yuv_sse2.c revision fa39824bb690c5806358871f46940d0450973d8a
1// Copyright 2014 Google Inc. All Rights Reserved. 2// 3// Use of this source code is governed by a BSD-style license 4// that can be found in the COPYING file in the root of the source 5// tree. An additional intellectual property rights grant can be found 6// in the file PATENTS. All contributing project authors may 7// be found in the AUTHORS file in the root of the source tree. 8// ----------------------------------------------------------------------------- 9// 10// YUV->RGB conversion functions 11// 12// Author: Skal (pascal.massimino@gmail.com) 13 14#include "./yuv.h" 15 16#if defined(WEBP_USE_SSE2) 17 18#include "./common_sse2.h" 19#include <stdlib.h> 20#include <emmintrin.h> 21 22//----------------------------------------------------------------------------- 23// Convert spans of 32 pixels to various RGB formats for the fancy upsampler. 24 25// These constants are 14b fixed-point version of ITU-R BT.601 constants. 26// R = (19077 * y + 26149 * v - 14234) >> 6 27// G = (19077 * y - 6419 * u - 13320 * v + 8708) >> 6 28// B = (19077 * y + 33050 * u - 17685) >> 6 29static void ConvertYUV444ToRGB(const __m128i* const Y0, 30 const __m128i* const U0, 31 const __m128i* const V0, 32 __m128i* const R, 33 __m128i* const G, 34 __m128i* const B) { 35 const __m128i k19077 = _mm_set1_epi16(19077); 36 const __m128i k26149 = _mm_set1_epi16(26149); 37 const __m128i k14234 = _mm_set1_epi16(14234); 38 // 33050 doesn't fit in a signed short: only use this with unsigned arithmetic 39 const __m128i k33050 = _mm_set1_epi16((short)33050); 40 const __m128i k17685 = _mm_set1_epi16(17685); 41 const __m128i k6419 = _mm_set1_epi16(6419); 42 const __m128i k13320 = _mm_set1_epi16(13320); 43 const __m128i k8708 = _mm_set1_epi16(8708); 44 45 const __m128i Y1 = _mm_mulhi_epu16(*Y0, k19077); 46 47 const __m128i R0 = _mm_mulhi_epu16(*V0, k26149); 48 const __m128i R1 = _mm_sub_epi16(Y1, k14234); 49 const __m128i R2 = _mm_add_epi16(R1, R0); 50 51 const __m128i G0 = _mm_mulhi_epu16(*U0, k6419); 52 const __m128i G1 = _mm_mulhi_epu16(*V0, k13320); 53 const __m128i G2 = _mm_add_epi16(Y1, k8708); 54 const __m128i G3 = _mm_add_epi16(G0, G1); 55 const __m128i G4 = _mm_sub_epi16(G2, G3); 56 57 // be careful with the saturated *unsigned* arithmetic here! 58 const __m128i B0 = _mm_mulhi_epu16(*U0, k33050); 59 const __m128i B1 = _mm_adds_epu16(B0, Y1); 60 const __m128i B2 = _mm_subs_epu16(B1, k17685); 61 62 // use logical shift for B2, which can be larger than 32767 63 *R = _mm_srai_epi16(R2, 6); // range: [-14234, 30815] 64 *G = _mm_srai_epi16(G4, 6); // range: [-10953, 27710] 65 *B = _mm_srli_epi16(B2, 6); // range: [0, 34238] 66} 67 68// Load the bytes into the *upper* part of 16b words. That's "<< 8", basically. 69static WEBP_INLINE __m128i Load_HI_16(const uint8_t* src) { 70 const __m128i zero = _mm_setzero_si128(); 71 return _mm_unpacklo_epi8(zero, _mm_loadl_epi64((const __m128i*)src)); 72} 73 74// Load and replicate the U/V samples 75static WEBP_INLINE __m128i Load_UV_HI_8(const uint8_t* src) { 76 const __m128i zero = _mm_setzero_si128(); 77 const __m128i tmp0 = _mm_cvtsi32_si128(*(const uint32_t*)src); 78 const __m128i tmp1 = _mm_unpacklo_epi8(zero, tmp0); 79 return _mm_unpacklo_epi16(tmp1, tmp1); // replicate samples 80} 81 82// Convert 32 samples of YUV444 to R/G/B 83static void YUV444ToRGB(const uint8_t* const y, 84 const uint8_t* const u, 85 const uint8_t* const v, 86 __m128i* const R, __m128i* const G, __m128i* const B) { 87 const __m128i Y0 = Load_HI_16(y), U0 = Load_HI_16(u), V0 = Load_HI_16(v); 88 ConvertYUV444ToRGB(&Y0, &U0, &V0, R, G, B); 89} 90 91// Convert 32 samples of YUV420 to R/G/B 92static void YUV420ToRGB(const uint8_t* const y, 93 const uint8_t* const u, 94 const uint8_t* const v, 95 __m128i* const R, __m128i* const G, __m128i* const B) { 96 const __m128i Y0 = Load_HI_16(y), U0 = Load_UV_HI_8(u), V0 = Load_UV_HI_8(v); 97 ConvertYUV444ToRGB(&Y0, &U0, &V0, R, G, B); 98} 99 100// Pack R/G/B/A results into 32b output. 