2 * Copyright 2011 The LibYuv Project Authors. All rights reserved.
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
11 #ifndef INCLUDE_LIBYUV_ROW_H_ // NOLINT
12 #define INCLUDE_LIBYUV_ROW_H_
14 #include <stdlib.h> // For malloc.
16 #include "libyuv/basic_types.h"
18 #if defined(__native_client__)
19 #include "ppapi/c/pp_macros.h" // For PPAPI_RELEASE
27 #define IS_ALIGNED(p, a) (!((uintptr_t)(p) & ((a) - 1)))
30 #define align_buffer_64(var, size) \
31 uint8* var##_mem = reinterpret_cast<uint8*>(malloc((size) + 63)); \
32 uint8* var = reinterpret_cast<uint8*> \
33 ((reinterpret_cast<intptr_t>(var##_mem) + 63) & ~63)
35 #define align_buffer_64(var, size) \
36 uint8* var##_mem = (uint8*)(malloc((size) + 63)); /* NOLINT */ \
37 uint8* var = (uint8*)(((intptr_t)(var##_mem) + 63) & ~63) /* NOLINT */
40 #define free_aligned_buffer_64(var) \
44 #if defined(__pnacl__) || defined(__CLR_VER) || defined(COVERAGE_ENABLED) || \
45 defined(TARGET_IPHONE_SIMULATOR) || \
46 (defined(_MSC_VER) && defined(__clang__))
47 #define LIBYUV_DISABLE_X86
49 // True if compiling for SSSE3 as a requirement.
50 #if defined(__SSSE3__) || (defined(_M_IX86_FP) && (_M_IX86_FP >= 3))
51 #define LIBYUV_SSSE3_ONLY
54 // Enable for NaCL pepper 33 for bundle and AVX2 support.
55 #if defined(__native_client__) && PPAPI_RELEASE >= 33
58 #if defined(__native_client__) && defined(__arm__) && PPAPI_RELEASE < 37
59 #define LIBYUV_DISABLE_NEON
62 // The following are available on all x86 platforms:
63 #if !defined(LIBYUV_DISABLE_X86) && \
64 (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__))
66 #define HAS_ARGBADDROW_SSE2
67 #define HAS_ARGBAFFINEROW_SSE2
68 #define HAS_ARGBATTENUATEROW_SSSE3
69 #define HAS_ARGBBLENDROW_SSSE3
70 #define HAS_ARGBCOLORMATRIXROW_SSSE3
71 #define HAS_ARGBCOLORTABLEROW_X86
72 #define HAS_ARGBCOPYALPHAROW_SSE2
73 #define HAS_ARGBCOPYYTOALPHAROW_SSE2
74 #define HAS_ARGBGRAYROW_SSSE3
75 #define HAS_ARGBLUMACOLORTABLEROW_SSSE3
76 #define HAS_ARGBMIRRORROW_SSSE3
77 #define HAS_ARGBMULTIPLYROW_SSE2
78 #define HAS_ARGBPOLYNOMIALROW_SSE2
79 #define HAS_ARGBQUANTIZEROW_SSE2
80 #define HAS_ARGBSEPIAROW_SSSE3
81 #define HAS_ARGBSHADEROW_SSE2
82 #define HAS_ARGBSUBTRACTROW_SSE2
83 #define HAS_ARGBTOUVROW_SSSE3
84 #define HAS_ARGBUNATTENUATEROW_SSE2
85 #define HAS_COMPUTECUMULATIVESUMROW_SSE2
86 #define HAS_CUMULATIVESUMTOAVERAGEROW_SSE2
87 #define HAS_INTERPOLATEROW_SSE2
88 #define HAS_INTERPOLATEROW_SSSE3
89 #define HAS_RGBCOLORTABLEROW_X86
90 #define HAS_SOBELROW_SSE2
91 #define HAS_SOBELTOPLANEROW_SSE2
92 #define HAS_SOBELXROW_SSE2
93 #define HAS_SOBELXYROW_SSE2
94 #define HAS_SOBELYROW_SSE2
97 #define HAS_ABGRTOUVROW_SSSE3
98 #define HAS_ABGRTOYROW_SSSE3
99 #define HAS_ARGB1555TOARGBROW_SSE2
100 #define HAS_ARGB4444TOARGBROW_SSE2
101 #define HAS_ARGBSHUFFLEROW_SSE2
102 #define HAS_ARGBSHUFFLEROW_SSSE3
103 #define HAS_ARGBTOARGB1555ROW_SSE2
104 #define HAS_ARGBTOARGB4444ROW_SSE2
105 #define HAS_ARGBTOBAYERGGROW_SSE2
106 #define HAS_ARGBTOBAYERROW_SSSE3
107 #define HAS_ARGBTORAWROW_SSSE3
108 #define HAS_ARGBTORGB24ROW_SSSE3
109 #define HAS_ARGBTORGB565ROW_SSE2
110 #define HAS_ARGBTOUV422ROW_SSSE3
111 #define HAS_ARGBTOUV444ROW_SSSE3
112 #define HAS_ARGBTOUVJROW_SSSE3
113 #define HAS_ARGBTOYJROW_SSSE3
114 #define HAS_ARGBTOYROW_SSSE3
115 #define HAS_BGRATOUVROW_SSSE3
116 #define HAS_BGRATOYROW_SSSE3
117 #define HAS_COPYROW_ERMS
118 #define HAS_COPYROW_SSE2
119 #define HAS_COPYROW_X86
120 #define HAS_HALFROW_SSE2
121 #define HAS_I400TOARGBROW_SSE2
122 #define HAS_I411TOARGBROW_SSSE3
123 #define HAS_I422TOARGB1555ROW_SSSE3
124 #define HAS_I422TOABGRROW_SSSE3
125 #define HAS_I422TOARGB1555ROW_SSSE3
126 #define HAS_I422TOARGB4444ROW_SSSE3
127 #define HAS_I422TOARGBROW_SSSE3
128 #define HAS_I422TOBGRAROW_SSSE3
129 #define HAS_I422TORAWROW_SSSE3
130 #define HAS_I422TORGB24ROW_SSSE3
131 #define HAS_I422TORGB565ROW_SSSE3
132 #define HAS_I422TORGBAROW_SSSE3
133 #define HAS_I422TOUYVYROW_SSE2
134 #define HAS_I422TOYUY2ROW_SSE2
135 #define HAS_I444TOARGBROW_SSSE3
136 #define HAS_MERGEUVROW_SSE2
137 #define HAS_MIRRORROW_SSE2
138 #define HAS_MIRRORROW_SSSE3
139 #define HAS_MIRRORROW_UV_SSSE3
140 #define HAS_MIRRORUVROW_SSSE3
141 #define HAS_NV12TOARGBROW_SSSE3
142 #define HAS_NV12TORGB565ROW_SSSE3
143 #define HAS_NV21TOARGBROW_SSSE3
144 #define HAS_NV21TORGB565ROW_SSSE3
145 #define HAS_RAWTOARGBROW_SSSE3
146 #define HAS_RAWTOYROW_SSSE3
147 #define HAS_RGB24TOARGBROW_SSSE3
148 #define HAS_RGB24TOYROW_SSSE3
149 #define HAS_RGB565TOARGBROW_SSE2
150 #define HAS_RGBATOUVROW_SSSE3
151 #define HAS_RGBATOYROW_SSSE3
152 #define HAS_SETROW_X86
153 #define HAS_SPLITUVROW_SSE2
154 #define HAS_UYVYTOARGBROW_SSSE3
155 #define HAS_UYVYTOUV422ROW_SSE2
156 #define HAS_UYVYTOUVROW_SSE2
157 #define HAS_UYVYTOYROW_SSE2
158 #define HAS_YTOARGBROW_SSE2
159 #define HAS_YUY2TOARGBROW_SSSE3
160 #define HAS_YUY2TOUV422ROW_SSE2
161 #define HAS_YUY2TOUVROW_SSE2
162 #define HAS_YUY2TOYROW_SSE2
165 // The following are available on x64 Visual C:
166 #if !defined(LIBYUV_DISABLE_X86) && defined (_M_X64)
167 #define HAS_I422TOARGBROW_SSSE3
170 // GCC >= 4.7.0 required for AVX2.
171 #if defined(__GNUC__) && (defined(__x86_64__) || defined(__i386__))
172 #if (__GNUC__ > 4) || (__GNUC__ == 4 && (__GNUC_MINOR__ >= 7))
173 #define GCC_HAS_AVX2 1
174 #endif // GNUC >= 4.7
177 // clang >= 3.4.0 required for AVX2.
178 #if defined(__clang__) && (defined(__x86_64__) || defined(__i386__))
179 #if (__clang_major__ > 3) || (__clang_major__ == 3 && (__clang_minor__ >= 4))
180 #define CLANG_HAS_AVX2 1
181 #endif // clang >= 3.4
184 // Visual C 2012 required for AVX2.
