1 /* 2 * Copyright 2011 The LibYuv Project Authors. All rights reserved. 3 * 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. 9 */ 10 11 #ifndef INCLUDE_LIBYUV_ROW_H_ // NOLINT 12 #define INCLUDE_LIBYUV_ROW_H_ 13 14 #include <stdlib.h> // For malloc. 15 16 #include "libyuv/basic_types.h" 17 18 #ifdef __cplusplus 19 namespace libyuv { 20 extern "C" { 21 #endif 22 23 #define IS_ALIGNED(p, a) (!((uintptr_t)(p) & ((a) - 1))) 24 25 #ifdef __cplusplus 26 #define align_buffer_64(var, size) \ 27 uint8* var##_mem = reinterpret_cast<uint8*>(malloc((size) + 63)); \ 28 uint8* var = reinterpret_cast<uint8*> \ 29 ((reinterpret_cast<intptr_t>(var##_mem) + 63) & ~63) 30 #else 31 #define align_buffer_64(var, size) \ 32 uint8* var##_mem = (uint8*)(malloc((size) + 63)); /* NOLINT */ \ 33 uint8* var = (uint8*)(((intptr_t)(var##_mem) + 63) & ~63) /* NOLINT */ 34 #endif 35 36 #define free_aligned_buffer_64(var) \ 37 free(var##_mem); \ 38 var = 0 39 40 #if defined(__pnacl__) || defined(__CLR_VER) || \ 41 (defined(__i386__) && !defined(__SSE2__)) 42 #define LIBYUV_DISABLE_X86 43 #endif 44 // MemorySanitizer does not support assembly code yet. http://crbug.com/344505 45 #if defined(__has_feature) 46 #if __has_feature(memory_sanitizer) 47 #define LIBYUV_DISABLE_X86 48 #endif 49 #endif 50 // True if compiling for SSSE3 as a requirement. 51 #if defined(__SSSE3__) || (defined(_M_IX86_FP) && (_M_IX86_FP >= 3)) 52 #define LIBYUV_SSSE3_ONLY 53 #endif 54 55 #if defined(__native_client__) 56 #define LIBYUV_DISABLE_NEON 57 #endif 58 // clang >= 3.5.0 required for Arm64. 59 #if defined(__clang__) && defined(__aarch64__) && !defined(LIBYUV_DISABLE_NEON) 60 #if (__clang_major__ < 3) || (__clang_major__ == 3 && (__clang_minor__ < 5)) 61 #define LIBYUV_DISABLE_NEON 62 #endif // clang >= 3.5 63 #endif // __clang__ 64 65 // GCC >= 4.7.0 required for AVX2. 66 #if defined(__GNUC__) && (defined(__x86_64__) || defined(__i386__)) 67 #if (__GNUC__ > 4) || (__GNUC__ == 4 && (__GNUC_MINOR__ >= 7)) 68 #define GCC_HAS_AVX2 1 69 #endif // GNUC >= 4.7 70 #endif // __GNUC__ 71 72 // clang >= 3.4.0 required for AVX2. 73 #if defined(__clang__) && (defined(__x86_64__) || defined(__i386__)) 74 #if (__clang_major__ > 3) || (__clang_major__ == 3 && (__clang_minor__ >= 4)) 75 #define CLANG_HAS_AVX2 1 76 #endif // clang >= 3.4 77 #endif // __clang__ 78 79 // Visual C 2012 required for AVX2. 80 #if defined(_M_IX86) && !defined(__clang__) && \ 81 defined(_MSC_VER) && _MSC_VER >= 1700 82 #define VISUALC_HAS_AVX2 1 83 #endif // VisualStudio >= 2012 84 85 // The following are available on all x86 platforms: 86 #if !defined(LIBYUV_DISABLE_X86) && \ 87 (defined(_M_IX86) || defined(__x86_64__) || defined(__i386__)) 88 // Conversions: 89 #define HAS_ABGRTOUVROW_SSSE3 90 #define HAS_ABGRTOYROW_SSSE3 91 #define HAS_ARGB1555TOARGBROW_SSE2 92 #define HAS_ARGB4444TOARGBROW_SSE2 93 #define HAS_ARGBSETROW_X86 94 #define HAS_ARGBSHUFFLEROW_SSE2 95 #define HAS_ARGBSHUFFLEROW_SSSE3 96 #define HAS_ARGBTOARGB1555ROW_SSE2 97 #define HAS_ARGBTOARGB4444ROW_SSE2 98 #define HAS_ARGBTORAWROW_SSSE3 99 #define HAS_ARGBTORGB24ROW_SSSE3 100 #define HAS_ARGBTORGB565DITHERROW_SSE2 101 #define HAS_ARGBTORGB565ROW_SSE2 102 #define HAS_ARGBTOUV444ROW_SSSE3 103 #define HAS_ARGBTOUVJROW_SSSE3 104 #define HAS_ARGBTOUVROW_SSSE3 105 #define HAS_ARGBTOYJROW_SSSE3 106 #define HAS_ARGBTOYROW_SSSE3 107 #define HAS_ARGBEXTRACTALPHAROW_SSE2 108 #define HAS_BGRATOUVROW_SSSE3 109 #define HAS_BGRATOYROW_SSSE3 110 #define HAS_COPYROW_ERMS 111 #define HAS_COPYROW_SSE2 112 #define HAS_H422TOARGBROW_SSSE3 113 #define HAS_I400TOARGBROW_SSE2 114 #define HAS_I422TOARGB1555ROW_SSSE3 115 #define HAS_I422TOARGB4444ROW_SSSE3 116 #define HAS_I422TOARGBROW_SSSE3 117 #define HAS_I422TORGB24ROW_SSSE3 118 #define HAS_I422TORGB565ROW_SSSE3 119 #define HAS_I422TORGBAROW_SSSE3 120 #define HAS_I422TOUYVYROW_SSE2 121 #define HAS_I422TOYUY2ROW_SSE2 122 #define HAS_I444TOARGBROW_SSSE3 123 #define HAS_J400TOARGBROW_SSE2 124 #define HAS_J422TOARGBROW_SSSE3 125 #define HAS_MERGEUVROW_SSE2 126 #define HAS_MIRRORROW_SSSE3 127 #define HAS_MIRRORUVROW_SSSE3 128 #define HAS_NV12TOARGBROW_SSSE3 129 #define HAS_NV12TORGB565ROW_SSSE3 130 #define HAS_NV21TOARGBROW_SSSE3 131 #define HAS_RAWTOARGBROW_SSSE3 132 #define HAS_RAWTORGB24ROW_SSSE3 133 #define HAS_RAWTOYROW_SSSE3 134 #define HAS_RGB24TOARGBROW_SSSE3 135 #define HAS_RGB24TOYROW_SSSE3 136 #define HAS_RGB565TOARGBROW_SSE2 137 #define HAS_RGBATOUVROW_SSSE3 138 #define HAS_RGBATOYROW_SSSE3 139 #define HAS_SETROW_ERMS 140 #define HAS_SETROW_X86 141 #define HAS_SPLITUVROW_SSE2 142 #define HAS_UYVYTOARGBROW_SSSE3 143 #define HAS_UYVYTOUV422ROW_SSE2 144 #define HAS_UYVYTOUVROW_SSE2 145 #define HAS_UYVYTOYROW_SSE2 146 #define HAS_YUY2TOARGBROW_SSSE3 147 #define HAS_YUY2TOUV422ROW_SSE2 148 #define HAS_YUY2TOUVROW_SSE2 149 #define HAS_YUY2TOYROW_SSE2 150 151 // Effects: 152 #define HAS_ARGBADDROW_SSE2 153 #define HAS_ARGBAFFINEROW_SSE2 154 #define HAS_ARGBATTENUATEROW_SSSE3 155 #define HAS_ARGBBLENDROW_SSSE3 156 #define HAS_ARGBCOLORMATRIXROW_SSSE3 157 #define HAS_ARGBCOLORTABLEROW_X86 158 #define HAS_ARGBCOPYALPHAROW_SSE2 159 #define HAS_ARGBCOPYYTOALPHAROW_SSE2 160 #define HAS_ARGBGRAYROW_SSSE3 161 #define HAS_ARGBLUMACOLORTABLEROW_SSSE3 162 #define HAS_ARGBMIRRORROW_SSE2 163 #define HAS_ARGBMULTIPLYROW_SSE2 164 #define HAS_ARGBPOLYNOMIALROW_SSE2 165 #define HAS_ARGBQUANTIZEROW_SSE2 166 #define HAS_ARGBSEPIAROW_SSSE3 167 #define HAS_ARGBSHADEROW_SSE2 168 #define HAS_ARGBSUBTRACTROW_SSE2 169 #define HAS_ARGBUNATTENUATEROW_SSE2 170 #define HAS_BLENDPLANEROW_SSSE3 171 #define HAS_COMPUTECUMULATIVESUMROW_SSE2 172 #define HAS_CUMULATIVESUMTOAVERAGEROW_SSE2 173 #define HAS_INTERPOLATEROW_SSSE3 174 #define HAS_RGBCOLORTABLEROW_X86 175 #define HAS_SOBELROW_SSE2 176 #define HAS_SOBELTOPLANEROW_SSE2 177 #define HAS_SOBELXROW_SSE2 178 #define HAS_SOBELXYROW_SSE2 179 #define HAS_SOBELYROW_SSE2 180 181 // The following functions fail on gcc/clang 32 bit with fpic and framepointer. 182 // caveat: clangcl uses row_win.cc which works. 183 #if defined(NDEBUG) || !(defined(_DEBUG) && defined(__i386__)) || \ 184 !defined(__i386__) || defined(_MSC_VER) 185 // TODO(fbarchard): fix build error on x86 debug 186 // https://code.google.com/p/libyuv/issues/detail?id=524 187 #define HAS_I411TOARGBROW_SSSE3 188 // TODO(fbarchard): fix build error on android_full_debug=1 189 // https://code.google.com/p/libyuv/issues/detail?id=517 190 #define HAS_I422ALPHATOARGBROW_SSSE3 191 #endif 192 #endif 193 194 // The following are available on all x86 platforms, but 195 // require VS2012, clang 3.4 or gcc 4.7. 196 // The code supports NaCL but requires a new compiler and validator. 197 #if !defined(LIBYUV_DISABLE_X86) && (defined(VISUALC_HAS_AVX2) || \ 198 defined(CLANG_HAS_AVX2) || defined(GCC_HAS_AVX2)) 199 #define HAS_ARGBCOPYALPHAROW_AVX2 200 #define HAS_ARGBCOPYYTOALPHAROW_AVX2 201 #define HAS_ARGBMIRRORROW_AVX2 202 #define HAS_ARGBPOLYNOMIALROW_AVX2 203 #define HAS_ARGBSHUFFLEROW_AVX2 204 #define HAS_ARGBTORGB565DITHERROW_AVX2 205 #define HAS_ARGBTOUVJROW_AVX2 206 #define HAS_ARGBTOUVROW_AVX2 207 #define HAS_ARGBTOYJROW_AVX2 208 #define HAS_ARGBTOYROW_AVX2 209 #define HAS_COPYROW_AVX 210 #define HAS_H422TOARGBROW_AVX2 211 #define HAS_I400TOARGBROW_AVX2 212 #if !(defined(_DEBUG) && defined(__i386__)) 213 // TODO(fbarchard): fix build error on android_full_debug=1 214 // https://code.google.com/p/libyuv/issues/detail?id=517 215 #define HAS_I422ALPHATOARGBROW_AVX2 216 #endif 217 #define HAS_I411TOARGBROW_AVX2 218 #define HAS_I422TOARGB1555ROW_AVX2 219 #define HAS_I422TOARGB4444ROW_AVX2 220 #define HAS_I422TOARGBROW_AVX2 221 #define HAS_I422TORGB24ROW_AVX2 222 #define HAS_I422TORGB565ROW_AVX2 223 #define HAS_I422TORGBAROW_AVX2 224 #define HAS_I444TOARGBROW_AVX2 225 #define HAS_INTERPOLATEROW_AVX2 226 #define HAS_J422TOARGBROW_AVX2 227 #define HAS_MERGEUVROW_AVX2 228 #define HAS_MIRRORROW_AVX2 229 #define HAS_NV12TOARGBROW_AVX2 230 #define HAS_NV12TORGB565ROW_AVX2 231 #define HAS_NV21TOARGBROW_AVX2 232 #define HAS_SPLITUVROW_AVX2 233 #define HAS_UYVYTOARGBROW_AVX2 234 #define HAS_UYVYTOUV422ROW_AVX2 235 #define HAS_UYVYTOUVROW_AVX2 236 #define HAS_UYVYTOYROW_AVX2 237 #define HAS_YUY2TOARGBROW_AVX2 238 #define HAS_YUY2TOUV422ROW_AVX2 239 #define HAS_YUY2TOUVROW_AVX2 240 #define HAS_YUY2TOYROW_AVX2 241 242 // Effects: 243 #define HAS_ARGBADDROW_AVX2 244 #define HAS_ARGBATTENUATEROW_AVX2 245 #define HAS_ARGBMULTIPLYROW_AVX2 246 #define HAS_ARGBSUBTRACTROW_AVX2 247 #define HAS_ARGBUNATTENUATEROW_AVX2 248 #define HAS_BLENDPLANEROW_AVX2 249 #endif 250 251 // The following are available for AVX2 Visual C and clangcl 32 bit: 252 // TODO(fbarchard): Port to gcc. 253 #if !defined(LIBYUV_DISABLE_X86) && defined(_M_IX86) && \ 254 (defined(VISUALC_HAS_AVX2) || defined(CLANG_HAS_AVX2)) 255 #define HAS_ARGB1555TOARGBROW_AVX2 256 #define HAS_ARGB4444TOARGBROW_AVX2 257 #define HAS_ARGBTOARGB1555ROW_AVX2 258 #define HAS_ARGBTOARGB4444ROW_AVX2 259 #define HAS_ARGBTORGB565ROW_AVX2 260 #define HAS_J400TOARGBROW_AVX2 261 #define HAS_RGB565TOARGBROW_AVX2 262 #endif 263 264 // The following are also available on x64 Visual C. 265 #if !defined(LIBYUV_DISABLE_X86) && defined(_MSC_VER) && defined(_M_X64) && \ 266 (!defined(__clang__) || defined(__SSSE3__)) 267 #define HAS_I422ALPHATOARGBROW_SSSE3 268 #define HAS_I422TOARGBROW_SSSE3 269 #endif 270 271 // The following are available on Neon platforms: 272 #if !defined(LIBYUV_DISABLE_NEON) && \ 273 (defined(__aarch64__) || defined(__ARM_NEON__) || defined(LIBYUV_NEON)) 274 #define HAS_ABGRTOUVROW_NEON 275 #define HAS_ABGRTOYROW_NEON 276 #define HAS_ARGB1555TOARGBROW_NEON 277 #define HAS_ARGB1555TOUVROW_NEON 278 #define HAS_ARGB1555TOYROW_NEON 279 #define HAS_ARGB4444TOARGBROW_NEON 280 #define HAS_ARGB4444TOUVROW_NEON 281 #define HAS_ARGB4444TOYROW_NEON 282 #define HAS_ARGBSETROW_NEON 283 #define HAS_ARGBTOARGB1555ROW_NEON 284 #define HAS_ARGBTOARGB4444ROW_NEON 285 #define HAS_ARGBTORAWROW_NEON 286 #define HAS_ARGBTORGB24ROW_NEON 287 #define HAS_ARGBTORGB565DITHERROW_NEON 288 #define HAS_ARGBTORGB565ROW_NEON 289 #define