101static WEBP_INLINE void PackAndStore4(const __m128i* const R, 102 const __m128i* const G, 103 const __m128i* const B, 104 const __m128i* const A, 105 uint8_t* const dst) { 106 const __m128i rb = _mm_packus_epi16(*R, *B); 107 const __m128i ga = _mm_packus_epi16(*G, *A); 108 const __m128i rg = _mm_unpacklo_epi8(rb, ga); 109 const __m128i ba = _mm_unpackhi_epi8(rb, ga); 110 const __m128i RGBA_lo = _mm_unpacklo_epi16(rg, ba); 111 const __m128i RGBA_hi = _mm_unpackhi_epi16(rg, ba); 112 _mm_storeu_si128((__m128i*)(dst + 0), RGBA_lo); 113 _mm_storeu_si128((__m128i*)(dst + 16), RGBA_hi); 114} 115 116// Pack R/G/B/A results into 16b output. 117static WEBP_INLINE void PackAndStore4444(const __m128i* const R, 118 const __m128i* const G, 119 const __m128i* const B, 120 const __m128i* const A, 121 uint8_t* const dst) { 122#if !defined(WEBP_SWAP_16BIT_CSP) 123 const __m128i rg0 = _mm_packus_epi16(*R, *G); 124 const __m128i ba0 = _mm_packus_epi16(*B, *A); 125#else 126 const __m128i rg0 = _mm_packus_epi16(*B, *A); 127 const __m128i ba0 = _mm_packus_epi16(*R, *G); 128#endif 129 const __m128i mask_0xf0 = _mm_set1_epi8(0xf0); 130 const __m128i rb1 = _mm_unpacklo_epi8(rg0, ba0); // rbrbrbrbrb... 131 const __m128i ga1 = _mm_unpackhi_epi8(rg0, ba0); // gagagagaga... 132 const __m128i rb2 = _mm_and_si128(rb1, mask_0xf0); 133 const __m128i ga2 = _mm_srli_epi16(_mm_and_si128(ga1, mask_0xf0), 4); 134 const __m128i rgba4444 = _mm_or_si128(rb2, ga2); 135 _mm_storeu_si128((__m128i*)dst, rgba4444); 136} 137 138// Pack R/G/B results into 16b output. 139static WEBP_INLINE void PackAndStore565(const __m128i* const R, 140 const __m128i* const G, 141 const __m128i* const B, 142 uint8_t* const dst) { 143 const __m128i r0 = _mm_packus_epi16(*R, *R); 144 const __m128i g0 = _mm_packus_epi16(*G, *G); 145 const __m128i b0 = _mm_packus_epi16(*B, *B); 146 const __m128i r1 = _mm_and_si128(r0, _mm_set1_epi8(0xf8)); 147 const __m128i b1 = _mm_and_si128(_mm_srli_epi16(b0, 3), _mm_set1_epi8(0x1f)); 148 const __m128i g1 = _mm_srli_epi16(_mm_and_si128(g0, _mm_set1_epi8(0xe0)), 5); 149 const __m128i g2 = _mm_slli_epi16(_mm_and_si128(g0, _mm_set1_epi8(0x1c)), 3); 150 const __m128i rg = _mm_or_si128(r1, g1); 151 const __m128i gb = _mm_or_si128(g2, b1); 152#if !defined(WEBP_SWAP_16BIT_CSP) 153 const __m128i rgb565 = _mm_unpacklo_epi8(rg, gb); 154#else 155 const __m128i rgb565 = _mm_unpacklo_epi8(gb, rg); 156#endif 157 _mm_storeu_si128((__m128i*)dst, rgb565); 158} 159 160// Pack the planar buffers 161// rrrr... rrrr... gggg... gggg... bbbb... bbbb.... 162// triplet by triplet in the output buffer rgb as rgbrgbrgbrgb ... 163static WEBP_INLINE void PlanarTo24b(__m128i* const in0, __m128i* const in1, 164 __m128i* const in2, __m128i* const in3, 165 __m128i* const in4, __m128i* const in5, 166 uint8_t* const rgb) { 167 // The input is 6 registers of sixteen 8b but for the sake of explanation, 168 // let's take 6 registers of four 8b values. 169 // To pack, we will keep taking one every two 8b integer and move it 170 // around as follows: 171 // Input: 172 // r0r1r2r3 | r4r5r6r7 | g0g1g2g3 | g4g5g6g7 | b0b1b2b3 | b4b5b6b7 173 // Split the 6 registers in two sets of 3 registers: the first set as the even 174 // 8b bytes, the second the odd ones: 175 // r0r2r4r6 | g0g2g4g6 | b0b2b4b6 | r1r3r5r7 | g1g3g5g7 | b1b3b5b7 176 // Repeat the same permutations twice more: 177 // r0r4g0g4 | b0b4r1r5 | g1g5b1b5 | r2r6g2g6 | b2b6r3r7 | g3g7b3b7 178 // r0g0b0r1 | g1b1r2g2 | b2r3g3b3 | r4g4b4r5 | g5b5r6g6 | b6r7g7b7 179 VP8PlanarTo24b(in0, in1, in2, in3, in4, in5); 180 181 _mm_storeu_si128((__m128i*)(rgb + 0), *in0); 182 _mm_storeu_si128((__m128i*)(rgb + 16), *in1); 183 _mm_storeu_si128((__m128i*)(rgb + 32), *in2); 184 _mm_storeu_si128((__m128i*)(rgb + 48), *in3); 185 _mm_storeu_si128((__m128i*)(rgb + 64), *in4); 186 _mm_storeu_si128((__m128i*)(rgb + 80), *in5); 187} 188 189void VP8YuvToRgba32(const uint8_t* y, const uint8_t* u, const uint8_t* v, 190 uint8_t* dst) { 191 const __m128i kAlpha = _mm_set1_epi16(255); 