__simdi_16madd( i, i, temp ); \
 8738+    stbir__simdi_32shr( i, i, 16 ); \
 8739+}
 8740+
 8741+#define stbir__simdi_table_lookup2( v0,v1, table ) \
 8742+{ \
 8743+  stbir__simdi_u32 temp0,temp1; \
 8744+  temp0.m128i_i128 = v0; \
 8745+  temp1.m128i_i128 = v1; \
 8746+  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]]; \
 8747+  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]]; \
 8748+  v0 = temp0.m128i_i128; \
 8749+  v1 = temp1.m128i_i128; \
 8750+}
 8751+
 8752+#define stbir__simdi_table_lookup3( v0,v1,v2, table ) \
 8753+{ \
 8754+  stbir__simdi_u32 temp0,temp1,temp2; \
 8755+  temp0.m128i_i128 = v0; \
 8756+  temp1.m128i_i128 = v1; \
 8757+  temp2.m128i_i128 = v2; \
 8758+  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]]; \
 8759+  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]]; \
 8760+  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]]; \
 8761+  v0 = temp0.m128i_i128; \
 8762+  v1 = temp1.m128i_i128; \
 8763+  v2 = temp2.m128i_i128; \
 8764+}
 8765+
 8766+#define stbir__simdi_table_lookup4( v0,v1,v2,v3, table ) \
 8767+{ \
 8768+  stbir__simdi_u32 temp0,temp1,temp2,temp3; \
 8769+  temp0.m128i_i128 = v0; \
 8770+  temp1.m128i_i128 = v1; \
 8771+  temp2.m128i_i128 = v2; \
 8772+  temp3.m128i_i128 = v3; \
 8773+  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]]; \
 8774+  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]]; \
 8775+  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]]; \
 8776+  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]]; \
 8777+  v0 = temp0.m128i_i128; \
 8778+  v1 = temp1.m128i_i128; \
 8779+  v2 = temp2.m128i_i128; \
 8780+  v3 = temp3.m128i_i128; \
 8781+}
 8782+
 8783+static void STBIR__CODER_NAME( stbir__encode_uint8_srgb )( void * outputp, int width_times_channels, float const * encode )
 8784+{
 8785+  unsigned char STBIR_SIMD_STREAMOUT_PTR( * ) output = (unsigned char*) outputp;
 8786+  unsigned char * end_output = ( (unsigned char*) output ) + width_times_channels;
 8787+
 8788+  #ifdef STBIR_SIMD
 8789+
 8790+  if ( width_times_channels >= 16 )
 8791+  {
 8792+    float const * end_encode_m16 = encode + width_times_channels - 16;
 8793+    end_output -= 16;
 8794+    STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 8795+    for(;;)
 8796+    {
 8797+      stbir__simdf f0, f1, f2, f3;
 8798+      stbir__simdi i0, i1, i2, i3;
 8799+      STBIR_SIMD_NO_UNROLL(encode);
 8800+
 8801+      stbir__simdf_load4_transposed( f0, f1, f2, f3, encode );
 8802+
 8803+      stbir__min_max_shift20( i0, f0 );
 8804+      stbir__min_max_shift20( i1, f1 );
 8805+      stbir__min_max_shift20( i2, f2 );
 8806+      stbir__min_max_shift20( i3, f3 );
 8807+
 8808+      stbir__simdi_table_lookup4( i0, i1, i2, i3, ( fp32_to_srgb8_tab4 - (127-13)*8 ) );
 8809+
 8810+      stbir__linear_to_srgb_finish( i0, f0 );
 8811+      stbir__linear_to_srgb_finish( i1, f1 );
 8812+      stbir__linear_to_srgb_finish( i2, f2 );
 8813+      stbir__linear_to_srgb_finish( i3, f3 );
 8814+
 8815+      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) );
 8816+
 8817+      encode += 16;
 8818+      output += 16;
 8819+      if ( output <= end_output )
 8820+        continue;
 8821+      if ( output == ( end_output + 16 ) )
 8822+        break;
 8823+      output = end_output; // backup and do last couple
 8824+      encode = end_encode_m16;
 8825+    }
 8826+    return;
 8827+  }
 8828+  #endif
 8829+
 8830+  // try to do blocks of 4 when you can
 8831+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 8832+  output += 4;
 8833+  STBIR_SIMD_NO_UNROLL_LOOP_START
 8834+  while ( output <= end_output )
 8835+  {
 8836+    STBIR_SIMD_NO_UNROLL(encode);
 8837+
 8838+    output[0-4] = stbir__linear_to_srgb_uchar( encode[stbir__encode_order0] );
 8839+    output[1-4] = stbir__linear_to_srgb_uchar( encode[stbir__encode_order1] );
 8840+    output[2-4] = stbir__linear_to_srgb_uchar( encode[stbir__encode_order2] );
 8841+    output[3-4] = stbir__linear_to_srgb_uchar( encode[stbir__encode_order3] );
 8842+
 8843+    output += 4;
 8844+    encode += 4;
 8845+  }
 8846+  output -= 4;
 8847+  #endif
 8848+
 8849+  // do the remnants
 8850+  #if stbir__coder_min_num < 4
 8851+  STBIR_NO_UNROLL_LOOP_START
 8852+  while( output < end_output )
 8853+  {
 8854+    STBIR_NO_UNROLL(encode);
 8855+    output[0] = stbir__linear_to_srgb_uchar( encode[stbir__encode_order0] );
 8856+    #if stbir__coder_min_num >= 2
 8857+    output[1] = stbir__linear_to_srgb_uchar( encode[stbir__encode_order1] );
 8858+    #endif
 8859+    #if stbir__coder_min_num >= 3
 8860+    output[2] = stbir__linear_to_srgb_uchar( encode[stbir__encode_order2] );
 8861+    #endif
 8862+    output += stbir__coder_min_num;
 8863+    encode += stbir__coder_min_num;
 8864+  }
 8865+  #endif
 8866+}
 8867+
 8868+#if ( stbir__coder_min_num == 4 ) || ( ( stbir__coder_min_num == 1 ) && ( !defined(stbir__decode_swizzle) ) )
 8869+
 8870+static float * STBIR__CODER_NAME(stbir__decode_uint8_srgb4_linearalpha)( float * decodep, int width_times_channels, void const * inputp )
 8871+{
 8872+  float STBIR_STREAMOUT_PTR( * ) decode = decodep;
 8873+  float * decode_end = (float*) decode + width_times_channels;
 8874+  unsigned char const * input = (unsigned char const *)inputp;
 8875+
 8876+  do {
 8877+    decode[0] = stbir__srgb_uchar_to_linear_float[ input[stbir__decode_order0] ];
 8878+    decode[1] = stbir__srgb_uchar_to_linear_float[ input[stbir__decode_order1] ];
 8879+    decode[2] = stbir__srgb_uchar_to_linear_float[ input[stbir__decode_order2] ];
 8880+    decode[3] = ( (float) input[stbir__decode_order3] ) * stbir__max_uint8_as_float_inverted;
 8881+    input += 4;
 8882+    decode += 4;
 8883+  } while( decode < decode_end );
 8884+  return decode_end;
 8885+}
 8886+
 8887+
 8888+static void STBIR__CODER_NAME( stbir__encode_uint8_srgb4_linearalpha )( void * outputp, int width_times_channels, float const * encode )
 8889+{
 8890+  unsigned char STBIR_SIMD_STREAMOUT_PTR( * ) output = (unsigned char*) outputp;
 8891+  unsigned char * end_output = ( (unsigned char*) output ) + width_times_channels;
 8892+
 8893+  #ifdef STBIR_SIMD
 8894+
 8895+  if ( width_times_channels >= 16 )
 8896+  {
 8897+    float const * end_encode_m16 = encode + width_times_channels - 16;
 8898+    end_output -= 16;
 8899+    STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 8900+    for(;;)
 8901+    {
 8902+      stbir__simdf f0, f1, f2, f3;
 8903+      stbir__simdi i0, i1, i2, i3;
 8904+
 8905+      STBIR_SIMD_NO_UNROLL(encode);
 8906+      stbir__simdf_load4_transposed( f0, f1, f2, f3, encode );
 8907+
 8908+      stbir__min_max_shift20( i0, f0 );
 8909+      stbir__min_max_shift20( i1, f1 );
 8910+      stbir__min_max_shift20( i2, f2 );
 8911+      stbir__scale_and_convert( i3, f3 );
 8912+
 8913+      stbir__simdi_table_lookup3( i0, i1, i2, ( fp32_to_srgb8_tab4 - (127-13)*8 ) );
 8914+
 8915+      stbir__linear_to_srgb_finish( i0, f0 );
 8916+      stbir__linear_to_srgb_finish( i1, f1 );
 8917+      stbir__linear_to_srgb_finish( i2, f2 );
 8918+
 8919+      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) );
 8920+
 8921+      output += 16;
 8922+      encode += 16;
 8923+
 8924+      if ( output <= end_output )
 8925+        continue;
 8926+      if ( output == ( end_output + 16 ) )
 8927+        break;
 8928+      output = end_output; // backup and do last couple
 8929+      encode = end_encode_m16;
 8930+    }
 8931+    return;
 8932+  }
 8933+  #endif
 8934+
 8935+  STBIR_SIMD_NO_UNROLL_LOOP_START
 8936+  do {
 8937+    float f;
 8938+    STBIR_SIMD_NO_UNROLL(encode);
 8939+
 8940+    output[stbir__decode_order0] = stbir__linear_to_srgb_uchar( encode[0] );
 8941+    output[stbir__decode_order1] = stbir__linear_to_srgb_uchar( encode[1] );
 8942+    output[stbir__decode_order2] = stbir__linear_to_srgb_uchar( encode[2] );
 8943+
 8944+    f = encode[3] * stbir__max_uint8_as_float + 0.5f;
 8945+    STBIR_CLAMP(f, 0, 255);
 8946+    output[stbir__decode_order3] = (unsigned char) f;
 8947+
 8948+    output += 4;
 8949+    encode += 4;
 8950+  } while( output < end_output );
 8951+}
 8952+
 8953+#endif
 8954+
 8955+#if ( stbir__coder_min_num == 2 ) || ( ( stbir__coder_min_num == 1 ) && ( !defined(stbir__decode_swizzle) ) )
 8956+
 8957+static float * STBIR__CODER_NAME(stbir__decode_uint8_srgb2_linearalpha)( float * decodep, int width_times_channels, void const * inputp )
 8958+{
 8959+  float STBIR_STREAMOUT_PTR( * ) decode = decodep;
 8960+  float * decode_end = (float*) decode + width_times_channels;
 8961+  unsigned char const * input = (unsigned char const *)inputp;
 8962+
 8963+  decode += 4;
 8964+  while( decode <= decode_end )
 8965+  {
 8966+    decode[0-4] = stbir__srgb_uchar_to_linear_float[ input[stbir__decode_order0] ];
 8967+    decode[1-4] = ( (float) input[stbir__decode_order1] ) * stbir__max_uint8_as_float_inverted;
 8968+    decode[2-4] = stbir__srgb_uchar_to_linear_float[ input[stbir__decode_order0+2] ];
 8969+    decode[3-4] = ( (float) input[stbir__decode_order1+2] ) * stbir__max_uint8_as_float_inverted;
 8970+    input += 4;
 8971+    decode += 4;
 8972+  }
 8973+  decode -= 4;
 8974+  if( decode < decode_end )
 8975+  {
 8976+    decode[0] = stbir__srgb_uchar_to_linear_float[ stbir__decode_order0 ];
 8977+    decode[1] = ( (float) input[stbir__decode_order1] ) * stbir__max_uint8_as_float_inverted;
 8978+  }
 8979+  return decode_end;
 8980+}
 8981+
 8982+static void STBIR__CODER_NAME( stbir__encode_uint8_srgb2_linearalpha )( void * outputp, int width_times_channels, float const * encode )
 8983+{
 8984+  unsigned char STBIR_SIMD_STREAMOUT_PTR( * ) output = (unsigned char*) outputp;
 8985+  unsigned char * end_output = ( (unsigned char*) output ) + width_times_channels;
 8986+
 8987+  #ifdef STBIR_SIMD
 8988+
 8989+  if ( width_times_channels >= 16 )
 8990+  {
 8991+    float const * end_encode_m16 = encode + width_times_channels - 16;
 8992+    end_output -= 16;
 8993+    STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 8994+    for(;;)
 8995+    {
 8996+      stbir__simdf f0, f1, f2, f3;
 8997+      stbir__simdi i0, i1, i2, i3;
 8998+
 8999+      STBIR_SIMD_NO_UNROLL(encode);