185 #if defined(_M_IX86) && defined(_MSC_VER) && _MSC_VER >= 1700
186 #define VISUALC_HAS_AVX2 1
187 #endif // VisualStudio >= 2012
189 // The following are available on all x86 platforms, but
190 // require VS2012, clang 3.4 or gcc 4.7.
191 // The code supports NaCL but requires a new compiler and validator.
192 #if !defined(LIBYUV_DISABLE_X86) && (defined(VISUALC_HAS_AVX2) || \
193 defined(CLANG_HAS_AVX2) || defined(GCC_HAS_AVX2))
195 #define HAS_ARGBPOLYNOMIALROW_AVX2
196 #define HAS_ARGBSHUFFLEROW_AVX2
197 #define HAS_ARGBCOPYALPHAROW_AVX2
198 #define HAS_ARGBCOPYYTOALPHAROW_AVX2
201 // The following are require VS2012.
202 // TODO(fbarchard): Port to gcc.
203 #if !defined(LIBYUV_DISABLE_X86) && defined(VISUALC_HAS_AVX2)
204 #define HAS_ARGBTOUVROW_AVX2
205 #define HAS_ARGBTOYJROW_AVX2
206 #define HAS_ARGBTOYROW_AVX2
207 #define HAS_HALFROW_AVX2
208 #define HAS_I422TOARGBROW_AVX2
209 #define HAS_INTERPOLATEROW_AVX2
210 #define HAS_MERGEUVROW_AVX2
211 #define HAS_MIRRORROW_AVX2
212 #define HAS_SPLITUVROW_AVX2
213 #define HAS_UYVYTOUV422ROW_AVX2
214 #define HAS_UYVYTOUVROW_AVX2
215 #define HAS_UYVYTOYROW_AVX2
216 #define HAS_YUY2TOUV422ROW_AVX2
217 #define HAS_YUY2TOUVROW_AVX2
218 #define HAS_YUY2TOYROW_AVX2
221 #define HAS_ARGBADDROW_AVX2
222 #define HAS_ARGBATTENUATEROW_AVX2
223 #define HAS_ARGBMIRRORROW_AVX2
224 #define HAS_ARGBMULTIPLYROW_AVX2
225 #define HAS_ARGBSUBTRACTROW_AVX2
226 #define HAS_ARGBUNATTENUATEROW_AVX2
227 #endif // defined(VISUALC_HAS_AVX2)
229 // The following are Yasm x86 only:
230 // TODO(fbarchard): Port AVX2 to inline.
231 #if !defined(LIBYUV_DISABLE_X86) && defined(HAVE_YASM)
232 (defined(_M_IX86) || defined(_M_X64) || \
233 defined(__x86_64__) || defined(__i386__))
234 #define HAS_MERGEUVROW_AVX2
235 #define HAS_MERGEUVROW_MMX
236 #define HAS_SPLITUVROW_AVX2
237 #define HAS_SPLITUVROW_MMX
238 #define HAS_UYVYTOYROW_AVX2
239 #define HAS_UYVYTOYROW_MMX
240 #define HAS_YUY2TOYROW_AVX2
241 #define HAS_YUY2TOYROW_MMX
244 // The following are disabled when SSSE3 is available:
245 #if !defined(LIBYUV_DISABLE_X86) && \
246 (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)) && \
247 !defined(LIBYUV_SSSE3_ONLY)
248 #define HAS_ARGBBLENDROW_SSE2
249 #define HAS_ARGBATTENUATEROW_SSE2
250 #define HAS_MIRRORROW_SSE2
253 // The following are available on Neon platforms:
254 #if !defined(LIBYUV_DISABLE_NEON) && \
255 (defined(__ARM_NEON__) || defined(LIBYUV_NEON))
256 #define HAS_ABGRTOUVROW_NEON
257 #define HAS_ABGRTOYROW_NEON
258 #define HAS_ARGB1555TOARGBROW_NEON
259 #define HAS_ARGB1555TOUVROW_NEON
260 #define HAS_ARGB1555TOYROW_NEON
261 #define HAS_ARGB4444TOARGBROW_NEON
262 #define HAS_ARGB4444TOUVROW_NEON
263 #define HAS_ARGB4444TOYROW_NEON
264 #define HAS_ARGBTOARGB1555ROW_NEON
265 #define HAS_ARGBTOARGB4444ROW_NEON
266 #define HAS_ARGBTOBAYERROW_NEON
267 #define HAS_ARGBTOBAYERGGROW_NEON
268 #define HAS_ARGBTORAWROW_NEON
269 #define HAS_ARGBTORGB24ROW_NEON
270 #define HAS_ARGBTORGB565ROW_NEON
271 #define HAS_ARGBTOUV411ROW_NEON
272 #define HAS_ARGBTOUV422ROW_NEON
273 #define HAS_ARGBTOUV444ROW_NEON
274 #define HAS_ARGBTOUVROW_NEON
275 #define HAS_ARGBTOUVJROW_NEON
276 #define HAS_ARGBTOYROW_NEON
277 #define HAS_ARGBTOYJROW_NEON
278 #define HAS_BGRATOUVROW_NEON
279 #define HAS_BGRATOYROW_NEON
280 #define HAS_COPYROW_NEON
281 #define HAS_HALFROW_NEON
282 #define HAS_I400TOARGBROW_NEON
283 #define HAS_I411TOARGBROW_NEON
284 #define HAS_I422TOABGRROW_NEON
285 #define HAS_I422TOARGB1555ROW_NEON
286 #define HAS_I422TOARGB4444ROW_NEON
287 #define HAS_I422TOARGBROW_NEON
288 #define HAS_I422TOBGRAROW_NEON
289 #define HAS_I422TORAWROW_NEON
290 #define HAS_I422TORGB24ROW_NEON
291 #define HAS_I422TORGB565ROW_NEON
292 #define HAS_I422TORGBAROW_NEON
293 #define HAS_I422TOUYVYROW_NEON
294 #define HAS_I422TOYUY2ROW_NEON
295 #define HAS_I444TOARGBROW_NEON
296 #define HAS_MERGEUVROW_NEON
297 #define HAS_MIRRORROW_NEON
298 #define HAS_MIRRORUVROW_NEON
299 #define HAS_NV12TOARGBROW_NEON
300 #define HAS_NV12TORGB565ROW_NEON
301 #define HAS_NV21TOARGBROW_NEON
302 #define HAS_NV21TORGB565ROW_NEON
303 #define HAS_RAWTOARGBROW_NEON
304 #define HAS_RAWTOUVROW_NEON
305 #define HAS_RAWTOYROW_NEON
306 #define HAS_RGB24TOARGBROW_NEON
307 #define HAS_RGB24TOUVROW_NEON
308 #define HAS_RGB24TOYROW_NEON
309 #define HAS_RGB565TOARGBROW_NEON
310 #define HAS_RGB565TOUVROW_NEON
311 #define HAS_RGB565TOYROW_NEON
312 #define HAS_RGBATOUVROW_NEON
313 #define HAS_RGBATOYROW_NEON
314 #define HAS_SETROW_NEON
315 #define HAS_SPLITUVROW_NEON
316 #define HAS_UYVYTOARGBROW_NEON
317 #define HAS_UYVYTOUV422ROW_NEON
318 #define HAS_UYVYTOUVROW_NEON
319 #define HAS_UYVYTOYROW_NEON
320 #define HAS_YTOARGBROW_NEON
321 #define HAS_YUY2TOARGBROW_NEON
322 #define HAS_YUY2TOUV422ROW_NEON
323 #define HAS_YUY2TOUVROW_NEON
324 #define HAS_YUY2TOYROW_NEON
327 #define HAS_ARGBADDROW_NEON
328 #define HAS_ARGBATTENUATEROW_NEON
329 #define HAS_ARGBBLENDROW_NEON
330 #define HAS_ARGBGRAYROW_NEON
331 #define HAS_ARGBMIRRORROW_NEON
332 #define HAS_ARGBMULTIPLYROW_NEON
333 #define HAS_ARGBQUANTIZEROW_NEON
334 #define HAS_ARGBSEPIAROW_NEON
335 #define HAS_ARGBSHADEROW_NEON
336 #define HAS_ARGBSUBTRACTROW_NEON
337 #define HAS_SOBELROW_NEON
338 #define HAS_SOBELTOPLANEROW_NEON
339 #define HAS_SOBELXYROW_NEON
340 #define HAS_SOBELXROW_NEON
341 #define HAS_SOBELYROW_NEON
342 #define HAS_INTERPOLATEROW_NEON
343 // TODO(fbarchard): Investigate neon unittest failure.