192 int n; 193 for (n = 0; n < 32; n += 8, dst += 32) { 194 __m128i R, G, B; 195 YUV444ToRGB(y + n, u + n, v + n, &R, &G, &B); 196 PackAndStore4(&R, &G, &B, &kAlpha, dst); 197 } 198} 199 200void VP8YuvToBgra32(const uint8_t* y, const uint8_t* u, const uint8_t* v, 201 uint8_t* dst) { 202 const __m128i kAlpha = _mm_set1_epi16(255); 203 int n; 204 for (n = 0; n < 32; n += 8, dst += 32) { 205 __m128i R, G, B; 206 YUV444ToRGB(y + n, u + n, v + n, &R, &G, &B); 207 PackAndStore4(&B, &G, &R, &kAlpha, dst); 208 } 209} 210 211void VP8YuvToArgb32(const uint8_t* y, const uint8_t* u, const uint8_t* v, 212 uint8_t* dst) { 213 const __m128i kAlpha = _mm_set1_epi16(255); 214 int n; 215 for (n = 0; n < 32; n += 8, dst += 32) { 216 __m128i R, G, B; 217 YUV444ToRGB(y + n, u + n, v + n, &R, &G, &B); 218 PackAndStore4(&kAlpha, &R, &G, &B, dst); 219 } 220} 221 222void VP8YuvToRgba444432(const uint8_t* y, const uint8_t* u, const uint8_t* v, 223 uint8_t* dst) { 224 const __m128i kAlpha = _mm_set1_epi16(255); 225 int n; 226 for (n = 0; n < 32; n += 8, dst += 16) { 227 __m128i R, G, B; 228 YUV444ToRGB(y + n, u + n, v + n, &R, &G, &B); 229 PackAndStore4444(&R, &G, &B, &kAlpha, dst); 230 } 231} 232 233void VP8YuvToRgb56532(const uint8_t* y, const uint8_t* u, const uint8_t* v, 234 uint8_t* dst) { 235 int n; 236 for (n = 0; n < 32; n += 8, dst += 16) { 237 __m128i R, G, B; 238 YUV444ToRGB(y + n, u + n, v + n, &R, &G, &B); 239 PackAndStore565(&R, &G, &B, dst); 240 } 241} 242 243void VP8YuvToRgb32(const uint8_t* y, const uint8_t* u, const uint8_t* v, 244 uint8_t* dst) { 245 __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3; 246 __m128i rgb0, rgb1, rgb2, rgb3, rgb4, rgb5; 247 248 YUV444ToRGB(y + 0, u + 0, v + 0, &R0, &G0, &B0); 249 YUV444ToRGB(y + 8, u + 8, v + 8, &R1, &G1, &B1); 250 YUV444ToRGB(y + 16, u + 16, v + 16, &R2, &G2, &B2); 251 YUV444ToRGB(y + 24, u + 24, v + 24, &R3, &G3, &B3); 252 253 // Cast to 8b and store as RRRRGGGGBBBB. 254 rgb0 = _mm_packus_epi16(R0, R1); 255 rgb1 = _mm_packus_epi16(R2, R3); 256 rgb2 = _mm_packus_epi16(G0, G1); 257 rgb3 = _mm_packus_epi16(G2, G3); 258 rgb4 = _mm_packus_epi16(B0, B1); 259 rgb5 = _mm_packus_epi16(B2, B3); 260 261 // Pack as RGBRGBRGBRGB. 262 PlanarTo24b(&rgb0, &rgb1, &rgb2, &rgb3, &rgb4, &rgb5, dst); 263} 264 265void VP8YuvToBgr32(const uint8_t* y, const uint8_t* u, const uint8_t* v, 266 uint8_t* dst) { 267 __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3; 268 __m128i bgr0, bgr1, bgr2, bgr3, bgr4, bgr5; 269 270 YUV444ToRGB(y + 0, u + 0, v + 0, &R0, &G0, &B0); 271 YUV444ToRGB(y + 8, u + 8, v + 8, &R1, &G1, &B1); 272 YUV444ToRGB(y + 16, u + 16, v + 16, &R2, &G2, &B2); 273 YUV444ToRGB(y + 24, u + 24, v + 24, &R3, &G3, &B3); 274 275 // Cast to 8b and store as BBBBGGGGRRRR. 276 bgr0 = _mm_packus_epi16(B0, B1); 277 bgr1 = _mm_packus_epi16(B2, B3); 278 bgr2 = _mm_packus_epi16(G0, G1); 279 bgr3 = _mm_packus_epi16(G2, G3); 280 bgr4 = _mm_packus_epi16(R0, R1); 281 bgr5= _mm_packus_epi16(R2, R3); 282 283 // Pack as BGRBGRBGRBGR. 284 PlanarTo24b(&bgr0, &bgr1, &bgr2, &bgr3, &bgr4, &bgr5, dst); 285} 286 287//----------------------------------------------------------------------------- 288// Arbitrary-length row conversion functions 289 290static void YuvToRgbaRow(const uint8_t* y, const uint8_t* u, const uint8_t* v, 291 uint8_t* dst, int len) { 292 const __m128i kAlpha = _mm_set1_epi16(255); 293 int n; 294 for (n = 0; n + 8 <= len; n += 8, dst += 32) { 295 __m128i R, G, B; 296 YUV420ToRGB(y, u, v, &R, &G, &B); 297 PackAndStore4(&R, &G, &B, &kAlpha, dst); 298 y += 8; 299 u += 4; 300 v += 4; 301 } 302 for (; n < len; ++n) { // Finish off 303 VP8YuvToRgba(y[0], u[0], v[0], dst); 304 dst += 4; 305 y += 1; 306 u += (n & 1); 307 v += (n & 1); 308 } 309} 310 311static void YuvToBgraRow(const uint8_t* y, const uint8_t* u, const uint8_t* v, 312 uint8_t* dst, int len) { 313 const __m128i kAlpha = _mm_set1_epi16(255); 314 int n; 315 for (n = 0; n + 8 <= len; n += 8, dst += 32) { 316 __m128i R, G, B; 317 YUV420ToRGB(y, u, v, &R, &G, &B); 