HAS_ARGBTOUV411ROW_NEON 290 #define HAS_ARGBTOUV444ROW_NEON 291 #define HAS_ARGBTOUVJROW_NEON 292 #define HAS_ARGBTOUVROW_NEON 293 #define HAS_ARGBTOYJROW_NEON 294 #define HAS_ARGBTOYROW_NEON 295 #define HAS_ARGBEXTRACTALPHAROW_NEON 296 #define HAS_BGRATOUVROW_NEON 297 #define HAS_BGRATOYROW_NEON 298 #define HAS_COPYROW_NEON 299 #define HAS_I400TOARGBROW_NEON 300 #define HAS_I411TOARGBROW_NEON 301 #define HAS_I422ALPHATOARGBROW_NEON 302 #define HAS_I422TOARGB1555ROW_NEON 303 #define HAS_I422TOARGB4444ROW_NEON 304 #define HAS_I422TOARGBROW_NEON 305 #define HAS_I422TORGB24ROW_NEON 306 #define HAS_I422TORGB565ROW_NEON 307 #define HAS_I422TORGBAROW_NEON 308 #define HAS_I422TOUYVYROW_NEON 309 #define HAS_I422TOYUY2ROW_NEON 310 #define HAS_I444TOARGBROW_NEON 311 #define HAS_J400TOARGBROW_NEON 312 #define HAS_MERGEUVROW_NEON 313 #define HAS_MIRRORROW_NEON 314 #define HAS_MIRRORUVROW_NEON 315 #define HAS_NV12TOARGBROW_NEON 316 #define HAS_NV12TORGB565ROW_NEON 317 #define HAS_NV21TOARGBROW_NEON 318 #define HAS_RAWTOARGBROW_NEON 319 #define HAS_RAWTORGB24ROW_NEON 320 #define HAS_RAWTOUVROW_NEON 321 #define HAS_RAWTOYROW_NEON 322 #define HAS_RGB24TOARGBROW_NEON 323 #define HAS_RGB24TOUVROW_NEON 324 #define HAS_RGB24TOYROW_NEON 325 #define HAS_RGB565TOARGBROW_NEON 326 #define HAS_RGB565TOUVROW_NEON 327 #define HAS_RGB565TOYROW_NEON 328 #define HAS_RGBATOUVROW_NEON 329 #define HAS_RGBATOYROW_NEON 330 #define HAS_SETROW_NEON 331 #define HAS_SPLITUVROW_NEON 332 #define HAS_UYVYTOARGBROW_NEON 333 #define HAS_UYVYTOUV422ROW_NEON 334 #define HAS_UYVYTOUVROW_NEON 335 #define HAS_UYVYTOYROW_NEON 336 #define HAS_YUY2TOARGBROW_NEON 337 #define HAS_YUY2TOUV422ROW_NEON 338 #define HAS_YUY2TOUVROW_NEON 339 #define HAS_YUY2TOYROW_NEON 340 341 // Effects: 342 #define HAS_ARGBADDROW_NEON 343 #define HAS_ARGBATTENUATEROW_NEON 344 #define HAS_ARGBBLENDROW_NEON 345 #define HAS_ARGBCOLORMATRIXROW_NEON 346 #define HAS_ARGBGRAYROW_NEON 347 #define HAS_ARGBMIRRORROW_NEON 348 #define HAS_ARGBMULTIPLYROW_NEON 349 #define HAS_ARGBQUANTIZEROW_NEON 350 #define HAS_ARGBSEPIAROW_NEON 351 #define HAS_ARGBSHADEROW_NEON 352 #define HAS_ARGBSHUFFLEROW_NEON 353 #define HAS_ARGBSUBTRACTROW_NEON 354 #define HAS_INTERPOLATEROW_NEON 355 #define HAS_SOBELROW_NEON 356 #define HAS_SOBELTOPLANEROW_NEON 357 #define HAS_SOBELXROW_NEON 358 #define HAS_SOBELXYROW_NEON 359 #define HAS_SOBELYROW_NEON 360 #endif 361 362 // The following are available on Mips platforms: 363 #if !defined(LIBYUV_DISABLE_MIPS) && defined(__mips__) && \ 364 (_MIPS_SIM == _MIPS_SIM_ABI32) && (__mips_isa_rev < 6) 365 #define HAS_COPYROW_MIPS 366 #if defined(__mips_dsp) && (__mips_dsp_rev >= 2) 367 #define HAS_I422TOARGBROW_DSPR2 368 #define HAS_INTERPOLATEROW_DSPR2 369 #define HAS_MIRRORROW_DSPR2 370 #define HAS_MIRRORUVROW_DSPR2 371 #define HAS_SPLITUVROW_DSPR2 372 #endif 373 #endif 374 375 #if defined(_MSC_VER) && !defined(__CLR_VER) && !defined(__clang__) 376 #define SIMD_ALIGNED(var) __declspec(align(16)) var 377 #define SIMD_ALIGNED32(var) __declspec(align(64)) var 378 typedef __declspec(align(16)) int16 vec16[8]; 379 typedef __declspec(align(16)) int32 vec32[4]; 380 typedef __declspec(align(16)) int8 vec8[16]; 381 typedef __declspec(align(16)) uint16 uvec16[8]; 382 typedef __declspec(align(16)) uint32 uvec32[4]; 383 typedef __declspec(align(16)) uint8 uvec8[16]; 384 typedef __declspec(align(32)) int16 lvec16[16]; 385 typedef __declspec(align(32)) int32 lvec32[8]; 386 typedef __declspec(align(32)) int8 lvec8[32]; 387 typedef __declspec(align(32)) uint16 ulvec16[16]; 388 typedef __declspec(align(32)) uint32 ulvec32[8]; 389 typedef __declspec(align(32)) uint8 ulvec8[32]; 390 #elif !defined(__pnacl__) && (defined(__GNUC__) || defined(__clang__)) 391 // Caveat GCC 4.2 to 4.7 have a known issue using vectors with const. 392 #define SIMD_ALIGNED(var) var __attribute__((aligned(16))) 393 #define SIMD_ALIGNED32(var) var __attribute__((aligned(64))) 394 typedef int16 __attribute__((vector_size(16))) vec16; 395 typedef int32 __attribute__((vector_size(16))) vec32; 396 typedef int8 __attribute__((vector_size(16))) vec8; 397 typedef uint16 __attribute__((vector_size(16))) uvec16; 398 typedef uint32 __attribute__((vector_size(16))) uvec32; 399 typedef uint8 __attribute__((vector_size(16))) uvec8; 400 typedef int16 __attribute__((vector_size(32))) lvec16; 401 typedef int32 __attribute__((vector_size(32))) lvec32; 402 typedef int8 __attribute__((vector_size(32))) lvec8; 403 typedef uint16 __attribute__((vector_size(32))) ulvec16; 404 typedef uint32 __attribute__((vector_size(32))) ulvec32; 405 typedef uint8 __attribute__((vector_size(32))) ulvec8; 406 #else 407 #define SIMD_ALIGNED(var) var 408 #define SIMD_ALIGNED32(var) var 409 typedef int16 vec16[8]; 410 typedef int32 vec32[4]; 411 typedef int8 vec8[16]; 412 typedef uint16 uvec16[8]; 413 typedef uint32 uvec32[4]; 414 typedef uint8 uvec8[16]; 415 typedef int16 lvec16[16]; 416 typedef int32 lvec32[8]; 417 typedef int8 lvec8[32]; 418 typedef uint16 ulvec16[16]; 419 typedef uint32 ulvec32[8]; 420 typedef uint8 ulvec8[32]; 421 #endif 422 423 #if defined(__aarch64__) 424 // This struct is for Arm64 color conversion. 425 struct YuvConstants { 426 uvec16 kUVToRB; 427 uvec16 kUVToRB2; 428 uvec16 kUVToG; 429 uvec16 kUVToG2; 430 vec16 kUVBiasBGR; 431 vec32 kYToRgb; 432 }; 433 #elif defined(__arm__) 434 // This struct is for ArmV7 color conversion. 435 struct YuvConstants { 436 uvec8 kUVToRB; 437 uvec8 kUVToG; 438 vec16 kUVBiasBGR; 439 vec32 kYToRgb; 440 }; 441 #else 442 // This struct is for Intel color conversion. 443 struct YuvConstants { 444 lvec8 kUVToB; 445 lvec8 kUVToG; 446 lvec8 kUVToR; 447 lvec16 kUVBiasB; 448 lvec16 kUVBiasG; 449 lvec16 kUVBiasR; 450 lvec16 kYToRgb; 451 }; 452 453 // Offsets into YuvConstants structure 454 #define KUVTOB 0 455 #define KUVTOG 32 456 #define KUVTOR 64 457 #define KUVBIASB 96 458 #define KUVBIASG 128 459 #define KUVBIASR 160 460 #define KYTORGB 192 461 #endif 462 463 // Conversion matrix for YUV to RGB 464 extern const struct YuvConstants kYuvI601Constants; // BT.601 465 extern const struct YuvConstants kYuvJPEGConstants; // JPeg color space 466 extern const struct YuvConstants kYuvH709Constants; // BT.709 467 468 // Conversion matrix for YVU to BGR 469 extern const struct YuvConstants kYvuI601Constants; // BT.601 470 extern const struct YuvConstants kYvuJPEGConstants; // JPeg color space 471 extern const struct YuvConstants kYvuH709Constants; // BT.709 472 473 #if defined(__APPLE__) || defined(__x86_64__) || defined(__llvm__) 474 #define OMITFP 475 #else 476 #define OMITFP __attribute__((optimize("omit-frame-pointer"))) 477 #endif 478 479 // NaCL macros for GCC x86 and x64. 480 #if defined(__native_client__) 481 #define LABELALIGN ".p2align 5\n" 482 #else 483 #define LABELALIGN 484 #endif 485 #if defined(__native_client__) && defined(__x86_64__) 486 // r14 is used for MEMOP macros. 487 #define NACL_R14 "r14", 488 #define BUNDLELOCK ".bundle_lock\n" 489 #define BUNDLEUNLOCK ".bundle_unlock\n" 490 #define MEMACCESS(base) "%%nacl:(%%r15,%q" #base ")" 491 #define MEMACCESS2(offset, base) "%%nacl:" #offset "(%%r15,%q" #base ")" 492 #define MEMLEA(offset, base) #offset "(%q" #base ")" 493 #define MEMLEA3(offset, index, scale) \ 494 #offset "(,%q" #index "," #scale ")" 495 #define MEMLEA4(offset, base, index, scale) \ 496 #offset "(%q" #base ",%q" #index "," #scale ")" 497 #define MEMMOVESTRING(s, d) "%%nacl:(%q" #s "),%%nacl:(%q" #d "), %%r15" 498 #define MEMSTORESTRING(reg, d) "%%" #reg ",%%nacl:(%q" #d "), %%r15" 499 #define MEMOPREG(opcode, offset, base, index, scale, reg) \ 500 BUNDLELOCK \ 501 "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \ 502 #opcode " (%%r15,%%r14),%%" #reg "\n" \ 503 BUNDLEUNLOCK 504 #define MEMOPMEM(opcode, reg, offset, base, index, scale) \ 505 BUNDLELOCK \ 506 "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \ 507 #opcode " %%" #reg ",(%%r15,%%r14)\n" \ 508 BUNDLEUNLOCK 509 #define MEMOPARG(opcode, offset, base, index, scale, arg) \ 510 BUNDLELOCK \ 511 "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \ 512 #opcode " (%%r15,%%r14),%" #arg "\n" \ 513 BUNDLEUNLOCK 514 #define VMEMOPREG(opcode, offset, base, index, scale, reg1, reg2) \ 515 BUNDLELOCK \ 516 "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \ 517 #opcode " (%%r15,%%r14),%%" #reg1 ",%%" #reg2 "\n" \ 518 BUNDLEUNLOCK 519 #define VEXTOPMEM(op, sel, reg, offset, base, index, scale) \ 520 BUNDLELOCK \ 521 "lea " #offset "(%q" #base ",%q" #index "," #scale "),%%r14d\n" \ 522 #op " $" #sel ",%%" #reg ",(%%r15,%%r14)\n" \ 523 BUNDLEUNLOCK 524 #else // defined(__native_client__) && defined(__x86_64__) 525 #define NACL_R14 526 #define BUNDLEALIGN 527 #define MEMACCESS(base) "(%" #base ")" 528 #define MEMACCESS2(offset, base) #offset "(%" #base ")" 529 #define MEMLEA(offset, base) #offset "(%" #base ")" 530 #define MEMLEA3(offset, index, scale) \ 531 #offset "(,%" #index "," #scale ")" 532 #define MEMLEA4(offset, base, index, scale) \ 533 #offset "(%" #base ",%" #index "," #scale ")" 534 #define MEMMOVESTRING(s, d) 535 #define MEMSTORESTRING(reg, d) 536 #define MEMOPREG(opcode, offset, base, index, scale, reg) \ 537 #opcode " " #offset "(%" #base ",%" #index "," #scale "),%%" #reg "\n" 538 #define MEMOPMEM(opcode, reg, offset, base, index, scale) \ 539 #opcode " %%" #reg ","#offset "(%" #base ",%" #index "," #scale ")\n" 540 #define MEMOPARG(opcode, offset, base, index, scale, arg) \ 541 #opcode " " #offset "(%" #base ",%" #index "," #scale "),%" #arg "\n" 542 #define VMEMOPREG(opcode, offset, base, index, scale, reg1, reg2) \ 543 #opcode " " #offset "(%" #base ",%" #index "," #scale "),%%" #reg1 ",%%" \ 544 #reg2 "\n" 545 #define VEXTOPMEM(op, sel, reg, offset, base, index, scale) \ 546 #op " $" #sel ",%%" #reg ","#offset "(%" #base ",%" #index "," #scale ")\n" 547 #endif // defined(__native_client__) && defined(__x86_64__) 548 549 #if defined(__arm__) || defined(__aarch64__) 550 #undef MEMACCESS 551 #if defined(__native_client__) 552 #define MEMACCESS(base) ".p2align 3\nbic %" #base ", #0xc0000000\n" 553 #else 554 #define