 9000+      stbir__simdf_load4_transposed( f0, f1, f2, f3, encode );
 9001+
 9002+      stbir__min_max_shift20( i0, f0 );
 9003+      stbir__scale_and_convert( i1, f1 );
 9004+      stbir__min_max_shift20( i2, f2 );
 9005+      stbir__scale_and_convert( i3, f3 );
 9006+
 9007+      stbir__simdi_table_lookup2( i0, i2, ( fp32_to_srgb8_tab4 - (127-13)*8 ) );
 9008+
 9009+      stbir__linear_to_srgb_finish( i0, f0 );
 9010+      stbir__linear_to_srgb_finish( i2, f2 );
 9011+
 9012+      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) );
 9013+
 9014+      output += 16;
 9015+      encode += 16;
 9016+      if ( output <= end_output )
 9017+        continue;
 9018+      if ( output == ( end_output + 16 ) )
 9019+        break;
 9020+      output = end_output; // backup and do last couple
 9021+      encode = end_encode_m16;
 9022+    }
 9023+    return;
 9024+  }
 9025+  #endif
 9026+
 9027+  STBIR_SIMD_NO_UNROLL_LOOP_START
 9028+  do {
 9029+    float f;
 9030+    STBIR_SIMD_NO_UNROLL(encode);
 9031+
 9032+    output[stbir__decode_order0] = stbir__linear_to_srgb_uchar( encode[0] );
 9033+
 9034+    f = encode[1] * stbir__max_uint8_as_float + 0.5f;
 9035+    STBIR_CLAMP(f, 0, 255);
 9036+    output[stbir__decode_order1] = (unsigned char) f;
 9037+
 9038+    output += 2;
 9039+    encode += 2;
 9040+  } while( output < end_output );
 9041+}
 9042+
 9043+#endif
 9044+
 9045+static float * STBIR__CODER_NAME(stbir__decode_uint16_linear_scaled)( float * decodep, int width_times_channels, void const * inputp )
 9046+{
 9047+  float STBIR_STREAMOUT_PTR( * ) decode = decodep;
 9048+  float * decode_end = (float*) decode + width_times_channels;
 9049+  unsigned short const * input = (unsigned short const *)inputp;
 9050+
 9051+  #ifdef STBIR_SIMD
 9052+  unsigned short const * end_input_m8 = input + width_times_channels - 8;
 9053+  if ( width_times_channels >= 8 )
 9054+  {
 9055+    decode_end -= 8;
 9056+    STBIR_NO_UNROLL_LOOP_START_INF_FOR
 9057+    for(;;)
 9058+    {
 9059+      #ifdef STBIR_SIMD8
 9060+      stbir__simdi i; stbir__simdi8 o;
 9061+      stbir__simdf8 of;
 9062+      STBIR_NO_UNROLL(decode);
 9063+      stbir__simdi_load( i, input );
 9064+      stbir__simdi8_expand_u16_to_u32( o, i );
 9065+      stbir__simdi8_convert_i32_to_float( of, o );
 9066+      stbir__simdf8_mult( of, of, STBIR_max_uint16_as_float_inverted8);
 9067+      stbir__decode_simdf8_flip( of );
 9068+      stbir__simdf8_store( decode + 0, of );
 9069+      #else
 9070+      stbir__simdi i, o0, o1;
 9071+      stbir__simdf of0, of1;
 9072+      STBIR_NO_UNROLL(decode);
 9073+      stbir__simdi_load( i, input );
 9074+      stbir__simdi_expand_u16_to_u32( o0,o1,i );
 9075+      stbir__simdi_convert_i32_to_float( of0, o0 );
 9076+      stbir__simdi_convert_i32_to_float( of1, o1 );
 9077+      stbir__simdf_mult( of0, of0, STBIR__CONSTF(STBIR_max_uint16_as_float_inverted) );
 9078+      stbir__simdf_mult( of1, of1, STBIR__CONSTF(STBIR_max_uint16_as_float_inverted));
 9079+      stbir__decode_simdf4_flip( of0 );
 9080+      stbir__decode_simdf4_flip( of1 );
 9081+      stbir__simdf_store( decode + 0,  of0 );
 9082+      stbir__simdf_store( decode + 4,  of1 );
 9083+      #endif
 9084+      decode += 8;
 9085+      input += 8;
 9086+      if ( decode <= decode_end )
 9087+        continue;
 9088+      if ( decode == ( decode_end + 8 ) )
 9089+        break;
 9090+      decode = decode_end; // backup and do last couple
 9091+      input = end_input_m8;
 9092+    }
 9093+    return decode_end + 8;
 9094+  }
 9095+  #endif
 9096+
 9097+  // try to do blocks of 4 when you can
 9098+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9099+  decode += 4;
 9100+  STBIR_SIMD_NO_UNROLL_LOOP_START
 9101+  while( decode <= decode_end )
 9102+  {
 9103+    STBIR_SIMD_NO_UNROLL(decode);
 9104+    decode[0-4] = ((float)(input[stbir__decode_order0])) * stbir__max_uint16_as_float_inverted;
 9105+    decode[1-4] = ((float)(input[stbir__decode_order1])) * stbir__max_uint16_as_float_inverted;
 9106+    decode[2-4] = ((float)(input[stbir__decode_order2])) * stbir__max_uint16_as_float_inverted;
 9107+    decode[3-4] = ((float)(input[stbir__decode_order3])) * stbir__max_uint16_as_float_inverted;
 9108+    decode += 4;
 9109+    input += 4;
 9110+  }
 9111+  decode -= 4;
 9112+  #endif
 9113+
 9114+  // do the remnants
 9115+  #if stbir__coder_min_num < 4
 9116+  STBIR_NO_UNROLL_LOOP_START
 9117+  while( decode < decode_end )
 9118+  {
 9119+    STBIR_NO_UNROLL(decode);
 9120+    decode[0] = ((float)(input[stbir__decode_order0])) * stbir__max_uint16_as_float_inverted;
 9121+    #if stbir__coder_min_num >= 2
 9122+    decode[1] = ((float)(input[stbir__decode_order1])) * stbir__max_uint16_as_float_inverted;
 9123+    #endif
 9124+    #if stbir__coder_min_num >= 3
 9125+    decode[2] = ((float)(input[stbir__decode_order2])) * stbir__max_uint16_as_float_inverted;
 9126+    #endif
 9127+    decode += stbir__coder_min_num;
 9128+    input += stbir__coder_min_num;
 9129+  }
 9130+  #endif
 9131+  return decode_end;
 9132+}
 9133+
 9134+
 9135+static void STBIR__CODER_NAME(stbir__encode_uint16_linear_scaled)( void * outputp, int width_times_channels, float const * encode )
 9136+{
 9137+  unsigned short STBIR_SIMD_STREAMOUT_PTR( * ) output = (unsigned short*) outputp;
 9138+  unsigned short * end_output = ( (unsigned short*) output ) + width_times_channels;
 9139+
 9140+  #ifdef STBIR_SIMD
 9141+  {
 9142+    if ( width_times_channels >= stbir__simdfX_float_count*2 )
 9143+    {
 9144+      float const * end_encode_m8 = encode + width_times_channels - stbir__simdfX_float_count*2;
 9145+      end_output -= stbir__simdfX_float_count*2;
 9146+      STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 9147+      for(;;)
 9148+      {
 9149+        stbir__simdfX e0, e1;
 9150+        stbir__simdiX i;
 9151+        STBIR_SIMD_NO_UNROLL(encode);
 9152+        stbir__simdfX_madd_mem( e0, STBIR_simd_point5X, STBIR_max_uint16_as_floatX, encode );
 9153+        stbir__simdfX_madd_mem( e1, STBIR_simd_point5X, STBIR_max_uint16_as_floatX, encode+stbir__simdfX_float_count );
 9154+        stbir__encode_simdfX_unflip( e0 );
 9155+        stbir__encode_simdfX_unflip( e1 );
 9156+        stbir__simdfX_pack_to_words( i, e0, e1 );
 9157+        stbir__simdiX_store( output, i );
 9158+        encode += stbir__simdfX_float_count*2;
 9159+        output += stbir__simdfX_float_count*2;
 9160+        if ( output <= end_output )
 9161+          continue;
 9162+        if ( output == ( end_output + stbir__simdfX_float_count*2 ) )
 9163+          break;
 9164+        output = end_output;     // backup and do last couple
 9165+        encode = end_encode_m8;
 9166+      }
 9167+      return;
 9168+    }
 9169+  }
 9170+
 9171+  // try to do blocks of 4 when you can
 9172+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9173+  output += 4;
 9174+  STBIR_NO_UNROLL_LOOP_START
 9175+  while( output <= end_output )
 9176+  {
 9177+    stbir__simdf e;
 9178+    stbir__simdi i;
 9179+    STBIR_NO_UNROLL(encode);
 9180+    stbir__simdf_load( e, encode );
 9181+    stbir__simdf_madd( e, STBIR__CONSTF(STBIR_simd_point5), STBIR__CONSTF(STBIR_max_uint16_as_float), e );
 9182+    stbir__encode_simdf4_unflip( e );
 9183+    stbir__simdf_pack_to_8words( i, e, e );  // only use first 4
 9184+    stbir__simdi_store2( output-4, i );
 9185+    output += 4;
 9186+    encode += 4;
 9187+  }
 9188+  output -= 4;
 9189+  #endif
 9190+
 9191+  // do the remnants
 9192+  #if stbir__coder_min_num < 4
 9193+  STBIR_NO_UNROLL_LOOP_START
 9194+  while( output < end_output )
 9195+  {
 9196+    stbir__simdf e;
 9197+    STBIR_NO_UNROLL(encode);
 9198+    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 );
 9199+    #if stbir__coder_min_num >= 2
 9200+    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 );
 9201+    #endif
 9202+    #if stbir__coder_min_num >= 3
 9203+    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 );
 9204+    #endif
 9205+    output += stbir__coder_min_num;
 9206+    encode += stbir__coder_min_num;
 9207+  }
 9208+  #endif
 9209+
 9210+  #else
 9211+
 9212+  // try to do blocks of 4 when you can
 9213+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9214+  output += 4;
 9215+  STBIR_SIMD_NO_UNROLL_LOOP_START
 9216+  while( output <= end_output )
 9217+  {
 9218+    float f;
 9219+    STBIR_SIMD_NO_UNROLL(encode);
 9220+    f = encode[stbir__encode_order0] * stbir__max_uint16_as_float + 0.5f; STBIR_CLAMP(f, 0, 65535); output[0-4] = (unsigned short)f;
 9221+    f = encode[stbir__encode_order1] * stbir__max_uint16_as_float + 0.5f; STBIR_CLAMP(f, 0, 65535); output[1-4] = (unsigned short)f;
 9222+    f = encode[stbir__encode_order2] * stbir__max_uint16_as_float + 0.5f; STBIR_CLAMP(f, 0, 65535); output[2-4] = (unsigned short)f;
 9223+    f = encode[stbir__encode_order3] * stbir__max_uint16_as_float + 0.5f; STBIR_CLAMP(f, 0, 65535); output[3-4] = (unsigned short)f;
 9224+    output += 4;
 9225+    encode += 4;
 9226+  }
 9227+  output -= 4;
 9228+  #endif
 9229+
 9230+  // do the remnants
 9231+  #if stbir__coder_min_num < 4
 9232+  STBIR_NO_UNROLL_LOOP_START
 9233+  while( output < end_output )
 9234+  {
 9235+    float f;
 9236+    STBIR_NO_UNROLL(encode);
 9237+    f = encode[stbir__encode_order0] * stbir__max_uint16_as_float + 0.5f; STBIR_CLAMP(f, 0, 65535); output[0] = (unsigned short)f;
 9238+    #if stbir__coder_min_num >= 2
 9239+    f = encode[stbir__encode_order1] * stbir__max_uint16_as_float + 0.5f; STBIR_CLAMP(f, 0, 65535); output[1] = (unsigned short)f;
 9240+    #endif
 9241+    #if stbir__coder_min_num >= 3
 9242+    f = encode[stbir__encode_order2] * stbir__max_uint16_as_float + 0.5f; STBIR_CLAMP(f, 0, 65535); output[2] = (unsigned short)f;
 9243+    #endif
 9244+    output += stbir__coder_min_num;
 9245+    encode += stbir__coder_min_num;
 9246+  }
 9247+  #endif
 9248+  #endif
 9249+}
 9250+
 9251+static float * STBIR__CODER_NAME(stbir__decode_uint16_linear)( float * decodep, int width_times_channels, void const * inputp )
 9252+{
 9253+  float STBIR_STREAMOUT_PTR( * ) decode = decodep;
 9254+  float * decode_end = (float*) decode + width_times_channels;
 9255+  unsigned short const * input = (unsigned short const *)inputp;