318 PackAndStore4(&B, &G, &R, &kAlpha, dst); 319 y += 8; 320 u += 4; 321 v += 4; 322 } 323 for (; n < len; ++n) { // Finish off 324 VP8YuvToBgra(y[0], u[0], v[0], dst); 325 dst += 4; 326 y += 1; 327 u += (n & 1); 328 v += (n & 1); 329 } 330} 331 332static void YuvToArgbRow(const uint8_t* y, const uint8_t* u, const uint8_t* v, 333 uint8_t* dst, int len) { 334 const __m128i kAlpha = _mm_set1_epi16(255); 335 int n; 336 for (n = 0; n + 8 <= len; n += 8, dst += 32) { 337 __m128i R, G, B; 338 YUV420ToRGB(y, u, v, &R, &G, &B); 339 PackAndStore4(&kAlpha, &R, &G, &B, dst); 340 y += 8; 341 u += 4; 342 v += 4; 343 } 344 for (; n < len; ++n) { // Finish off 345 VP8YuvToArgb(y[0], u[0], v[0], dst); 346 dst += 4; 347 y += 1; 348 u += (n & 1); 349 v += (n & 1); 350 } 351} 352 353static void YuvToRgbRow(const uint8_t* y, const uint8_t* u, const uint8_t* v, 354 uint8_t* dst, int len) { 355 int n; 356 for (n = 0; n + 32 <= len; n += 32, dst += 32 * 3) { 357 __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3; 358 __m128i rgb0, rgb1, rgb2, rgb3, rgb4, rgb5; 359 360 YUV420ToRGB(y + 0, u + 0, v + 0, &R0, &G0, &B0); 361 YUV420ToRGB(y + 8, u + 4, v + 4, &R1, &G1, &B1); 362 YUV420ToRGB(y + 16, u + 8, v + 8, &R2, &G2, &B2); 363 YUV420ToRGB(y + 24, u + 12, v + 12, &R3, &G3, &B3); 364 365 // Cast to 8b and store as RRRRGGGGBBBB. 366 rgb0 = _mm_packus_epi16(R0, R1); 367 rgb1 = _mm_packus_epi16(R2, R3); 368 rgb2 = _mm_packus_epi16(G0, G1); 369 rgb3 = _mm_packus_epi16(G2, G3); 370 rgb4 = _mm_packus_epi16(B0, B1); 371 rgb5 = _mm_packus_epi16(B2, B3); 372 373 // Pack as RGBRGBRGBRGB. 374 PlanarTo24b(&rgb0, &rgb1, &rgb2, &rgb3, &rgb4, &rgb5, dst); 375 376 y += 32; 377 u += 16; 378 v += 16; 379 } 380 for (; n < len; ++n) { // Finish off 381 VP8YuvToRgb(y[0], u[0], v[0], dst); 382 dst += 3; 383 y += 1; 384 u += (n & 1); 385 v += (n & 1); 386 } 387} 388 389static void YuvToBgrRow(const uint8_t* y, const uint8_t* u, const uint8_t* v, 390 uint8_t* dst, int len) { 391 int n; 392 for (n = 0; n + 32 <= len; n += 32, dst += 32 * 3) { 393 __m128i R0, R1, R2, R3, G0, G1, G2, G3, B0, B1, B2, B3; 394 __m128i bgr0, bgr1, bgr2, bgr3, bgr4, bgr5; 395 396 YUV420ToRGB(y + 0, u + 0, v + 0, &R0, &G0, &B0); 397 YUV420ToRGB(y + 8, u + 4, v + 4, &R1, &G1, &B1); 398 YUV420ToRGB(y + 16, u + 8, v + 8, &R2, &G2, &B2); 399 YUV420ToRGB(y + 24, u + 12, v + 12, &R3, &G3, &B3); 400 401 // Cast to 8b and store as BBBBGGGGRRRR. 402 bgr0 = _mm_packus_epi16(B0, B1); 403 bgr1 = _mm_packus_epi16(B2, B3); 404 bgr2 = _mm_packus_epi16(G0, G1); 405 bgr3 = _mm_packus_epi16(G2, G3); 406 bgr4 = _mm_packus_epi16(R0, R1); 407 bgr5 = _mm_packus_epi16(R2, R3); 408 409 // Pack as BGRBGRBGRBGR. 410 PlanarTo24b(&bgr0, &bgr1, &bgr2, &bgr3, &bgr4, &bgr5, dst); 411 412 y += 32; 413 u += 16; 414 v += 16; 415 } 416 for (; n < len; ++n) { // Finish off 417 VP8YuvToBgr(y[0], u[0], v[0], dst); 418 dst += 3; 419 y += 1; 420 u += (n & 1); 421 v += (n & 1); 422 } 423} 424 425//------------------------------------------------------------------------------ 426// Entry point 427 428extern void WebPInitSamplersSSE2(void); 429 430WEBP_TSAN_IGNORE_FUNCTION void WebPInitSamplersSSE2(void) { 431 WebPSamplers[MODE_RGB] = YuvToRgbRow; 432 WebPSamplers[MODE_RGBA] = YuvToRgbaRow; 433 WebPSamplers[MODE_BGR] = YuvToBgrRow; 434 WebPSamplers[MODE_BGRA] = YuvToBgraRow; 435 WebPSamplers[MODE_ARGB] = YuvToArgbRow; 436} 437 438//------------------------------------------------------------------------------ 439// RGB24/32 -> YUV converters 440 441// Load eight 16b-words from *src. 442#define LOAD_16(src) _mm_loadu_si128((const __m128i*)(src)) 443// Store either 16b-words into *dst 444#define STORE_16(V, dst) _mm_storeu_si128((__m128i*)(dst), (V)) 445 446// Function that inserts a value of the second half of the in buffer in between 447// every two char of the first half. 448static WEBP_INLINE void RGB24PackedToPlanarHelper( 449 const __m128i* const in /*in[6]*/, __m128i* const out /*out[6]*/) { 450 out[0] = _mm_unpacklo_epi8(in[0], in[3]); 451 out[1] = _mm_unpackhi_epi8(in[0], in[3]); 452 out[2] = _mm_unpacklo_epi8(in[1], in[4]); 453 out[3] = _mm_unpackhi_epi8(in[1], in[4]); 454 out[4] = _mm_unpacklo_epi8(in[2], in[5]); 455 out[5] = _mm_unpackhi_epi8(in[2], in[5]); 456} 457 458// Unpack the 8b input rgbrgbrgbrgb ... as contiguous registers: 459// rrrr... rrrr... gggg... gggg... bbbb... bbbb.... 460// Similar to PlanarTo24bHelper(), but in reverse order. 