MEMACCESS(base) 555 #endif 556 #endif 557 558 void I444ToARGBRow_NEON(const uint8* src_y, 559 const uint8* src_u, 560 const uint8* src_v, 561 uint8* dst_argb, 562 const struct YuvConstants* yuvconstants, 563 int width); 564 void I422ToARGBRow_NEON(const uint8* src_y, 565 const uint8* src_u, 566 const uint8* src_v, 567 uint8* dst_argb, 568 const struct YuvConstants* yuvconstants, 569 int width); 570 void I422AlphaToARGBRow_NEON(const uint8* y_buf, 571 const uint8* u_buf, 572 const uint8* v_buf, 573 const uint8* a_buf, 574 uint8* dst_argb, 575 const struct YuvConstants* yuvconstants, 576 int width); 577 void I422ToARGBRow_NEON(const uint8* src_y, 578 const uint8* src_u, 579 const uint8* src_v, 580 uint8* dst_argb, 581 const struct YuvConstants* yuvconstants, 582 int width); 583 void I411ToARGBRow_NEON(const uint8* src_y, 584 const uint8* src_u, 585 const uint8* src_v, 586 uint8* dst_argb, 587 const struct YuvConstants* yuvconstants, 588 int width); 589 void I422ToRGBARow_NEON(const uint8* src_y, 590 const uint8* src_u, 591 const uint8* src_v, 592 uint8* dst_rgba, 593 const struct YuvConstants* yuvconstants, 594 int width); 595 void I422ToRGB24Row_NEON(const uint8* src_y, 596 const uint8* src_u, 597 const uint8* src_v, 598 uint8* dst_rgb24, 599 const struct YuvConstants* yuvconstants, 600 int width); 601 void I422ToRGB565Row_NEON(const uint8* src_y, 602 const uint8* src_u, 603 const uint8* src_v, 604 uint8* dst_rgb565, 605 const struct YuvConstants* yuvconstants, 606 int width); 607 void I422ToARGB1555Row_NEON(const uint8* src_y, 608 const uint8* src_u, 609 const uint8* src_v, 610 uint8* dst_argb1555, 611 const struct YuvConstants* yuvconstants, 612 int width); 613 void I422ToARGB4444Row_NEON(const uint8* src_y, 614 const uint8* src_u, 615 const uint8* src_v, 616 uint8* dst_argb4444, 617 const struct YuvConstants* yuvconstants, 618 int width); 619 void NV12ToARGBRow_NEON(const uint8* src_y, 620 const uint8* src_uv, 621 uint8* dst_argb, 622 const struct YuvConstants* yuvconstants, 623 int width); 624 void NV12ToRGB565Row_NEON(const uint8* src_y, 625 const uint8* src_uv, 626 uint8* dst_rgb565, 627 const struct YuvConstants* yuvconstants, 628 int width); 629 void NV21ToARGBRow_NEON(const uint8* src_y, 630 const uint8* src_vu, 631 uint8* dst_argb, 632 const struct YuvConstants* yuvconstants, 633 int width); 634 void YUY2ToARGBRow_NEON(const uint8* src_yuy2, 635 uint8* dst_argb, 636 const struct YuvConstants* yuvconstants, 637 int width); 638 void UYVYToARGBRow_NEON(const uint8* src_uyvy, 639 uint8* dst_argb, 640 const struct YuvConstants* yuvconstants, 641 int width); 642 643 void ARGBToYRow_AVX2(const uint8* src_argb, uint8* dst_y, int width); 644 void ARGBToYRow_Any_AVX2(const uint8* src_argb, uint8* dst_y, int width); 645 void ARGBToYRow_SSSE3(const uint8* src_argb, uint8* dst_y, int width); 646 void ARGBToYJRow_AVX2(const uint8* src_argb, uint8* dst_y, int width); 647 void ARGBToYJRow_Any_AVX2(const uint8* src_argb, uint8* dst_y, int width); 648 void ARGBToYJRow_SSSE3(const uint8* src_argb, uint8* dst_y, int width); 649 void BGRAToYRow_SSSE3(const uint8* src_bgra, uint8* dst_y, int width); 650 void ABGRToYRow_SSSE3(const uint8* src_abgr, uint8* dst_y, int width); 651 void RGBAToYRow_SSSE3(const uint8* src_rgba, uint8* dst_y, int width); 652 void RGB24ToYRow_SSSE3(const uint8* src_rgb24, uint8* dst_y, int width); 653 void RAWToYRow_SSSE3(const uint8* src_raw, uint8* dst_y, int width); 654 void ARGBToYRow_NEON(const uint8* src_argb, uint8* dst_y, int width); 655 void ARGBToYJRow_NEON(const uint8* src_argb, uint8* dst_y, int width); 656 void ARGBToUV444Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v, 657 int width); 658 void ARGBToUV411Row_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v, 659 int width); 660 void ARGBToUVRow_NEON(const uint8* src_argb, int src_stride_argb, 661 uint8* dst_u, uint8* dst_v, int width); 662 void ARGBToUVJRow_NEON(const uint8* src_argb, int src_stride_argb, 663 uint8* dst_u, uint8* dst_v, int width); 664 void BGRAToUVRow_NEON(const uint8* src_bgra, int src_stride_bgra, 665 uint8* dst_u, uint8* dst_v, int width); 666 void ABGRToUVRow_NEON(const uint8* src_abgr, int src_stride_abgr, 667 uint8* dst_u, uint8* dst_v, int width); 668 void RGBAToUVRow_NEON(const uint8* src_rgba, int src_stride_rgba, 669 uint8* dst_u, uint8* dst_v, int width); 670 void RGB24ToUVRow_NEON(const uint8* src_rgb24, int src_stride_rgb24, 671 uint8* dst_u, uint8* dst_v, int width); 672 void RAWToUVRow_NEON(const uint8* src_raw, int src_stride_raw, 673 uint8* dst_u, uint8* dst_v, int width); 674 void RGB565ToUVRow_NEON(const uint8* src_rgb565, int src_stride_rgb565, 675 uint8* dst_u, uint8* dst_v, int width); 676 void ARGB1555ToUVRow_NEON(const uint8* src_argb1555, int src_stride_argb1555, 677 uint8* dst_u, uint8* dst_v, int width); 678 void ARGB4444ToUVRow_NEON(const uint8* src_argb4444, int src_stride_argb4444, 679 uint8* dst_u, uint8* dst_v, int width); 680 void BGRAToYRow_NEON(const uint8* src_bgra, uint8* dst_y, int width); 681 void ABGRToYRow_NEON(const uint8* src_abgr, uint8* dst_y, int width); 682 void RGBAToYRow_NEON(const uint8* src_rgba, uint8* dst_y, int width); 683 void RGB24ToYRow_NEON(const uint8* src_rgb24, uint8* dst_y, int width); 684 void RAWToYRow_NEON(const uint8* src_raw, uint8* dst_y, int width); 685 void RGB565ToYRow_NEON(const uint8* src_rgb565, uint8* dst_y, int width); 686 void ARGB1555ToYRow_NEON(const uint8* src_argb1555, uint8* dst_y, int width); 687 void ARGB4444ToYRow_NEON(const uint8* src_argb4444, uint8* dst_y, int width); 688 void ARGBToYRow_C(const uint8* src_argb, uint8* dst_y, int width); 689 void ARGBToYJRow_C(const uint8* src_argb, uint8* dst_y, int width); 690 void BGRAToYRow_C(const uint8* src_bgra, uint8* dst_y, int width); 691 void ABGRToYRow_C(const uint8* src_abgr, uint8* dst_y, int width); 692 void RGBAToYRow_C(const uint8* src_rgba, uint8* dst_y, int width); 693 void RGB24ToYRow_C(const uint8* src_rgb24, uint8* dst_y, int width); 694 void RAWToYRow_C(const uint8* src_raw, uint8* dst_y, int width); 695 void RGB565ToYRow_C(const uint8* src_rgb565, uint8* dst_y, int width); 696 void ARGB1555ToYRow_C(const uint8* src_argb1555, uint8* dst_y, int width); 697 void ARGB4444ToYRow_C(const uint8* src_argb4444, uint8* dst_y, int width); 698 void ARGBToYRow_Any_SSSE3(const uint8* src_argb, uint8* dst_y, int width); 699 void ARGBToYJRow_Any_SSSE3(const uint8* src_argb, uint8* dst_y, int width); 700 void BGRAToYRow_Any_SSSE3(const uint8* src_bgra, uint8* dst_y, int width); 701 void ABGRToYRow_Any_SSSE3(const uint8* src_abgr, uint8* dst_y, int width); 702 void RGBAToYRow_Any_SSSE3(const uint8* src_rgba, uint8* dst_y, int width); 703 void RGB24ToYRow_Any_SSSE3(const uint8* src_rgb24, uint8* dst_y, int width); 704 void RAWToYRow_Any_SSSE3(const uint8* src_raw, uint8* dst_y, int width); 705 void ARGBToYRow_Any_NEON(const uint8* src_argb, uint8* dst_y, int width); 706 void ARGBToYJRow_Any_NEON(const uint8* src_argb, uint8* dst_y, int width); 707 void BGRAToYRow_Any_NEON(const uint8* src_bgra, uint8* dst_y, int width); 708 void ABGRToYRow_Any_NEON(const uint8* src_abgr, uint8* dst_y, int width); 709 void RGBAToYRow_Any_NEON(const uint8* src_rgba, uint8* dst_y, int width); 710 void RGB24ToYRow_Any_NEON(const uint8* src_rgb24, uint8* dst_y, int width); 711 void RAWToYRow_Any_NEON(const uint8* src_raw, uint8* dst_y, int width); 712 void RGB565ToYRow_Any_NEON(const uint8* src_rgb565, uint8* dst_y, int width); 713 void ARGB1555ToYRow_Any_NEON(const uint8* src_argb1555, uint8* dst_y, 714 int width); 715 void ARGB4444ToYRow_Any_NEON(const uint8* src_argb4444, uint8* dst_y, 716 int width); 717 718 void ARGBToUVRow_AVX2(const uint8* src_argb, int src_stride_argb, 719 uint8* dst_u, uint8* dst_v, int width); 720 void ARGBToUVJRow_AVX2(const uint8* src_argb, int src_stride_argb, 721 uint8* dst_u, uint8* dst_v, int width); 722 void ARGBToUVRow_SSSE3(const uint8* src_argb, int src_stride_argb, 723 uint8* dst_u, uint8* dst_v, int width); 724 void ARGBToUVJRow_SSSE3(const uint8* src_argb, int src_stride_argb, 725 uint8* dst_u, uint8* dst_v, int width); 726 void BGRAToUVRow_SSSE3(const uint8* src_bgra, int src_stride_bgra, 727 uint8* dst_u, uint8* dst_v, int width); 728 void ABGRToUVRow_SSSE3(const uint8* src_abgr, int src_stride_abgr, 729 uint8* dst_u, uint8* dst_v, int width); 730 void RGBAToUVRow_SSSE3(const uint8* src_rgba, int src_stride_rgba, 731 uint8* dst_u, uint8* dst_v, int width); 732 void ARGBToUVRow_Any_AVX2(const uint8* src_argb, int src_stride_argb, 733 uint8* dst_u, uint8* dst_v, int width); 734 void ARGBToUVJRow_Any_AVX2(const uint8* src_argb, int src_stride_argb, 735 uint8* dst_u, uint8* dst_v, int width); 736 void ARGBToUVRow_Any_SSSE3(const uint8* src_argb, int src_stride_argb, 737 uint8* dst_u, uint8* dst_v, int width); 738 void ARGBToUVJRow_Any_SSSE3(const uint8* src_argb, int src_stride_argb, 739 uint8* dst_u, uint8* dst_v, int width); 740 void BGRAToUVRow_Any_SSSE3(const uint8* src_bgra, int src_stride_bgra, 741 uint8* dst_u, uint8* dst_v, int width); 742 void ABGRToUVRow_Any_SSSE3(const uint8* src_abgr, int src_stride_abgr, 743 uint8* dst_u, uint8* dst_v, int width); 744 void RGBAToUVRow_Any_SSSE3(const uint8* src_rgba, int src_stride_rgba, 745 uint8* dst_u, uint8* dst_v, int width); 746 void ARGBToUV444Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v, 747 int width); 748 void ARGBToUV411Row_Any_NEON(const uint8* src_argb, uint8* dst_u, uint8* dst_v, 749 int width); 750 void ARGBToUVRow_Any_NEON(const uint8* src_argb, int src_stride_argb, 751 uint8* dst_u, uint8* dst_v, int width); 752 void ARGBToUVJRow_Any_NEON(const uint8* src_argb, int src_stride_argb, 753 uint8* dst_u, uint8* dst_v, int width); 754 void BGRAToUVRow_Any_NEON(const uint8* src_bgra, int src_stride_bgra, 755 uint8* dst_u, uint8* dst_v, int width); 756 void ABGRToUVRow_Any_NEON(const uint8* src_abgr, int src_stride_abgr, 757 uint8* dst_u, uint8* dst_v, int width); 758 void RGBAToUVRow_Any_NEON(const uint8* src_rgba, int src_stride_rgba, 759 uint8* dst_u, uint8* dst_v, int width); 760 void RGB24ToUVRow_Any_NEON(const