344 // #define HAS_ARGBCOLORMATRIXROW_NEON
347 // The following are available on Mips platforms:
348 #if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips__) && \
349 (_MIPS_SIM == _MIPS_SIM_ABI32)
350 #define HAS_COPYROW_MIPS
351 #if defined(__mips_dsp) && (__mips_dsp_rev >= 2)
352 #define HAS_I422TOABGRROW_MIPS_DSPR2
353 #define HAS_I422TOARGBROW_MIPS_DSPR2
354 #define HAS_I422TOBGRAROW_MIPS_DSPR2
355 #define HAS_INTERPOLATEROWS_MIPS_DSPR2
356 #define HAS_MIRRORROW_MIPS_DSPR2
357 #define HAS_MIRRORUVROW_MIPS_DSPR2
358 #define HAS_SPLITUVROW_MIPS_DSPR2
362 #if defined(_MSC_VER) && !defined(__CLR_VER)
363 #define SIMD_ALIGNED(var) __declspec(align(16)) var
364 typedef __declspec(align(16)) int16 vec16[8];
365 typedef __declspec(align(16)) int32 vec32[4];
366 typedef __declspec(align(16)) int8 vec8[16];
367 typedef __declspec(align(16)) uint16 uvec16[8];
368 typedef __declspec(align(16)) uint32 uvec32[4];
369 typedef __declspec(align(16)) uint8 uvec8[16];
370 typedef __declspec(align(32)) int16 lvec16[16];
371 typedef __declspec(align(32)) int32 lvec32[8];
372 typedef __declspec(align(32)) int8 lvec8[32];
373 typedef __declspec(align(32)) uint16 ulvec16[16];
374 typedef __declspec(align(32)) uint32 ulvec32[8];
375 typedef __declspec(align(32)) uint8 ulvec8[32];
377 #elif defined(__GNUC__)
378 // Caveat GCC 4.2 to 4.7 have a known issue using vectors with const.
379 #define SIMD_ALIGNED(var) var __attribute__((aligned(16)))
380 typedef int16 __attribute__((vector_size(16))) vec16;
381 typedef int32 __attribute__((vector_size(16))) vec32;
382 typedef int8 __attribute__((vector_size(16))) vec8;
383 typedef uint16 __attribute__((vector_size(16))) uvec16;
384 typedef uint32 __attribute__((vector_size(16))) uvec32;
385 typedef uint8 __attribute__((vector_size(16))) uvec8;
387 #define SIMD_ALIGNED(var) var
388 typedef int16 vec16[8];
389 typedef int32 vec32[4];
390 typedef int8 vec8[16];
391 typedef uint16 uvec16[8];
392 typedef uint32 uvec32[4];
393 typedef uint8 uvec8[16];
396 #if defined(__APPLE__) || defined(__x86_64__) || defined(__llvm__)
399 #define OMITFP __attribute__((optimize("omit-frame-pointer")))
402 // NaCL macros for GCC x86 and x64.
404 // TODO(nfullagar): When pepper_33 toolchain is distributed, default to
405 // NEW_BINUTILS and remove all BUNDLEALIGN occurances.
406 #if defined(__native_client__)
407 #define LABELALIGN ".p2align 5\n"
409 #define LABELALIGN ".p2align 2\n"
411 #if defined(__native_client__) && defined(__x86_64__)
412 #if defined(NEW_BINUTILS)
413 #define BUNDLELOCK ".bundle_lock\n"
414 #define BUNDLEUNLOCK ".bundle_unlock\n"
415 #define BUNDLEALIGN "\n"
417 #define BUNDLELOCK "\n"
418 #define BUNDLEUNLOCK "\n"
419 #define BUNDLEALIGN ".p2align 5\n"
421 #define MEMACCESS(base) "%%nacl:(%%r15,%q" #base ")"
422 #define MEMACCESS2(offset, base) "%%nacl:" #offset "(%%r15,%q" #base ")"
423 #define MEMLEA(offset, base) #offset "(%q" #base ")"
424 #define MEMLEA3(offset, index, scale) \
425 #offset "(,%q" #index "," #scale ")"
426 #define MEMLEA4(offset, base, index, scale) \
427 #offset "(%q" #base ",%q" #index "," #scale ")"
428 #define MEMMOVESTRING(s, d) "%%nacl:(%q" #s "),%%nacl:(%q" #d "), %%r15"
429 #define MEMSTORESTRING(reg, d) "%%" #reg ",%%nacl:(%q" #d "), %%r15"
430 #define MEMOPREG(opcode, offset, base, index, scale, reg) \
432 "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
433 #opcode " (%%r15,%%r14),%%" #reg "\n" \
435 #define MEMOPMEM(opcode, reg, offset, base, index, scale) \
437 "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
438 #opcode " %%" #reg ",(%%r15,%%r14)\n" \
440 #define MEMOPARG(opcode, offset, base, index, scale, arg) \
442 "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \
443 #opcode " (%%r15,%%r14),%" #arg "\n" \
445 #else // defined(__native_client__) && defined(__x86_64__)
446 #define BUNDLEALIGN "\n"
447 #define MEMACCESS(base) "(%" #base ")"
448 #define MEMACCESS2(offset, base) #offset "(%" #base ")"
449 #define MEMLEA(offset, base) #offset "(%" #base ")"
450 #define MEMLEA3(offset, index, scale) \
451 #offset "(,%" #index "," #scale ")"
452 #define MEMLEA4(offset, base, index, scale) \
453 #offset "(%" #base ",%" #index "," #scale ")"
454 #define MEMMOVESTRING(s, d)
455 #define MEMSTORESTRING(reg, d)
456 #define MEMOPREG(opcode, offset, base, index, scale, reg) \
457 #opcode " " #offset "(%" #base ",%" #index "," #scale "),%%" #reg "\n"
458 #define MEMOPMEM(opcode, reg, offset, base, index, scale) \
459 #opcode " %%" #reg ","#offset "(%" #base ",%" #index "," #scale ")\n"
460 #define MEMOPARG(opcode, offset, base, index, scale, arg) \
461 #opcode " " #offset "(%" #base ",%" #index "," #scale "),%" #arg "\n"
462 #endif // defined(__native_client__) && defined(__x86_64__)
466 #if defined(__native_client__)
467 #define MEMACCESS(base) ".p2align 3\nbic %" #base ", #0xc0000000\n"
469 #define MEMACCESS(base) "\n"
473 void I444ToARGBRow_NEON(const uint8* src_y,
478 void I422ToARGBRow_NEON(const uint8* src_y,
483 void I411ToARGBRow_NEON(const uint8* src_y,
488 void I422ToBGRARow_NEON(const uint8* src_y,
493 void I422ToABGRRow_NEON(const uint8* src_y,
498 void I422ToRGBARow_NEON(const uint8* src_y,
503 void I422ToRGB24Row_NEON(const uint8* src_y,
508 void I422ToRAWRow_NEON(const uint8* src_y,
513 void I422ToRGB565Row_NEON(const uint8* src_y,
518 void I422ToARGB1555Row_NEON(const uint8* src_y,
523 void I422ToARGB4444Row_NEON(const uint8* src_y,
528 void NV12ToARGBRow_NEON(const uint8* src_y,
532 void NV21ToARGBRow_NEON(const uint8* src_y,
536 void NV12ToRGB565Row_NEON(const uint8* src_y,
540 void NV21ToRGB565Row_NEON(const uint8* src_y,
544 void YUY2ToARGBRow_NEON(const uint8* src_yuy2,
547 void UYVYToARGBRow_NEON(const uint8* src_uyvy,
551 void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
552 void ARGBToYRow_Any_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
553 void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
554 void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
555 void ARGBToYJRow_Any_AVX2(const uint8* src_argb, uint8* dst_y, int pix);
556 void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
557 void BGRAToYRow_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix);
558 void ABGRToYRow_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix);
559 void RGBAToYRow_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix);
560 void RGB24ToYRow_SSSE3(const uint8* src_rgb24, uint8* dst_y, int pix);
561 void RAWToYRow_SSSE3(const uint8* src_raw, uint8* dst_y, int pix);
562 void ARGBToYRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
563 void ARGBToYJRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
564 void BGRAToYRow_Unaligned_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix);
565 void ABGRToYRow_Unaligned_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix);
566 void RGBAToYRow_Unaligned_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix);
567 void RGB24ToYRow_Unaligned_SSSE3(const uint8* src_rgb24, uint8* dst_y, int pix);
568 void RAWToYRow_Unaligned_SSSE3(const uint8* src_raw, uint8* dst_y, int pix);
569 void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int pix);
570 void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int pix);
571 void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
573 void ARGBToUV422Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
575 void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
577 void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb,
578 uint8* dst_u, uint8* dst_v, int pix);
579 void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb,
580 uint8* dst_u, uint8* dst_v, int pix);
581 void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra,
582 uint8* dst_u, uint8* dst_v, int pix);
583 void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr,
584 uint8* dst_u, uint8* dst_v, int pix);
585 void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba,
586 uint8* dst_u, uint8* dst_v, int pix);
587 void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24,
588 uint8* dst_u, uint8* dst_v, int pix);
589 void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw,
590 uint8* dst_u, uint8* dst_v, int pix);
591 void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565,
592 uint8* dst_u, uint8* dst_v, int pix);
593 void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555,
594 uint8* dst_u, uint8* dst_v, int pix);
595 void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444,
596 uint8* dst_u, uint8* dst_v, int pix);
597 void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int pix);
598 void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int pix);
599 void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int pix);
600 void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int pix);
601 void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int pix);
602 void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int pix);