 9256+
 9257+  #ifdef STBIR_SIMD
 9258+  unsigned short const * end_input_m8 = input + width_times_channels - 8;
 9259+  if ( width_times_channels >= 8 )
 9260+  {
 9261+    decode_end -= 8;
 9262+    STBIR_NO_UNROLL_LOOP_START_INF_FOR
 9263+    for(;;)
 9264+    {
 9265+      #ifdef STBIR_SIMD8
 9266+      stbir__simdi i; stbir__simdi8 o;
 9267+      stbir__simdf8 of;
 9268+      STBIR_NO_UNROLL(decode);
 9269+      stbir__simdi_load( i, input );
 9270+      stbir__simdi8_expand_u16_to_u32( o, i );
 9271+      stbir__simdi8_convert_i32_to_float( of, o );
 9272+      stbir__decode_simdf8_flip( of );
 9273+      stbir__simdf8_store( decode + 0, of );
 9274+      #else
 9275+      stbir__simdi i, o0, o1;
 9276+      stbir__simdf of0, of1;
 9277+      STBIR_NO_UNROLL(decode);
 9278+      stbir__simdi_load( i, input );
 9279+      stbir__simdi_expand_u16_to_u32( o0, o1, i );
 9280+      stbir__simdi_convert_i32_to_float( of0, o0 );
 9281+      stbir__simdi_convert_i32_to_float( of1, o1 );
 9282+      stbir__decode_simdf4_flip( of0 );
 9283+      stbir__decode_simdf4_flip( of1 );
 9284+      stbir__simdf_store( decode + 0,  of0 );
 9285+      stbir__simdf_store( decode + 4,  of1 );
 9286+      #endif
 9287+      decode += 8;
 9288+      input += 8;
 9289+      if ( decode <= decode_end )
 9290+        continue;
 9291+      if ( decode == ( decode_end + 8 ) )
 9292+        break;
 9293+      decode = decode_end; // backup and do last couple
 9294+      input = end_input_m8;
 9295+    }
 9296+    return decode_end + 8;
 9297+  }
 9298+  #endif
 9299+
 9300+  // try to do blocks of 4 when you can
 9301+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9302+  decode += 4;
 9303+  STBIR_SIMD_NO_UNROLL_LOOP_START
 9304+  while( decode <= decode_end )
 9305+  {
 9306+    STBIR_SIMD_NO_UNROLL(decode);
 9307+    decode[0-4] = ((float)(input[stbir__decode_order0]));
 9308+    decode[1-4] = ((float)(input[stbir__decode_order1]));
 9309+    decode[2-4] = ((float)(input[stbir__decode_order2]));
 9310+    decode[3-4] = ((float)(input[stbir__decode_order3]));
 9311+    decode += 4;
 9312+    input += 4;
 9313+  }
 9314+  decode -= 4;
 9315+  #endif
 9316+
 9317+  // do the remnants
 9318+  #if stbir__coder_min_num < 4
 9319+  STBIR_NO_UNROLL_LOOP_START
 9320+  while( decode < decode_end )
 9321+  {
 9322+    STBIR_NO_UNROLL(decode);
 9323+    decode[0] = ((float)(input[stbir__decode_order0]));
 9324+    #if stbir__coder_min_num >= 2
 9325+    decode[1] = ((float)(input[stbir__decode_order1]));
 9326+    #endif
 9327+    #if stbir__coder_min_num >= 3
 9328+    decode[2] = ((float)(input[stbir__decode_order2]));
 9329+    #endif
 9330+    decode += stbir__coder_min_num;
 9331+    input += stbir__coder_min_num;
 9332+  }
 9333+  #endif
 9334+  return decode_end;
 9335+}
 9336+
 9337+static void STBIR__CODER_NAME(stbir__encode_uint16_linear)( void * outputp, int width_times_channels, float const * encode )
 9338+{
 9339+  unsigned short STBIR_SIMD_STREAMOUT_PTR( * ) output = (unsigned short*) outputp;
 9340+  unsigned short * end_output = ( (unsigned short*) output ) + width_times_channels;
 9341+
 9342+  #ifdef STBIR_SIMD
 9343+  {
 9344+    if ( width_times_channels >= stbir__simdfX_float_count*2 )
 9345+    {
 9346+      float const * end_encode_m8 = encode + width_times_channels - stbir__simdfX_float_count*2;
 9347+      end_output -= stbir__simdfX_float_count*2;
 9348+      STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 9349+      for(;;)
 9350+      {
 9351+        stbir__simdfX e0, e1;
 9352+        stbir__simdiX i;
 9353+        STBIR_SIMD_NO_UNROLL(encode);
 9354+        stbir__simdfX_add_mem( e0, STBIR_simd_point5X, encode );
 9355+        stbir__simdfX_add_mem( e1, STBIR_simd_point5X, encode+stbir__simdfX_float_count );
 9356+        stbir__encode_simdfX_unflip( e0 );
 9357+        stbir__encode_simdfX_unflip( e1 );
 9358+        stbir__simdfX_pack_to_words( i, e0, e1 );
 9359+        stbir__simdiX_store( output, i );
 9360+        encode += stbir__simdfX_float_count*2;
 9361+        output += stbir__simdfX_float_count*2;
 9362+        if ( output <= end_output )
 9363+          continue;
 9364+        if ( output == ( end_output + stbir__simdfX_float_count*2 ) )
 9365+          break;
 9366+        output = end_output; // backup and do last couple
 9367+        encode = end_encode_m8;
 9368+      }
 9369+      return;
 9370+    }
 9371+  }
 9372+
 9373+  // try to do blocks of 4 when you can
 9374+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9375+  output += 4;
 9376+  STBIR_NO_UNROLL_LOOP_START
 9377+  while( output <= end_output )
 9378+  {
 9379+    stbir__simdf e;
 9380+    stbir__simdi i;
 9381+    STBIR_NO_UNROLL(encode);
 9382+    stbir__simdf_load( e, encode );
 9383+    stbir__simdf_add( e, STBIR__CONSTF(STBIR_simd_point5), e );
 9384+    stbir__encode_simdf4_unflip( e );
 9385+    stbir__simdf_pack_to_8words( i, e, e );  // only use first 4
 9386+    stbir__simdi_store2( output-4, i );
 9387+    output += 4;
 9388+    encode += 4;
 9389+  }
 9390+  output -= 4;
 9391+  #endif
 9392+
 9393+  #else
 9394+
 9395+  // try to do blocks of 4 when you can
 9396+  #if  stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9397+  output += 4;
 9398+  STBIR_SIMD_NO_UNROLL_LOOP_START
 9399+  while( output <= end_output )
 9400+  {
 9401+    float f;
 9402+    STBIR_SIMD_NO_UNROLL(encode);
 9403+    f = encode[stbir__encode_order0] + 0.5f; STBIR_CLAMP(f, 0, 65535); output[0-4] = (unsigned short)f;
 9404+    f = encode[stbir__encode_order1] + 0.5f; STBIR_CLAMP(f, 0, 65535); output[1-4] = (unsigned short)f;
 9405+    f = encode[stbir__encode_order2] + 0.5f; STBIR_CLAMP(f, 0, 65535); output[2-4] = (unsigned short)f;
 9406+    f = encode[stbir__encode_order3] + 0.5f; STBIR_CLAMP(f, 0, 65535); output[3-4] = (unsigned short)f;
 9407+    output += 4;
 9408+    encode += 4;
 9409+  }
 9410+  output -= 4;
 9411+  #endif
 9412+
 9413+  #endif
 9414+
 9415+  // do the remnants
 9416+  #if stbir__coder_min_num < 4
 9417+  STBIR_NO_UNROLL_LOOP_START
 9418+  while( output < end_output )
 9419+  {
 9420+    float f;
 9421+    STBIR_NO_UNROLL(encode);
 9422+    f = encode[stbir__encode_order0] + 0.5f; STBIR_CLAMP(f, 0, 65535); output[0] = (unsigned short)f;
 9423+    #if stbir__coder_min_num >= 2
 9424+    f = encode[stbir__encode_order1] + 0.5f; STBIR_CLAMP(f, 0, 65535); output[1] = (unsigned short)f;
 9425+    #endif
 9426+    #if stbir__coder_min_num >= 3
 9427+    f = encode[stbir__encode_order2] + 0.5f; STBIR_CLAMP(f, 0, 65535); output[2] = (unsigned short)f;
 9428+    #endif
 9429+    output += stbir__coder_min_num;
 9430+    encode += stbir__coder_min_num;
 9431+  }
 9432+  #endif
 9433+}
 9434+
 9435+static float * STBIR__CODER_NAME(stbir__decode_half_float_linear)( float * decodep, int width_times_channels, void const * inputp )
 9436+{
 9437+  float STBIR_STREAMOUT_PTR( * ) decode = decodep;
 9438+  float * decode_end = (float*) decode + width_times_channels;
 9439+  stbir__FP16 const * input = (stbir__FP16 const *)inputp;
 9440+
 9441+  #ifdef STBIR_SIMD
 9442+  if ( width_times_channels >= 8 )
 9443+  {
 9444+    stbir__FP16 const * end_input_m8 = input + width_times_channels - 8;
 9445+    decode_end -= 8;
 9446+    STBIR_NO_UNROLL_LOOP_START_INF_FOR
 9447+    for(;;)
 9448+    {
 9449+      STBIR_NO_UNROLL(decode);
 9450+
 9451+      stbir__half_to_float_SIMD( decode, input );
 9452+      #ifdef stbir__decode_swizzle
 9453+      #ifdef STBIR_SIMD8
 9454+      {
 9455+        stbir__simdf8 of;
 9456+        stbir__simdf8_load( of, decode );
 9457+        stbir__decode_simdf8_flip( of );
 9458+        stbir__simdf8_store( decode, of );
 9459+      }
 9460+      #else
 9461+      {
 9462+        stbir__simdf of0,of1;
 9463+        stbir__simdf_load( of0, decode );
 9464+        stbir__simdf_load( of1, decode+4 );
 9465+        stbir__decode_simdf4_flip( of0 );
 9466+        stbir__decode_simdf4_flip( of1 );
 9467+        stbir__simdf_store( decode, of0 );
 9468+        stbir__simdf_store( decode+4, of1 );
 9469+      }
 9470+      #endif
 9471+      #endif
 9472+      decode += 8;
 9473+      input += 8;
 9474+      if ( decode <= decode_end )
 9475+        continue;
 9476+      if ( decode == ( decode_end + 8 ) )
 9477+        break;
 9478+      decode = decode_end; // backup and do last couple
 9479+      input = end_input_m8;
 9480+    }
 9481+    return decode_end + 8;
 9482+  }
 9483+  #endif
 9484+
 9485+  // try to do blocks of 4 when you can
 9486+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9487+  decode += 4;
 9488+  STBIR_SIMD_NO_UNROLL_LOOP_START
 9489+  while( decode <= decode_end )
 9490+  {
 9491+    STBIR_SIMD_NO_UNROLL(decode);
 9492+    decode[0-4] = stbir__half_to_float(input[stbir__decode_order0]);
 9493+    decode[1-4] = stbir__half_to_float(input[stbir__decode_order1]);
 9494+    decode[2-4] = stbir__half_to_float(input[stbir__decode_order2]);
 9495+    decode[3-4] = stbir__half_to_float(input[stbir__decode_order3]);
 9496+    decode += 4;
 9497+    input += 4;
 9498+  }
 9499+  decode -= 4;
 9500+  #endif
 9501+
 9502+  // do the remnants
 9503+  #if stbir__coder_min_num < 4
 9504+  STBIR_NO_UNROLL_LOOP_START
 9505+  while( decode < decode_end )
 9506+  {
 9507+    STBIR_NO_UNROLL(decode);
 9508+    decode[0] = stbir__half_to_float(input[stbir__decode_order0]);
 9509+    #if stbir__coder_min_num >= 2
 9510+    decode[1] = stbir__half_to_float(input[stbir__decode_order1]);
 9511+    #endif
 9512+    #if stbir__coder_min_num >= 3
 9513+    decode[2] = stbir__half_to_float(input[stbir__decode_order2]);
 9514+    #endif
 9515+    decode += stbir__coder_min_num;
 9516+    input += stbir__coder_min_num;
 9517+  }
 9518+  #endif
 9519+  return decode_end;
 9520+}
 9521+
 9522+static void STBIR__CODER_NAME( stbir__encode_half_float_linear )( void * outputp, int width_times_channels, float const * encode )
 9523+{
 9524+  stbir__FP16 STBIR_SIMD_STREAMOUT_PTR( * ) output = (stbir__FP16*) outputp;
 9525+  stbir__FP16 * end_output = ( (stbir__FP16*) output ) + width_times_channels;
 9526+
 9527+  #ifdef STBIR_SIMD
 9528+  if ( width_times_channels >= 8 )
 9529+  {