uint8* src_rgb24, int src_stride_rgb24, 761 uint8* dst_u, uint8* dst_v, int width); 762 void RAWToUVRow_Any_NEON(const uint8* src_raw, int src_stride_raw, 763 uint8* dst_u, uint8* dst_v, int width); 764 void RGB565ToUVRow_Any_NEON(const uint8* src_rgb565, int src_stride_rgb565, 765 uint8* dst_u, uint8* dst_v, int width); 766 void ARGB1555ToUVRow_Any_NEON(const uint8* src_argb1555, 767 int src_stride_argb1555, 768 uint8* dst_u, uint8* dst_v, int width); 769 void ARGB4444ToUVRow_Any_NEON(const uint8* src_argb4444, 770 int src_stride_argb4444, 771 uint8* dst_u, uint8* dst_v, int width); 772 void ARGBToUVRow_C(const uint8* src_argb, int src_stride_argb, 773 uint8* dst_u, uint8* dst_v, int width); 774 void ARGBToUVJRow_C(const uint8* src_argb, int src_stride_argb, 775 uint8* dst_u, uint8* dst_v, int width); 776 void BGRAToUVRow_C(const uint8* src_bgra, int src_stride_bgra, 777 uint8* dst_u, uint8* dst_v, int width); 778 void ABGRToUVRow_C(const uint8* src_abgr, int src_stride_abgr, 779 uint8* dst_u, uint8* dst_v, int width); 780 void RGBAToUVRow_C(const uint8* src_rgba, int src_stride_rgba, 781 uint8* dst_u, uint8* dst_v, int width); 782 void RGB24ToUVRow_C(const uint8* src_rgb24, int src_stride_rgb24, 783 uint8* dst_u, uint8* dst_v, int width); 784 void RAWToUVRow_C(const uint8* src_raw, int src_stride_raw, 785 uint8* dst_u, uint8* dst_v, int width); 786 void RGB565ToUVRow_C(const uint8* src_rgb565, int src_stride_rgb565, 787 uint8* dst_u, uint8* dst_v, int width); 788 void ARGB1555ToUVRow_C(const uint8* src_argb1555, int src_stride_argb1555, 789 uint8* dst_u, uint8* dst_v, int width); 790 void ARGB4444ToUVRow_C(const uint8* src_argb4444, int src_stride_argb4444, 791 uint8* dst_u, uint8* dst_v, int width); 792 793 void ARGBToUV444Row_SSSE3(const uint8* src_argb, 794 uint8* dst_u, uint8* dst_v, int width); 795 void ARGBToUV444Row_Any_SSSE3(const uint8* src_argb, 796 uint8* dst_u, uint8* dst_v, int width); 797 798 void ARGBToUV444Row_C(const uint8* src_argb, 799 uint8* dst_u, uint8* dst_v, int width); 800 void ARGBToUV411Row_C(const uint8* src_argb, 801 uint8* dst_u, uint8* dst_v, int width); 802 803 void MirrorRow_AVX2(const uint8* src, uint8* dst, int width); 804 void MirrorRow_SSSE3(const uint8* src, uint8* dst, int width); 805 void MirrorRow_NEON(const uint8* src, uint8* dst, int width); 806 void MirrorRow_DSPR2(const uint8* src, uint8* dst, int width); 807 void MirrorRow_C(const uint8* src, uint8* dst, int width); 808 void MirrorRow_Any_AVX2(const uint8* src, uint8* dst, int width); 809 void MirrorRow_Any_SSSE3(const uint8* src, uint8* dst, int width); 810 void MirrorRow_Any_SSE2(const uint8* src, uint8* dst, int width); 811 void MirrorRow_Any_NEON(const uint8* src, uint8* dst, int width); 812 813 void MirrorUVRow_SSSE3(const uint8* src_uv, uint8* dst_u, uint8* dst_v, 814 int width); 815 void MirrorUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v, 816 int width); 817 void MirrorUVRow_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, 818 int width); 819 void MirrorUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width); 820 821 void ARGBMirrorRow_AVX2(const uint8* src, uint8* dst, int width); 822 void ARGBMirrorRow_SSE2(const uint8* src, uint8* dst, int width); 823 void ARGBMirrorRow_NEON(const uint8* src, uint8* dst, int width); 824 void ARGBMirrorRow_C(const uint8* src, uint8* dst, int width); 825 void ARGBMirrorRow_Any_AVX2(const uint8* src, uint8* dst, int width); 826 void ARGBMirrorRow_Any_SSE2(const uint8* src, uint8* dst, int width); 827 void ARGBMirrorRow_Any_NEON(const uint8* src, uint8* dst, int width); 828 829 void SplitUVRow_C(const uint8* src_uv, uint8* dst_u, uint8* dst_v, int width); 830 void SplitUVRow_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, 831 int width); 832 void SplitUVRow_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, 833 int width); 834 void SplitUVRow_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v, 835 int width); 836 void SplitUVRow_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, 837 int width); 838 void SplitUVRow_Any_SSE2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, 839 int width); 840 void SplitUVRow_Any_AVX2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, 841 int width); 842 void SplitUVRow_Any_NEON(const uint8* src_uv, uint8* dst_u, uint8* dst_v, 843 int width); 844 void SplitUVRow_Any_DSPR2(const uint8* src_uv, uint8* dst_u, uint8* dst_v, 845 int width); 846 847 void MergeUVRow_C(const uint8* src_u, const uint8* src_v, uint8* dst_uv, 848 int width); 849 void MergeUVRow_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv, 850 int width); 851 void MergeUVRow_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv, 852 int width); 853 void MergeUVRow_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv, 854 int width); 855 void MergeUVRow_Any_SSE2(const uint8* src_u, const uint8* src_v, uint8* dst_uv, 856 int width); 857 void MergeUVRow_Any_AVX2(const uint8* src_u, const uint8* src_v, uint8* dst_uv, 858 int width); 859 void MergeUVRow_Any_NEON(const uint8* src_u, const uint8* src_v, uint8* dst_uv, 860 int width); 861 862 void CopyRow_SSE2(const uint8* src, uint8* dst, int count); 863 void CopyRow_AVX(const uint8* src, uint8* dst, int count); 864 void CopyRow_ERMS(const uint8* src, uint8* dst, int count); 865 void CopyRow_NEON(const uint8* src, uint8* dst, int count); 866 void CopyRow_MIPS(const uint8* src, uint8* dst, int count); 867 void CopyRow_C(const uint8* src, uint8* dst, int count); 868 void CopyRow_Any_SSE2(const uint8* src, uint8* dst, int count); 869 void CopyRow_Any_AVX(const uint8* src, uint8* dst, int count); 870 void CopyRow_Any_NEON(const uint8* src, uint8* dst, int count); 871 872 void CopyRow_16_C(const uint16* src, uint16* dst, int count); 873 874 void ARGBCopyAlphaRow_C(const uint8* src_argb, uint8* dst_argb, int width); 875 void ARGBCopyAlphaRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width); 876 void ARGBCopyAlphaRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width); 877 void ARGBCopyAlphaRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb, 878 int width); 879 void ARGBCopyAlphaRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb, 880 int width); 881 882 void ARGBExtractAlphaRow_C(const uint8* src_argb, uint8* dst_a, int width); 883 void ARGBExtractAlphaRow_SSE2(const uint8* src_argb, uint8* dst_a, int width); 884 void ARGBExtractAlphaRow_NEON(const uint8* src_argb, uint8* dst_a, int width); 885 void ARGBExtractAlphaRow_Any_SSE2(const uint8* src_argb, uint8* dst_a, 886 int width); 887 void ARGBExtractAlphaRow_Any_NEON(const uint8* src_argb, uint8* dst_a, 888 int width); 889 890 void ARGBCopyYToAlphaRow_C(const uint8* src_y, uint8* dst_argb, int width); 891 void ARGBCopyYToAlphaRow_SSE2(const uint8* src_y, uint8* dst_argb, int width); 892 void ARGBCopyYToAlphaRow_AVX2(const uint8* src_y, uint8* dst_argb, int width); 893 void ARGBCopyYToAlphaRow_Any_SSE2(const uint8* src_y, uint8* dst_argb, 894 int width); 895 void ARGBCopyYToAlphaRow_Any_AVX2(const uint8* src_y, uint8* dst_argb, 896 int width); 897 898 void SetRow_C(uint8* dst, uint8 v8, int count); 899 void SetRow_X86(uint8* dst, uint8 v8, int count); 900 void SetRow_ERMS(uint8* dst, uint8 v8, int count); 901 void SetRow_NEON(uint8* dst, uint8 v8, int count); 902 void SetRow_Any_X86(uint8* dst, uint8 v8, int count); 903 void SetRow_Any_NEON(uint8* dst, uint8 v8, int count); 904 905 void ARGBSetRow_C(uint8* dst_argb, uint32 v32, int count); 906 void ARGBSetRow_X86(uint8* dst_argb, uint32 v32, int count); 907 void ARGBSetRow_NEON(uint8* dst_argb, uint32 v32, int count); 908 void ARGBSetRow_Any_NEON(uint8* dst_argb, uint32 v32, int count); 909 910 // ARGBShufflers for BGRAToARGB etc. 911 void ARGBShuffleRow_C(const uint8* src_argb, uint8* dst_argb, 912 const uint8* shuffler, int width); 913 void ARGBShuffleRow_SSE2(const uint8* src_argb, uint8* dst_argb, 914 const uint8* shuffler, int width); 915 void ARGBShuffleRow_SSSE3(const uint8* src_argb, uint8* dst_argb, 916 const uint8* shuffler, int width); 917 void ARGBShuffleRow_AVX2(const uint8* src_argb, uint8* dst_argb, 918 const uint8* shuffler, int width); 919 void ARGBShuffleRow_NEON(const uint8* src_argb, uint8* dst_argb, 920 const uint8* shuffler, int width); 921 void ARGBShuffleRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb, 922 const uint8* shuffler, int width); 923 void ARGBShuffleRow_Any_SSSE3(const uint8* src_argb, uint8* dst_argb, 924 const uint8* shuffler, int width); 925 void ARGBShuffleRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb, 926 const uint8* shuffler, int width); 927 void ARGBShuffleRow_Any_NEON(const uint8* src_argb, uint8* dst_argb, 928 const uint8* shuffler, int width); 929 930 void RGB24ToARGBRow_SSSE3(const uint8* src_rgb24, uint8* dst_argb, int width); 931 void RAWToARGBRow_SSSE3(const uint8* src_raw, uint8* dst_argb, int width); 932 void RAWToRGB24Row_SSSE3(const uint8* src_raw, uint8* dst_rgb24, int width); 933 void RGB565ToARGBRow_SSE2(const uint8* src_rgb565, uint8* dst_argb, int width); 934 void ARGB1555ToARGBRow_SSE2(const uint8* src_argb1555, uint8* dst_argb, 935 int width); 936 void ARGB4444ToARGBRow_SSE2(const uint8* src_argb4444, uint8* dst_argb, 937 int width); 938 void RGB565ToARGBRow_AVX2(const uint8* src_rgb565, uint8* dst_argb, int width); 939 void ARGB1555ToARGBRow_AVX2(const uint8* src_argb1555, uint8* dst_argb, 940 int width); 941 void ARGB4444ToARGBRow_AVX2(const uint8* src_argb4444, uint8* dst_argb, 942 int width); 943 944 void RGB24ToARGBRow_NEON(const uint8* src_rgb24, uint8* dst_argb, int width); 945 void RAWToARGBRow_NEON(const uint8* src_raw, uint8* dst_argb, int width); 946 void RAWToRGB24Row_NEON(const uint8* src_raw, uint8* dst_rgb24, int width); 947 void RGB565ToARGBRow_NEON(const uint8* src_rgb565, uint8* dst_argb, int width); 948 void ARGB1555ToARGBRow_NEON(const uint8* src_argb1555, uint8* dst_argb, 949 int width); 950 void ARGB4444ToARGBRow_NEON(const uint8* src_argb4444, uint8* dst_argb, 951 int width); 952 void RGB24ToARGBRow_C(const uint8* src_rgb24, uint8* dst_argb, int width); 953 