461static WEBP_INLINE void RGB24PackedToPlanar(const uint8_t* const rgb, 462 __m128i* const out /*out[6]*/) { 463 __m128i tmp[6]; 464 tmp[0] = _mm_loadu_si128((const __m128i*)(rgb + 0)); 465 tmp[1] = _mm_loadu_si128((const __m128i*)(rgb + 16)); 466 tmp[2] = _mm_loadu_si128((const __m128i*)(rgb + 32)); 467 tmp[3] = _mm_loadu_si128((const __m128i*)(rgb + 48)); 468 tmp[4] = _mm_loadu_si128((const __m128i*)(rgb + 64)); 469 tmp[5] = _mm_loadu_si128((const __m128i*)(rgb + 80)); 470 471 RGB24PackedToPlanarHelper(tmp, out); 472 RGB24PackedToPlanarHelper(out, tmp); 473 RGB24PackedToPlanarHelper(tmp, out); 474 RGB24PackedToPlanarHelper(out, tmp); 475 RGB24PackedToPlanarHelper(tmp, out); 476} 477 478// Convert 8 packed ARGB to r[], g[], b[] 479static WEBP_INLINE void RGB32PackedToPlanar(const uint32_t* const argb, 480 __m128i* const rgb /*in[6]*/) { 481 const __m128i zero = _mm_setzero_si128(); 482 __m128i a0 = LOAD_16(argb + 0); 483 __m128i a1 = LOAD_16(argb + 4); 484 __m128i a2 = LOAD_16(argb + 8); 485 __m128i a3 = LOAD_16(argb + 12); 486 VP8L32bToPlanar(&a0, &a1, &a2, &a3); 487 rgb[0] = _mm_unpacklo_epi8(a1, zero); 488 rgb[1] = _mm_unpackhi_epi8(a1, zero); 489 rgb[2] = _mm_unpacklo_epi8(a2, zero); 490 rgb[3] = _mm_unpackhi_epi8(a2, zero); 491 rgb[4] = _mm_unpacklo_epi8(a3, zero); 492 rgb[5] = _mm_unpackhi_epi8(a3, zero); 493} 494 495// This macro computes (RG * MULT_RG + GB * MULT_GB + ROUNDER) >> DESCALE_FIX 496// It's a macro and not a function because we need to use immediate values with 497// srai_epi32, e.g. 498#define TRANSFORM(RG_LO, RG_HI, GB_LO, GB_HI, MULT_RG, MULT_GB, \ 499 ROUNDER, DESCALE_FIX, OUT) do { \ 500 const __m128i V0_lo = _mm_madd_epi16(RG_LO, MULT_RG); \ 501 const __m128i V0_hi = _mm_madd_epi16(RG_HI, MULT_RG); \ 502 const __m128i V1_lo = _mm_madd_epi16(GB_LO, MULT_GB); \ 503 const __m128i V1_hi = _mm_madd_epi16(GB_HI, MULT_GB); \ 504 const __m128i V2_lo = _mm_add_epi32(V0_lo, V1_lo); \ 505 const __m128i V2_hi = _mm_add_epi32(V0_hi, V1_hi); \ 506 const __m128i V3_lo = _mm_add_epi32(V2_lo, ROUNDER); \ 507 const __m128i V3_hi = _mm_add_epi32(V2_hi, ROUNDER); \ 508 const __m128i V5_lo = _mm_srai_epi32(V3_lo, DESCALE_FIX); \ 509 const __m128i V5_hi = _mm_srai_epi32(V3_hi, DESCALE_FIX); \ 510 (OUT) = _mm_packs_epi32(V5_lo, V5_hi); \ 511} while (0) 512 513#define MK_CST_16(A, B) _mm_set_epi16((B), (A), (B), (A), (B), (A), (B), (A)) 514static WEBP_INLINE void ConvertRGBToY(const __m128i* const R, 515 const __m128i* const G, 516 const __m128i* const B, 517 __m128i* const Y) { 518 const __m128i kRG_y = MK_CST_16(16839, 33059 - 16384); 519 const __m128i kGB_y = MK_CST_16(16384, 6420); 520 const __m128i kHALF_Y = _mm_set1_epi32((16 << YUV_FIX) + YUV_HALF); 521 522 const __m128i RG_lo = _mm_unpacklo_epi16(*R, *G); 523 const __m128i RG_hi = _mm_unpackhi_epi16(*R, *G); 524 const __m128i GB_lo = _mm_unpacklo_epi16(*G, *B); 525 const __m128i GB_hi = _mm_unpackhi_epi16(*G, *B); 526 TRANSFORM(RG_lo, RG_hi, GB_lo, GB_hi, kRG_y, kGB_y, kHALF_Y, YUV_FIX, *Y); 527} 528 529static WEBP_INLINE void ConvertRGBToUV(const __m128i* const R, 530 const __m128i* const G, 531 const __m128i* const B, 532 __m128i* const U, __m128i* const V) { 533 const __m128i kRG_u = MK_CST_16(-9719, -19081); 534 const __m128i kGB_u = MK_CST_16(0, 28800); 535 const __m128i kRG_v = MK_CST_16(28800, 0); 536 const __m128i kGB_v = MK_CST_16(-24116, -4684); 537 const __m128i kHALF_UV = _mm_set1_epi32(((128 << YUV_FIX) + YUV_HALF) << 2); 538 539 const __m128i RG_lo = _mm_unpacklo_epi16(*R, *G); 540 const __m128i RG_hi = _mm_unpackhi_epi16(*R, *G); 541 const __m128i GB_lo = _mm_unpacklo_epi16(*G, *B); 542 const __m128i GB_hi = _mm_unpackhi_epi16(*G, *B); 543 TRANSFORM(RG_lo, RG_hi, GB_lo, GB_hi, kRG_u, kGB_u, 544 kHALF_UV, YUV_FIX + 2, *U); 545 TRANSFORM(RG_lo, RG_hi, GB_lo, GB_hi, kRG_v, kGB_v, 546 kHALF_UV, YUV_FIX + 2, *V); 547} 548 549#undef MK_CST_16 550#undef TRANSFORM 551 552static void ConvertRGB24ToY(const uint8_t* rgb, uint8_t* y, int width) { 553 const int max_width = width & ~31; 554 int i; 555 for (i = 0; i < max_width; rgb += 3 * 16 * 2) { 556 __m128i rgb_plane[6]; 557 int j; 558 559 RGB24PackedToPlanar(rgb, rgb_plane); 560 561 for (j = 0; j < 2; ++j, i += 16) { 562 const __m128i zero = _mm_setzero_si128(); 563 __m128i r, g, b, Y0, Y1; 564 565 // Convert to 16-bit Y. 566 r = _mm_unpacklo_epi8(rgb_plane[0 + j], zero); 567 g = _mm_unpacklo_epi8(rgb_plane[2 + j], zero); 568 b = _mm_unpacklo_epi8(rgb_plane[4 + j], zero); 569 ConvertRGBToY(&r, &g, &b, &Y0); 570 571 // Convert to 16-bit Y. 572 r = _mm_unpackhi_epi8(rgb_plane[0 + j], zero); 573 g = _mm_unpackhi_epi8(rgb_plane[2 + j], zero); 574 b = _mm_unpackhi_epi8(rgb_plane[4 + j], zero); 575 ConvertRGBToY(&r, &g, &b, &Y1); 576 577 // Cast to 8-bit and store. 578 STORE_16(_mm_packus_epi16(Y0, Y1), y + i); 579 } 580 } 581 for (; i < width; ++i, rgb += 3) { // left-over 582 y[i] = VP8RGBToY(rgb[0], rgb[1], rgb[2], YUV_HALF); 583 } 584} 585 586static void ConvertBGR24ToY(const uint8_t* bgr, uint8_t* y, int width) { 587 const int max_width = width & ~31; 588 int i; 589 for (i = 0; i < max_width; bgr += 3 * 16 * 2) { 590 __m128i bgr_plane[6]; 591 int j; 592 593 RGB24PackedToPlanar(bgr, bgr_plane); 594 595 for (j = 0; j < 2; ++j, i += 16) { 596 const __m128i zero = _mm_setzero_si128(); 597 __m128i r, g, b, Y0, Y1; 598 599 // Convert to 16-bit Y. 600 b = _mm_unpacklo_epi8(bgr_plane[0 + j], zero); 601 g = _mm_unpacklo_epi8(bgr_plane[2 + j], zero); 602 r = _mm_unpacklo_epi8(bgr_plane[4 + j], zero); 603 ConvertRGBToY(&r, &g, &b, &Y0); 604 605 // Convert to 16-bit Y. 606 b = _mm_unpackhi_epi8(bgr_plane[0 + j], zero); 607 g = _mm_unpackhi_epi8(bgr_plane[2 + j], zero); 608 r = _mm_unpackhi_epi8(bgr_plane[4 + j], zero); 609 ConvertRGBToY(&r, &g, &b, &Y1); 610 611 // Cast to 8-bit and store. 612 STORE_16(_mm_packus_epi16(Y0, Y1), y + i); 613 } 614 } 615 for (; i < width; ++i, bgr += 3) { // left-over 616 y[i] = VP8RGBToY(bgr[2], bgr[1], bgr[0], YUV_HALF); 617 } 618} 619 620static void ConvertARGBToY(const uint32_t* argb, uint8_t* y, int width) { 621 const int max_width = width & ~15; 622 int i; 623 for (i = 0; i < max_width; i += 16) { 624 __m128i Y0, Y1, rgb[6]; 625 RGB32PackedToPlanar(&argb[i], rgb); 626 ConvertRGBToY(&rgb[0], &rgb[2], &rgb[4], &Y0); 627 ConvertRGBToY(&rgb[1], &rgb[3], &rgb[5], &Y1); 628 STORE_16(_mm_packus_epi16(Y0, Y1), y + i); 629 } 630 for (; i < width; ++i) { // left-over 631 const uint32_t p = argb[i]; 632 y[i] = VP8RGBToY((p >> 16) & 0xff, (p >> 8) & 0xff, (p >> 0) & 0xff, 633 YUV_HALF); 634 } 635} 636 637// Horizontal add (doubled) of two 16b values, result is 16b. 638// in: A | B | C | D | ... -> out: 2*(A+B) | 2*(C+D) | ... 639static void HorizontalAddPack(const __m128i* const A, const __m128i* const B, 640 __m128i* const out) { 641 const __m128i k2 = _mm_set1_epi16(2); 642 const __m128i C = _mm_madd_epi16(*A, k2); 643 const __m128i D = _mm_madd_epi16(*B, k2); 644 *out = _mm_packs_epi32(C, D); 645} 646 647static void ConvertARGBToUV(const uint32_t* argb, uint8_t* u, uint8_t* v, 648 int src_width, int do_store) { 649 const int max_width = src_width & ~31; 650 int i; 651 for (i = 0; i < max_width; i += 32, u += 16, v += 16) { 652 __m128i rgb[6], U0, V0, U1, V1; 653 RGB32PackedToPlanar(&argb[i], rgb); 654 HorizontalAddPack(&rgb[0], &rgb[1], &rgb[0]); 655 HorizontalAddPack(&rgb[2], &rgb[3], &rgb[2]); 656 HorizontalAddPack(&rgb[4], &rgb[5], &rgb[4]); 657 ConvertRGBToUV(&rgb[0], &rgb[2], &rgb[4], &U0, &V0); 658 659 RGB32PackedToPlanar(&argb[i + 16], rgb); 660 HorizontalAddPack(&rgb[0], &rgb[1], &rgb[0]); 661 HorizontalAddPack(&rgb[2], &rgb[3], &rgb[2]); 