 9530+    float const * end_encode_m8 = encode + width_times_channels - 8;
 9531+    end_output -= 8;
 9532+    STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 9533+    for(;;)
 9534+    {
 9535+      STBIR_SIMD_NO_UNROLL(encode);
 9536+      #ifdef stbir__decode_swizzle
 9537+      #ifdef STBIR_SIMD8
 9538+      {
 9539+        stbir__simdf8 of;
 9540+        stbir__simdf8_load( of, encode );
 9541+        stbir__encode_simdf8_unflip( of );
 9542+        stbir__float_to_half_SIMD( output, (float*)&of );
 9543+      }
 9544+      #else
 9545+      {
 9546+        stbir__simdf of[2];
 9547+        stbir__simdf_load( of[0], encode );
 9548+        stbir__simdf_load( of[1], encode+4 );
 9549+        stbir__encode_simdf4_unflip( of[0] );
 9550+        stbir__encode_simdf4_unflip( of[1] );
 9551+        stbir__float_to_half_SIMD( output, (float*)of );
 9552+      }
 9553+      #endif
 9554+      #else
 9555+      stbir__float_to_half_SIMD( output, encode );
 9556+      #endif
 9557+      encode += 8;
 9558+      output += 8;
 9559+      if ( output <= end_output )
 9560+        continue;
 9561+      if ( output == ( end_output + 8 ) )
 9562+        break;
 9563+      output = end_output; // backup and do last couple
 9564+      encode = end_encode_m8;
 9565+    }
 9566+    return;
 9567+  }
 9568+  #endif
 9569+
 9570+  // try to do blocks of 4 when you can
 9571+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9572+  output += 4;
 9573+  STBIR_SIMD_NO_UNROLL_LOOP_START
 9574+  while( output <= end_output )
 9575+  {
 9576+    STBIR_SIMD_NO_UNROLL(output);
 9577+    output[0-4] = stbir__float_to_half(encode[stbir__encode_order0]);
 9578+    output[1-4] = stbir__float_to_half(encode[stbir__encode_order1]);
 9579+    output[2-4] = stbir__float_to_half(encode[stbir__encode_order2]);
 9580+    output[3-4] = stbir__float_to_half(encode[stbir__encode_order3]);
 9581+    output += 4;
 9582+    encode += 4;
 9583+  }
 9584+  output -= 4;
 9585+  #endif
 9586+
 9587+  // do the remnants
 9588+  #if stbir__coder_min_num < 4
 9589+  STBIR_NO_UNROLL_LOOP_START
 9590+  while( output < end_output )
 9591+  {
 9592+    STBIR_NO_UNROLL(output);
 9593+    output[0] = stbir__float_to_half(encode[stbir__encode_order0]);
 9594+    #if stbir__coder_min_num >= 2
 9595+    output[1] = stbir__float_to_half(encode[stbir__encode_order1]);
 9596+    #endif
 9597+    #if stbir__coder_min_num >= 3
 9598+    output[2] = stbir__float_to_half(encode[stbir__encode_order2]);
 9599+    #endif
 9600+    output += stbir__coder_min_num;
 9601+    encode += stbir__coder_min_num;
 9602+  }
 9603+  #endif
 9604+}
 9605+
 9606+static float * STBIR__CODER_NAME(stbir__decode_float_linear)( float * decodep, int width_times_channels, void const * inputp )
 9607+{
 9608+  #ifdef stbir__decode_swizzle
 9609+  float STBIR_STREAMOUT_PTR( * ) decode = decodep;
 9610+  float * decode_end = (float*) decode + width_times_channels;
 9611+  float const * input = (float const *)inputp;
 9612+
 9613+  #ifdef STBIR_SIMD
 9614+  if ( width_times_channels >= 16 )
 9615+  {
 9616+    float const * end_input_m16 = input + width_times_channels - 16;
 9617+    decode_end -= 16;
 9618+    STBIR_NO_UNROLL_LOOP_START_INF_FOR
 9619+    for(;;)
 9620+    {
 9621+      STBIR_NO_UNROLL(decode);
 9622+      #ifdef stbir__decode_swizzle
 9623+      #ifdef STBIR_SIMD8
 9624+      {
 9625+        stbir__simdf8 of0,of1;
 9626+        stbir__simdf8_load( of0, input );
 9627+        stbir__simdf8_load( of1, input+8 );
 9628+        stbir__decode_simdf8_flip( of0 );
 9629+        stbir__decode_simdf8_flip( of1 );
 9630+        stbir__simdf8_store( decode, of0 );
 9631+        stbir__simdf8_store( decode+8, of1 );
 9632+      }
 9633+      #else
 9634+      {
 9635+        stbir__simdf of0,of1,of2,of3;
 9636+        stbir__simdf_load( of0, input );
 9637+        stbir__simdf_load( of1, input+4 );
 9638+        stbir__simdf_load( of2, input+8 );
 9639+        stbir__simdf_load( of3, input+12 );
 9640+        stbir__decode_simdf4_flip( of0 );
 9641+        stbir__decode_simdf4_flip( of1 );
 9642+        stbir__decode_simdf4_flip( of2 );
 9643+        stbir__decode_simdf4_flip( of3 );
 9644+        stbir__simdf_store( decode, of0 );
 9645+        stbir__simdf_store( decode+4, of1 );
 9646+        stbir__simdf_store( decode+8, of2 );
 9647+        stbir__simdf_store( decode+12, of3 );
 9648+      }
 9649+      #endif
 9650+      #endif
 9651+      decode += 16;
 9652+      input += 16;
 9653+      if ( decode <= decode_end )
 9654+        continue;
 9655+      if ( decode == ( decode_end + 16 ) )
 9656+        break;
 9657+      decode = decode_end; // backup and do last couple
 9658+      input = end_input_m16;
 9659+    }
 9660+    return decode_end + 16;
 9661+  }
 9662+  #endif
 9663+
 9664+  // try to do blocks of 4 when you can
 9665+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9666+  decode += 4;
 9667+  STBIR_SIMD_NO_UNROLL_LOOP_START
 9668+  while( decode <= decode_end )
 9669+  {
 9670+    STBIR_SIMD_NO_UNROLL(decode);
 9671+    decode[0-4] = input[stbir__decode_order0];
 9672+    decode[1-4] = input[stbir__decode_order1];
 9673+    decode[2-4] = input[stbir__decode_order2];
 9674+    decode[3-4] = input[stbir__decode_order3];
 9675+    decode += 4;
 9676+    input += 4;
 9677+  }
 9678+  decode -= 4;
 9679+  #endif
 9680+
 9681+  // do the remnants
 9682+  #if stbir__coder_min_num < 4
 9683+  STBIR_NO_UNROLL_LOOP_START
 9684+  while( decode < decode_end )
 9685+  {
 9686+    STBIR_NO_UNROLL(decode);
 9687+    decode[0] = input[stbir__decode_order0];
 9688+    #if stbir__coder_min_num >= 2
 9689+    decode[1] = input[stbir__decode_order1];
 9690+    #endif
 9691+    #if stbir__coder_min_num >= 3
 9692+    decode[2] = input[stbir__decode_order2];
 9693+    #endif
 9694+    decode += stbir__coder_min_num;
 9695+    input += stbir__coder_min_num;
 9696+  }
 9697+  #endif
 9698+  return decode_end;
 9699+
 9700+  #else
 9701+
 9702+  if ( (void*)decodep != inputp )
 9703+    STBIR_MEMCPY( decodep, inputp, width_times_channels * sizeof( float ) );
 9704+
 9705+  return decodep + width_times_channels;
 9706+
 9707+  #endif
 9708+}
 9709+
 9710+static void STBIR__CODER_NAME( stbir__encode_float_linear )( void * outputp, int width_times_channels, float const * encode )
 9711+{
 9712+  #if !defined( STBIR_FLOAT_HIGH_CLAMP ) && !defined(STBIR_FLOAT_LO_CLAMP) && !defined(stbir__decode_swizzle)
 9713+
 9714+  if ( (void*)outputp != (void*) encode )
 9715+    STBIR_MEMCPY( outputp, encode, width_times_channels * sizeof( float ) );
 9716+
 9717+  #else
 9718+
 9719+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = (float*) outputp;
 9720+  float * end_output = ( (float*) output ) + width_times_channels;
 9721+
 9722+  #ifdef STBIR_FLOAT_HIGH_CLAMP
 9723+  #define stbir_scalar_hi_clamp( v ) if ( v > STBIR_FLOAT_HIGH_CLAMP ) v = STBIR_FLOAT_HIGH_CLAMP;
 9724+  #else
 9725+  #define stbir_scalar_hi_clamp( v )
 9726+  #endif
 9727+  #ifdef STBIR_FLOAT_LOW_CLAMP
 9728+  #define stbir_scalar_lo_clamp( v ) if ( v < STBIR_FLOAT_LOW_CLAMP ) v = STBIR_FLOAT_LOW_CLAMP;
 9729+  #else
 9730+  #define stbir_scalar_lo_clamp( v )
 9731+  #endif
 9732+
 9733+  #ifdef STBIR_SIMD
 9734+
 9735+  #ifdef STBIR_FLOAT_HIGH_CLAMP
 9736+  const stbir__simdfX high_clamp = stbir__simdf_frepX(STBIR_FLOAT_HIGH_CLAMP);
 9737+  #endif
 9738+  #ifdef STBIR_FLOAT_LOW_CLAMP
 9739+  const stbir__simdfX low_clamp = stbir__simdf_frepX(STBIR_FLOAT_LOW_CLAMP);
 9740+  #endif
 9741+
 9742+  if ( width_times_channels >= ( stbir__simdfX_float_count * 2 ) )
 9743+  {
 9744+    float const * end_encode_m8 = encode + width_times_channels - ( stbir__simdfX_float_count * 2 );
 9745+    end_output -= ( stbir__simdfX_float_count * 2 );
 9746+    STBIR_SIMD_NO_UNROLL_LOOP_START_INF_FOR
 9747+    for(;;)
 9748+    {
 9749+      stbir__simdfX e0, e1;
 9750+      STBIR_SIMD_NO_UNROLL(encode);
 9751+      stbir__simdfX_load( e0, encode );
 9752+      stbir__simdfX_load( e1, encode+stbir__simdfX_float_count );
 9753+#ifdef STBIR_FLOAT_HIGH_CLAMP
 9754+      stbir__simdfX_min( e0, e0, high_clamp );
 9755+      stbir__simdfX_min( e1, e1, high_clamp );
 9756+#endif
 9757+#ifdef STBIR_FLOAT_LOW_CLAMP
 9758+      stbir__simdfX_max( e0, e0, low_clamp );
 9759+      stbir__simdfX_max( e1, e1, low_clamp );
 9760+#endif
 9761+      stbir__encode_simdfX_unflip( e0 );
 9762+      stbir__encode_simdfX_unflip( e1 );
 9763+      stbir__simdfX_store( output, e0 );
 9764+      stbir__simdfX_store( output+stbir__simdfX_float_count, e1 );
 9765+      encode += stbir__simdfX_float_count * 2;
 9766+      output += stbir__simdfX_float_count * 2;
 9767+      if ( output < end_output )
 9768+        continue;
 9769+      if ( output == ( end_output + ( stbir__simdfX_float_count * 2 ) ) )
 9770+        break;
 9771+      output = end_output; // backup and do last couple
 9772+      encode = end_encode_m8;
 9773+    }
 9774+    return;
 9775+  }
 9776+
 9777+  // try to do blocks of 4 when you can
 9778+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9779+  output += 4;
 9780+  STBIR_NO_UNROLL_LOOP_START
 9781+  while( output <= end_output )
 9782+  {
 9783+    stbir__simdf e0;
 9784+    STBIR_NO_UNROLL(encode);
 9785+    stbir__simdf_load( e0, encode );
 9786+#ifdef STBIR_FLOAT_HIGH_CLAMP
 9787+    stbir__simdf_min( e0, e0, high_clamp );
 9788+#endif
 9789+#ifdef STBIR_FLOAT_LOW_CLAMP
 9790+    stbir__simdf_max( e0, e0, low_clamp );
 9791+#endif
 9792+    stbir__encode_simdf4_unflip( e0 );
 9793+    stbir__simdf_store( output-4, e0 );
 9794+    output += 4;
 9795+    encode += 4;
 9796+  }
 9797+  output -= 4;
 9798+  #endif
 9799+
 9800+  #else
 9801+
 9802+  // try to do blocks of 4 when you can
 9803+  #if stbir__coder_min_num != 3 // doesn't divide cleanly by four
 9804+  output += 4;
 9805+  STBIR_SIMD_NO_UNROLL_LOOP_START
 9806+  while( output <= end_output )
 9807+  {
 9808+    float e;
 9809+    STBIR_SIMD_NO_UNROLL(encode);
 9810+    e = encode[ stbir__encode_order0 ]; stbir_scalar_hi_clamp( e ); stbir_scalar_lo_clamp( e ); output[0-4] = e;
 9811+    e = encode[ stbir__encode_order1 ]; stbir_scalar_hi_clamp( e ); stbir_scalar_lo_clamp( e ); output[1-4] = e;
