forked from nothings/stb
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathstb_image_resize2.h
10325 lines (8888 loc) · 430 KB
/
stb_image_resize2.h
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
775
776
777
778
779
780
781
782
783
784
785
786
787
788
789
790
791
792
793
794
795
796
797
798
799
800
801
802
803
804
805
806
807
808
809
810
811
812
813
814
815
816
817
818
819
820
821
822
823
824
825
826
827
828
829
830
831
832
833
834
835
836
837
838
839
840
841
842
843
844
845
846
847
848
849
850
851
852
853
854
855
856
857
858
859
860
861
862
863
864
865
866
867
868
869
870
871
872
873
874
875
876
877
878
879
880
881
882
883
884
885
886
887
888
889
890
891
892
893
894
895
896
897
898
899
900
901
902
903
904
905
906
907
908
909
910
911
912
913
914
915
916
917
918
919
920
921
922
923
924
925
926
927
928
929
930
931
932
933
934
935
936
937
938
939
940
941
942
943
944
945
946
947
948
949
950
951
952
953
954
955
956
957
958
959
960
961
962
963
964
965
966
967
968
969
970
971
972
973
974
975
976
977
978
979
980
981
982
983
984
985
986
987
988
989
990
991
992
993
994
995
996
997
998
999
1000
/* stb_image_resize2 - v2.04 - public domain image resizing
by Jeff Roberts (v2) and Jorge L Rodriguez
http://github.com/nothings/stb
Can be threaded with the extended API. SSE2, AVX, Neon and WASM SIMD support. Only
scaling and translation is supported, no rotations or shears.
COMPILING & LINKING
In one C/C++ file that #includes this file, do this:
#define STB_IMAGE_RESIZE_IMPLEMENTATION
before the #include. That will create the implementation in that file.
PORTING FROM VERSION 1
The API has changed. You can continue to use the old version of stb_image_resize.h,
which is available in the "deprecated/" directory.
If you're using the old simple-to-use API, porting is straightforward.
(For more advanced APIs, read the documentation.)
stbir_resize_uint8():
- call `stbir_resize_uint8_linear`, cast channel count to `stbir_pixel_layout`
stbir_resize_float():
- call `stbir_resize_float_linear`, cast channel count to `stbir_pixel_layout`
stbir_resize_uint8_srgb():
- function name is unchanged
- cast channel count to `stbir_pixel_layout`
- above is sufficient unless your image has alpha and it's not RGBA/BGRA
- in that case, follow the below instructions for stbir_resize_uint8_srgb_edgemode
stbir_resize_uint8_srgb_edgemode()
- switch to the "medium complexity" API
- stbir_resize(), very similar API but a few more parameters:
- pixel_layout: cast channel count to `stbir_pixel_layout`
- data_type: STBIR_TYPE_UINT8_SRGB
- edge: unchanged (STBIR_EDGE_WRAP, etc.)
- filter: STBIR_FILTER_DEFAULT
- which channel is alpha is specified in stbir_pixel_layout, see enum for details
EASY API CALLS:
Easy API downsamples w/Mitchell filter, upsamples w/cubic interpolation, clamps to edge.
stbir_resize_uint8_srgb( input_pixels, input_w, input_h, input_stride_in_bytes,
output_pixels, output_w, output_h, output_stride_in_bytes,
pixel_layout_enum )
stbir_resize_uint8_linear( input_pixels, input_w, input_h, input_stride_in_bytes,
output_pixels, output_w, output_h, output_stride_in_bytes,
pixel_layout_enum )
stbir_resize_float_linear( input_pixels, input_w, input_h, input_stride_in_bytes,
output_pixels, output_w, output_h, output_stride_in_bytes,
pixel_layout_enum )
If you pass NULL or zero for the output_pixels, we will allocate the output buffer
for you and return it from the function (free with free() or STBIR_FREE).
As a special case, XX_stride_in_bytes of 0 means packed continuously in memory.
API LEVELS
There are three levels of API - easy-to-use, medium-complexity and extended-complexity.
See the "header file" section of the source for API documentation.
ADDITIONAL DOCUMENTATION
MEMORY ALLOCATION
By default, we use malloc and free for memory allocation. To override the
memory allocation, before the implementation #include, add a:
#define STBIR_MALLOC(size,user_data) ...
#define STBIR_FREE(ptr,user_data) ...
Each resize makes exactly one call to malloc/free (unless you use the
extended API where you can do one allocation for many resizes). Under
address sanitizer, we do separate allocations to find overread/writes.
PERFORMANCE
This library was written with an emphasis on performance. When testing
stb_image_resize with RGBA, the fastest mode is STBIR_4CHANNEL with
STBIR_TYPE_UINT8 pixels and CLAMPed edges (which is what many other resize
libs do by default). Also, make sure SIMD is turned on of course (default
for 64-bit targets). Avoid WRAP edge mode if you want the fastest speed.
This library also comes with profiling built-in. If you define STBIR_PROFILE,
you can use the advanced API and get low-level profiling information by
calling stbir_resize_extended_profile_info() or stbir_resize_split_profile_info()
after a resize.
SIMD
Most of the routines have optimized SSE2, AVX, NEON and WASM versions.