void RAWToARGBRow_C(const uint8* src_raw, uint8* dst_argb, int width); 954 void RAWToRGB24Row_C(const uint8* src_raw, uint8* dst_rgb24, int width); 955 void RGB565ToARGBRow_C(const uint8* src_rgb, uint8* dst_argb, int width); 956 void ARGB1555ToARGBRow_C(const uint8* src_argb, uint8* dst_argb, int width); 957 void ARGB4444ToARGBRow_C(const uint8* src_argb, uint8* dst_argb, int width); 958 void RGB24ToARGBRow_Any_SSSE3(const uint8* src_rgb24, uint8* dst_argb, 959 int width); 960 void RAWToARGBRow_Any_SSSE3(const uint8* src_raw, uint8* dst_argb, int width); 961 void RAWToRGB24Row_Any_SSSE3(const uint8* src_raw, uint8* dst_rgb24, int width); 962 963 void RGB565ToARGBRow_Any_SSE2(const uint8* src_rgb565, uint8* dst_argb, 964 int width); 965 void ARGB1555ToARGBRow_Any_SSE2(const uint8* src_argb1555, uint8* dst_argb, 966 int width); 967 void ARGB4444ToARGBRow_Any_SSE2(const uint8* src_argb4444, uint8* dst_argb, 968 int width); 969 void RGB565ToARGBRow_Any_AVX2(const uint8* src_rgb565, uint8* dst_argb, 970 int width); 971 void ARGB1555ToARGBRow_Any_AVX2(const uint8* src_argb1555, uint8* dst_argb, 972 int width); 973 void ARGB4444ToARGBRow_Any_AVX2(const uint8* src_argb4444, uint8* dst_argb, 974 int width); 975 976 void RGB24ToARGBRow_Any_NEON(const uint8* src_rgb24, uint8* dst_argb, 977 int width); 978 void RAWToARGBRow_Any_NEON(const uint8* src_raw, uint8* dst_argb, int width); 979 void RAWToRGB24Row_Any_NEON(const uint8* src_raw, uint8* dst_rgb24, int width); 980 void RGB565ToARGBRow_Any_NEON(const uint8* src_rgb565, uint8* dst_argb, 981 int width); 982 void ARGB1555ToARGBRow_Any_NEON(const uint8* src_argb1555, uint8* dst_argb, 983 int width); 984 void ARGB4444ToARGBRow_Any_NEON(const uint8* src_argb4444, uint8* dst_argb, 985 int width); 986 987 void ARGBToRGB24Row_SSSE3(const uint8* src_argb, uint8* dst_rgb, int width); 988 void ARGBToRAWRow_SSSE3(const uint8* src_argb, uint8* dst_rgb, int width); 989 void ARGBToRGB565Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int width); 990 void ARGBToARGB1555Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int width); 991 void ARGBToARGB4444Row_SSE2(const uint8* src_argb, uint8* dst_rgb, int width); 992 993 void ARGBToRGB565DitherRow_C(const uint8* src_argb, uint8* dst_rgb, 994 const uint32 dither4, int width); 995 void ARGBToRGB565DitherRow_SSE2(const uint8* src_argb, uint8* dst_rgb, 996 const uint32 dither4, int width); 997 void ARGBToRGB565DitherRow_AVX2(const uint8* src_argb, uint8* dst_rgb, 998 const uint32 dither4, int width); 999 1000 void ARGBToRGB565Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int width); 1001 void ARGBToARGB1555Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int width); 1002 void ARGBToARGB4444Row_AVX2(const uint8* src_argb, uint8* dst_rgb, int width); 1003 1004 void ARGBToRGB24Row_NEON(const uint8* src_argb, uint8* dst_rgb, int width); 1005 void ARGBToRAWRow_NEON(const uint8* src_argb, uint8* dst_rgb, int width); 1006 void ARGBToRGB565Row_NEON(const uint8* src_argb, uint8* dst_rgb, int width); 1007 void ARGBToARGB1555Row_NEON(const uint8* src_argb, uint8* dst_rgb, int width); 1008 void ARGBToARGB4444Row_NEON(const uint8* src_argb, uint8* dst_rgb, int width); 1009 void ARGBToRGB565DitherRow_NEON(const uint8* src_argb, uint8* dst_rgb, 1010 const uint32 dither4, int width); 1011 1012 void ARGBToRGBARow_C(const uint8* src_argb, uint8* dst_rgb, int width); 1013 void ARGBToRGB24Row_C(const uint8* src_argb, uint8* dst_rgb, int width); 1014 void ARGBToRAWRow_C(const uint8* src_argb, uint8* dst_rgb, int width); 1015 void ARGBToRGB565Row_C(const uint8* src_argb, uint8* dst_rgb, int width); 1016 void ARGBToARGB1555Row_C(const uint8* src_argb, uint8* dst_rgb, int width); 1017 void ARGBToARGB4444Row_C(const uint8* src_argb, uint8* dst_rgb, int width); 1018 1019 void J400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int width); 1020 void J400ToARGBRow_AVX2(const uint8* src_y, uint8* dst_argb, int width); 1021 void J400ToARGBRow_NEON(const uint8* src_y, uint8* dst_argb, int width); 1022 void J400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int width); 1023 void J400ToARGBRow_Any_SSE2(const uint8* src_y, uint8* dst_argb, int width); 1024 void J400ToARGBRow_Any_AVX2(const uint8* src_y, uint8* dst_argb, int width); 1025 void J400ToARGBRow_Any_NEON(const uint8* src_y, uint8* dst_argb, int width); 1026 1027 void I444ToARGBRow_C(const uint8* src_y, 1028 const uint8* src_u, 1029 const uint8* src_v, 1030 uint8* dst_argb, 1031 const struct YuvConstants* yuvconstants, 1032 int width); 1033 void I422ToARGBRow_C(const uint8* src_y, 1034 const uint8* src_u, 1035 const uint8* src_v, 1036 uint8* dst_argb, 1037 const struct YuvConstants* yuvconstants, 1038 int width); 1039 void I422ToARGBRow_C(const uint8* src_y, 1040 const uint8* src_u, 1041 const uint8* src_v, 1042 uint8* dst_argb, 1043 const struct YuvConstants* yuvconstants, 1044 int width); 1045 void I422AlphaToARGBRow_C(const uint8* y_buf, 1046 const uint8* u_buf, 1047 const uint8* v_buf, 1048 const uint8* a_buf, 1049 uint8* dst_argb, 1050 const struct YuvConstants* yuvconstants, 1051 int width); 1052 void I411ToARGBRow_C(const uint8* src_y, 1053 const uint8* src_u, 1054 const uint8* src_v, 1055 uint8* dst_argb, 1056 const struct YuvConstants* yuvconstants, 1057 int width); 1058 void NV12ToARGBRow_C(const uint8* src_y, 1059 const uint8* src_uv, 1060 uint8* dst_argb, 1061 const struct YuvConstants* yuvconstants, 1062 int width); 1063 void NV12ToRGB565Row_C(const uint8* src_y, 1064 const uint8* src_uv, 1065 uint8* dst_argb, 1066 const struct YuvConstants* yuvconstants, 1067 int width); 1068 void NV21ToARGBRow_C(const uint8* src_y, 1069 const uint8* src_uv, 1070 uint8* dst_argb, 1071 const struct YuvConstants* yuvconstants, 1072 int width); 1073 void YUY2ToARGBRow_C(const uint8* src_yuy2, 1074 uint8* dst_argb, 1075 const struct YuvConstants* yuvconstants, 1076 int width); 1077 void UYVYToARGBRow_C(const uint8* src_uyvy, 1078 uint8* dst_argb, 1079 const struct YuvConstants* yuvconstants, 1080 int width); 1081 void I422ToRGBARow_C(const uint8* src_y, 1082 const uint8* src_u, 1083 const uint8* src_v, 1084 uint8* dst_rgba, 1085 const struct YuvConstants* yuvconstants, 1086 int width); 1087 void I422ToRGB24Row_C(const uint8* src_y, 1088 const uint8* src_u, 1089 const uint8* src_v, 1090 uint8* dst_rgb24, 1091 const struct YuvConstants* yuvconstants, 1092 int width); 1093 void I422ToARGB4444Row_C(const uint8* src_y, 1094 const uint8* src_u, 1095 const uint8* src_v, 1096 uint8* dst_argb4444, 1097 const struct YuvConstants* yuvconstants, 1098 int width); 1099 void I422ToARGB1555Row_C(const uint8* src_y, 1100 const uint8* src_u, 1101 const uint8* src_v, 1102 uint8* dst_argb4444, 1103 const struct YuvConstants* yuvconstants, 1104 int width); 1105 void I422ToRGB565Row_C(const uint8* src_y, 1106 const uint8* src_u, 1107 const uint8* src_v, 1108 uint8* dst_rgb565, 1109 const struct YuvConstants* yuvconstants, 1110 int width); 1111 void I422ToARGBRow_AVX2(const uint8* src_y, 1112 const uint8* src_u, 1113 const uint8* src_v, 1114 uint8* dst_argb, 1115 const struct YuvConstants* yuvconstants, 1116 int width); 1117 void I422ToARGBRow_AVX2(const uint8* src_y, 1118 const uint8* src_u, 1119 const uint8* src_v, 1120 uint8* dst_argb, 1121 const struct YuvConstants* yuvconstants, 1122 int width); 1123 void I422ToRGBARow_AVX2(const uint8* src_y, 1124 const uint8* src_u, 1125 const uint8* src_v, 1126 uint8* dst_argb, 1127 const struct YuvConstants* yuvconstants, 1128 int width); 1129 void I444ToARGBRow_SSSE3(const uint8* src_y, 1130 const uint8* src_u, 1131 const uint8* src_v, 1132 uint8* dst_argb, 1133 const struct YuvConstants* yuvconstants, 1134 int width); 1135 void I444ToARGBRow_AVX2(const uint8* src_y, 1136 const uint8* src_u, 1137 const uint8* src_v, 1138 uint8* dst_argb, 1139 const struct YuvConstants* yuvconstants, 1140 int width); 1141 void I444ToARGBRow_SSSE3(const uint8* src_y, 1142 const uint8* src_u, 1143 const uint8* src_v, 1144 uint8* dst_argb, 1145 const struct YuvConstants* yuvconstants, 1146 int width); 1147 void I444ToARGBRow_AVX2(const uint8* src_y, 1148 const uint8* src_u, 1149 const uint8* src_v, 1150 uint8* dst_argb, 1151 const struct YuvConstants* yuvconstants, 1152 int width); 1153 void I422ToARGBRow_SSSE3(const uint8* src_y, 1154 const uint8* src_u, 1155 const uint8* src_v, 1156 uint8* dst_argb, 1157 const struct YuvConstants* yuvconstants, 1158 int width); 1159 void I422AlphaToARGBRow_SSSE3(const uint8* y_buf, 1160 const uint8* u_buf, 1161 const uint8* v_buf, 1162 const uint8* a_buf, 1163 uint8* dst_argb, 1164 const struct YuvConstants* yuvconstants, 1165 int width); 1166 void I422AlphaToARGBRow_AVX2(const uint8* y_buf, 1167 const uint8* u_buf, 1168 const uint8* v_buf, 1169 const uint8* a_buf, 1170 uint8* dst_argb, 1171 const struct YuvConstants* yuvconstants, 1172 int width); 1173 void I422ToARGBRow_SSSE3(const uint8* src_y, 1174 const uint8* src_u, 1175 const uint8* src_v, 1176 uint8* dst_argb, 1177 const struct YuvConstants* yuvconstants, 1178 int width); 1179 void I411ToARGBRow_SSSE3(const uint8* src_y, 1180 const uint8* src_u, 1181 const uint8* src_v, 1182 uint8* dst_argb, 1183 const struct YuvConstants* yuvconstants, 1184 int width); 1185 void I411ToARGBRow_AVX2(const uint8* src_y, 1186 const uint8* src_u, 1187 const uint8* src_v, 1188 uint8* dst_argb, 1189 const struct YuvConstants* yuvconstants, 1190 int width); 1191 void NV12ToARGBRow_SSSE3(const uint8* src_y, 1192 const uint8* src_uv, 1193 uint8* dst_argb, 1194 const struct YuvConstants* yuvconstants, 1195 int width); 1196 void NV12ToARGBRow_AVX2(const uint8* src_y, 1197 const uint8* src_uv, 1198 uint8* dst_argb, 1199 const struct YuvConstants* yuvconstants, 1200 int width); 1201 void NV12ToRGB565Row_SSSE3(const uint8* src_y, 1202 const uint8* src_uv, 1203 uint8* dst_argb, 1204 const struct YuvConstants* yuvconstants, 1205 int width); 1206 void NV12ToRGB565Row_AVX2(const uint8* src_y, 1207 const uint8* src_uv, 1208 uint8* dst_argb, 1209 const struct YuvConstants* yuvconstants, 1210 int width); 1211 void NV21ToARGBRow_SSSE3(const