603 void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int pix);
604 void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int pix);
605 void ARGBToYRow_C(const uint8* src_argb, uint8* dst_y, int pix);
606 void ARGBToYJRow_C(const uint8* src_argb, uint8* dst_y, int pix);
607 void BGRAToYRow_C(const uint8* src_bgra, uint8* dst_y, int pix);
608 void ABGRToYRow_C(const uint8* src_abgr, uint8* dst_y, int pix);
609 void RGBAToYRow_C(const uint8* src_rgba, uint8* dst_y, int pix);
610 void RGB24ToYRow_C(const uint8* src_rgb24, uint8* dst_y, int pix);
611 void RAWToYRow_C(const uint8* src_raw, uint8* dst_y, int pix);
612 void RGB565ToYRow_C(const uint8* src_rgb565, uint8* dst_y, int pix);
613 void ARGB1555ToYRow_C(const uint8* src_argb1555, uint8* dst_y, int pix);
614 void ARGB4444ToYRow_C(const uint8* src_argb4444, uint8* dst_y, int pix);
615 void ARGBToYRow_Any_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
616 void ARGBToYJRow_Any_SSSE3(const uint8* src_argb, uint8* dst_y, int pix);
617 void BGRAToYRow_Any_SSSE3(const uint8* src_bgra, uint8* dst_y, int pix);
618 void ABGRToYRow_Any_SSSE3(const uint8* src_abgr, uint8* dst_y, int pix);
619 void RGBAToYRow_Any_SSSE3(const uint8* src_rgba, uint8* dst_y, int pix);
620 void RGB24ToYRow_Any_SSSE3(const uint8* src_rgb24, uint8* dst_y, int pix);
621 void RAWToYRow_Any_SSSE3(const uint8* src_raw, uint8* dst_y, int pix);
622 void ARGBToYRow_Any_NEON(const uint8* src_argb, uint8* dst_y, int pix);
623 void ARGBToYJRow_Any_NEON(const uint8* src_argb, uint8* dst_y, int pix);
624 void BGRAToYRow_Any_NEON(const uint8* src_bgra, uint8* dst_y, int pix);
625 void ABGRToYRow_Any_NEON(const uint8* src_abgr, uint8* dst_y, int pix);
626 void RGBAToYRow_Any_NEON(const uint8* src_rgba, uint8* dst_y, int pix);
627 void RGB24ToYRow_Any_NEON(const uint8* src_rgb24, uint8* dst_y, int pix);
628 void RAWToYRow_Any_NEON(const uint8* src_raw, uint8* dst_y, int pix);
629 void RGB565ToYRow_Any_NEON(const uint8* src_rgb565, uint8* dst_y, int pix);
630 void ARGB1555ToYRow_Any_NEON(const uint8* src_argb1555, uint8* dst_y, int pix);
631 void ARGB4444ToYRow_Any_NEON(const uint8* src_argb4444, uint8* dst_y, int pix);
633 void ARGBToUVRow_AVX2(const uint8* src_argb, int src_stride_argb,
634 uint8* dst_u, uint8* dst_v, int width);
635 void ARGBToUVRow_Any_AVX2(const uint8* src_argb, int src_stride_argb,
636 uint8* dst_u, uint8* dst_v, int width);
637 void ARGBToUVRow_SSSE3(const uint8* src_argb, int src_stride_argb,
638 uint8* dst_u, uint8* dst_v, int width);
639 void ARGBToUVJRow_SSSE3(const uint8* src_argb, int src_stride_argb,
640 uint8* dst_u, uint8* dst_v, int width);
641 void BGRAToUVRow_SSSE3(const uint8* src_bgra, int src_stride_bgra,
642 uint8* dst_u, uint8* dst_v, int width);
643 void ABGRToUVRow_SSSE3(const uint8* src_abgr, int src_stride_abgr,
644 uint8* dst_u, uint8* dst_v, int width);
645 void RGBAToUVRow_SSSE3(const uint8* src_rgba, int src_stride_rgba,
646 uint8* dst_u, uint8* dst_v, int width);
647 void ARGBToUVRow_Unaligned_SSSE3(const uint8* src_argb, int src_stride_argb,
648 uint8* dst_u, uint8* dst_v, int width);
649 void ARGBToUVJRow_Unaligned_SSSE3(const uint8* src_argb, int src_stride_argb,
650 uint8* dst_u, uint8* dst_v, int width);
651 void BGRAToUVRow_Unaligned_SSSE3(const uint8* src_bgra, int src_stride_bgra,
652 uint8* dst_u, uint8* dst_v, int width);
653 void ABGRToUVRow_Unaligned_SSSE3(const uint8* src_abgr, int src_stride_abgr,
654 uint8* dst_u, uint8* dst_v, int width);
655 void RGBAToUVRow_Unaligned_SSSE3(const uint8* src_rgba, int src_stride_rgba,
656 uint8* dst_u, uint8* dst_v, int width);
657 void ARGBToUVRow_Any_SSSE3(const uint8* src_argb, int src_stride_argb,
658 uint8* dst_u, uint8* dst_v, int width);
659 void ARGBToUVJRow_Any_SSSE3(const uint8* src_argb, int src_stride_argb,
660 uint8* dst_u, uint8* dst_v, int width);
661 void BGRAToUVRow_Any_SSSE3(const uint8* src_bgra, int src_stride_bgra,
662 uint8* dst_u, uint8* dst_v, int width);
663 void ABGRToUVRow_Any_SSSE3(const uint8* src_abgr, int src_stride_abgr,
664 uint8* dst_u, uint8* dst_v, int width);
665 void RGBAToUVRow_Any_SSSE3(const uint8* src_rgba, int src_stride_rgba,
666 uint8* dst_u, uint8* dst_v, int width);
667 void ARGBToUV444Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
669 void ARGBToUV422Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
671 void ARGBToUV411Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v,
673 void ARGBToUVRow_Any_NEON(const uint8* src_argb, int src_stride_argb,
674 uint8* dst_u, uint8* dst_v, int pix);
675 void ARGBToUVJRow_Any_NEON(const uint8* src_argb, int src_stride_argb,
676 uint8* dst_u, uint8* dst_v, int pix);
677 void BGRAToUVRow_Any_NEON(const uint8* src_bgra, int src_stride_bgra,
678 uint8* dst_u, uint8* dst_v, int pix);
679 void ABGRToUVRow_Any_NEON(const uint8* src_abgr, int src_stride_abgr,
680 uint8* dst_u, uint8* dst_v, int pix);
681 void RGBAToUVRow_Any_NEON(const uint8* src_rgba, int src_stride_rgba,
682 uint8* dst_u, uint8* dst_v, int pix);
683 void RGB24ToUVRow_Any_NEON(const uint8* src_rgb24, int src_stride_rgb24,
684 uint8* dst_u, uint8* dst_v, int pix);
685 void RAWToUVRow_Any_NEON(const uint8* src_raw, int src_stride_raw,
686 uint8* dst_u, uint8* dst_v, int pix);
687 void RGB565ToUVRow_Any_NEON(const uint8* src_rgb565, int src_stride_rgb565,
688 uint8* dst_u, uint8* dst_v, int pix);
689 void ARGB1555ToUVRow_Any_NEON(const uint8* src_argb1555,
690 int src_stride_argb1555,
691 uint8* dst_u, uint8* dst_v, int pix);
692 void ARGB4444ToUVRow_Any_NEON(const uint8* src_argb4444,
693 int src_stride_argb4444,
694 uint8* dst_u, uint8* dst_v, int pix);
695 void ARGBToUVRow_C(const uint8* src_argb, int src_stride_argb,
696 uint8* dst_u, uint8* dst_v, int width);
697 void ARGBToUVJRow_C(const uint8* src_argb, int src_stride_argb,
698 uint8* dst_u, uint8* dst_v, int width);
699 void BGRAToUVRow_C(const uint8* src_bgra, int src_stride_bgra,
700 uint8* dst_u, uint8* dst_v, int width);
701 void ABGRToUVRow_C(const uint8* src_abgr, int src_stride_abgr,
702 uint8* dst_u, uint8* dst_v, int width);
703 void RGBAToUVRow_C(const uint8* src_rgba, int src_stride_rgba,
704 uint8* dst_u, uint8* dst_v, int width);
705 void RGB24ToUVRow_C(const uint8* src_rgb24, int src_stride_rgb24,
706 uint8* dst_u, uint8* dst_v, int width);
707 void RAWToUVRow_C(const uint8* src_raw, int src_stride_raw,
708 uint8* dst_u, uint8* dst_v, int width);
709 void RGB565ToUVRow_C(const uint8* src_rgb565, int src_stride_rgb565,
710 uint8* dst_u, uint8* dst_v, int width);
711 void ARGB1555ToUVRow_C(const uint8* src_argb1555, int src_stride_argb1555,
712 uint8* dst_u, uint8* dst_v, int width);
713 void ARGB4444ToUVRow_C(const uint8* src_argb4444, int src_stride_argb4444,
714 uint8* dst_u, uint8* dst_v, int width);
716 void ARGBToUV444Row_SSSE3(const uint8* src_argb,
717 uint8* dst_u, uint8* dst_v, int width);
718 void ARGBToUV444Row_Unaligned_SSSE3(const uint8* src_argb,
719 uint8* dst_u, uint8* dst_v, int width);
720 void ARGBToUV444Row_Any_SSSE3(const uint8* src_argb,
721 uint8* dst_u, uint8* dst_v, int width);
723 void ARGBToUV422Row_SSSE3(const uint8* src_argb,
724 uint8* dst_u, uint8* dst_v, int width);
725 void ARGBToUV422Row_Unaligned_SSSE3(const uint8* src_argb,
726 uint8* dst_u, uint8* dst_v, int width);
727 void ARGBToUV422Row_Any_SSSE3(const uint8* src_argb,
728 uint8* dst_u, uint8* dst_v, int width);
730 void ARGBToUV444Row_C(const uint8* src_argb,
731 uint8* dst_u, uint8* dst_v, int width);
732 void ARGBToUV422Row_C(const uint8* src_argb,
733 uint8* dst_u, uint8* dst_v, int width);
734 void ARGBToUV411Row_C(const uint8* src_argb,
735 uint8* dst_u, uint8* dst_v, int width);
737 void MirrorRow_AVX2(const uint8* src, uint8* dst, int width);
738 void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width);
739 void MirrorRow_SSE2(const uint8* src, uint8* dst, int width);
740 void MirrorRow_NEON(const uint8* src, uint8* dst, int width);
741 void MirrorRow_MIPS_DSPR2(const uint8* src, uint8* dst, int width);
742 void MirrorRow_C(const uint8* src, uint8* dst, int width);
744 void MirrorUVRow_SSSE3(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
746 void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
748 void MirrorUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
750 void MirrorUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
753 void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width);
754 void ARGBMirrorRow_SSSE3(const uint8* src, uint8* dst, int width);
755 void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width);
756 void ARGBMirrorRow_C(const uint8* src, uint8* dst, int width);
758 void SplitUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
759 void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
760 void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
761 void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int pix);