On Microsoft compilers, we automatically turn on SIMD for 64-bit x64 and
ARM; for 32-bit x86 and ARM, you select SIMD mode by defining STBIR_SSE2 or
STBIR_NEON. For AVX and AVX2, we auto-select it by detecting the /arch:AVX
or /arch:AVX2 switches. You can also always manually turn SSE2, AVX or AVX2
support on by defining STBIR_SSE2, STBIR_AVX or STBIR_AVX2.
On Linux, SSE2 and Neon is on by default for 64-bit x64 or ARM64. For 32-bit,
we select x86 SIMD mode by whether you have -msse2, -mavx or -mavx2 enabled
on the command line. For 32-bit ARM, you must pass -mfpu=neon-vfpv4 for both
clang and GCC, but GCC also requires an additional -mfp16-format=ieee to
automatically enable NEON.
On x86 platforms, you can also define STBIR_FP16C to turn on FP16C instructions
for converting back and forth to half-floats. This is autoselected when we
are using AVX2. Clang and GCC also require the -mf16c switch. ARM always uses
the built-in half float hardware NEON instructions.
You can also tell us to use multiply-add instructions with STBIR_USE_FMA.
Because x86 doesn't always have fma, we turn it off by default to maintain
determinism across all platforms. If you don't care about non-FMA determinism
and are willing to restrict yourself to more recent x86 CPUs (around the AVX
timeframe), then fma will give you around a 15% speedup.
You can force off SIMD in all cases by defining STBIR_NO_SIMD. You can turn
off AVX or AVX2 specifically with STBIR_NO_AVX or STBIR_NO_AVX2. AVX is 10%
to 40% faster, and AVX2 is generally another 12%.
ALPHA CHANNEL
Most of the resizing functions provide the ability to control how the alpha
channel of an image is processed.
When alpha represents transparency, it is important that when combining
colors with filtering, the pixels should not be treated equally; they
should use a weighted average based on their alpha values. For example,
if a pixel is 1% opaque bright green and another pixel is 99% opaque
black and you average them, the average will be 50% opaque, but the
unweighted average and will be a middling green color, while the weighted
average will be nearly black. This means the unweighted version introduced
green energy that didn't exist in the source image.
(If you want to know why this makes sense, you can work out the math for
the following: consider what happens if you alpha composite a source image
over a fixed color and then average the output, vs. if you average the
source image pixels and then composite that over the same fixed color.
Only the weighted average produces the same result as the ground truth
composite-then-average result.)
Therefore, it is in general best to "alpha weight" the pixels when applying
filters to them. This essentially means multiplying the colors by the alpha
values before combining them, and then dividing by the alpha value at the
end.
The computer graphics industry introduced a technique called "premultiplied
alpha" or "associated alpha" in which image colors are stored in image files
already multiplied by their alpha. This saves some math when compositing,
and also avoids the need to divide by the alpha at the end (which is quite
inefficient). However, while premultiplied alpha is common in the movie CGI
industry, it is not commonplace in other industries like videogames, and most
consumer file formats are generally expected to contain not-premultiplied
colors. For example, Photoshop saves PNG files "unpremultiplied", and web
browsers like Chrome and Firefox expect PNG images to be unpremultiplied.
Note that there are three possibilities that might describe your image
and resize expectation:
1. images are not premultiplied, alpha weighting is desired
2. images are not premultiplied, alpha weighting is not desired
3. images are premultiplied
Both case #2 and case #3 require the exact same math: no alpha weighting
should be applied or removed. Only case 1 requires extra math operations;
the other two cases can be handled identically.
stb_image_resize expects case #1 by default, applying alpha weighting to
images, expecting the input images to be unpremultiplied. This is what the
COLOR+ALPHA buffer types tell the resizer to do.
When you use the pixel layouts STBIR_RGBA, STBIR_BGRA, STBIR_ARGB,
STBIR_ABGR, STBIR_RX, or STBIR_XR you are telling us that the pixels are
non-premultiplied. In these cases, the resizer will alpha weight the colors
(effectively creating the premultiplied image), do the filtering, and then
convert back to non-premult on exit.
When you use the pixel layouts STBIR_RGBA_PM, STBIR_RGBA_PM, STBIR_RGBA_PM,
STBIR_RGBA_PM, STBIR_RX_PM or STBIR_XR_PM, you are telling that the pixels
ARE premultiplied. In this case, the resizer doesn't have to do the
premultipling - it can filter directly on the input. This about twice as
fast as the non-premultiplied case, so it's the right option if your data is
already setup correctly.
When you use the pixel layout STBIR_4CHANNEL or STBIR_2CHANNEL, you are
telling us that there is no channel that represents transparency; it may be
RGB and some unrelated fourth channel that has been stored in the alpha
channel, but it is actually not alpha. No special processing will be
performed.
The difference between the generic 4 or 2 channel layouts, and the
specialized _PM versions is with the _PM versions you are telling us that
the data *is* alpha, just don't premultiply it. That's important when
using SRGB pixel formats, we need to know where the alpha is, because
it is converted linearly (rather than with the SRGB converters).
Because alpha weighting produces the same effect as premultiplying, you
even have the option with non-premultiplied inputs to let the resizer
produce a premultiplied output. Because the intially computed alpha-weighted
output image is effectively premultiplied, this is actually more performant
than the normal path which un-premultiplies the output image as a final step.