uint8* src_y, 1212 const uint8* src_uv, 1213 uint8* dst_argb, 1214 const struct YuvConstants* yuvconstants, 1215 int width); 1216 void NV21ToARGBRow_AVX2(const uint8* src_y, 1217 const uint8* src_uv, 1218 uint8* dst_argb, 1219 const struct YuvConstants* yuvconstants, 1220 int width); 1221 void YUY2ToARGBRow_SSSE3(const uint8* src_yuy2, 1222 uint8* dst_argb, 1223 const struct YuvConstants* yuvconstants, 1224 int width); 1225 void UYVYToARGBRow_SSSE3(const uint8* src_uyvy, 1226 uint8* dst_argb, 1227 const struct YuvConstants* yuvconstants, 1228 int width); 1229 void YUY2ToARGBRow_AVX2(const uint8* src_yuy2, 1230 uint8* dst_argb, 1231 const struct YuvConstants* yuvconstants, 1232 int width); 1233 void UYVYToARGBRow_AVX2(const uint8* src_uyvy, 1234 uint8* dst_argb, 1235 const struct YuvConstants* yuvconstants, 1236 int width); 1237 void I422ToRGBARow_SSSE3(const uint8* src_y, 1238 const uint8* src_u, 1239 const uint8* src_v, 1240 uint8* dst_rgba, 1241 const struct YuvConstants* yuvconstants, 1242 int width); 1243 void I422ToARGB4444Row_SSSE3(const uint8* src_y, 1244 const uint8* src_u, 1245 const uint8* src_v, 1246 uint8* dst_argb, 1247 const struct YuvConstants* yuvconstants, 1248 int width); 1249 void I422ToARGB4444Row_AVX2(const uint8* src_y, 1250 const uint8* src_u, 1251 const uint8* src_v, 1252 uint8* dst_argb, 1253 const struct YuvConstants* yuvconstants, 1254 int width); 1255 void I422ToARGB1555Row_SSSE3(const uint8* src_y, 1256 const uint8* src_u, 1257 const uint8* src_v, 1258 uint8* dst_argb, 1259 const struct YuvConstants* yuvconstants, 1260 int width); 1261 void I422ToARGB1555Row_AVX2(const uint8* src_y, 1262 const uint8* src_u, 1263 const uint8* src_v, 1264 uint8* dst_argb, 1265 const struct YuvConstants* yuvconstants, 1266 int width); 1267 void I422ToRGB565Row_SSSE3(const uint8* src_y, 1268 const uint8* src_u, 1269 const uint8* src_v, 1270 uint8* dst_argb, 1271 const struct YuvConstants* yuvconstants, 1272 int width); 1273 void I422ToRGB565Row_AVX2(const uint8* src_y, 1274 const uint8* src_u, 1275 const uint8* src_v, 1276 uint8* dst_argb, 1277 const struct YuvConstants* yuvconstants, 1278 int width); 1279 void I422ToRGB24Row_SSSE3(const uint8* src_y, 1280 const uint8* src_u, 1281 const uint8* src_v, 1282 uint8* dst_rgb24, 1283 const struct YuvConstants* yuvconstants, 1284 int width); 1285 void I422ToRGB24Row_AVX2(const uint8* src_y, 1286 const uint8* src_u, 1287 const uint8* src_v, 1288 uint8* dst_rgb24, 1289 const struct YuvConstants* yuvconstants, 1290 int width); 1291 void I422ToARGBRow_Any_AVX2(const uint8* src_y, 1292 const uint8* src_u, 1293 const uint8* src_v, 1294 uint8* dst_argb, 1295 const struct YuvConstants* yuvconstants, 1296 int width); 1297 void I422ToRGBARow_Any_AVX2(const uint8* src_y, 1298 const uint8* src_u, 1299 const uint8* src_v, 1300 uint8* dst_argb, 1301 const struct YuvConstants* yuvconstants, 1302 int width); 1303 void I444ToARGBRow_Any_SSSE3(const uint8* src_y, 1304 const uint8* src_u, 1305 const uint8* src_v, 1306 uint8* dst_argb, 1307 const struct YuvConstants* yuvconstants, 1308 int width); 1309 void I444ToARGBRow_Any_AVX2(const uint8* src_y, 1310 const uint8* src_u, 1311 const uint8* src_v, 1312 uint8* dst_argb, 1313 const struct YuvConstants* yuvconstants, 1314 int width); 1315 void I422ToARGBRow_Any_SSSE3(const uint8* src_y, 1316 const uint8* src_u, 1317 const uint8* src_v, 1318 uint8* dst_argb, 1319 const struct YuvConstants* yuvconstants, 1320 int width); 1321 void I422AlphaToARGBRow_Any_SSSE3(const uint8* y_buf, 1322 const uint8* u_buf, 1323 const uint8* v_buf, 1324 const uint8* a_buf, 1325 uint8* dst_argb, 1326 const struct YuvConstants* yuvconstants, 1327 int width); 1328 void I422AlphaToARGBRow_Any_AVX2(const uint8* y_buf, 1329 const uint8* u_buf, 1330 const uint8* v_buf, 1331 const uint8* a_buf, 1332 uint8* dst_argb, 1333 const struct YuvConstants* yuvconstants, 1334 int width); 1335 void I411ToARGBRow_Any_SSSE3(const uint8* src_y, 1336 const uint8* src_u, 1337 const uint8* src_v, 1338 uint8* dst_argb, 1339 const struct YuvConstants* yuvconstants, 1340 int width); 1341 void I411ToARGBRow_Any_AVX2(const uint8* src_y, 1342 const uint8* src_u, 1343 const uint8* src_v, 1344 uint8* dst_argb, 1345 const struct YuvConstants* yuvconstants, 1346 int width); 1347 void NV12ToARGBRow_Any_SSSE3(const uint8* src_y, 1348 const uint8* src_uv, 1349 uint8* dst_argb, 1350 const struct YuvConstants* yuvconstants, 1351 int width); 1352 void NV12ToARGBRow_Any_AVX2(const uint8* src_y, 1353 const uint8* src_uv, 1354 uint8* dst_argb, 1355 const struct YuvConstants* yuvconstants, 1356 int width); 1357 void NV21ToARGBRow_Any_SSSE3(const uint8* src_y, 1358 const uint8* src_vu, 1359 uint8* dst_argb, 1360 const struct YuvConstants* yuvconstants, 1361 int width); 1362 void NV21ToARGBRow_Any_AVX2(const uint8* src_y, 1363 const uint8* src_vu, 1364 uint8* dst_argb, 1365 const struct YuvConstants* yuvconstants, 1366 int width); 1367 void NV12ToRGB565Row_Any_SSSE3(const uint8* src_y, 1368 const uint8* src_uv, 1369 uint8* dst_argb, 1370 const struct YuvConstants* yuvconstants, 1371 int width); 1372 void NV12ToRGB565Row_Any_AVX2(const uint8* src_y, 1373 const uint8* src_uv, 1374 uint8* dst_argb, 1375 const struct YuvConstants* yuvconstants, 1376 int width); 1377 void YUY2ToARGBRow_Any_SSSE3(const uint8* src_yuy2, 1378 uint8* dst_argb, 1379 const struct YuvConstants* yuvconstants, 1380 int width); 1381 void UYVYToARGBRow_Any_SSSE3(const uint8* src_uyvy, 1382 uint8* dst_argb, 1383 const struct YuvConstants* yuvconstants, 1384 int width); 1385 void YUY2ToARGBRow_Any_AVX2(const uint8* src_yuy2, 1386 uint8* dst_argb, 1387 const struct YuvConstants* yuvconstants, 1388 int width); 1389 void UYVYToARGBRow_Any_AVX2(const uint8* src_uyvy, 1390 uint8* dst_argb, 1391 const struct YuvConstants* yuvconstants, 1392 int width); 1393 void I422ToRGBARow_Any_SSSE3(const uint8* src_y, 1394 const uint8* src_u, 1395 const uint8* src_v, 1396 uint8* dst_rgba, 1397 const struct YuvConstants* yuvconstants, 1398 int width); 1399 void I422ToARGB4444Row_Any_SSSE3(const uint8* src_y, 1400 const uint8* src_u, 1401 const uint8* src_v, 1402 uint8* dst_rgba, 1403 const struct YuvConstants* yuvconstants, 1404 int width); 1405 void I422ToARGB4444Row_Any_AVX2(const uint8* src_y, 1406 const uint8* src_u, 1407 const uint8* src_v, 1408 uint8* dst_rgba, 1409 const struct YuvConstants* yuvconstants, 1410 int width); 1411 void I422ToARGB1555Row_Any_SSSE3(const uint8* src_y, 1412 const uint8* src_u, 1413 const uint8* src_v, 1414 uint8* dst_rgba, 1415 const struct YuvConstants* yuvconstants, 1416 int width); 1417 void I422ToARGB1555Row_Any_AVX2(const uint8* src_y, 1418 const uint8* src_u, 1419 const uint8* src_v, 1420 uint8* dst_rgba, 1421 const struct YuvConstants* yuvconstants, 1422 int width); 1423 void I422ToRGB565Row_Any_SSSE3(const uint8* src_y, 1424 const uint8* src_u, 1425 const uint8* src_v, 1426 uint8* dst_rgba, 1427 const struct YuvConstants* yuvconstants, 1428 int width); 1429 void I422ToRGB565Row_Any_AVX2(const uint8* src_y, 1430 const uint8* src_u, 1431 const uint8* src_v, 1432 uint8* dst_rgba, 1433 const struct YuvConstants* yuvconstants, 1434 int width); 1435 void I422ToRGB24Row_Any_SSSE3(const uint8* src_y, 1436 const uint8* src_u, 1437 const uint8* src_v, 1438 uint8* dst_argb, 1439 const struct YuvConstants* yuvconstants, 1440 int width); 1441 void I422ToRGB24Row_Any_AVX2(const uint8* src_y, 1442 const uint8* src_u, 1443 const uint8* src_v, 1444 uint8* dst_argb, 1445 const struct YuvConstants* yuvconstants, 1446 int width); 1447 1448 void I400ToARGBRow_C(const uint8* src_y, uint8* dst_argb, int width); 1449 void I400ToARGBRow_SSE2(const uint8* src_y, uint8* dst_argb, int width); 1450 void I400ToARGBRow_AVX2(const uint8* src_y, uint8* dst_argb, int width); 1451 void I400ToARGBRow_NEON(const uint8* src_y, uint8* dst_argb, int width); 1452 void I400ToARGBRow_Any_SSE2(const uint8* src_y, uint8* dst_argb, int width); 1453 void I400ToARGBRow_Any_AVX2(const uint8* src_y, uint8* dst_argb, int width); 1454 void I400ToARGBRow_Any_NEON(const uint8* src_y, uint8* dst_argb, int width); 1455 1456 // ARGB preattenuated alpha blend. 1457 void ARGBBlendRow_SSSE3(const uint8* src_argb, const uint8* src_argb1, 1458 uint8* dst_argb, int width); 1459 void ARGBBlendRow_NEON(const uint8* src_argb, const uint8* src_argb1, 1460 uint8* dst_argb, int width); 1461 void ARGBBlendRow_C(const uint8* src_argb, const uint8* src_argb1, 1462 uint8* dst_argb, int width); 1463 1464 // Unattenuated planar alpha blend. 1465 void BlendPlaneRow_SSSE3(const uint8* src0, const uint8* src1, 1466 const uint8* alpha, uint8* dst, int width); 1467 void BlendPlaneRow_Any_SSSE3(const uint8* src0, const uint8* src1, 1468 const uint8* alpha, uint8* dst, int width); 1469 void BlendPlaneRow_AVX2(const uint8* src0, const uint8* src1, 1470 const uint8* alpha, uint8* dst, int width); 1471 void BlendPlaneRow_Any_AVX2(const uint8* src0, const uint8* src1, 1472 const uint8* alpha, uint8* dst, int width); 1473 void BlendPlaneRow_C(const uint8* src0, const uint8* src1, 1474 const uint8* alpha, uint8* dst, int width); 1475 1476 // ARGB multiply images. Same API as Blend, but these require 1477 // pointer and width alignment for SSE2. 1478 void ARGBMultiplyRow_C(const uint8* src_argb, const uint8* src_argb1, 1479 uint8* dst_argb, int width); 1480 void ARGBMultiplyRow_SSE2(const uint8* src_argb, const uint8* src_argb1, 1481 uint8* dst_argb, int width); 1482 void ARGBMultiplyRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1, 1483 uint8* dst_argb, int width); 1484 void ARGBMultiplyRow_AVX2(const uint8* src_argb, const uint8* src_argb1, 1485 uint8* dst_argb, int width); 1486 void ARGBMultiplyRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1, 1487 uint8* dst_argb, int width); 1488 void ARGBMultiplyRow_NEON(const uint8* src_argb, const uint8* src_argb1, 1489 uint8* dst_argb, int width); 1490 void ARGBMultiplyRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1, 1491 uint8* dst_argb, int width); 1492 1493 // ARGB add images. 