 9812+    e = encode[ stbir__encode_order2 ]; stbir_scalar_hi_clamp( e ); stbir_scalar_lo_clamp( e ); output[2-4] = e;
 9813+    e = encode[ stbir__encode_order3 ]; stbir_scalar_hi_clamp( e ); stbir_scalar_lo_clamp( e ); output[3-4] = e;
 9814+    output += 4;
 9815+    encode += 4;
 9816+  }
 9817+  output -= 4;
 9818+
 9819+  #endif
 9820+
 9821+  #endif
 9822+
 9823+  // do the remnants
 9824+  #if stbir__coder_min_num < 4
 9825+  STBIR_NO_UNROLL_LOOP_START
 9826+  while( output < end_output )
 9827+  {
 9828+    float e;
 9829+    STBIR_NO_UNROLL(encode);
 9830+    e = encode[ stbir__encode_order0 ]; stbir_scalar_hi_clamp( e ); stbir_scalar_lo_clamp( e ); output[0] = e;
 9831+    #if stbir__coder_min_num >= 2
 9832+    e = encode[ stbir__encode_order1 ]; stbir_scalar_hi_clamp( e ); stbir_scalar_lo_clamp( e ); output[1] = e;
 9833+    #endif
 9834+    #if stbir__coder_min_num >= 3
 9835+    e = encode[ stbir__encode_order2 ]; stbir_scalar_hi_clamp( e ); stbir_scalar_lo_clamp( e ); output[2] = e;
 9836+    #endif
 9837+    output += stbir__coder_min_num;
 9838+    encode += stbir__coder_min_num;
 9839+  }
 9840+  #endif
 9841+
 9842+  #endif
 9843+}
 9844+
 9845+#undef stbir__decode_suffix
 9846+#undef stbir__decode_simdf8_flip
 9847+#undef stbir__decode_simdf4_flip
 9848+#undef stbir__decode_order0
 9849+#undef stbir__decode_order1
 9850+#undef stbir__decode_order2
 9851+#undef stbir__decode_order3
 9852+#undef stbir__encode_order0
 9853+#undef stbir__encode_order1
 9854+#undef stbir__encode_order2
 9855+#undef stbir__encode_order3
 9856+#undef stbir__encode_simdf8_unflip
 9857+#undef stbir__encode_simdf4_unflip
 9858+#undef stbir__encode_simdfX_unflip
 9859+#undef STBIR__CODER_NAME
 9860+#undef stbir__coder_min_num
 9861+#undef stbir__decode_swizzle
 9862+#undef stbir_scalar_hi_clamp
 9863+#undef stbir_scalar_lo_clamp
 9864+#undef STB_IMAGE_RESIZE_DO_CODERS
 9865+
 9866+#elif defined( STB_IMAGE_RESIZE_DO_VERTICALS)
 9867+
 9868+#ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 9869+#define STBIR_chans( start, end ) STBIR_strs_join14(start,STBIR__vertical_channels,end,_cont)
 9870+#else
 9871+#define STBIR_chans( start, end ) STBIR_strs_join1(start,STBIR__vertical_channels,end)
 9872+#endif
 9873+
 9874+#if STBIR__vertical_channels >= 1
 9875+#define stbIF0( code ) code
 9876+#else
 9877+#define stbIF0( code )
 9878+#endif
 9879+#if STBIR__vertical_channels >= 2
 9880+#define stbIF1( code ) code
 9881+#else
 9882+#define stbIF1( code )
 9883+#endif
 9884+#if STBIR__vertical_channels >= 3
 9885+#define stbIF2( code ) code
 9886+#else
 9887+#define stbIF2( code )
 9888+#endif
 9889+#if STBIR__vertical_channels >= 4
 9890+#define stbIF3( code ) code
 9891+#else
 9892+#define stbIF3( code )
 9893+#endif
 9894+#if STBIR__vertical_channels >= 5
 9895+#define stbIF4( code ) code
 9896+#else
 9897+#define stbIF4( code )
 9898+#endif
 9899+#if STBIR__vertical_channels >= 6
 9900+#define stbIF5( code ) code
 9901+#else
 9902+#define stbIF5( code )
 9903+#endif
 9904+#if STBIR__vertical_channels >= 7
 9905+#define stbIF6( code ) code
 9906+#else
 9907+#define stbIF6( code )
 9908+#endif
 9909+#if STBIR__vertical_channels >= 8
 9910+#define stbIF7( code ) code
 9911+#else
 9912+#define stbIF7( code )
 9913+#endif
 9914+
 9915+static void STBIR_chans( stbir__vertical_scatter_with_,_coeffs)( float ** outputs, float const * vertical_coefficients, float const * input, float const * input_end )
 9916+{
 9917+  stbIF0( float STBIR_SIMD_STREAMOUT_PTR( * ) output0 = outputs[0]; float c0s = vertical_coefficients[0]; )
 9918+  stbIF1( float STBIR_SIMD_STREAMOUT_PTR( * ) output1 = outputs[1]; float c1s = vertical_coefficients[1]; )
 9919+  stbIF2( float STBIR_SIMD_STREAMOUT_PTR( * ) output2 = outputs[2]; float c2s = vertical_coefficients[2]; )
 9920+  stbIF3( float STBIR_SIMD_STREAMOUT_PTR( * ) output3 = outputs[3]; float c3s = vertical_coefficients[3]; )
 9921+  stbIF4( float STBIR_SIMD_STREAMOUT_PTR( * ) output4 = outputs[4]; float c4s = vertical_coefficients[4]; )
 9922+  stbIF5( float STBIR_SIMD_STREAMOUT_PTR( * ) output5 = outputs[5]; float c5s = vertical_coefficients[5]; )
 9923+  stbIF6( float STBIR_SIMD_STREAMOUT_PTR( * ) output6 = outputs[6]; float c6s = vertical_coefficients[6]; )
 9924+  stbIF7( float STBIR_SIMD_STREAMOUT_PTR( * ) output7 = outputs[7]; float c7s = vertical_coefficients[7]; )
 9925+
 9926+  #ifdef STBIR_SIMD
 9927+  {
 9928+    stbIF0(stbir__simdfX c0 = stbir__simdf_frepX( c0s ); )
 9929+    stbIF1(stbir__simdfX c1 = stbir__simdf_frepX( c1s ); )
 9930+    stbIF2(stbir__simdfX c2 = stbir__simdf_frepX( c2s ); )
 9931+    stbIF3(stbir__simdfX c3 = stbir__simdf_frepX( c3s ); )
 9932+    stbIF4(stbir__simdfX c4 = stbir__simdf_frepX( c4s ); )
 9933+    stbIF5(stbir__simdfX c5 = stbir__simdf_frepX( c5s ); )
 9934+    stbIF6(stbir__simdfX c6 = stbir__simdf_frepX( c6s ); )
 9935+    stbIF7(stbir__simdfX c7 = stbir__simdf_frepX( c7s ); )
 9936+    STBIR_SIMD_NO_UNROLL_LOOP_START
 9937+    while ( ( (char*)input_end - (char*) input ) >= (16*stbir__simdfX_float_count) )
 9938+    {
 9939+      stbir__simdfX o0, o1, o2, o3, r0, r1, r2, r3;
 9940+      STBIR_SIMD_NO_UNROLL(output0);
 9941+
 9942+      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) );
 9943+
 9944+      #ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
 9945+      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) );
 9946+              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 );
 9947+              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 ); )
 9948+      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) );
 9949+              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 );
 9950+              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 ); )
 9951+      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) );
 9952+              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 );
 9953+              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 ); )
 9954+      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) );
 9955+              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 );
 9956+              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 ); )
 9957+      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) );
 9958+              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 );
 9959+              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 ); )
 9960+      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) );
 9961+              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 );
 9962+              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 ); )
 9963+      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) );
 9964+              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 );
 9965+              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 ); )
 9966+      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) );
 9967+              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 );
 9968+              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 ); )
 9969+      #else
 9970+      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 );
 9971+              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 ); )
 9972+      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 );
 9973+              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 ); )
 9974+      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 );
 9975+              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 ); )
 9976+      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 );
 9977+              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 ); )
 9978+      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 );
 9979+              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 ); )
 9980+      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 );
 9981+              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 ); )
 9982+      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 );
 9983+              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 ); )
 9984+      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 );
 9985+              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 ); )
 9986+      #endif
 9987+
 9988+      input += (4*stbir__simdfX_float_count);
 9989+      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); )
 9990+    }
 9991+    STBIR_SIMD_NO_UNROLL_LOOP_START
 9992+    while ( ( (char*)input_end - (char*) input ) >= 16 )
 9993+    {
 9994+      stbir__simdf o0, r0;
 9995+      STBIR_SIMD_NO_UNROLL(output0);
 9996+
 9997+      stbir__simdf_load( r0, input );
 9998+
 9999+      #ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
10000+      stbIF0( stbir__simdf_load( o0, output0 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c0 ) );  stbir__simdf_store( output0, o0 ); )
10001+      stbIF1( stbir__simdf_load( o0, output1 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c1 ) );  stbir__simdf_store( output1, o0 ); )
10002+      stbIF2( stbir__simdf_load( o0, output2 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c2 ) );  stbir__simdf_store( output2, o0 ); )
10003+      stbIF3( stbir__simdf_load( o0, output3 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c3 ) );  stbir__simdf_store( output3, o0 ); )
10004+      stbIF4( stbir__simdf_load( o0, output4 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c4 ) );  stbir__simdf_store( output4, o0 ); )
10005+      stbIF5( stbir__simdf_load( o0, output5 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c5 ) );  stbir__simdf_store( output5, o0 ); )
10006+      stbIF6( stbir__simdf_load( o0, output6 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c6 ) );  stbir__simdf_store( output6, o0 ); )
10007+      stbIF7( stbir__simdf_load( o0, output7 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c7 ) );  stbir__simdf_store( output7, o0 ); )
10008+      #else
10009+      stbIF0( stbir__simdf_mult( o0, r0, stbir__if_simdf8_cast_to_simdf4( c0 ) );   stbir__simdf_store( output0, o0 ); )
10010+      stbIF1( stbir__simdf_mult( o0, r0, stbir__if_simdf8_cast_to_simdf4( c1 ) );   stbir__simdf_store( output1, o0 ); )