762 void SplitUVRow_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
764 void SplitUVRow_Unaligned_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
766 void SplitUVRow_Unaligned_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u,
767 uint8* dst_v, int pix);
768 void SplitUVRow_Any_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
770 void SplitUVRow_Any_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
772 void SplitUVRow_Any_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
774 void SplitUVRow_Any_MIPS_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v,
777 void MergeUVRow_C(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
779 void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
781 void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
783 void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
785 void MergeUVRow_Unaligned_SSE2(const uint8* src_u, const uint8* src_v,
786 uint8* dst_uv, int width);
787 void MergeUVRow_Any_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
789 void MergeUVRow_Any_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
791 void MergeUVRow_Any_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv,
794 void CopyRow_SSE2(const uint8* src, uint8* dst, int count);
795 void CopyRow_ERMS(const uint8* src, uint8* dst, int count);
796 void CopyRow_X86(const uint8* src, uint8* dst, int count);
797 void CopyRow_NEON(const uint8* src, uint8* dst, int count);
798 void CopyRow_MIPS(const uint8* src, uint8* dst, int count);
799 void CopyRow_C(const uint8* src, uint8* dst, int count);
801 void CopyRow_16_C(const uint16* src, uint16* dst, int count);
803 void ARGBCopyAlphaRow_C(const uint8* src_argb, uint8* dst_argb, int width);
804 void ARGBCopyAlphaRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
805 void ARGBCopyAlphaRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
807 void ARGBCopyYToAlphaRow_C(const uint8* src_y, uint8* dst_argb, int width);
808 void ARGBCopyYToAlphaRow_SSE2(const uint8* src_y, uint8* dst_argb, int width);
809 void ARGBCopyYToAlphaRow_AVX2(const uint8* src_y, uint8* dst_argb, int width);
811 void SetRow_X86(uint8* dst, uint32 v32, int count);
812 void ARGBSetRows_X86(uint8* dst, uint32 v32, int width,
813 int dst_stride, int height);
814 void SetRow_NEON(uint8* dst, uint32 v32, int count);
815 void ARGBSetRows_NEON(uint8* dst, uint32 v32, int width,
816 int dst_stride, int height);
817 void SetRow_C(uint8* dst, uint32 v32, int count);
818 void ARGBSetRows_C(uint8* dst, uint32 v32, int width, int dst_stride,
821 // ARGBShufflers for BGRAToARGB etc.
822 void ARGBShuffleRow_C(const uint8* src_argb, uint8* dst_argb,
823 const uint8* shuffler, int pix);
824 void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb,
825 const uint8* shuffler, int pix);
826 void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
827 const uint8* shuffler, int pix);
828 void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb,
829 const uint8* shuffler, int pix);
830 void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb,
831 const uint8* shuffler, int pix);
832 void ARGBShuffleRow_Unaligned_SSSE3(const uint8* src_argb, uint8* dst_argb,
833 const uint8* shuffler, int pix);
834 void ARGBShuffleRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
835 const uint8* shuffler, int pix);
836 void ARGBShuffleRow_Any_SSSE3(const uint8* src_argb, uint8* dst_argb,
837 const uint8* shuffler, int pix);
838 void ARGBShuffleRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
839 const uint8* shuffler, int pix);
840 void ARGBShuffleRow_Any_NEON(const uint8* src_argb, uint8* dst_argb,
841 const uint8* shuffler, int pix);
843 void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix);
844 void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix);
845 void RGB565ToARGBRow_SSE2(const uint8* src_rgb565, uint8* dst_argb, int pix);
846 void ARGB1555ToARGBRow_SSE2(const uint8* src_argb1555, uint8* dst_argb,
848 void ARGB4444ToARGBRow_SSE2(const uint8* src_argb4444, uint8* dst_argb,
851 void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix);
852 void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int pix);
853 void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int pix);
854 void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb,
856 void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb,
858 void RGB24ToARGBRow_C(const uint8* src_rgb24, uint8* dst_argb, int pix);
859 void RAWToARGBRow_C(const uint8* src_raw, uint8* dst_argb, int pix);
860 void RGB565ToARGBRow_C(const uint8* src_rgb, uint8* dst_argb, int pix);
861 void ARGB1555ToARGBRow_C(const uint8* src_argb, uint8* dst_argb, int pix);
862 void ARGB4444ToARGBRow_C(const uint8* src_argb, uint8* dst_argb, int pix);
863 void RGB24ToARGBRow_Any_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int pix);
864 void RAWToARGBRow_Any_SSSE3(const uint8* src_raw, uint8* dst_argb, int pix);
865 void RGB565ToARGBRow_Any_SSE2(const uint8* src_rgb565, uint8* dst_argb,
867 void ARGB1555ToARGBRow_Any_SSE2(const uint8* src_argb1555, uint8* dst_argb,
869 void ARGB4444ToARGBRow_Any_SSE2(const uint8* src_argb4444, uint8* dst_argb,
871 void RGB24ToARGBRow_Any_NEON(const uint8* src_rgb24, uint8* dst_argb, int pix);
872 void RAWToARGBRow_Any_NEON(const uint8* src_raw, uint8* dst_argb, int pix);
873 void RGB565ToARGBRow_Any_NEON(const uint8* src_rgb565, uint8* dst_argb,
875 void ARGB1555ToARGBRow_Any_NEON(const uint8* src_argb1555, uint8* dst_argb,
877 void ARGB4444ToARGBRow_Any_NEON(const uint8* src_argb4444, uint8* dst_argb,
880 void ARGBToRGB24Row_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
881 void ARGBToRAWRow_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
882 void ARGBToRGB565Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
883 void ARGBToARGB1555Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
884 void ARGBToARGB4444Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
886 void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
887 void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
888 void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
889 void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
890 void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
892 void ARGBToRGBARow_C(const uint8* src_argb, uint8* dst_rgb, int pix);
893 void ARGBToRGB24Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
894 void ARGBToRAWRow_C(const uint8* src_argb, uint8* dst_rgb, int pix);
895 void ARGBToRGB565Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
896 void ARGBToARGB1555Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
897 void ARGBToARGB4444Row_C(const uint8* src_argb, uint8* dst_rgb, int pix);
899 void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
900 void I400ToARGBRow_Unaligned_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
901 void I400ToARGBRow_NEON(const uint8* src_y, uint8* dst_argb, int pix);
902 void I400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int pix);
903 void I400ToARGBRow_Any_SSE2(const uint8* src_y, uint8* dst_argb, int pix);
904 void I400ToARGBRow_Any_NEON(const uint8* src_y, uint8* dst_argb, int pix);
906 void I444ToARGBRow_C(const uint8* src_y,
911 void I422ToARGBRow_C(const uint8* src_y,
916 void I411ToARGBRow_C(const uint8* src_y,
921 void NV12ToARGBRow_C(const uint8* src_y,
925 void NV21ToRGB565Row_C(const uint8* src_y,
929 void NV12ToRGB565Row_C(const uint8* src_y,
933 void NV21ToARGBRow_C(const uint8* src_y,
937 void YUY2ToARGBRow_C(const uint8* src_yuy2,
940 void UYVYToARGBRow_C(const uint8* src_uyvy,
943 void I422ToBGRARow_C(const uint8* src_y,
948 void I422ToABGRRow_C(const uint8* src_y,
953 void I422ToRGBARow_C(const uint8* src_y,
958 void I422ToRGB24Row_C(const uint8* src_y,
963 void I422ToRAWRow_C(const uint8* src_y,
968 void I422ToARGB4444Row_C(const uint8* src_y,
973 void I422ToARGB1555Row_C(const uint8* src_y,
978 void I422ToRGB565Row_C(const uint8* src_y,
983 void YToARGBRow_C(const uint8* src_y,
986 void I422ToARGBRow_AVX2(const uint8* src_y,
991 void I444ToARGBRow_SSSE3(const uint8* src_y,
996 void I422ToARGBRow_SSSE3(const uint8* src_y,
1001 void I411ToARGBRow_SSSE3(const uint8* src_y,
1006 void NV12ToARGBRow_SSSE3(const uint8* src_y,
1007 const uint8* src_uv,
1010 void NV21ToARGBRow_SSSE3(const uint8* src_y,
1011 const uint8* src_vu,
1014 void NV12ToRGB565Row_SSSE3(const uint8* src_y,
1015 const uint8* src_uv,
1018 void NV21ToRGB565Row_SSSE3(const uint8* src_y,
1019 const uint8* src_vu,
1022 void YUY2ToARGBRow_SSSE3(const uint8* src_yuy2,
1025 void UYVYToARGBRow_SSSE3(const uint8* src_uyvy,
1028 void I422ToBGRARow_SSSE3(const uint8* src_y,
1033 void I422ToABGRRow_SSSE3(const uint8* src_y,
1038 void I422ToRGBARow_SSSE3(const uint8* src_y,
1043 void I422ToARGB4444Row_SSSE3(const uint8* src_y,
1048 void I422ToARGB1555Row_SSSE3(const uint8* src_y,
1053 void I422ToRGB565Row_SSSE3(const uint8* src_y,
1058 // RGB24/RAW are unaligned.