1494 void ARGBAddRow_C(const uint8* src_argb, const uint8* src_argb1, 1495 uint8* dst_argb, int width); 1496 void ARGBAddRow_SSE2(const uint8* src_argb, const uint8* src_argb1, 1497 uint8* dst_argb, int width); 1498 void ARGBAddRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1, 1499 uint8* dst_argb, int width); 1500 void ARGBAddRow_AVX2(const uint8* src_argb, const uint8* src_argb1, 1501 uint8* dst_argb, int width); 1502 void ARGBAddRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1, 1503 uint8* dst_argb, int width); 1504 void ARGBAddRow_NEON(const uint8* src_argb, const uint8* src_argb1, 1505 uint8* dst_argb, int width); 1506 void ARGBAddRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1, 1507 uint8* dst_argb, int width); 1508 1509 // ARGB subtract images. Same API as Blend, but these require 1510 // pointer and width alignment for SSE2. 1511 void ARGBSubtractRow_C(const uint8* src_argb, const uint8* src_argb1, 1512 uint8* dst_argb, int width); 1513 void ARGBSubtractRow_SSE2(const uint8* src_argb, const uint8* src_argb1, 1514 uint8* dst_argb, int width); 1515 void ARGBSubtractRow_Any_SSE2(const uint8* src_argb, const uint8* src_argb1, 1516 uint8* dst_argb, int width); 1517 void ARGBSubtractRow_AVX2(const uint8* src_argb, const uint8* src_argb1, 1518 uint8* dst_argb, int width); 1519 void ARGBSubtractRow_Any_AVX2(const uint8* src_argb, const uint8* src_argb1, 1520 uint8* dst_argb, int width); 1521 void ARGBSubtractRow_NEON(const uint8* src_argb, const uint8* src_argb1, 1522 uint8* dst_argb, int width); 1523 void ARGBSubtractRow_Any_NEON(const uint8* src_argb, const uint8* src_argb1, 1524 uint8* dst_argb, int width); 1525 1526 void ARGBToRGB24Row_Any_SSSE3(const uint8* src_argb, uint8* dst_rgb, int width); 1527 void ARGBToRAWRow_Any_SSSE3(const uint8* src_argb, uint8* dst_rgb, int width); 1528 void ARGBToRGB565Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, int width); 1529 void ARGBToARGB1555Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, 1530 int width); 1531 void ARGBToARGB4444Row_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, 1532 int width); 1533 1534 void ARGBToRGB565DitherRow_Any_SSE2(const uint8* src_argb, uint8* dst_rgb, 1535 const uint32 dither4, int width); 1536 void ARGBToRGB565DitherRow_Any_AVX2(const uint8* src_argb, uint8* dst_rgb, 1537 const uint32 dither4, int width); 1538 1539 void ARGBToRGB565Row_Any_AVX2(const uint8* src_argb, uint8* dst_rgb, int width); 1540 void ARGBToARGB1555Row_Any_AVX2(const uint8* src_argb, uint8* dst_rgb, 1541 int width); 1542 void ARGBToARGB4444Row_Any_AVX2(const uint8* src_argb, uint8* dst_rgb, 1543 int width); 1544 1545 void ARGBToRGB24Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int width); 1546 void ARGBToRAWRow_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int width); 1547 void ARGBToRGB565Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, int width); 1548 void ARGBToARGB1555Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, 1549 int width); 1550 void ARGBToARGB4444Row_Any_NEON(const uint8* src_argb, uint8* dst_rgb, 1551 int width); 1552 void ARGBToRGB565DitherRow_Any_NEON(const uint8* src_argb, uint8* dst_rgb, 1553 const uint32 dither4, int width); 1554 1555 void I444ToARGBRow_Any_NEON(const uint8* src_y, 1556 const uint8* src_u, 1557 const uint8* src_v, 1558 uint8* dst_argb, 1559 const struct YuvConstants* yuvconstants, 1560 int width); 1561 void I422ToARGBRow_Any_NEON(const uint8* src_y, 1562 const uint8* src_u, 1563 const uint8* src_v, 1564 uint8* dst_argb, 1565 const struct YuvConstants* yuvconstants, 1566 int width); 1567 void I422AlphaToARGBRow_Any_NEON(const uint8* src_y, 1568 const uint8* src_u, 1569 const uint8* src_v, 1570 const uint8* src_a, 1571 uint8* dst_argb, 1572 const struct YuvConstants* yuvconstants, 1573 int width); 1574 void I411ToARGBRow_Any_NEON(const uint8* src_y, 1575 const uint8* src_u, 1576 const uint8* src_v, 1577 uint8* dst_argb, 1578 const struct YuvConstants* yuvconstants, 1579 int width); 1580 void I422ToRGBARow_Any_NEON(const uint8* src_y, 1581 const uint8* src_u, 1582 const uint8* src_v, 1583 uint8* dst_argb, 1584 const struct YuvConstants* yuvconstants, 1585 int width); 1586 void I422ToRGB24Row_Any_NEON(const uint8* src_y, 1587 const uint8* src_u, 1588 const uint8* src_v, 1589 uint8* dst_argb, 1590 const struct YuvConstants* yuvconstants, 1591 int width); 1592 void I422ToARGB4444Row_Any_NEON(const uint8* src_y, 1593 const uint8* src_u, 1594 const uint8* src_v, 1595 uint8* dst_argb, 1596 const struct YuvConstants* yuvconstants, 1597 int width); 1598 void I422ToARGB1555Row_Any_NEON(const uint8* src_y, 1599 const uint8* src_u, 1600 const uint8* src_v, 1601 uint8* dst_argb, 1602 const struct YuvConstants* yuvconstants, 1603 int width); 1604 void I422ToRGB565Row_Any_NEON(const uint8* src_y, 1605 const uint8* src_u, 1606 const uint8* src_v, 1607 uint8* dst_argb, 1608 const struct YuvConstants* yuvconstants, 1609 int width); 1610 void NV12ToARGBRow_Any_NEON(const uint8* src_y, 1611 const uint8* src_uv, 1612 uint8* dst_argb, 1613 const struct YuvConstants* yuvconstants, 1614 int width); 1615 void NV21ToARGBRow_Any_NEON(const uint8* src_y, 1616 const uint8* src_vu, 1617 uint8* dst_argb, 1618 const struct YuvConstants* yuvconstants, 1619 int width); 1620 void NV12ToRGB565Row_Any_NEON(const uint8* src_y, 1621 const uint8* src_uv, 1622 uint8* dst_argb, 1623 const struct YuvConstants* yuvconstants, 1624 int width); 1625 void YUY2ToARGBRow_Any_NEON(const uint8* src_yuy2, 1626 uint8* dst_argb, 1627 const struct YuvConstants* yuvconstants, 1628 int width); 1629 void UYVYToARGBRow_Any_NEON(const uint8* src_uyvy, 1630 uint8* dst_argb, 1631 const struct YuvConstants* yuvconstants, 1632 int width); 1633 void I422ToARGBRow_DSPR2(const uint8* src_y, 1634 const uint8* src_u, 1635 const uint8* src_v, 1636 uint8* dst_argb, 1637 const struct YuvConstants* yuvconstants, 1638 int width); 1639 void I422ToARGBRow_DSPR2(const uint8* src_y, 1640 const uint8* src_u, 1641 const uint8* src_v, 1642 uint8* dst_argb, 1643 const struct YuvConstants* yuvconstants, 1644 int width); 1645 1646 void YUY2ToYRow_AVX2(const uint8* src_yuy2, uint8* dst_y, int width); 1647 void YUY2ToUVRow_AVX2(const uint8* src_yuy2, int stride_yuy2, 1648 uint8* dst_u, uint8* dst_v, int width); 1649 void YUY2ToUV422Row_AVX2(const uint8* src_yuy2, 1650 uint8* dst_u, uint8* dst_v, int width); 1651 void YUY2ToYRow_SSE2(const uint8* src_yuy2, uint8* dst_y, int width); 1652 void YUY2ToUVRow_SSE2(const uint8* src_yuy2, int stride_yuy2, 1653 uint8* dst_u, uint8* dst_v, int width); 1654 void YUY2ToUV422Row_SSE2(const uint8* src_yuy2, 1655 uint8* dst_u, uint8* dst_v, int width); 1656 void YUY2ToYRow_NEON(const uint8* src_yuy2, uint8* dst_y, int width); 1657 void YUY2ToUVRow_NEON(const uint8* src_yuy2, int stride_yuy2, 1658 uint8* dst_u, uint8* dst_v, int width); 1659 void YUY2ToUV422Row_NEON(const uint8* src_yuy2, 1660 uint8* dst_u, uint8* dst_v, int width); 1661 void YUY2ToYRow_C(const uint8* src_yuy2, uint8* dst_y, int width); 1662 void YUY2ToUVRow_C(const uint8* src_yuy2, int stride_yuy2, 1663 uint8* dst_u, uint8* dst_v, int width); 1664 void YUY2ToUV422Row_C(const uint8* src_yuy2, 1665 uint8* dst_u, uint8* dst_v, int width); 1666 void YUY2ToYRow_Any_AVX2(const uint8* src_yuy2, uint8* dst_y, int width); 1667 void YUY2ToUVRow_Any_AVX2(const uint8* src_yuy2, int stride_yuy2, 1668 uint8* dst_u, uint8* dst_v, int width); 1669 void YUY2ToUV422Row_Any_AVX2(const uint8* src_yuy2, 1670 uint8* dst_u, uint8* dst_v, int width); 1671 void YUY2ToYRow_Any_SSE2(const uint8* src_yuy2, uint8* dst_y, int width); 1672 void YUY2ToUVRow_Any_SSE2(const uint8* src_yuy2, int stride_yuy2, 1673 uint8* dst_u, uint8* dst_v, int width); 1674 void YUY2ToUV422Row_Any_SSE2(const uint8* src_yuy2, 1675 uint8* dst_u, uint8* dst_v, int width); 1676 void YUY2ToYRow_Any_NEON(const uint8* src_yuy2, uint8* dst_y, int width); 1677 void YUY2ToUVRow_Any_NEON(const uint8* src_yuy2, int stride_yuy2, 1678 uint8* dst_u, uint8* dst_v, int width); 1679 void YUY2ToUV422Row_Any_NEON(const uint8* src_yuy2, 1680 uint8* dst_u, uint8* dst_v, int width); 1681 void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int width); 1682 void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy, 1683 uint8* dst_u, uint8* dst_v, int width); 1684 void UYVYToUV422Row_AVX2(const uint8* src_uyvy, 1685 uint8* dst_u, uint8* dst_v, int width); 1686 void UYVYToYRow_SSE2(const uint8* src_uyvy, uint8* dst_y, int width); 1687 void UYVYToUVRow_SSE2(const uint8* src_uyvy, int stride_uyvy, 1688 uint8* dst_u, uint8* dst_v, int width); 1689 void UYVYToUV422Row_SSE2(const uint8* src_uyvy, 1690 uint8* dst_u, uint8* dst_v, int width); 1691 void UYVYToYRow_AVX2(const uint8* src_uyvy, uint8* dst_y, int width); 1692 void UYVYToUVRow_AVX2(const uint8* src_uyvy, int stride_uyvy, 1693 uint8* dst_u, uint8* dst_v, int width); 1694 void UYVYToUV422Row_AVX2(const uint8* src_uyvy, 1695 uint8* dst_u, uint8* dst_v, int width); 1696 void UYVYToYRow_NEON(const uint8* src_uyvy, uint8* dst_y, int width); 1697 void UYVYToUVRow_NEON(const uint8* src_uyvy, int stride_uyvy, 1698 uint8* dst_u, uint8* dst_v, int width); 1699 void UYVYToUV422Row_NEON(const uint8* src_uyvy, 1700 uint8* dst_u, uint8* dst_v, int width); 1701 1702 void UYVYToYRow_C(const uint8* src_uyvy, uint8* dst_y, int width); 1703 void UYVYToUVRow_C(const uint8* src_uyvy, int stride_uyvy, 1704 uint8* dst_u, uint8* dst_v, int width); 1705 void UYVYToUV422Row_C(const uint8* src_uyvy, 1706 uint8* dst_u, uint8* dst_v, int width); 1707 void UYVYToYRow_Any_AVX2(const uint8* src_uyvy, uint8* dst_y, int width); 1708 void UYVYToUVRow_Any_AVX2(const uint8* src_uyvy, int stride_uyvy, 1709 uint8* dst_u, uint8* dst_v, int width); 1710 void UYVYToUV422Row_Any_AVX2(const uint8* src_uyvy, 1711 uint8* dst_u, uint8* dst_v, int width); 1712 void UYVYToYRow_Any_SSE2(const uint8* src_uyvy, uint8* dst_y, int width); 1713 void UYVYToUVRow_Any_SSE2(const uint8* src_uyvy, int stride_uyvy, 1714 uint8* dst_u, uint8* dst_v, int width); 1715 void UYVYToUV422Row_Any_SSE2(const uint8* src_uyvy, 1716 uint8* dst_u, uint8* dst_v, int width); 1717 void