Finally, when converting both in and out of non-premulitplied space (for
example, when using STBIR_RGBA), we go to somewhat heroic measures to
ensure that areas with zero alpha value pixels get something reasonable
in the RGB values. If you don't care about the RGB values of zero alpha
pixels, you can call the stbir_set_non_pm_alpha_speed_over_quality()
function - this runs a premultiplied resize about 25% faster. That said,
when you really care about speed, using premultiplied pixels for both in
and out (STBIR_RGBA_PM, etc) much faster than both of these premultiplied
options.
PIXEL LAYOUT CONVERSION
The resizer can convert from some pixel layouts to others. When using the
stbir_set_pixel_layouts(), you can, for example, specify STBIR_RGBA
on input, and STBIR_ARGB on output, and it will re-organize the channels
during the resize. Currently, you can only convert between two pixel
layouts with the same number of channels.
DETERMINISM
We commit to being deterministic (from x64 to ARM to scalar to SIMD, etc).
This requires compiling with fast-math off (using at least /fp:precise).
Also, you must turn off fp-contracting (which turns mult+adds into fmas)!
We attempt to do this with pragmas, but with Clang, you usually want to add
-ffp-contract=off to the command line as well.
For 32-bit x86, you must use SSE and SSE2 codegen for determinism. That is,
if the scalar x87 unit gets used at all, we immediately lose determinism.
On Microsoft Visual Studio 2008 and earlier, from what we can tell there is
no way to be deterministic in 32-bit x86 (some x87 always leaks in, even
with fp:strict). On 32-bit x86 GCC, determinism requires both -msse2 and
-fpmath=sse.
Note that we will not be deterministic with float data containing NaNs -
the NaNs will propagate differently on different SIMD and platforms.
If you turn on STBIR_USE_FMA, then we will be deterministic with other
fma targets, but we will differ from non-fma targets (this is unavoidable,
because a fma isn't simply an add with a mult - it also introduces a
rounding difference compared to non-fma instruction sequences.
FLOAT PIXEL FORMAT RANGE
Any range of values can be used for the non-alpha float data that you pass
in (0 to 1, -1 to 1, whatever). However, if you are inputting float values
but *outputting* bytes or shorts, you must use a range of 0 to 1 so that we
scale back properly. The alpha channel must also be 0 to 1 for any format
that does premultiplication prior to resizing.
Note also that with float output, using filters with negative lobes, the
output filtered values might go slightly out of range. You can define
STBIR_FLOAT_LOW_CLAMP and/or STBIR_FLOAT_HIGH_CLAMP to specify the range
to clamp to on output, if that's important.
MAX/MIN SCALE FACTORS
The input pixel resolutions are in integers, and we do the internal pointer
resolution in size_t sized integers. However, the scale ratio from input
resolution to output resolution is calculated in float form. This means
the effective possible scale ratio is limited to 24 bits (or 16 million
to 1). As you get close to the size of the float resolution (again, 16
million pixels wide or high), you might start seeing float inaccuracy
issues in general in the pipeline. If you have to do extreme resizes,
you can usually do this is multiple stages (using float intermediate
buffers).
FLIPPED IMAGES
Stride is just the delta from one scanline to the next. This means you can
use a negative stride to handle inverted images (point to the final
scanline and use a negative stride). You can invert the input or output,
using negative strides.
DEFAULT FILTERS
For functions which don't provide explicit control over what filters to
use, you can change the compile-time defaults with:
#define STBIR_DEFAULT_FILTER_UPSAMPLE STBIR_FILTER_something
#define STBIR_DEFAULT_FILTER_DOWNSAMPLE STBIR_FILTER_something
See stbir_filter in the header-file section for the list of filters.
NEW FILTERS
A number of 1D filter kernels are supplied. For a list of supported
filters, see the stbir_filter enum. You can install your own filters by
using the stbir_set_filter_callbacks function.
PROGRESS
For interactive use with slow resize operations, you can use the the
scanline callbacks in the extended API. It would have to be a *very* large
image resample to need progress though - we're very fast.
CEIL and FLOOR
In scalar mode, the only functions we use from math.h are ceilf and floorf,
but if you have your own versions, you can define the STBIR_CEILF(v) and
STBIR_FLOORF(v) macros and we'll use them instead. In SIMD, we just use
our own versions.
ASSERT
Define STBIR_ASSERT(boolval) to override assert() and not use assert.h
FUTURE TODOS
* For polyphase integral filters, we just memcpy the coeffs to dupe
them, but we should indirect and use the same coeff memory.
* Add pixel layout conversions for sensible different channel counts
(maybe, 1->3/4, 3->4, 4->1, 3->1).
* For SIMD encode and decode scanline routines, do any pre-aligning
for bad input/output buffer alignments and pitch?
* For very wide scanlines, we should we do vertical strips to stay within
L2 cache. Maybe do chunks of 1K pixels at a time. There would be
some pixel reconversion, but probably dwarfed by things falling out
of cache. Probably also something possible with alternating between
scattering and gathering at high resize scales?
* Rewrite the coefficient generator to do many at once.
* AVX-512 vertical kernels - worried about downclocking here.
* Convert the reincludes to macros when we know they aren't changing.