662 HorizontalAddPack(&rgb[4], &rgb[5], &rgb[4]); 663 ConvertRGBToUV(&rgb[0], &rgb[2], &rgb[4], &U1, &V1); 664 665 U0 = _mm_packus_epi16(U0, U1); 666 V0 = _mm_packus_epi16(V0, V1); 667 if (!do_store) { 668 const __m128i prev_u = LOAD_16(u); 669 const __m128i prev_v = LOAD_16(v); 670 U0 = _mm_avg_epu8(U0, prev_u); 671 V0 = _mm_avg_epu8(V0, prev_v); 672 } 673 STORE_16(U0, u); 674 STORE_16(V0, v); 675 } 676 if (i < src_width) { // left-over 677 WebPConvertARGBToUV_C(argb + i, u, v, src_width - i, do_store); 678 } 679} 680 681// Convert 16 packed ARGB 16b-values to r[], g[], b[] 682static WEBP_INLINE void RGBA32PackedToPlanar_16b(const uint16_t* const rgbx, 683 __m128i* const r, 684 __m128i* const g, 685 __m128i* const b) { 686 const __m128i in0 = LOAD_16(rgbx + 0); // r0 | g0 | b0 |x| r1 | g1 | b1 |x 687 const __m128i in1 = LOAD_16(rgbx + 8); // r2 | g2 | b2 |x| r3 | g3 | b3 |x 688 const __m128i in2 = LOAD_16(rgbx + 16); // r4 | ... 689 const __m128i in3 = LOAD_16(rgbx + 24); // r6 | ... 690 // column-wise transpose 691 const __m128i A0 = _mm_unpacklo_epi16(in0, in1); 692 const __m128i A1 = _mm_unpackhi_epi16(in0, in1); 693 const __m128i A2 = _mm_unpacklo_epi16(in2, in3); 694 const __m128i A3 = _mm_unpackhi_epi16(in2, in3); 695 const __m128i B0 = _mm_unpacklo_epi16(A0, A1); // r0 r1 r2 r3 | g0 g1 .. 696 const __m128i B1 = _mm_unpackhi_epi16(A0, A1); // b0 b1 b2 b3 | x x x x 697 const __m128i B2 = _mm_unpacklo_epi16(A2, A3); // r4 r5 r6 r7 | g4 g5 .. 698 const __m128i B3 = _mm_unpackhi_epi16(A2, A3); // b4 b5 b6 b7 | x x x x 699 *r = _mm_unpacklo_epi64(B0, B2); 700 *g = _mm_unpackhi_epi64(B0, B2); 701 *b = _mm_unpacklo_epi64(B1, B3); 702} 703 704static void ConvertRGBA32ToUV(const uint16_t* rgb, 705 uint8_t* u, uint8_t* v, int width) { 706 const int max_width = width & ~15; 707 const uint16_t* const last_rgb = rgb + 4 * max_width; 708 while (rgb < last_rgb) { 709 __m128i r, g, b, U0, V0, U1, V1; 710 RGBA32PackedToPlanar_16b(rgb + 0, &r, &g, &b); 711 ConvertRGBToUV(&r, &g, &b, &U0, &V0); 712 RGBA32PackedToPlanar_16b(rgb + 32, &r, &g, &b); 713 ConvertRGBToUV(&r, &g, &b, &U1, &V1); 714 STORE_16(_mm_packus_epi16(U0, U1), u); 715 STORE_16(_mm_packus_epi16(V0, V1), v); 716 u += 16; 717 v += 16; 718 rgb += 2 * 32; 719 } 720 if (max_width < width) { // left-over 721 WebPConvertRGBA32ToUV_C(rgb, u, v, width - max_width); 722 } 723} 724 725//------------------------------------------------------------------------------ 726 727extern void WebPInitConvertARGBToYUVSSE2(void); 728 729WEBP_TSAN_IGNORE_FUNCTION void WebPInitConvertARGBToYUVSSE2(void) { 730 WebPConvertARGBToY = ConvertARGBToY; 731 WebPConvertARGBToUV = ConvertARGBToUV; 732 733 WebPConvertRGB24ToY = ConvertRGB24ToY; 734 WebPConvertBGR24ToY = ConvertBGR24ToY; 735 736 WebPConvertRGBA32ToUV = ConvertRGBA32ToUV; 737} 738 739//------------------------------------------------------------------------------ 740 741#define MAX_Y ((1 << 10) - 1) // 10b precision over 16b-arithmetic 742static uint16_t clip_y(int v) { 743 return (v < 0) ? 0 : (v > MAX_Y) ? MAX_Y : (uint16_t)v; 744} 745 746static uint64_t SharpYUVUpdateY_SSE2(const uint16_t* ref, const uint16_t* src, 747 uint16_t* dst, int len) { 748 uint64_t diff = 0; 749 uint32_t tmp[4]; 750 int i; 751 const __m128i zero = _mm_setzero_si128(); 752 const __m128i max = _mm_set1_epi16(MAX_Y); 753 const __m128i one = _mm_set1_epi16(1); 754 __m128i sum = zero; 755 756 for (i = 0; i + 8 <= len; i += 8) { 757 const __m128i A = _mm_loadu_si128((const __m128i*)(ref + i)); 758 const __m128i B = _mm_loadu_si128((const __m128i*)(src + i)); 759 const __m128i C = _mm_loadu_si128((const __m128i*)(dst + i)); 760 const __m128i D = _mm_sub_epi16(A, B); // diff_y 761 const __m128i E = _mm_cmpgt_epi16(zero, D); // sign (-1 or 0) 762 const __m128i F = _mm_add_epi16(C, D); // new_y 763 const __m128i G = _mm_or_si128(E, one); // -1 or 1 764 const __m128i H = _mm_max_epi16(_mm_min_epi16(F, max), zero); 765 const __m128i I = _mm_madd_epi16(D, G); // sum(abs(...)) 