UYVYToYRow_Any_NEON(const uint8* src_uyvy, uint8* dst_y, int width); 1718 void UYVYToUVRow_Any_NEON(const uint8* src_uyvy, int stride_uyvy, 1719 uint8* dst_u, uint8* dst_v, int width); 1720 void UYVYToUV422Row_Any_NEON(const uint8* src_uyvy, 1721 uint8* dst_u, uint8* dst_v, int width); 1722 1723 void I422ToYUY2Row_C(const uint8* src_y, 1724 const uint8* src_u, 1725 const uint8* src_v, 1726 uint8* dst_yuy2, int width); 1727 void I422ToUYVYRow_C(const uint8* src_y, 1728 const uint8* src_u, 1729 const uint8* src_v, 1730 uint8* dst_uyvy, int width); 1731 void I422ToYUY2Row_SSE2(const uint8* src_y, 1732 const uint8* src_u, 1733 const uint8* src_v, 1734 uint8* dst_yuy2, int width); 1735 void I422ToUYVYRow_SSE2(const uint8* src_y, 1736 const uint8* src_u, 1737 const uint8* src_v, 1738 uint8* dst_uyvy, int width); 1739 void I422ToYUY2Row_Any_SSE2(const uint8* src_y, 1740 const uint8* src_u, 1741 const uint8* src_v, 1742 uint8* dst_yuy2, int width); 1743 void I422ToUYVYRow_Any_SSE2(const uint8* src_y, 1744 const uint8* src_u, 1745 const uint8* src_v, 1746 uint8* dst_uyvy, int width); 1747 void I422ToYUY2Row_NEON(const uint8* src_y, 1748 const uint8* src_u, 1749 const uint8* src_v, 1750 uint8* dst_yuy2, int width); 1751 void I422ToUYVYRow_NEON(const uint8* src_y, 1752 const uint8* src_u, 1753 const uint8* src_v, 1754 uint8* dst_uyvy, int width); 1755 void I422ToYUY2Row_Any_NEON(const uint8* src_y, 1756 const uint8* src_u, 1757 const uint8* src_v, 1758 uint8* dst_yuy2, int width); 1759 void I422ToUYVYRow_Any_NEON(const uint8* src_y, 1760 const uint8* src_u, 1761 const uint8* src_v, 1762 uint8* dst_uyvy, int width); 1763 1764 // Effects related row functions. 1765 void ARGBAttenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width); 1766 void ARGBAttenuateRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width); 1767 void ARGBAttenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width); 1768 void ARGBAttenuateRow_NEON(const uint8* src_argb, uint8* dst_argb, int width); 1769 void ARGBAttenuateRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb, 1770 int width); 1771 void ARGBAttenuateRow_Any_SSSE3(const uint8* src_argb, uint8* dst_argb, 1772 int width); 1773 void ARGBAttenuateRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb, 1774 int width); 1775 void ARGBAttenuateRow_Any_NEON(const uint8* src_argb, uint8* dst_argb, 1776 int width); 1777 1778 // Inverse table for unattenuate, shared by C and SSE2. 1779 extern const uint32 fixed_invtbl8[256]; 1780 void ARGBUnattenuateRow_C(const uint8* src_argb, uint8* dst_argb, int width); 1781 void ARGBUnattenuateRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width); 1782 void ARGBUnattenuateRow_AVX2(const uint8* src_argb, uint8* dst_argb, int width); 1783 void ARGBUnattenuateRow_Any_SSE2(const uint8* src_argb, uint8* dst_argb, 1784 int width); 1785 void ARGBUnattenuateRow_Any_AVX2(const uint8* src_argb, uint8* dst_argb, 1786 int width); 1787 1788 void ARGBGrayRow_C(const uint8* src_argb, uint8* dst_argb, int width); 1789 void ARGBGrayRow_SSSE3(const uint8* src_argb, uint8* dst_argb, int width); 1790 void ARGBGrayRow_NEON(const uint8* src_argb, uint8* dst_argb, int width); 1791 1792 void ARGBSepiaRow_C(uint8* dst_argb, int width); 1793 void ARGBSepiaRow_SSSE3(uint8* dst_argb, int width); 1794 void ARGBSepiaRow_NEON(uint8* dst_argb, int width); 1795 1796 void ARGBColorMatrixRow_C(const uint8* src_argb, uint8* dst_argb, 1797 const int8* matrix_argb, int width); 1798 void ARGBColorMatrixRow_SSSE3(const uint8* src_argb, uint8* dst_argb, 1799 const int8* matrix_argb, int width); 1800 void ARGBColorMatrixRow_NEON(const uint8* src_argb, uint8* dst_argb, 1801 const int8* matrix_argb, int width); 1802 1803 void ARGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width); 1804 void ARGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width); 1805 1806 void RGBColorTableRow_C(uint8* dst_argb, const uint8* table_argb, int width); 1807 void RGBColorTableRow_X86(uint8* dst_argb, const uint8* table_argb, int width); 1808 1809 void ARGBQuantizeRow_C(uint8* dst_argb, int scale, int interval_size, 1810 int interval_offset, int width); 1811 void ARGBQuantizeRow_SSE2(uint8* dst_argb, int scale, int interval_size, 1812 int interval_offset, int width); 1813 void ARGBQuantizeRow_NEON(uint8* dst_argb, int scale, int interval_size, 1814 int interval_offset, int width); 1815 1816 void ARGBShadeRow_C(const uint8* src_argb, uint8* dst_argb, int width, 1817 uint32 value); 1818 void ARGBShadeRow_SSE2(const uint8* src_argb, uint8* dst_argb, int width, 1819 uint32 value); 1820 void ARGBShadeRow_NEON(const uint8* src_argb, uint8* dst_argb, int width, 1821 uint32 value); 1822 1823 // Used for blur. 1824 void CumulativeSumToAverageRow_SSE2(const int32* topleft, const int32* botleft, 1825 int width, int area, uint8* dst, int count); 1826 void ComputeCumulativeSumRow_SSE2(const uint8* row, int32* cumsum, 1827 const int32* previous_cumsum, int width); 1828 1829 void CumulativeSumToAverageRow_C(const int32* topleft, const int32* botleft, 1830 int width, int area, uint8* dst, int count); 1831 void ComputeCumulativeSumRow_C(const uint8* row, int32* cumsum, 1832 const int32* previous_cumsum, int width); 1833 1834 LIBYUV_API 1835 void ARGBAffineRow_C(const uint8* src_argb, int src_argb_stride, 1836 uint8* dst_argb, const float* uv_dudv, int width); 1837 LIBYUV_API 1838 void ARGBAffineRow_SSE2(const uint8* src_argb, int src_argb_stride, 1839 uint8* dst_argb, const float* uv_dudv, int width); 1840 1841 // Used for I420Scale, ARGBScale, and ARGBInterpolate. 1842 void InterpolateRow_C(uint8* dst_ptr, const uint8* src_ptr, 1843 ptrdiff_t src_stride_ptr, 1844 int width, int source_y_fraction); 1845 void InterpolateRow_SSSE3(uint8* dst_ptr, const uint8* src_ptr, 1846 ptrdiff_t src_stride_ptr, int width, 1847 int source_y_fraction); 1848 void InterpolateRow_AVX2(uint8* dst_ptr, const uint8* src_ptr, 1849 ptrdiff_t src_stride_ptr, int width, 1850 int source_y_fraction); 1851 void InterpolateRow_NEON(uint8* dst_ptr, const uint8* src_ptr, 1852 ptrdiff_t src_stride_ptr, int width, 1853 int source_y_fraction); 1854 void InterpolateRow_DSPR2(uint8* dst_ptr, const uint8* src_ptr, 1855 ptrdiff_t src_stride_ptr, int width, 1856 int source_y_fraction); 1857 void InterpolateRow_Any_NEON(uint8* dst_ptr, const uint8* src_ptr, 1858 ptrdiff_t src_stride_ptr, int width, 1859 int source_y_fraction); 1860 void InterpolateRow_Any_SSSE3(uint8* dst_ptr, const uint8* src_ptr, 1861 ptrdiff_t src_stride_ptr, int width, 1862 int source_y_fraction); 1863 void InterpolateRow_Any_AVX2(uint8* dst_ptr, const uint8* src_ptr, 1864 ptrdiff_t src_stride_ptr, int width, 1865 int source_y_fraction); 1866 void InterpolateRow_Any_DSPR2(uint8* dst_ptr, const uint8* src_ptr, 1867 ptrdiff_t src_stride_ptr, int width, 1868 int source_y_fraction); 1869 1870 void InterpolateRow_16_C(uint16* dst_ptr, const uint16* src_ptr, 1871 ptrdiff_t src_stride_ptr, 1872 int width, int source_y_fraction); 1873 1874 // Sobel images. 1875 void SobelXRow_C(const uint8* src_y0, const uint8* src_y1, const uint8* src_y2, 1876 uint8* dst_sobelx, int width); 1877 void SobelXRow_SSE2(const uint8* src_y0, const uint8* src_y1, 1878 const uint8* src_y2, uint8* dst_sobelx, int width); 1879 void SobelXRow_NEON(const uint8* src_y0, const uint8* src_y1, 1880 const uint8* src_y2, uint8* dst_sobelx, int width); 1881 void SobelYRow_C(const uint8* src_y0, const uint8* src_y1, 1882 uint8* dst_sobely, int width); 1883 void SobelYRow_SSE2(const uint8* src_y0, const uint8* src_y1, 1884 uint8* dst_sobely, int width); 1885 void SobelYRow_NEON(const uint8* src_y0, const uint8* src_y1, 1886 uint8* dst_sobely, int width); 1887 void SobelRow_C(const uint8* src_sobelx, const uint8* src_sobely, 1888 uint8* dst_argb, int width); 1889 void SobelRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely, 1890 uint8* dst_argb, int width); 1891 void SobelRow_NEON(const uint8* src_sobelx, const uint8* src_sobely, 1892 uint8* dst_argb, int width); 1893 void SobelToPlaneRow_C(const uint8* src_sobelx, const uint8* src_sobely, 1894 uint8* dst_y, int width); 1895 void SobelToPlaneRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely, 1896 uint8* dst_y, int width); 1897 void SobelToPlaneRow_NEON(const uint8* src_sobelx, const uint8* src_sobely, 1898 uint8* dst_y, int width); 1899 void SobelXYRow_C(const uint8* src_sobelx, const uint8* src_sobely, 1900 uint8* dst_argb, int width); 1901 void SobelXYRow_SSE2(const uint8* src_sobelx, const uint8* src_sobely, 1902 uint8* dst_argb, int width); 1903 void SobelXYRow_NEON(const uint8* src_sobelx, const uint8* src_sobely, 1904 uint8* dst_argb, int width); 1905 void SobelRow_Any_SSE2(const uint8* src_sobelx, const uint8* src_sobely, 1906 uint8* dst_argb, int width); 1907 void SobelRow_Any_NEON(const uint8* src_sobelx, const uint8* src_sobely, 1908 uint8* dst_argb, int width); 1909 void SobelToPlaneRow_Any_SSE2(const uint8* src_sobelx, const uint8* src_sobely, 1910 uint8* dst_y, int width); 1911 void SobelToPlaneRow_Any_NEON(const uint8* src_sobelx, const uint8* src_sobely, 1912 uint8* dst_y, int width); 1913 void SobelXYRow_Any_SSE2(const uint8* src_sobelx, const uint8* src_sobely, 1914 uint8* dst_argb, int width); 1915 void SobelXYRow_Any_NEON(const uint8* src_sobelx, const uint8* src_sobely, 1916 uint8* dst_argb, int width); 1917 1918 void ARGBPolynomialRow_C(const uint8* src_argb, 1919 uint8* dst_argb, const float* poly, 1920 int width); 1921 void ARGBPolynomialRow_SSE2(const uint8* src_argb, 1922 uint8* dst_argb, const float* poly, 1923 int width); 1924 void ARGBPolynomialRow_AVX2(const uint8* src_argb, 1925 uint8* dst_argb, const float* poly, 1926 int width); 1927 1928 void ARGBLumaColorTableRow_C(const uint8* src_argb, uint8* dst_argb, int width, 1929 const uint8* luma, uint32 lumacoeff); 1930 void ARGBLumaColorTableRow_SSSE3(const uint8* src_argb, uint8* dst_argb, 1931 int width, 1932 const uint8* luma, uint32 lumacoeff); 1933 1934 #ifdef __cplusplus 1935 } // extern "C" 1936 } // namespace libyuv 1937 #endif 1938 1939 #endif // INCLUDE_LIBYUV_ROW_H_ NOLINT 1940