* Experiment with pivoting the horizontal and always using the
vertical filters (which are faster, but perhaps not enough to overcome
the pivot cost and the extra memory touches). Need to buffer the whole
image so have to balance memory use.
* Most of our code is internally function pointers, should we compile
all the SIMD stuff always and dynamically dispatch?
CONTRIBUTORS
Jeff Roberts: 2.0 implementation, optimizations, SIMD
Martins Mozeiko: NEON simd, WASM simd, clang and GCC whisperer.
Fabian Giesen: half float and srgb converters
Sean Barrett: API design, optimizations
Jorge L Rodriguez: Original 1.0 implementation
Aras Pranckevicius: bugfixes for 1.0
Nathan Reed: warning fixes for 1.0
REVISIONS
2.04 (2023-11-17) Fix for rare AVX bug, shadowed symbol (thanks Nikola Smiljanic).
2.03 (2023-11-01) ASAN and TSAN warnings fixed, minor tweaks.
2.00 (2023-10-10) mostly new source: new api, optimizations, simd, vertical-first, etc
(2x-5x faster without simd, 4x-12x faster with simd)
(in some cases, 20x to 40x faster - resizing to very small for example)
0.96 (2019-03-04) fixed warnings
0.95 (2017-07-23) fixed warnings
0.94 (2017-03-18) fixed warnings
0.93 (2017-03-03) fixed bug with certain combinations of heights
0.92 (2017-01-02) fix integer overflow on large (>2GB) images
0.91 (2016-04-02) fix warnings; fix handling of subpixel regions
0.90 (2014-09-17) first released version
LICENSE
See end of file for license information.
*/
#if !defined(STB_IMAGE_RESIZE_DO_HORIZONTALS) && !defined(STB_IMAGE_RESIZE_DO_VERTICALS) && !defined(STB_IMAGE_RESIZE_DO_CODERS) // for internal re-includes
#ifndef STBIR_INCLUDE_STB_IMAGE_RESIZE2_H
#define STBIR_INCLUDE_STB_IMAGE_RESIZE2_H
#include <stddef.h>
#ifdef _MSC_VER
typedef unsigned char stbir_uint8;
typedef unsigned short stbir_uint16;
typedef unsigned int stbir_uint32;
typedef unsigned __int64 stbir_uint64;
#else
#include <stdint.h>
typedef uint8_t stbir_uint8;
typedef uint16_t stbir_uint16;
typedef uint32_t stbir_uint32;
typedef uint64_t stbir_uint64;
#endif
#ifdef _M_IX86_FP
#if ( _M_IX86_FP >= 1 )
#ifndef STBIR_SSE
#define STBIR_SSE
#endif
#endif
#endif
#if defined(_x86_64) || defined( __x86_64__ ) || defined( _M_X64 ) || defined(__x86_64) || defined(_M_AMD64) || defined(__SSE2__) || defined(STBIR_SSE) || defined(STBIR_SSE2)
#ifndef STBIR_SSE2
#define STBIR_SSE2
#endif
#if defined(__AVX__) || defined(STBIR_AVX2)
#ifndef STBIR_AVX
#ifndef STBIR_NO_AVX
#define STBIR_AVX
#endif
#endif
#endif
#if defined(__AVX2__) || defined(STBIR_AVX2)
#ifndef STBIR_NO_AVX2
#ifndef STBIR_AVX2
#define STBIR_AVX2
#endif
#if defined( _MSC_VER ) && !defined(__clang__)
#ifndef STBIR_FP16C // FP16C instructions are on all AVX2 cpus, so we can autoselect it here on microsoft - clang needs -m16c
#define STBIR_FP16C
#endif
#endif
#endif
#endif
#ifdef __F16C__
#ifndef STBIR_FP16C // turn on FP16C instructions if the define is set (for clang and gcc)
#define STBIR_FP16C
#endif
#endif
#endif
#if defined( _M_ARM64 ) || defined( __aarch64__ ) || defined( __arm64__ ) || defined(_M_ARM) || (__ARM_NEON_FP & 4) != 0 && __ARM_FP16_FORMAT_IEEE != 0
#ifndef STBIR_NEON
#define STBIR_NEON
#endif
#endif
#if defined(_M_ARM)
#ifdef STBIR_USE_FMA
#undef STBIR_USE_FMA // no FMA for 32-bit arm on MSVC
#endif
#endif
#if defined(__wasm__) && defined(__wasm_simd128__)
#ifndef STBIR_WASM
#define STBIR_WASM
#endif
#endif
#ifndef STBIRDEF
#ifdef STB_IMAGE_RESIZE_STATIC
#define STBIRDEF static
#else
#ifdef __cplusplus
#define STBIRDEF extern "C"
#else
#define STBIRDEF extern
#endif
#endif
#endif
//////////////////////////////////////////////////////////////////////////////
//// start "header file" ///////////////////////////////////////////////////
//
// Easy-to-use API:
//
// * stride is the offset between successive rows of image data
// in memory, in bytes. specify 0 for packed continuously in memory
// * colorspace is linear or sRGB as specified by function name
// * Uses the default filters
// * Uses edge mode clamped
// * returned result is 1 for success or 0 in case of an error.