10011+      stbIF2( stbir__simdf_mult( o0, r0, stbir__if_simdf8_cast_to_simdf4( c2 ) );   stbir__simdf_store( output2, o0 ); )
10012+      stbIF3( stbir__simdf_mult( o0, r0, stbir__if_simdf8_cast_to_simdf4( c3 ) );   stbir__simdf_store( output3, o0 ); )
10013+      stbIF4( stbir__simdf_mult( o0, r0, stbir__if_simdf8_cast_to_simdf4( c4 ) );   stbir__simdf_store( output4, o0 ); )
10014+      stbIF5( stbir__simdf_mult( o0, r0, stbir__if_simdf8_cast_to_simdf4( c5 ) );   stbir__simdf_store( output5, o0 ); )
10015+      stbIF6( stbir__simdf_mult( o0, r0, stbir__if_simdf8_cast_to_simdf4( c6 ) );   stbir__simdf_store( output6, o0 ); )
10016+      stbIF7( stbir__simdf_mult( o0, r0, stbir__if_simdf8_cast_to_simdf4( c7 ) );   stbir__simdf_store( output7, o0 ); )
10017+      #endif
10018+
10019+      input += 4;
10020+      stbIF0( output0 += 4; ) stbIF1( output1 += 4; ) stbIF2( output2 += 4; ) stbIF3( output3 += 4; ) stbIF4( output4 += 4; ) stbIF5( output5 += 4; ) stbIF6( output6 += 4; ) stbIF7( output7 += 4; )
10021+    }
10022+  }
10023+  #else
10024+  STBIR_NO_UNROLL_LOOP_START
10025+  while ( ( (char*)input_end - (char*) input ) >= 16 )
10026+  {
10027+    float r0, r1, r2, r3;
10028+    STBIR_NO_UNROLL(input);
10029+
10030+    r0 = input[0], r1 = input[1], r2 = input[2], r3 = input[3];
10031+
10032+    #ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
10033+    stbIF0( output0[0] += ( r0 * c0s ); output0[1] += ( r1 * c0s ); output0[2] += ( r2 * c0s ); output0[3] += ( r3 * c0s ); )
10034+    stbIF1( output1[0] += ( r0 * c1s ); output1[1] += ( r1 * c1s ); output1[2] += ( r2 * c1s ); output1[3] += ( r3 * c1s ); )
10035+    stbIF2( output2[0] += ( r0 * c2s ); output2[1] += ( r1 * c2s ); output2[2] += ( r2 * c2s ); output2[3] += ( r3 * c2s ); )
10036+    stbIF3( output3[0] += ( r0 * c3s ); output3[1] += ( r1 * c3s ); output3[2] += ( r2 * c3s ); output3[3] += ( r3 * c3s ); )
10037+    stbIF4( output4[0] += ( r0 * c4s ); output4[1] += ( r1 * c4s ); output4[2] += ( r2 * c4s ); output4[3] += ( r3 * c4s ); )
10038+    stbIF5( output5[0] += ( r0 * c5s ); output5[1] += ( r1 * c5s ); output5[2] += ( r2 * c5s ); output5[3] += ( r3 * c5s ); )
10039+    stbIF6( output6[0] += ( r0 * c6s ); output6[1] += ( r1 * c6s ); output6[2] += ( r2 * c6s ); output6[3] += ( r3 * c6s ); )
10040+    stbIF7( output7[0] += ( r0 * c7s ); output7[1] += ( r1 * c7s ); output7[2] += ( r2 * c7s ); output7[3] += ( r3 * c7s ); )
10041+    #else
10042+    stbIF0( output0[0]  = ( r0 * c0s ); output0[1]  = ( r1 * c0s ); output0[2]  = ( r2 * c0s ); output0[3]  = ( r3 * c0s ); )
10043+    stbIF1( output1[0]  = ( r0 * c1s ); output1[1]  = ( r1 * c1s ); output1[2]  = ( r2 * c1s ); output1[3]  = ( r3 * c1s ); )
10044+    stbIF2( output2[0]  = ( r0 * c2s ); output2[1]  = ( r1 * c2s ); output2[2]  = ( r2 * c2s ); output2[3]  = ( r3 * c2s ); )
10045+    stbIF3( output3[0]  = ( r0 * c3s ); output3[1]  = ( r1 * c3s ); output3[2]  = ( r2 * c3s ); output3[3]  = ( r3 * c3s ); )
10046+    stbIF4( output4[0]  = ( r0 * c4s ); output4[1]  = ( r1 * c4s ); output4[2]  = ( r2 * c4s ); output4[3]  = ( r3 * c4s ); )
10047+    stbIF5( output5[0]  = ( r0 * c5s ); output5[1]  = ( r1 * c5s ); output5[2]  = ( r2 * c5s ); output5[3]  = ( r3 * c5s ); )
10048+    stbIF6( output6[0]  = ( r0 * c6s ); output6[1]  = ( r1 * c6s ); output6[2]  = ( r2 * c6s ); output6[3]  = ( r3 * c6s ); )
10049+    stbIF7( output7[0]  = ( r0 * c7s ); output7[1]  = ( r1 * c7s ); output7[2]  = ( r2 * c7s ); output7[3]  = ( r3 * c7s ); )
10050+    #endif
10051+
10052+    input += 4;
10053+    stbIF0( output0 += 4; ) stbIF1( output1 += 4; ) stbIF2( output2 += 4; ) stbIF3( output3 += 4; ) stbIF4( output4 += 4; ) stbIF5( output5 += 4; ) stbIF6( output6 += 4; ) stbIF7( output7 += 4; )
10054+  }
10055+  #endif
10056+  STBIR_NO_UNROLL_LOOP_START
10057+  while ( input < input_end )
10058+  {
10059+    float r = input[0];
10060+    STBIR_NO_UNROLL(output0);
10061+
10062+    #ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
10063+    stbIF0( output0[0] += ( r * c0s ); )
10064+    stbIF1( output1[0] += ( r * c1s ); )
10065+    stbIF2( output2[0] += ( r * c2s ); )
10066+    stbIF3( output3[0] += ( r * c3s ); )
10067+    stbIF4( output4[0] += ( r * c4s ); )
10068+    stbIF5( output5[0] += ( r * c5s ); )
10069+    stbIF6( output6[0] += ( r * c6s ); )
10070+    stbIF7( output7[0] += ( r * c7s ); )
10071+    #else
10072+    stbIF0( output0[0]  = ( r * c0s ); )
10073+    stbIF1( output1[0]  = ( r * c1s ); )
10074+    stbIF2( output2[0]  = ( r * c2s ); )
10075+    stbIF3( output3[0]  = ( r * c3s ); )
10076+    stbIF4( output4[0]  = ( r * c4s ); )
10077+    stbIF5( output5[0]  = ( r * c5s ); )
10078+    stbIF6( output6[0]  = ( r * c6s ); )
10079+    stbIF7( output7[0]  = ( r * c7s ); )
10080+    #endif
10081+
10082+    ++input;
10083+    stbIF0( ++output0; ) stbIF1( ++output1; ) stbIF2( ++output2; ) stbIF3( ++output3; ) stbIF4( ++output4; ) stbIF5( ++output5; ) stbIF6( ++output6; ) stbIF7( ++output7; )
10084+  }
10085+}
10086+
10087+static void STBIR_chans( stbir__vertical_gather_with_,_coeffs)( float * outputp, float const * vertical_coefficients, float const ** inputs, float const * input0_end )
10088+{
10089+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = outputp;
10090+
10091+  stbIF0( float const * input0 = inputs[0]; float c0s = vertical_coefficients[0]; )
10092+  stbIF1( float const * input1 = inputs[1]; float c1s = vertical_coefficients[1]; )
10093+  stbIF2( float const * input2 = inputs[2]; float c2s = vertical_coefficients[2]; )
10094+  stbIF3( float const * input3 = inputs[3]; float c3s = vertical_coefficients[3]; )
10095+  stbIF4( float const * input4 = inputs[4]; float c4s = vertical_coefficients[4]; )
10096+  stbIF5( float const * input5 = inputs[5]; float c5s = vertical_coefficients[5]; )
10097+  stbIF6( float const * input6 = inputs[6]; float c6s = vertical_coefficients[6]; )
10098+  stbIF7( float const * input7 = inputs[7]; float c7s = vertical_coefficients[7]; )
10099+
10100+#if ( STBIR__vertical_channels == 1 ) && !defined(STB_IMAGE_RESIZE_VERTICAL_CONTINUE)
10101+  // check single channel one weight
10102+  if ( ( c0s >= (1.0f-0.000001f) ) && ( c0s <= (1.0f+0.000001f) ) )
10103+  {
10104+    STBIR_MEMCPY( output, input0, (char*)input0_end - (char*)input0 );
10105+    return;
10106+  }
10107+#endif
10108+
10109+  #ifdef STBIR_SIMD
10110+  {
10111+    stbIF0(stbir__simdfX c0 = stbir__simdf_frepX( c0s ); )
10112+    stbIF1(stbir__simdfX c1 = stbir__simdf_frepX( c1s ); )
10113+    stbIF2(stbir__simdfX c2 = stbir__simdf_frepX( c2s ); )
10114+    stbIF3(stbir__simdfX c3 = stbir__simdf_frepX( c3s ); )
10115+    stbIF4(stbir__simdfX c4 = stbir__simdf_frepX( c4s ); )
10116+    stbIF5(stbir__simdfX c5 = stbir__simdf_frepX( c5s ); )
10117+    stbIF6(stbir__simdfX c6 = stbir__simdf_frepX( c6s ); )
10118+    stbIF7(stbir__simdfX c7 = stbir__simdf_frepX( c7s ); )
10119+
10120+    STBIR_SIMD_NO_UNROLL_LOOP_START
10121+    while ( ( (char*)input0_end - (char*) input0 ) >= (16*stbir__simdfX_float_count) )
10122+    {
10123+      stbir__simdfX o0, o1, o2, o3, r0, r1, r2, r3;
10124+      STBIR_SIMD_NO_UNROLL(output);
10125+
10126+      // prefetch four loop iterations ahead (doesn't affect much for small resizes, but helps with big ones)
10127+      stbIF0( stbir__prefetch( input0 + (16*stbir__simdfX_float_count) ); )
10128+      stbIF1( stbir__prefetch( input1 + (16*stbir__simdfX_float_count) ); )
10129+      stbIF2( stbir__prefetch( input2 + (16*stbir__simdfX_float_count) ); )
10130+      stbIF3( stbir__prefetch( input3 + (16*stbir__simdfX_float_count) ); )
10131+      stbIF4( stbir__prefetch( input4 + (16*stbir__simdfX_float_count) ); )
10132+      stbIF5( stbir__prefetch( input5 + (16*stbir__simdfX_float_count) ); )
10133+      stbIF6( stbir__prefetch( input6 + (16*stbir__simdfX_float_count) ); )
10134+      stbIF7( stbir__prefetch( input7 + (16*stbir__simdfX_float_count) ); )
10135+
10136+      #ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
10137+      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) );
10138+              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) );
10139+              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 ); )
10140+      #else
10141+      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) );
10142+              stbir__simdfX_mult( o0, r0, c0 );      stbir__simdfX_mult( o1, r1, c0 );                             stbir__simdfX_mult( o2, r2, c0 );                                 stbir__simdfX_mult( o3, r3, c0 );  )
10143+      #endif
10144+
10145+      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) );
10146+              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 ); )
10147+      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) );
10148+              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 ); )
10149+      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) );
10150+              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 ); )
10151+      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) );
10152+              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 ); )
10153+      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) );
10154+              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 ); )
10155+      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) );
10156+              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 ); )
10157+      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) );
10158+              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 ); )
10159+
10160+      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 );
10161+      output += (4*stbir__simdfX_float_count);
10162+      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); )
10163+    }
10164+
10165+    STBIR_SIMD_NO_UNROLL_LOOP_START
10166+    while ( ( (char*)input0_end - (char*) input0 ) >= 16 )
10167+    {
10168+      stbir__simdf o0, r0;
10169+      STBIR_SIMD_NO_UNROLL(output);
10170+
10171+      #ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
10172+      stbIF0( stbir__simdf_load( o0, output );   stbir__simdf_load( r0, input0 ); stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c0 ) ); )