766 _mm_storeu_si128((__m128i*)(dst + i), H); 767 sum = _mm_add_epi32(sum, I); 768 } 769 _mm_storeu_si128((__m128i*)tmp, sum); 770 diff = tmp[3] + tmp[2] + tmp[1] + tmp[0]; 771 for (; i < len; ++i) { 772 const int diff_y = ref[i] - src[i]; 773 const int new_y = (int)dst[i] + diff_y; 774 dst[i] = clip_y(new_y); 775 diff += (uint64_t)abs(diff_y); 776 } 777 return diff; 778} 779 780static void SharpYUVUpdateRGB_SSE2(const int16_t* ref, const int16_t* src, 781 int16_t* dst, int len) { 782 int i = 0; 783 for (i = 0; i + 8 <= len; i += 8) { 784 const __m128i A = _mm_loadu_si128((const __m128i*)(ref + i)); 785 const __m128i B = _mm_loadu_si128((const __m128i*)(src + i)); 786 const __m128i C = _mm_loadu_si128((const __m128i*)(dst + i)); 787 const __m128i D = _mm_sub_epi16(A, B); // diff_uv 788 const __m128i E = _mm_add_epi16(C, D); // new_uv 789 _mm_storeu_si128((__m128i*)(dst + i), E); 790 } 791 for (; i < len; ++i) { 792 const int diff_uv = ref[i] - src[i]; 793 dst[i] += diff_uv; 794 } 795} 796 797static void SharpYUVFilterRow_SSE2(const int16_t* A, const int16_t* B, int len, 798 const uint16_t* best_y, uint16_t* out) { 799 int i; 800 const __m128i kCst8 = _mm_set1_epi16(8); 801 const __m128i max = _mm_set1_epi16(MAX_Y); 802 const __m128i zero = _mm_setzero_si128(); 803 for (i = 0; i + 8 <= len; i += 8) { 804 const __m128i a0 = _mm_loadu_si128((const __m128i*)(A + i + 0)); 805 const __m128i a1 = _mm_loadu_si128((const __m128i*)(A + i + 1)); 806 const __m128i b0 = _mm_loadu_si128((const __m128i*)(B + i + 0)); 807 const __m128i b1 = _mm_loadu_si128((const __m128i*)(B + i + 1)); 808 const __m128i a0b1 = _mm_add_epi16(a0, b1); 809 const __m128i a1b0 = _mm_add_epi16(a1, b0); 810 const __m128i a0a1b0b1 = _mm_add_epi16(a0b1, a1b0); // A0+A1+B0+B1 811 const __m128i a0a1b0b1_8 = _mm_add_epi16(a0a1b0b1, kCst8); 812 const __m128i a0b1_2 = _mm_add_epi16(a0b1, a0b1); // 2*(A0+B1) 813 const __m128i a1b0_2 = _mm_add_epi16(a1b0, a1b0); // 2*(A1+B0) 814 const __m128i c0 = _mm_srai_epi16(_mm_add_epi16(a0b1_2, a0a1b0b1_8), 3); 815 const __m128i c1 = _mm_srai_epi16(_mm_add_epi16(a1b0_2, a0a1b0b1_8), 3); 816 const __m128i d0 = _mm_add_epi16(c1, a0); 817 const __m128i d1 = _mm_add_epi16(c0, a1); 818 const __m128i e0 = _mm_srai_epi16(d0, 1); 819 const __m128i e1 = _mm_srai_epi16(d1, 1); 820 const __m128i f0 = _mm_unpacklo_epi16(e0, e1); 821 const __m128i f1 = _mm_unpackhi_epi16(e0, e1); 822 const __m128i g0 = _mm_loadu_si128((const __m128i*)(best_y + 2 * i + 0)); 823 const __m128i g1 = _mm_loadu_si128((const __m128i*)(best_y + 2 * i + 8)); 824 const __m128i h0 = _mm_add_epi16(g0, f0); 825 const __m128i h1 = _mm_add_epi16(g1, f1); 826 const __m128i i0 = _mm_max_epi16(_mm_min_epi16(h0, max), zero); 827 const __m128i i1 = _mm_max_epi16(_mm_min_epi16(h1, max), zero); 828 _mm_storeu_si128((__m128i*)(out + 2 * i + 0), i0); 829 _mm_storeu_si128((__m128i*)(out + 2 * i + 8), i1); 830 } 831 for (; i < len; ++i) { 832 // (9 * A0 + 3 * A1 + 3 * B0 + B1 + 8) >> 4 = 833 // = (8 * A0 + 2 * (A1 + B0) + (A0 + A1 + B0 + B1 + 8)) >> 4 834 // We reuse the common sub-expressions. 835 const int a0b1 = A[i + 0] + B[i + 1]; 836 const int a1b0 = A[i + 1] + B[i + 0]; 837 const int a0a1b0b1 = a0b1 + a1b0 + 8; 838 const int v0 = (8 * A[i + 0] + 2 * a1b0 + a0a1b0b1) >> 4; 839 const int v1 = (8 * A[i + 1] + 2 * a0b1 + a0a1b0b1) >> 4; 840 out[2 * i + 0] = clip_y(best_y[2 * i + 0] + v0); 841 out[2 * i + 1] = clip_y(best_y[2 * i + 1] + v1); 842 } 843} 844 845#undef MAX_Y 846 847//------------------------------------------------------------------------------ 848 849extern void WebPInitSharpYUVSSE2(void); 850 851WEBP_TSAN_IGNORE_FUNCTION void WebPInitSharpYUVSSE2(void) { 852 WebPSharpYUVUpdateY = SharpYUVUpdateY_SSE2; 853 WebPSharpYUVUpdateRGB = SharpYUVUpdateRGB_SSE2; 854 WebPSharpYUVFilterRow = SharpYUVFilterRow_SSE2; 855} 856 857#else // !WEBP_USE_SSE2 858 859WEBP_DSP_INIT_STUB(WebPInitSamplersSSE2) 860WEBP_DSP_INIT_STUB(WebPInitConvertARGBToYUVSSE2) 861WEBP_DSP_INIT_STUB(WebPInitSharpYUVSSE2) 862 863#endif // WEBP_USE_SSE2 864