// stbir_pixel_layout specifies:
// number of channels
// order of channels
// whether color is premultiplied by alpha
// for back compatibility, you can cast the old channel count to an stbir_pixel_layout
typedef enum
{
STBIR_1CHANNEL = 1,
STBIR_2CHANNEL = 2,
STBIR_RGB = 3, // 3-chan, with order specified (for channel flipping)
STBIR_BGR = 0, // 3-chan, with order specified (for channel flipping)
STBIR_4CHANNEL = 5,
STBIR_RGBA = 4, // alpha formats, where alpha is NOT premultiplied into color channels
STBIR_BGRA = 6,
STBIR_ARGB = 7,
STBIR_ABGR = 8,
STBIR_RA = 9,
STBIR_AR = 10,
STBIR_RGBA_PM = 11, // alpha formats, where alpha is premultiplied into color channels
STBIR_BGRA_PM = 12,
STBIR_ARGB_PM = 13,
STBIR_ABGR_PM = 14,
STBIR_RA_PM = 15,
STBIR_AR_PM = 16,
STBIR_RGBA_NO_AW = 11, // alpha formats, where NO alpha weighting is applied at all!
STBIR_BGRA_NO_AW = 12, // these are just synonyms for the _PM flags (which also do
STBIR_ARGB_NO_AW = 13, // no alpha weighting). These names just make it more clear
STBIR_ABGR_NO_AW = 14, // for some folks).
STBIR_RA_NO_AW = 15,
STBIR_AR_NO_AW = 16,
} stbir_pixel_layout;
//===============================================================
// Simple-complexity API
//
// If output_pixels is NULL (0), then we will allocate the buffer and return it to you.
//--------------------------------
STBIRDEF unsigned char * stbir_resize_uint8_srgb( const unsigned char *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
unsigned char *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
stbir_pixel_layout pixel_type );
STBIRDEF unsigned char * stbir_resize_uint8_linear( const unsigned char *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
unsigned char *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
stbir_pixel_layout pixel_type );
STBIRDEF float * stbir_resize_float_linear( const float *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
float *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
stbir_pixel_layout pixel_type );
//===============================================================
//===============================================================
// Medium-complexity API
//
// This extends the easy-to-use API as follows:
//
// * Can specify the datatype - U8, U8_SRGB, U16, FLOAT, HALF_FLOAT
// * Edge wrap can selected explicitly
// * Filter can be selected explicitly
//--------------------------------
typedef enum
{
STBIR_EDGE_CLAMP = 0,
STBIR_EDGE_REFLECT = 1,
STBIR_EDGE_WRAP = 2, // this edge mode is slower and uses more memory
STBIR_EDGE_ZERO = 3,
} stbir_edge;
typedef enum
{
STBIR_FILTER_DEFAULT = 0, // use same filter type that easy-to-use API chooses
STBIR_FILTER_BOX = 1, // A trapezoid w/1-pixel wide ramps, same result as box for integer scale ratios
STBIR_FILTER_TRIANGLE = 2, // On upsampling, produces same results as bilinear texture filtering
STBIR_FILTER_CUBICBSPLINE = 3, // The cubic b-spline (aka Mitchell-Netrevalli with B=1,C=0), gaussian-esque
STBIR_FILTER_CATMULLROM = 4, // An interpolating cubic spline
STBIR_FILTER_MITCHELL = 5, // Mitchell-Netrevalli filter with B=1/3, C=1/3
STBIR_FILTER_POINT_SAMPLE = 6, // Simple point sampling
STBIR_FILTER_OTHER = 7, // User callback specified
} stbir_filter;
typedef enum
{
STBIR_TYPE_UINT8 = 0,
STBIR_TYPE_UINT8_SRGB = 1,
STBIR_TYPE_UINT8_SRGB_ALPHA = 2, // alpha channel, when present, should also be SRGB (this is very unusual)
STBIR_TYPE_UINT16 = 3,
STBIR_TYPE_FLOAT = 4,
STBIR_TYPE_HALF_FLOAT = 5
} stbir_datatype;
// medium api
STBIRDEF void * stbir_resize( const void *input_pixels , int input_w , int input_h, int input_stride_in_bytes,
void *output_pixels, int output_w, int output_h, int output_stride_in_bytes,
stbir_pixel_layout pixel_layout, stbir_datatype data_type,
stbir_edge edge, stbir_filter filter );
//===============================================================
//===============================================================
// Extended-complexity API
//
// This API exposes all resize functionality.
//
// * Separate filter types for each axis
// * Separate edge modes for each axis
// * Separate input and output data types
// * Can specify regions with subpixel correctness
// * Can specify alpha flags
// * Can specify a memory callback
// * Can specify a callback data type for pixel input and output
// * Can be threaded for a single resize
// * Can be used to resize many frames without recalculating the sampler info
//
// Use this API as follows:
// 1) Call the stbir_resize_init function on a local STBIR_RESIZE structure
// 2) Call any of the stbir_set functions
// 3) Optionally call stbir_build_samplers() if you are going to resample multiple times
// with the same input and output dimensions (like resizing video frames)
// 4) Resample by calling stbir_resize_extended().