1059 void I422ToRGB24Row_SSSE3(const uint8* src_y,
1064 void I422ToRAWRow_SSSE3(const uint8* src_y,
1070 void I444ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
1075 void I422ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
1080 void I411ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
1085 void NV12ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
1086 const uint8* src_uv,
1089 void NV21ToARGBRow_Unaligned_SSSE3(const uint8* src_y,
1090 const uint8* src_vu,
1093 void YUY2ToARGBRow_Unaligned_SSSE3(const uint8* src_yuy2,
1096 void UYVYToARGBRow_Unaligned_SSSE3(const uint8* src_uyvy,
1099 void I422ToBGRARow_Unaligned_SSSE3(const uint8* src_y,
1104 void I422ToABGRRow_Unaligned_SSSE3(const uint8* src_y,
1109 void I422ToRGBARow_Unaligned_SSSE3(const uint8* src_y,
1114 void I422ToARGBRow_Any_AVX2(const uint8* src_y,
1119 void I444ToARGBRow_Any_SSSE3(const uint8* src_y,
1124 void I422ToARGBRow_Any_SSSE3(const uint8* src_y,
1129 void I411ToARGBRow_Any_SSSE3(const uint8* src_y,
1134 void NV12ToARGBRow_Any_SSSE3(const uint8* src_y,
1135 const uint8* src_uv,
1138 void NV21ToARGBRow_Any_SSSE3(const uint8* src_y,
1139 const uint8* src_vu,
1142 void NV12ToRGB565Row_Any_SSSE3(const uint8* src_y,
1143 const uint8* src_uv,
1146 void NV21ToRGB565Row_Any_SSSE3(const uint8* src_y,
1147 const uint8* src_vu,
1150 void YUY2ToARGBRow_Any_SSSE3(const uint8* src_yuy2,
1153 void UYVYToARGBRow_Any_SSSE3(const uint8* src_uyvy,
1156 void I422ToBGRARow_Any_SSSE3(const uint8* src_y,
1161 void I422ToABGRRow_Any_SSSE3(const uint8* src_y,
1166 void I422ToRGBARow_Any_SSSE3(const uint8* src_y,
1171 void I422ToARGB4444Row_Any_SSSE3(const uint8* src_y,
1176 void I422ToARGB1555Row_Any_SSSE3(const uint8* src_y,
1181 void I422ToRGB565Row_Any_SSSE3(const uint8* src_y,
1186 // RGB24/RAW are unaligned.
1187 void I422ToRGB24Row_Any_SSSE3(const uint8* src_y,
1192 void I422ToRAWRow_Any_SSSE3(const uint8* src_y,
1197 void YToARGBRow_SSE2(const uint8* src_y,
1200 void YToARGBRow_NEON(const uint8* src_y,
1203 void YToARGBRow_Any_SSE2(const uint8* src_y,
1206 void YToARGBRow_Any_NEON(const uint8* src_y,
1210 // ARGB preattenuated alpha blend.
1211 void ARGBBlendRow_SSSE3(const uint8* src_argb, const uint8* src_argb1,
1212 uint8* dst_argb, int width);
1213 void ARGBBlendRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
1214 uint8* dst_argb, int width);
1215 void ARGBBlendRow_NEON(const uint8* src_argb, const uint8* src_argb1,
1216 uint8* dst_argb, int width);
1217 void ARGBBlendRow_C(const uint8* src_argb, const uint8* src_argb1,
1218 uint8* dst_argb, int width);
1220 // ARGB multiply images. Same API as Blend, but these require
1221 // pointer and width alignment for SSE2.
1222 void ARGBMultiplyRow_C(const uint8* src_argb, const uint8* src_argb1,
1223 uint8* dst_argb, int width);
1224 void ARGBMultiplyRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
1225 uint8* dst_argb, int width);
1226 void ARGBMultiplyRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
1227 uint8* dst_argb, int width);
1228 void ARGBMultiplyRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
1229 uint8* dst_argb, int width);
1230 void ARGBMultiplyRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
1231 uint8* dst_argb, int width);
1232 void ARGBMultiplyRow_NEON(const uint8* src_argb, const uint8* src_argb1,
1233 uint8* dst_argb, int width);
1234 void ARGBMultiplyRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
1235 uint8* dst_argb, int width);
1238 void ARGBAddRow_C(const uint8* src_argb, const uint8* src_argb1,
1239 uint8* dst_argb, int width);
1240 void ARGBAddRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
1241 uint8* dst_argb, int width);
1242 void ARGBAddRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
1243 uint8* dst_argb, int width);
1244 void ARGBAddRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
1245 uint8* dst_argb, int width);
1246 void ARGBAddRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
1247 uint8* dst_argb, int width);
1248 void ARGBAddRow_NEON(const uint8* src_argb, const uint8* src_argb1,
1249 uint8* dst_argb, int width);
1250 void ARGBAddRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
1251 uint8* dst_argb, int width);
1253 // ARGB subtract images. Same API as Blend, but these require
1254 // pointer and width alignment for SSE2.