10173+      #else
10174+      stbIF0( stbir__simdf_load( r0, input0 );  stbir__simdf_mult( o0, r0, stbir__if_simdf8_cast_to_simdf4( c0 ) ); )
10175+      #endif
10176+      stbIF1( stbir__simdf_load( r0, input1 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c1 ) ); )
10177+      stbIF2( stbir__simdf_load( r0, input2 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c2 ) ); )
10178+      stbIF3( stbir__simdf_load( r0, input3 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c3 ) ); )
10179+      stbIF4( stbir__simdf_load( r0, input4 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c4 ) ); )
10180+      stbIF5( stbir__simdf_load( r0, input5 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c5 ) ); )
10181+      stbIF6( stbir__simdf_load( r0, input6 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c6 ) ); )
10182+      stbIF7( stbir__simdf_load( r0, input7 );  stbir__simdf_madd( o0, o0, r0, stbir__if_simdf8_cast_to_simdf4( c7 ) ); )
10183+
10184+      stbir__simdf_store( output, o0 );
10185+      output += 4;
10186+      stbIF0( input0 += 4; ) stbIF1( input1 += 4; ) stbIF2( input2 += 4; ) stbIF3( input3 += 4; ) stbIF4( input4 += 4; ) stbIF5( input5 += 4; ) stbIF6( input6 += 4; ) stbIF7( input7 += 4; )
10187+    }
10188+  }
10189+  #else
10190+  STBIR_NO_UNROLL_LOOP_START
10191+  while ( ( (char*)input0_end - (char*) input0 ) >= 16 )
10192+  {
10193+    float o0, o1, o2, o3;
10194+    STBIR_NO_UNROLL(output);
10195+    #ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
10196+    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; )
10197+    #else
10198+    stbIF0( o0  = input0[0] * c0s; o1  = input0[1] * c0s; o2  = input0[2] * c0s; o3  = input0[3] * c0s; )
10199+    #endif
10200+    stbIF1( o0 += input1[0] * c1s; o1 += input1[1] * c1s; o2 += input1[2] * c1s; o3 += input1[3] * c1s; )
10201+    stbIF2( o0 += input2[0] * c2s; o1 += input2[1] * c2s; o2 += input2[2] * c2s; o3 += input2[3] * c2s; )
10202+    stbIF3( o0 += input3[0] * c3s; o1 += input3[1] * c3s; o2 += input3[2] * c3s; o3 += input3[3] * c3s; )
10203+    stbIF4( o0 += input4[0] * c4s; o1 += input4[1] * c4s; o2 += input4[2] * c4s; o3 += input4[3] * c4s; )
10204+    stbIF5( o0 += input5[0] * c5s; o1 += input5[1] * c5s; o2 += input5[2] * c5s; o3 += input5[3] * c5s; )
10205+    stbIF6( o0 += input6[0] * c6s; o1 += input6[1] * c6s; o2 += input6[2] * c6s; o3 += input6[3] * c6s; )
10206+    stbIF7( o0 += input7[0] * c7s; o1 += input7[1] * c7s; o2 += input7[2] * c7s; o3 += input7[3] * c7s; )
10207+    output[0] = o0; output[1] = o1; output[2] = o2; output[3] = o3;
10208+    output += 4;
10209+    stbIF0( input0 += 4; ) stbIF1( input1 += 4; ) stbIF2( input2 += 4; ) stbIF3( input3 += 4; ) stbIF4( input4 += 4; ) stbIF5( input5 += 4; ) stbIF6( input6 += 4; ) stbIF7( input7 += 4; )
10210+  }
10211+  #endif
10212+  STBIR_NO_UNROLL_LOOP_START
10213+  while ( input0 < input0_end )
10214+  {
10215+    float o0;
10216+    STBIR_NO_UNROLL(output);
10217+    #ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
10218+    stbIF0( o0 = output[0] + input0[0] * c0s; )
10219+    #else
10220+    stbIF0( o0  = input0[0] * c0s; )
10221+    #endif
10222+    stbIF1( o0 += input1[0] * c1s; )
10223+    stbIF2( o0 += input2[0] * c2s; )
10224+    stbIF3( o0 += input3[0] * c3s; )
10225+    stbIF4( o0 += input4[0] * c4s; )
10226+    stbIF5( o0 += input5[0] * c5s; )
10227+    stbIF6( o0 += input6[0] * c6s; )
10228+    stbIF7( o0 += input7[0] * c7s; )
10229+    output[0] = o0;
10230+    ++output;
10231+    stbIF0( ++input0; ) stbIF1( ++input1; ) stbIF2( ++input2; ) stbIF3( ++input3; ) stbIF4( ++input4; ) stbIF5( ++input5; ) stbIF6( ++input6; ) stbIF7( ++input7; )
10232+  }
10233+}
10234+
10235+#undef stbIF0
10236+#undef stbIF1
10237+#undef stbIF2
10238+#undef stbIF3
10239+#undef stbIF4
10240+#undef stbIF5
10241+#undef stbIF6
10242+#undef stbIF7
10243+#undef STB_IMAGE_RESIZE_DO_VERTICALS
10244+#undef STBIR__vertical_channels
10245+#undef STB_IMAGE_RESIZE_DO_HORIZONTALS
10246+#undef STBIR_strs_join24
10247+#undef STBIR_strs_join14
10248+#undef STBIR_chans
10249+#ifdef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
10250+#undef STB_IMAGE_RESIZE_VERTICAL_CONTINUE
10251+#endif
10252+
10253+#else // !STB_IMAGE_RESIZE_DO_VERTICALS
10254+
10255+#define STBIR_chans( start, end ) STBIR_strs_join1(start,STBIR__horizontal_channels,end)
10256+
10257+#ifndef stbir__2_coeff_only
10258+#define stbir__2_coeff_only()             \
10259+    stbir__1_coeff_only();                \
10260+    stbir__1_coeff_remnant(1);
10261+#endif
10262+
10263+#ifndef stbir__2_coeff_remnant
10264+#define stbir__2_coeff_remnant( ofs )     \
10265+    stbir__1_coeff_remnant(ofs);          \
10266+    stbir__1_coeff_remnant((ofs)+1);
10267+#endif
10268+
10269+#ifndef stbir__3_coeff_only
10270+#define stbir__3_coeff_only()             \
10271+    stbir__2_coeff_only();                \
10272+    stbir__1_coeff_remnant(2);
10273+#endif
10274+
10275+#ifndef stbir__3_coeff_remnant
10276+#define stbir__3_coeff_remnant( ofs )     \
10277+    stbir__2_coeff_remnant(ofs);          \
10278+    stbir__1_coeff_remnant((ofs)+2);
10279+#endif
10280+
10281+#ifndef stbir__3_coeff_setup
10282+#define stbir__3_coeff_setup()
10283+#endif
10284+
10285+#ifndef stbir__4_coeff_start
10286+#define stbir__4_coeff_start()            \
10287+    stbir__2_coeff_only();                \
10288+    stbir__2_coeff_remnant(2);
10289+#endif
10290+
10291+#ifndef stbir__4_coeff_continue_from_4
10292+#define stbir__4_coeff_continue_from_4( ofs )     \
10293+    stbir__2_coeff_remnant(ofs);                  \
10294+    stbir__2_coeff_remnant((ofs)+2);
10295+#endif
10296+
10297+#ifndef stbir__store_output_tiny
10298+#define stbir__store_output_tiny stbir__store_output
10299+#endif
10300+
10301+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 )
10302+{
10303+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10304+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10305+  STBIR_SIMD_NO_UNROLL_LOOP_START
10306+  do {
10307+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10308+    float const * hc = horizontal_coefficients;
10309+    stbir__1_coeff_only();
10310+    stbir__store_output_tiny();
10311+  } while ( output < output_end );
10312+}
10313+
10314+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 )
10315+{
10316+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10317+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10318+  STBIR_SIMD_NO_UNROLL_LOOP_START
10319+  do {
10320+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10321+    float const * hc = horizontal_coefficients;
10322+    stbir__2_coeff_only();
10323+    stbir__store_output_tiny();
10324+  } while ( output < output_end );
10325+}
10326+
10327+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 )
10328+{
10329+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10330+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10331+  STBIR_SIMD_NO_UNROLL_LOOP_START
10332+  do {
10333+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10334+    float const * hc = horizontal_coefficients;
10335+    stbir__3_coeff_only();
10336+    stbir__store_output_tiny();
10337+  } while ( output < output_end );
10338+}
10339+
10340+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 )
10341+{
10342+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10343+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10344+  STBIR_SIMD_NO_UNROLL_LOOP_START
10345+  do {
10346+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10347+    float const * hc = horizontal_coefficients;
10348+    stbir__4_coeff_start();
10349+    stbir__store_output();
10350+  } while ( output < output_end );
10351+}
10352+
10353+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 )
10354+{
10355+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10356+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10357+  STBIR_SIMD_NO_UNROLL_LOOP_START
10358+  do {
10359+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10360+    float const * hc = horizontal_coefficients;
10361+    stbir__4_coeff_start();
10362+    stbir__1_coeff_remnant(4);
10363+    stbir__store_output();
10364+  } while ( output < output_end );
10365+}
10366+
10367+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 )
10368+{
10369+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10370+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10371+  STBIR_SIMD_NO_UNROLL_LOOP_START
10372+  do {
10373+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10374+    float const * hc = horizontal_coefficients;
10375+    stbir__4_coeff_start();
10376+    stbir__2_coeff_remnant(4);
10377+    stbir__store_output();
10378+  } while ( output < output_end );
10379+}
10380+
10381+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 )
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__3_coeff_setup();
10386+  STBIR_SIMD_NO_UNROLL_LOOP_START
10387+  do {
10388+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10389+    float const * hc = horizontal_coefficients;
10390+
10391+    stbir__4_coeff_start();
10392+    stbir__3_coeff_remnant(4);
10393+    stbir__store_output();
10394+  } while ( output < output_end );
10395+}
10396+
10397+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 )
10398+{
10399+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10400+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10401+  STBIR_SIMD_NO_UNROLL_LOOP_START
10402+  do {
10403+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10404+    float const * hc = horizontal_coefficients;
10405+    stbir__4_coeff_start();
10406+    stbir__4_coeff_continue_from_4(4);
10407+    stbir__store_output();
10408+  } while ( output < output_end );
10409+}
10410+
10411+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 )
10412+{
10413+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10414+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10415+  STBIR_SIMD_NO_UNROLL_LOOP_START
10416+  do {
10417+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10418+    float const * hc = horizontal_coefficients;
10419+    stbir__4_coeff_start();
10420+    stbir__4_coeff_continue_from_4(4);