// 5) Call stbir_free_samplers() if you called stbir_build_samplers()
//--------------------------------
// Types:
// INPUT CALLBACK: this callback is used for input scanlines
typedef void const * stbir_input_callback( void * optional_output, void const * input_ptr, int num_pixels, int x, int y, void * context );
// OUTPUT CALLBACK: this callback is used for output scanlines
typedef void stbir_output_callback( void const * output_ptr, int num_pixels, int y, void * context );
// callbacks for user installed filters
typedef float stbir__kernel_callback( float x, float scale, void * user_data ); // centered at zero
typedef float stbir__support_callback( float scale, void * user_data );
// internal structure with precomputed scaling
typedef struct stbir__info stbir__info;
typedef struct STBIR_RESIZE // use the stbir_resize_init and stbir_override functions to set these values for future compatibility
{
void * user_data;
void const * input_pixels;
int input_w, input_h;
double input_s0, input_t0, input_s1, input_t1;
stbir_input_callback * input_cb;
void * output_pixels;
int output_w, output_h;
int output_subx, output_suby, output_subw, output_subh;
stbir_output_callback * output_cb;
int input_stride_in_bytes;
int output_stride_in_bytes;
int splits;
int fast_alpha;
int needs_rebuild;
int called_alloc;
stbir_pixel_layout input_pixel_layout_public;
stbir_pixel_layout output_pixel_layout_public;
stbir_datatype input_data_type;
stbir_datatype output_data_type;
stbir_filter horizontal_filter, vertical_filter;
stbir_edge horizontal_edge, vertical_edge;
stbir__kernel_callback * horizontal_filter_kernel; stbir__support_callback * horizontal_filter_support;
stbir__kernel_callback * vertical_filter_kernel; stbir__support_callback * vertical_filter_support;
stbir__info * samplers;
} STBIR_RESIZE;
// extended complexity api
// First off, you must ALWAYS call stbir_resize_init on your resize structure before any of the other calls!
STBIRDEF void stbir_resize_init( STBIR_RESIZE * resize,
const void *input_pixels, int input_w, int input_h, int input_stride_in_bytes, // stride can be zero
void *output_pixels, int output_w, int output_h, int output_stride_in_bytes, // stride can be zero
stbir_pixel_layout pixel_layout, stbir_datatype data_type );
//===============================================================
// You can update these parameters any time after resize_init and there is no cost
//--------------------------------
STBIRDEF void stbir_set_datatypes( STBIR_RESIZE * resize, stbir_datatype input_type, stbir_datatype output_type );
STBIRDEF void stbir_set_pixel_callbacks( STBIR_RESIZE * resize, stbir_input_callback * input_cb, stbir_output_callback * output_cb ); // no callbacks by default
STBIRDEF void stbir_set_user_data( STBIR_RESIZE * resize, void * user_data ); // pass back STBIR_RESIZE* by default
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 );
//===============================================================
//===============================================================
// If you call any of these functions, you will trigger a sampler rebuild!
//--------------------------------
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
STBIRDEF int stbir_set_edgemodes( STBIR_RESIZE * resize, stbir_edge horizontal_edge, stbir_edge vertical_edge ); // CLAMP by default
STBIRDEF int stbir_set_filters( STBIR_RESIZE * resize, stbir_filter horizontal_filter, stbir_filter vertical_filter ); // STBIR_DEFAULT_FILTER_UPSAMPLE/DOWNSAMPLE by default
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 );
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)
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)
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)
// when inputting AND outputting non-premultiplied alpha pixels, we use a slower but higher quality technique
// that fills the zero alpha pixel's RGB values with something plausible. If you don't care about areas of
// zero alpha, you can call this function to get about a 25% speed improvement for STBIR_RGBA to STBIR_RGBA
// types of resizes.
STBIRDEF int stbir_set_non_pm_alpha_speed_over_quality( STBIR_RESIZE * resize, int non_pma_alpha_speed_over_quality );
//===============================================================
//===============================================================
// You can call build_samplers to prebuild all the internal data we need to resample.
// Then, if you call resize_extended many times with the same resize, you only pay the
// cost once.
// If you do call build_samplers, you MUST call free_samplers eventually.
//--------------------------------
// This builds the samplers and does one allocation
STBIRDEF int stbir_build_samplers( STBIR_RESIZE * resize );
// You MUST call this, if you call stbir_build_samplers or stbir_build_samplers_with_splits
STBIRDEF void stbir_free_samplers( STBIR_RESIZE * resize );
//===============================================================
// And this is the main function to perform the resize synchronously on one thread.
STBIRDEF int stbir_resize_extended( STBIR_RESIZE * resize );
//===============================================================
// Use these functions for multithreading.
// 1) You call stbir_build_samplers_with_splits first on the main thread
// 2) Then stbir_resize_with_split on each thread
// 3) stbir_free_samplers when done on the main thread
//--------------------------------
// This will build samplers for threading.
// You can pass in the number of threads you'd like to use (try_splits).
// It returns the number of splits (threads) that you can call it with.
/// It might be less if the image resize can't be split up that many ways.
STBIRDEF int stbir_build_samplers_with_splits( STBIR_RESIZE * resize, int try_splits );
// This function does a split of the resizing (you call this fuction for each
// split, on multiple threads). A split is a piece of the output resize pixel space.
// Note that you MUST call stbir_build_samplers_with_splits before stbir_resize_extended_split!
// Usually, you will always call stbir_resize_split with split_start as the thread_index
// and "1" for the split_count.