1255 void ARGBSubtractRow_C(const uint8* src_argb, const uint8* src_argb1,
1256 uint8* dst_argb, int width);
1257 void ARGBSubtractRow_SSE2(const uint8* src_argb, const uint8* src_argb1,
1258 uint8* dst_argb, int width);
1259 void ARGBSubtractRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1,
1260 uint8* dst_argb, int width);
1261 void ARGBSubtractRow_AVX2(const uint8* src_argb, const uint8* src_argb1,
1262 uint8* dst_argb, int width);
1263 void ARGBSubtractRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1,
1264 uint8* dst_argb, int width);
1265 void ARGBSubtractRow_NEON(const uint8* src_argb, const uint8* src_argb1,
1266 uint8* dst_argb, int width);
1267 void ARGBSubtractRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1,
1268 uint8* dst_argb, int width);
1270 void ARGBToRGB24Row_Any_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
1271 void ARGBToRAWRow_Any_SSSE3(const uint8* src_argb, uint8* dst_rgb, int pix);
1272 void ARGBToRGB565Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
1273 void ARGBToARGB1555Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
1274 void ARGBToARGB4444Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int pix);
1276 void ARGBToRGB24Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
1277 void ARGBToRAWRow_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
1278 void ARGBToRGB565Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
1279 void ARGBToARGB1555Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
1280 void ARGBToARGB4444Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int pix);
1282 void I444ToARGBRow_Any_NEON(const uint8* src_y,
1287 void I422ToARGBRow_Any_NEON(const uint8* src_y,
1292 void I411ToARGBRow_Any_NEON(const uint8* src_y,
1297 void I422ToBGRARow_Any_NEON(const uint8* src_y,
1302 void I422ToABGRRow_Any_NEON(const uint8* src_y,
1307 void I422ToRGBARow_Any_NEON(const uint8* src_y,
1312 void I422ToRGB24Row_Any_NEON(const uint8* src_y,
1317 void I422ToRAWRow_Any_NEON(const uint8* src_y,
1322 void I422ToARGB4444Row_Any_NEON(const uint8* src_y,
1327 void I422ToARGB1555Row_Any_NEON(const uint8* src_y,
1332 void I422ToRGB565Row_Any_NEON(const uint8* src_y,
1337 void NV12ToARGBRow_Any_NEON(const uint8* src_y,
1338 const uint8* src_uv,
1341 void NV21ToARGBRow_Any_NEON(const uint8* src_y,
1342 const uint8* src_uv,
1345 void NV12ToRGB565Row_Any_NEON(const uint8* src_y,
1346 const uint8* src_uv,
1349 void NV21ToRGB565Row_Any_NEON(const uint8* src_y,
1350 const uint8* src_uv,
1353 void YUY2ToARGBRow_Any_NEON(const uint8* src_yuy2,
1356 void UYVYToARGBRow_Any_NEON(const uint8* src_uyvy,
1359 void I422ToARGBRow_MIPS_DSPR2(const uint8* src_y,
1364 void I422ToBGRARow_MIPS_DSPR2(const uint8* src_y,
1369 void I422ToABGRRow_MIPS_DSPR2(const uint8* src_y,
1374 void I422ToARGBRow_MIPS_DSPR2(const uint8* src_y,
1379 void I422ToBGRARow_MIPS_DSPR2(const uint8* src_y,
1384 void I422ToABGRRow_MIPS_DSPR2(const uint8* src_y,
1390 void YUY2ToYRow_AVX2(const uint8* src_yuy2, uint8* dst_y, int pix);
1391 void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2,
1392 uint8* dst_u, uint8* dst_v, int pix);
1393 void YUY2ToUV422Row_AVX2(const uint8* src_yuy2,
1394 uint8* dst_u, uint8* dst_v, int pix);
1395 void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix);
1396 void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2,
1397 uint8* dst_u, uint8* dst_v, int pix);
1398 void YUY2ToUV422Row_SSE2(const uint8* src_yuy2,
1399 uint8* dst_u, uint8* dst_v, int pix);
1400 void YUY2ToYRow_Unaligned_SSE2(const uint8* src_yuy2,
1401 uint8* dst_y, int pix);
1402 void YUY2ToUVRow_Unaligned_SSE2(const uint8* src_yuy2, int stride_yuy2,
1403 uint8* dst_u, uint8* dst_v, int pix);
1404 void YUY2ToUV422Row_Unaligned_SSE2(const uint8* src_yuy2,
1405 uint8* dst_u, uint8* dst_v, int pix);
1406 void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int pix);
1407 void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2,
1408 uint8* dst_u, uint8* dst_v, int pix);
1409 void YUY2ToUV422Row_NEON(const uint8* src_yuy2,
1410 uint8* dst_u, uint8* dst_v, int pix);
1411 void YUY2ToYRow_C(const uint8* src_yuy2, uint8* dst_y, int pix);
1412 void YUY2ToUVRow_C(const uint8* src_yuy2, int stride_yuy2,
1413 uint8* dst_u, uint8* dst_v, int pix);
1414 void YUY2ToUV422Row_C(const uint8* src_yuy2,
1415 uint8* dst_u, uint8* dst_v, int pix);
1416 void YUY2ToYRow_Any_AVX2(const uint8* src_yuy2, uint8* dst_y, int pix);
1417 void YUY2ToUVRow_Any_AVX2(const uint8* src_yuy2, int stride_yuy2,
1418 uint8* dst_u, uint8* dst_v, int pix);
1419 void YUY2ToUV422Row_Any_AVX2(const uint8* src_yuy2,
1420 uint8* dst_u, uint8* dst_v, int pix);
1421 void YUY2ToYRow_Any_SSE2(const uint8* src_yuy2, uint8* dst_y, int pix);
1422 void YUY2ToUVRow_Any_SSE2(const uint8* src_yuy2, int stride_yuy2,
1423 uint8* dst_u, uint8* dst_v, int pix);
1424 void YUY2ToUV422Row_Any_SSE2(const uint8* src_yuy2,
1425 uint8* dst_u, uint8* dst_v, int pix);
1426 void YUY2ToYRow_Any_NEON(const uint8* src_yuy2, uint8* dst_y, int pix);
1427 void YUY2ToUVRow_Any_NEON(const uint8* src_yuy2, int stride_yuy2,
1428 uint8* dst_u, uint8* dst_v, int pix);
1429 void YUY2ToUV422Row_Any_NEON(const uint8* src_yuy2,
1430 uint8* dst_u, uint8* dst_v, int pix);
1431 void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix);
1432 void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
1433 uint8* dst_u, uint8* dst_v, int pix);
1434 void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
1435 uint8* dst_u, uint8* dst_v, int pix);
1436 void UYVYToYRow_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix);
1437 void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy,
1438 uint8* dst_u, uint8* dst_v, int pix);
1439 void UYVYToUV422Row_SSE2(const uint8* src_uyvy,
1440 uint8* dst_u, uint8* dst_v, int pix);
1441 void UYVYToYRow_Unaligned_SSE2(const uint8* src_uyvy,
1442 uint8* dst_y, int pix);
1443 void UYVYToUVRow_Unaligned_SSE2(const uint8* src_uyvy, int stride_uyvy,
1444 uint8* dst_u, uint8* dst_v, int pix);
1445 void UYVYToUV422Row_Unaligned_SSE2(const uint8* src_uyvy,
1446 uint8* dst_u, uint8* dst_v, int pix);
1447 void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix);
1448 void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy,
1449 uint8* dst_u, uint8* dst_v, int pix);
1450 void UYVYToUV422Row_AVX2(const uint8* src_uyvy,
1451 uint8* dst_u, uint8* dst_v, int pix);
1452 void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int pix);
1453 void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy,
1454 uint8* dst_u, uint8* dst_v, int pix);
1455 void UYVYToUV422Row_NEON(const uint8* src_uyvy,
1456 uint8* dst_u, uint8* dst_v, int pix);
1458 void UYVYToYRow_C(const uint8* src_uyvy, uint8* dst_y, int pix);
1459 void UYVYToUVRow_C(const uint8* src_uyvy, int stride_uyvy,
1460 uint8* dst_u, uint8* dst_v, int pix);
1461 void UYVYToUV422Row_C(const uint8* src_uyvy,
1462 uint8* dst_u, uint8* dst_v, int pix);
1463 void UYVYToYRow_Any_AVX2(const uint8* src_uyvy, uint8* dst_y, int pix);
1464 void UYVYToUVRow_Any_AVX2(const uint8* src_uyvy, int stride_uyvy,
1465 uint8* dst_u, uint8* dst_v, int pix);
1466 void UYVYToUV422Row_Any_AVX2(const uint8* src_uyvy,
1467 uint8* dst_u, uint8* dst_v, int pix);
1468 void UYVYToYRow_Any_SSE2(const uint8* src_uyvy, uint8* dst_y, int pix);
1469 void UYVYToUVRow_Any_SSE2(const uint8* src_uyvy, int stride_uyvy,
1470 uint8* dst_u, uint8* dst_v, int pix);
1471 void UYVYToUV422Row_Any_SSE2(const uint8* src_uyvy,
1472 uint8* dst_u, uint8* dst_v, int pix);
1473 void UYVYToYRow_Any_NEON(const uint8* src_uyvy, uint8* dst_y, int pix);
1474 void UYVYToUVRow_Any_NEON(const uint8* src_uyvy, int stride_uyvy,
1475 uint8* dst_u, uint8* dst_v, int pix);
1476 void UYVYToUV422Row_Any_NEON(const uint8* src_uyvy,
1477 uint8* dst_u, uint8* dst_v, int pix);
1479 void HalfRow_C(const uint8* src_uv, int src_uv_stride,
1480 uint8* dst_uv, int pix);
1481 void HalfRow_SSE2(const uint8* src_uv, int src_uv_stride,
1482 uint8* dst_uv, int pix);
1483 void HalfRow_AVX2(const uint8* src_uv, int src_uv_stride,
1484 uint8* dst_uv, int pix);
1485 void HalfRow_NEON(const uint8* src_uv, int src_uv_stride,
1486 uint8* dst_uv, int pix);
1488 void HalfRow_16_C(const uint16* src_uv, int src_uv_stride,
1489 uint16* dst_uv, int pix);
1491 void ARGBToBayerRow_C(const uint8* src_argb, uint8* dst_bayer,
1492 uint32 selector, int pix);
1493 void ARGBToBayerRow_SSSE3(const uint8* src_argb, uint8* dst_bayer,
1494 uint32 selector, int pix);
1495 void ARGBToBayerRow_NEON(const uint8* src_argb, uint8* dst_bayer,
1496 uint32 selector, int pix);
1497 void ARGBToBayerRow_Any_SSSE3(const uint8* src_argb, uint8* dst_bayer,
1498 uint32 selector, int pix);
1499 void ARGBToBayerRow_Any_NEON(const uint8* src_argb, uint8* dst_bayer,
1500 uint32 selector, int pix);
1501 void ARGBToBayerGGRow_C(const uint8* src_argb, uint8* dst_bayer,
1502 uint32 /* selector */, int pix);
1503 void ARGBToBayerGGRow_SSE2(const uint8* src_argb, uint8* dst_bayer,
1504 uint32 /* selector */, int pix);
1505 void ARGBToBayerGGRow_NEON(const uint8* src_argb, uint8* dst_bayer,
1506 uint32 /* selector */, int pix);
1507 void ARGBToBayerGGRow_Any_SSE2(const uint8* src_argb, uint8* dst_bayer,
1508 uint32 /* selector */, int pix);
1509 void ARGBToBayerGGRow_Any_NEON(const uint8* src_argb, uint8* dst_bayer,
1510 uint32 /* selector */, int pix);
1512 void I422ToYUY2Row_C(const uint8* src_y,
1515 uint8* dst_yuy2, int width);
1516 void I422ToUYVYRow_C(const uint8* src_y,
1519 uint8* dst_uyvy, int width);
1520 void I422ToYUY2Row_SSE2(const uint8* src_y,
1523 uint8* dst_yuy2, int width);
1524 void I422ToUYVYRow_SSE2(const uint8* src_y,
1527 uint8* dst_uyvy, int width);
1528 void I422ToYUY2Row_Any_SSE2(const uint8* src_y,
1531 uint8* dst_yuy2, int width);
1532 void I422ToUYVYRow_Any_SSE2(const uint8* src_y,
1535 uint8* dst_uyvy, int width);
1536 void I422ToYUY2Row_NEON(const uint8* src_y,
1539 uint8* dst_yuy2, int width);
1540 void I422ToUYVYRow_NEON(const uint8* src_y,
1543 uint8* dst_uyvy, int width);
1544 void I422ToYUY2Row_Any_NEON(const uint8* src_y,
1547 uint8* dst_yuy2, int width);
1548 void I422ToUYVYRow_Any_NEON(const uint8* src_y,
1551 uint8* dst_uyvy, int width);
1553 // Effects related row functions.