10421+    stbir__1_coeff_remnant(8);
10422+    stbir__store_output();
10423+  } while ( output < output_end );
10424+}
10425+
10426+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 )
10427+{
10428+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10429+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10430+  STBIR_SIMD_NO_UNROLL_LOOP_START
10431+  do {
10432+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10433+    float const * hc = horizontal_coefficients;
10434+    stbir__4_coeff_start();
10435+    stbir__4_coeff_continue_from_4(4);
10436+    stbir__2_coeff_remnant(8);
10437+    stbir__store_output();
10438+  } while ( output < output_end );
10439+}
10440+
10441+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 )
10442+{
10443+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10444+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10445+  stbir__3_coeff_setup();
10446+  STBIR_SIMD_NO_UNROLL_LOOP_START
10447+  do {
10448+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10449+    float const * hc = horizontal_coefficients;
10450+    stbir__4_coeff_start();
10451+    stbir__4_coeff_continue_from_4(4);
10452+    stbir__3_coeff_remnant(8);
10453+    stbir__store_output();
10454+  } while ( output < output_end );
10455+}
10456+
10457+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 )
10458+{
10459+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10460+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10461+  STBIR_SIMD_NO_UNROLL_LOOP_START
10462+  do {
10463+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10464+    float const * hc = horizontal_coefficients;
10465+    stbir__4_coeff_start();
10466+    stbir__4_coeff_continue_from_4(4);
10467+    stbir__4_coeff_continue_from_4(8);
10468+    stbir__store_output();
10469+  } while ( output < output_end );
10470+}
10471+
10472+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 )
10473+{
10474+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10475+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10476+  STBIR_SIMD_NO_UNROLL_LOOP_START
10477+  do {
10478+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10479+    int n = ( ( horizontal_contributors->n1 - horizontal_contributors->n0 + 1 ) - 4 + 3 ) >> 2;
10480+    float const * hc = horizontal_coefficients;
10481+
10482+    stbir__4_coeff_start();
10483+    STBIR_SIMD_NO_UNROLL_LOOP_START
10484+    do {
10485+      hc += 4;
10486+      decode += STBIR__horizontal_channels * 4;
10487+      stbir__4_coeff_continue_from_4( 0 );
10488+      --n;
10489+    } while ( n > 0 );
10490+    stbir__store_output();
10491+  } while ( output < output_end );
10492+}
10493+
10494+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 )
10495+{
10496+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10497+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10498+  STBIR_SIMD_NO_UNROLL_LOOP_START
10499+  do {
10500+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10501+    int n = ( ( horizontal_contributors->n1 - horizontal_contributors->n0 + 1 ) - 5 + 3 ) >> 2;
10502+    float const * hc = horizontal_coefficients;
10503+
10504+    stbir__4_coeff_start();
10505+    STBIR_SIMD_NO_UNROLL_LOOP_START
10506+    do {
10507+      hc += 4;
10508+      decode += STBIR__horizontal_channels * 4;
10509+      stbir__4_coeff_continue_from_4( 0 );
10510+      --n;
10511+    } while ( n > 0 );
10512+    stbir__1_coeff_remnant( 4 );
10513+    stbir__store_output();
10514+  } while ( output < output_end );
10515+}
10516+
10517+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 )
10518+{
10519+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10520+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10521+  STBIR_SIMD_NO_UNROLL_LOOP_START
10522+  do {
10523+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10524+    int n = ( ( horizontal_contributors->n1 - horizontal_contributors->n0 + 1 ) - 6 + 3 ) >> 2;
10525+    float const * hc = horizontal_coefficients;
10526+
10527+    stbir__4_coeff_start();
10528+    STBIR_SIMD_NO_UNROLL_LOOP_START
10529+    do {
10530+      hc += 4;
10531+      decode += STBIR__horizontal_channels * 4;
10532+      stbir__4_coeff_continue_from_4( 0 );
10533+      --n;
10534+    } while ( n > 0 );
10535+    stbir__2_coeff_remnant( 4 );
10536+
10537+    stbir__store_output();
10538+  } while ( output < output_end );
10539+}
10540+
10541+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 )
10542+{
10543+  float const * output_end = output_buffer + output_sub_size * STBIR__horizontal_channels;
10544+  float STBIR_SIMD_STREAMOUT_PTR( * ) output = output_buffer;
10545+  stbir__3_coeff_setup();
10546+  STBIR_SIMD_NO_UNROLL_LOOP_START
10547+  do {
10548+    float const * decode = decode_buffer + horizontal_contributors->n0 * STBIR__horizontal_channels;
10549+    int n = ( ( horizontal_contributors->n1 - horizontal_contributors->n0 + 1 ) - 7 + 3 ) >> 2;
10550+    float const * hc = horizontal_coefficients;
10551+
10552+    stbir__4_coeff_start();
10553+    STBIR_SIMD_NO_UNROLL_LOOP_START
10554+    do {
10555+      hc += 4;
10556+      decode += STBIR__horizontal_channels * 4;
10557+      stbir__4_coeff_continue_from_4( 0 );
10558+      --n;
10559+    } while ( n > 0 );
10560+    stbir__3_coeff_remnant( 4 );
10561+
10562+    stbir__store_output();
10563+  } while ( output < output_end );
10564+}
10565+
10566+static stbir__horizontal_gather_channels_func * STBIR_chans(stbir__horizontal_gather_,_channels_with_n_coeffs_funcs)[4]=
10567+{
10568+  STBIR_chans(stbir__horizontal_gather_,_channels_with_n_coeffs_mod0),
10569+  STBIR_chans(stbir__horizontal_gather_,_channels_with_n_coeffs_mod1),
10570+  STBIR_chans(stbir__horizontal_gather_,_channels_with_n_coeffs_mod2),
10571+  STBIR_chans(stbir__horizontal_gather_,_channels_with_n_coeffs_mod3),
10572+};
10573+
10574+static stbir__horizontal_gather_channels_func * STBIR_chans(stbir__horizontal_gather_,_channels_funcs)[12]=
10575+{
10576+  STBIR_chans(stbir__horizontal_gather_,_channels_with_1_coeff),
10577+  STBIR_chans(stbir__horizontal_gather_,_channels_with_2_coeffs),
10578+  STBIR_chans(stbir__horizontal_gather_,_channels_with_3_coeffs),
10579+  STBIR_chans(stbir__horizontal_gather_,_channels_with_4_coeffs),
10580+  STBIR_chans(stbir__horizontal_gather_,_channels_with_5_coeffs),
10581+  STBIR_chans(stbir__horizontal_gather_,_channels_with_6_coeffs),
10582+  STBIR_chans(stbir__horizontal_gather_,_channels_with_7_coeffs),
10583+  STBIR_chans(stbir__horizontal_gather_,_channels_with_8_coeffs),
10584+  STBIR_chans(stbir__horizontal_gather_,_channels_with_9_coeffs),
10585+  STBIR_chans(stbir__horizontal_gather_,_channels_with_10_coeffs),
10586+  STBIR_chans(stbir__horizontal_gather_,_channels_with_11_coeffs),
10587+  STBIR_chans(stbir__horizontal_gather_,_channels_with_12_coeffs),
10588+};
10589+
10590+#undef STBIR__horizontal_channels
10591+#undef STB_IMAGE_RESIZE_DO_HORIZONTALS
10592+#undef stbir__1_coeff_only
10593+#undef stbir__1_coeff_remnant
10594+#undef stbir__2_coeff_only
10595+#undef stbir__2_coeff_remnant
10596+#undef stbir__3_coeff_only
10597+#undef stbir__3_coeff_remnant
10598+#undef stbir__3_coeff_setup
10599+#undef stbir__4_coeff_start
10600+#undef stbir__4_coeff_continue_from_4
10601+#undef stbir__store_output
10602+#undef stbir__store_output_tiny
10603+#undef STBIR_chans
10604+
10605+#endif  // HORIZONALS
10606+
10607+#undef STBIR_strs_join2
10608+#undef STBIR_strs_join1
10609+
10610+#endif // STB_IMAGE_RESIZE_DO_HORIZONTALS/VERTICALS/CODERS
10611+
10612+/*
10613+------------------------------------------------------------------------------
10614+This software is available under 2 licenses -- choose whichever you prefer.
10615+------------------------------------------------------------------------------
10616+ALTERNATIVE A - MIT License
10617+Copyright (c) 2017 Sean Barrett
10618+Permission is hereby granted, free of charge, to any person obtaining a copy of
10619+this software and associated documentation files (the "Software"), to deal in
10620+the Software without restriction, including without limitation the rights to
10621+use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies
10622+of the Software, and to permit persons to whom the Software is furnished to do
10623+so, subject to the following conditions:
10624+The above copyright notice and this permission notice shall be included in all
10625+copies or substantial portions of the Software.
10626+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
10627+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
10628+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
10629+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
10630+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
10631+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
10632+SOFTWARE.
10633+------------------------------------------------------------------------------
10634+ALTERNATIVE B - Public Domain (www.unlicense.org)
10635+This is free and unencumbered software released into the public domain.
10636+Anyone is free to copy, modify, publish, use, compile, sell, or distribute this
10637+software, either in source code form or as a compiled binary, for any purpose,
10638+commercial or non-commercial, and by any means.
10639+In jurisdictions that recognize copyright laws, the author or authors of this
10640+software dedicate any and all copyright interest in the software to the public
10641+domain. We make this dedication for the benefit of the public at large and to
10642+the detriment of our heirs and successors. We intend this dedication to be an
10643+overt act of relinquishment in perpetuity of all present and future rights to
10644+this software under copyright law.
10645+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
10646+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
10647+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
10648+AUTHORS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER LIABILITY, WHETHER IN AN
10649+ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, OUT OF OR IN CONNECTION
10650+WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE SOFTWARE.
10651+------------------------------------------------------------------------------
10652+*/