// But, if you have a weird situation where you MIGHT want 8 threads, but sometimes
// only 4 threads, you can use 0,2,4,6 for the split_start's and use "2" for the
// split_count each time to turn in into a 4 thread resize. (This is unusual).
STBIRDEF int stbir_resize_extended_split( STBIR_RESIZE * resize, int split_start, int split_count );
//===============================================================
//===============================================================
// Pixel Callbacks info:
//--------------------------------
// The input callback is super flexible - it calls you with the input address
// (based on the stride and base pointer), it gives you an optional_output
// pointer that you can fill, or you can just return your own pointer into
// your own data.
//
// You can also do conversion from non-supported data types if necessary - in
// this case, you ignore the input_ptr and just use the x and y parameters to
// calculate your own input_ptr based on the size of each non-supported pixel.
// (Something like the third example below.)
//
// You can also install just an input or just an output callback by setting the
// callback that you don't want to zero.
//
// First example, progress: (getting a callback that you can monitor the progress):
// void const * my_callback( void * optional_output, void const * input_ptr, int num_pixels, int x, int y, void * context )
// {
// percentage_done = y / input_height;
// return input_ptr; // use buffer from call
// }
//
// Next example, copying: (copy from some other buffer or stream):
// void const * my_callback( void * optional_output, void const * input_ptr, int num_pixels, int x, int y, void * context )
// {
// CopyOrStreamData( optional_output, other_data_src, num_pixels * pixel_width_in_bytes );
// return optional_output; // return the optional buffer that we filled
// }
//
// Third example, input another buffer without copying: (zero-copy from other buffer):
// void const * my_callback( void * optional_output, void const * input_ptr, int num_pixels, int x, int y, void * context )
// {
// void * pixels = ( (char*) other_image_base ) + ( y * other_image_stride ) + ( x * other_pixel_width_in_bytes );
// return pixels; // return pointer to your data without copying
// }
//
//
// The output callback is considerably simpler - it just calls you so that you can dump
// out each scanline. You could even directly copy out to disk if you have a simple format
// like TGA or BMP. You can also convert to other output types here if you want.
//
// Simple example:
// void const * my_output( void * output_ptr, int num_pixels, int y, void * context )
// {
// percentage_done = y / output_height;
// fwrite( output_ptr, pixel_width_in_bytes, num_pixels, output_file );
// }
//===============================================================
//===============================================================
// optional built-in profiling API
//--------------------------------
#ifdef STBIR_PROFILE
typedef struct STBIR_PROFILE_INFO
{
stbir_uint64 total_clocks;
// how many clocks spent (of total_clocks) in the various resize routines, along with a string description
// there are "resize_count" number of zones
stbir_uint64 clocks[ 8 ];
char const ** descriptions;
// count of clocks and descriptions
stbir_uint32 count;
} STBIR_PROFILE_INFO;
// use after calling stbir_resize_extended (or stbir_build_samplers or stbir_build_samplers_with_splits)
STBIRDEF void stbir_resize_build_profile_info( STBIR_PROFILE_INFO * out_info, STBIR_RESIZE const * resize );
// use after calling stbir_resize_extended
STBIRDEF void stbir_resize_extended_profile_info( STBIR_PROFILE_INFO * out_info, STBIR_RESIZE const * resize );
// use after calling stbir_resize_extended_split
STBIRDEF void stbir_resize_split_profile_info( STBIR_PROFILE_INFO * out_info, STBIR_RESIZE const * resize, int split_start, int split_num );
//===============================================================
#endif
//// end header file /////////////////////////////////////////////////////
#endif // STBIR_INCLUDE_STB_IMAGE_RESIZE2_H
#if defined(STB_IMAGE_RESIZE_IMPLEMENTATION) || defined(STB_IMAGE_RESIZE2_IMPLEMENTATION)
#ifndef STBIR_ASSERT
#include <assert.h>
#define STBIR_ASSERT(x) assert(x)
#endif
#ifndef STBIR_MALLOC
#include <stdlib.h>
#define STBIR_MALLOC(size,user_data) ((void)(user_data), malloc(size))
#define STBIR_FREE(ptr,user_data) ((void)(user_data), free(ptr))
// (we used the comma operator to evaluate user_data, to avoid "unused parameter" warnings)
#endif
#ifdef _MSC_VER
#define stbir__inline __forceinline
#else
#define stbir__inline __inline__
// Clang address sanitizer
#if defined(__has_feature)
#if __has_feature(address_sanitizer) || __has_feature(memory_sanitizer)
#ifndef STBIR__SEPARATE_ALLOCATIONS
#define STBIR__SEPARATE_ALLOCATIONS
#endif
#endif
#endif
#endif
// GCC and MSVC
#if defined(__SANITIZE_ADDRESS__)
#ifndef STBIR__SEPARATE_ALLOCATIONS
#define STBIR__SEPARATE_ALLOCATIONS
#endif
#endif
// Always turn off automatic FMA use - use STBIR_USE_FMA if you want.
// Otherwise, this is a determinism disaster.