1554 void ARGBAttenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width);
1555 void ARGBAttenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
1556 void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width);
1557 void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
1558 void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width);
1559 void ARGBAttenuateRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
1561 void ARGBAttenuateRow_Any_SSSE3(const uint8* src_argb, uint8* dst_argb,
1563 void ARGBAttenuateRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
1565 void ARGBAttenuateRow_Any_NEON(const uint8* src_argb, uint8* dst_argb,
1568 // Inverse table for unattenuate, shared by C and SSE2.
1569 extern const uint32 fixed_invtbl8[256];
1570 void ARGBUnattenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width);
1571 void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width);
1572 void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width);
1573 void ARGBUnattenuateRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb,
1575 void ARGBUnattenuateRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb,
1578 void ARGBGrayRow_C(const uint8* src_argb, uint8* dst_argb, int width);
1579 void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width);
1580 void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width);
1582 void ARGBSepiaRow_C(uint8* dst_argb, int width);
1583 void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width);
1584 void ARGBSepiaRow_NEON(uint8* dst_argb, int width);
1586 void ARGBColorMatrixRow_C(const uint8* src_argb, uint8* dst_argb,
1587 const int8* matrix_argb, int width);
1588 void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
1589 const int8* matrix_argb, int width);
1590 void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb,
1591 const int8* matrix_argb, int width);
1593 void ARGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width);
1594 void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width);
1596 void RGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width);
1597 void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width);
1599 void ARGBQuantizeRow_C(uint8* dst_argb, int scale, int interval_size,
1600 int interval_offset, int width);
1601 void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size,
1602 int interval_offset, int width);
1603 void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size,
1604 int interval_offset, int width);
1606 void ARGBShadeRow_C(const uint8* src_argb, uint8* dst_argb, int width,
1608 void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width,
1610 void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width,
1614 void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft,
1615 int width, int area, uint8* dst, int count);
1616 void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum,
1617 const int32* previous_cumsum, int width);
1619 void CumulativeSumToAverageRow_C(const int32* topleft, const int32* botleft,
1620 int width, int area, uint8* dst, int count);
1621 void ComputeCumulativeSumRow_C(const uint8* row, int32* cumsum,
1622 const int32* previous_cumsum, int width);
1625 void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride,
1626 uint8* dst_argb, const float* uv_dudv, int width);
1628 void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride,
1629 uint8* dst_argb, const float* uv_dudv, int width);
1631 // Used for I420Scale, ARGBScale, and ARGBInterpolate.
1632 void InterpolateRow_C(uint8* dst_ptr, const uint8* src_ptr,
1633 ptrdiff_t src_stride_ptr,
1634 int width, int source_y_fraction);
1635 void InterpolateRow_SSE2(uint8* dst_ptr, const uint8* src_ptr,
1636 ptrdiff_t src_stride_ptr, int width,
1637 int source_y_fraction);
1638 void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
1639 ptrdiff_t src_stride_ptr, int width,
1640 int source_y_fraction);
1641 void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr,
1642 ptrdiff_t src_stride_ptr, int width,
1643 int source_y_fraction);
1644 void InterpolateRow_NEON(uint8* dst_ptr, const uint8* src_ptr,
1645 ptrdiff_t src_stride_ptr, int width,
1646 int source_y_fraction);
1647 void InterpolateRows_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
1648 ptrdiff_t src_stride_ptr, int width,
1649 int source_y_fraction);
1650 void InterpolateRow_Unaligned_SSE2(uint8* dst_ptr, const uint8* src_ptr,
1651 ptrdiff_t src_stride_ptr, int width,
1652 int source_y_fraction);
1653 void InterpolateRow_Unaligned_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
1654 ptrdiff_t src_stride_ptr, int width,
1655 int source_y_fraction);
1656 void InterpolateRow_Any_NEON(uint8* dst_ptr, const uint8* src_ptr,
1657 ptrdiff_t src_stride_ptr, int width,
1658 int source_y_fraction);
1659 void InterpolateRow_Any_SSE2(uint8* dst_ptr, const uint8* src_ptr,
1660 ptrdiff_t src_stride_ptr, int width,
1661 int source_y_fraction);
1662 void InterpolateRow_Any_SSSE3(uint8* dst_ptr, const uint8* src_ptr,
1663 ptrdiff_t src_stride_ptr, int width,
1664 int source_y_fraction);
1665 void InterpolateRow_Any_AVX2(uint8* dst_ptr, const uint8* src_ptr,
1666 ptrdiff_t src_stride_ptr, int width,
1667 int source_y_fraction);
1668 void InterpolateRows_Any_MIPS_DSPR2(uint8* dst_ptr, const uint8* src_ptr,
1669 ptrdiff_t src_stride_ptr, int width,
1670 int source_y_fraction);
1672 void InterpolateRow_16_C(uint16* dst_ptr, const uint16* src_ptr,
1673 ptrdiff_t src_stride_ptr,
1674 int width, int source_y_fraction);
1677 void SobelXRow_C(const uint8* src_y0, const uint8* src_y1, const uint8* src_y2,
1678 uint8* dst_sobelx, int width);
1679 void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1,
1680 const uint8* src_y2, uint8* dst_sobelx, int width);
1681 void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1,
1682 const uint8* src_y2, uint8* dst_sobelx, int width);
1683 void SobelYRow_C(const uint8* src_y0, const uint8* src_y1,
1684 uint8* dst_sobely, int width);
1685 void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1,
1686 uint8* dst_sobely, int width);
1687 void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1,
1688 uint8* dst_sobely, int width);
1689 void SobelRow_C(const uint8* src_sobelx, const uint8* src_sobely,
1690 uint8* dst_argb, int width);
1691 void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
1692 uint8* dst_argb, int width);
1693 void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
1694 uint8* dst_argb, int width);
1695 void SobelToPlaneRow_C(const uint8* src_sobelx, const uint8* src_sobely,
1696 uint8* dst_y, int width);
1697 void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
1698 uint8* dst_y, int width);
1699 void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
1700 uint8* dst_y, int width);
1701 void SobelXYRow_C(const uint8* src_sobelx, const uint8* src_sobely,
1702 uint8* dst_argb, int width);
1703 void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely,
1704 uint8* dst_argb, int width);
1705 void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely,
1706 uint8* dst_argb, int width);
1708 void ARGBPolynomialRow_C(const uint8* src_argb,
1709 uint8* dst_argb, const float* poly,
1711 void ARGBPolynomialRow_SSE2(const uint8* src_argb,
1712 uint8* dst_argb, const float* poly,
1714 void ARGBPolynomialRow_AVX2(const uint8* src_argb,
1715 uint8* dst_argb, const float* poly,
1718 void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width,
1719 const uint8* luma, uint32 lumacoeff);
1720 void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb,
1722 const uint8* luma, uint32 lumacoeff);
1726 } // namespace libyuv
1729 #endif // INCLUDE_LIBYUV_ROW_H_ NOLINT