#ifndef STBIR_DONT_CHANGE_FP_CONTRACT // override in case you don't want this behavior
#if defined(_MSC_VER) && !defined(__clang__)
#if _MSC_VER > 1200
#pragma fp_contract(off)
#endif
#elif defined(__GNUC__) && !defined(__clang__)
#pragma GCC optimize("fp-contract=off")
#else
#pragma STDC FP_CONTRACT OFF
#endif
#endif
#ifdef _MSC_VER
#define STBIR__UNUSED(v) (void)(v)
#else
#define STBIR__UNUSED(v) (void)sizeof(v)
#endif
#define STBIR__ARRAY_SIZE(a) (sizeof((a))/sizeof((a)[0]))
#ifndef STBIR_DEFAULT_FILTER_UPSAMPLE
#define STBIR_DEFAULT_FILTER_UPSAMPLE STBIR_FILTER_CATMULLROM
#endif
#ifndef STBIR_DEFAULT_FILTER_DOWNSAMPLE
#define STBIR_DEFAULT_FILTER_DOWNSAMPLE STBIR_FILTER_MITCHELL
#endif
#ifndef STBIR__HEADER_FILENAME
#define STBIR__HEADER_FILENAME "stb_image_resize2.h"
#endif
// the internal pixel layout enums are in a different order, so we can easily do range comparisons of types
// the public pixel layout is ordered in a way that if you cast num_channels (1-4) to the enum, you get something sensible
typedef enum
{
STBIRI_1CHANNEL = 0,
STBIRI_2CHANNEL = 1,
STBIRI_RGB = 2,
STBIRI_BGR = 3,
STBIRI_4CHANNEL = 4,
STBIRI_RGBA = 5,
STBIRI_BGRA = 6,
STBIRI_ARGB = 7,
STBIRI_ABGR = 8,
STBIRI_RA = 9,
STBIRI_AR = 10,
STBIRI_RGBA_PM = 11,
STBIRI_BGRA_PM = 12,
STBIRI_ARGB_PM = 13,
STBIRI_ABGR_PM = 14,
STBIRI_RA_PM = 15,
STBIRI_AR_PM = 16,
} stbir_internal_pixel_layout;
// define the public pixel layouts to not compile inside the implementation (to avoid accidental use)
#define STBIR_BGR bad_dont_use_in_implementation
#define STBIR_1CHANNEL STBIR_BGR
#define STBIR_2CHANNEL STBIR_BGR
#define STBIR_RGB STBIR_BGR
#define STBIR_RGBA STBIR_BGR
#define STBIR_4CHANNEL STBIR_BGR
#define STBIR_BGRA STBIR_BGR
#define STBIR_ARGB STBIR_BGR
#define STBIR_ABGR STBIR_BGR
#define STBIR_RA STBIR_BGR
#define STBIR_AR STBIR_BGR
#define STBIR_RGBA_PM STBIR_BGR
#define STBIR_BGRA_PM STBIR_BGR
#define STBIR_ARGB_PM STBIR_BGR
#define STBIR_ABGR_PM STBIR_BGR
#define STBIR_RA_PM STBIR_BGR
#define STBIR_AR_PM STBIR_BGR
// must match stbir_datatype
static unsigned char stbir__type_size[] = {
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
};
// When gathering, the contributors are which source pixels contribute.
// When scattering, the contributors are which destination pixels are contributed to.
typedef struct
{
int n0; // First contributing pixel
int n1; // Last contributing pixel
} stbir__contributors;
typedef struct
{
int lowest; // First sample index for whole filter
int highest; // Last sample index for whole filter
int widest; // widest single set of samples for an output
} stbir__filter_extent_info;
typedef struct
{
int n0; // First pixel of decode buffer to write to
int n1; // Last pixel of decode that will be written to
int pixel_offset_for_input; // Pixel offset into input_scanline
} stbir__span;
typedef struct stbir__scale_info
{
int input_full_size;
int output_sub_size;
float scale;
float inv_scale;
float pixel_shift; // starting shift in output pixel space (in pixels)
int scale_is_rational;
stbir_uint32 scale_numerator, scale_denominator;
} stbir__scale_info;
typedef struct
{
stbir__contributors * contributors;
float* coefficients;
stbir__contributors * gather_prescatter_contributors;
float * gather_prescatter_coefficients;
stbir__scale_info scale_info;
float support;
stbir_filter filter_enum;
stbir__kernel_callback * filter_kernel;
stbir__support_callback * filter_support;
stbir_edge edge;
int coefficient_width;
int filter_pixel_width;
int filter_pixel_margin;
int num_contributors;
int contributors_size;
int coefficients_size;
stbir__filter_extent_info extent_info;
int is_gather; // 0 = scatter, 1 = gather with scale >= 1, 2 = gather with scale < 1
int gather_prescatter_num_contributors;
int gather_prescatter_coefficient_width;
int gather_prescatter_contributors_size;
int gather_prescatter_coefficients_size;
} stbir__sampler;
typedef struct
{
stbir__contributors conservative;
int edge_sizes[2]; // this can be less than filter_pixel_margin, if the filter and scaling falls off
stbir__span spans[2]; // can be two spans, if doing input subrect with clamp mode WRAP
} stbir__extents;
typedef struct
{
#ifdef STBIR_PROFILE
union
{
struct { stbir_uint64 total, looping, vertical, horizontal, decode, encode, alpha, unalpha; } named;
stbir_uint64 array[8];
} profile;
stbir_uint64 * current_zone_excluded_ptr;