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// SSE2 variant of methods for lossless decoder 11// 12// Author: Skal (pascal.massimino@gmail.com) 13 14#include "./dsp.h" 15 16#include <assert.h> 17 18#if defined(WEBP_USE_SSE2) 19#include <emmintrin.h> 20#include "./lossless.h" 21 22//------------------------------------------------------------------------------ 23// Predictor Transform 24 25static WEBP_INLINE uint32_t ClampedAddSubtractFull(uint32_t c0, uint32_t c1, 26 uint32_t c2) { 27 const __m128i zero = _mm_setzero_si128(); 28 const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero); 29 const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero); 30 const __m128i C2 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero); 31 const __m128i V1 = _mm_add_epi16(C0, C1); 32 const __m128i V2 = _mm_sub_epi16(V1, C2); 33 const __m128i b = _mm_packus_epi16(V2, V2); 34 const uint32_t output = _mm_cvtsi128_si32(b); 35 return output; 36} 37 38static WEBP_INLINE uint32_t ClampedAddSubtractHalf(uint32_t c0, uint32_t c1, 39 uint32_t c2) { 40 const __m128i zero = _mm_setzero_si128(); 41 const __m128i C0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c0), zero); 42 const __m128i C1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c1), zero); 43 const __m128i B0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(c2), zero); 44 const __m128i avg = _mm_add_epi16(C1, C0); 45 const __m128i A0 = _mm_srli_epi16(avg, 1); 46 const __m128i A1 = _mm_sub_epi16(A0, B0); 47 const __m128i BgtA = _mm_cmpgt_epi16(B0, A0); 48 const __m128i A2 = _mm_sub_epi16(A1, BgtA); 49 const __m128i A3 = _mm_srai_epi16(A2, 1); 50 const __m128i A4 = _mm_add_epi16(A0, A3); 51 const __m128i A5 = _mm_packus_epi16(A4, A4); 52 const uint32_t output = _mm_cvtsi128_si32(A5); 53 return output; 54} 55 56static WEBP_INLINE uint32_t Select(uint32_t a, uint32_t b, uint32_t c) { 57 int pa_minus_pb; 58 const __m128i zero = _mm_setzero_si128(); 59 const __m128i A0 = _mm_cvtsi32_si128(a); 60 const __m128i B0 = _mm_cvtsi32_si128(b); 61 const __m128i C0 = _mm_cvtsi32_si128(c); 62 const __m128i AC0 = _mm_subs_epu8(A0, C0); 63 const __m128i CA0 = _mm_subs_epu8(C0, A0); 64 const __m128i BC0 = _mm_subs_epu8(B0, C0); 65 const __m128i CB0 = _mm_subs_epu8(C0, B0); 66 const __m128i AC = _mm_or_si128(AC0, CA0); 67 const __m128i BC = _mm_or_si128(BC0, CB0); 68 const __m128i pa = _mm_unpacklo_epi8(AC, zero); // |a - c| 69 const __m128i pb = _mm_unpacklo_epi8(BC, zero); // |b - c| 70 const __m128i diff = _mm_sub_epi16(pb, pa); 71 { 72 int16_t out[8]; 73 _mm_storeu_si128((__m128i*)out, diff); 74 pa_minus_pb = out[0] + out[1] + out[2] + out[3]; 75 } 76 return (pa_minus_pb <= 0) ? a : b; 77} 78 79static WEBP_INLINE __m128i Average2_128i(uint32_t a0, uint32_t a1) { 80 const __m128i zero = _mm_setzero_si128(); 81 const __m128i A0 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a0), zero); 82 const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a1), zero); 83 const __m128i sum = _mm_add_epi16(A1, A0); 84 const __m128i avg = _mm_srli_epi16(sum, 1); 85 return avg; 86} 87 88static WEBP_INLINE uint32_t Average2(uint32_t a0, uint32_t a1) { 89 const __m128i avg = Average2_128i(a0, a1); 90 const __m128i A2 = _mm_packus_epi16(avg, avg); 91 const uint32_t output = _mm_cvtsi128_si32(A2); 92 return output; 93} 94 95static WEBP_INLINE uint32_t Average3(uint32_t a0, uint32_t a1, uint32_t a2) { 96 const __m128i zero = _mm_setzero_si128(); 97 const __m128i avg1 = Average2_128i(a0, a2); 98 const __m128i A1 = _mm_unpacklo_epi8(_mm_cvtsi32_si128(a1), zero); 99 const __m128i sum = _mm_add_epi16(avg1, A1); 100 const __m128i avg2 = _mm_srli_epi16(sum, 1); 101 const __m128i A2 = _mm_packus_epi16(avg2, avg2); 102 const uint32_t output = _mm_cvtsi128_si32(A2); 103 return output; 104} 105 106static WEBP_INLINE uint32_t Average4(uint32_t a0, uint32_t a1, 107 uint32_t a2, uint32_t a3) { 108 const __m128i avg1 = Average2_128i(a0, a1); 109 const __m128i avg2 = Average2_128i(a2, a3); 110 const __m128i sum = _mm_add_epi16(avg2, avg1); 111 const __m128i avg3 = _mm_srli_epi16(sum, 1); 112 const __m128i A0 = _mm_packus_epi16(avg3, avg3); 113 const uint32_t output = _mm_cvtsi128_si32(A0); 114 return output; 115} 116 117static uint32_t Predictor5(uint32_t left, const uint32_t* const top) { 118 const uint32_t pred = Average3(left, top[0], top[1]); 119 return pred; 120} 121static uint32_t Predictor6(uint32_t left, const uint32_t* const top) { 122 const uint32_t pred = Average2(left, top[-1]); 123 return pred; 124} 125static uint32_t Predictor7(uint32_t left, const uint32_t* const top) { 126 const uint32_t pred = Average2(left, top[0]); 127 return pred; 128} 129static uint32_t Predictor8(uint32_t left, const uint32_t* const top) { 130 const uint32_t pred = Average2(top[-1], top[0]); 131 (void)left; 132 return pred; 133} 134static uint32_t Predictor9(uint32_t left, const uint32_t* const top) { 135 const uint32_t pred = Average2(top[0], top[1]); 136 (void)left; 137 return pred; 138} 139static uint32_t Predictor10(uint32_t left, const uint32_t* const top) { 140 const uint32_t pred = Average4(left, top[-1], top[0], top[1]); 141 return pred; 142} 143static uint32_t Predictor11(uint32_t left, const uint32_t* const top) { 144 const uint32_t pred = Select(top[0], left, top[-1]); 145 return pred; 146} 147static uint32_t Predictor12(uint32_t left, const uint32_t* const top) { 148 const uint32_t pred = ClampedAddSubtractFull(left, top[0], top[-1]); 149 return pred; 150} 151static uint32_t Predictor13(uint32_t left, const uint32_t* const top) { 152 const uint32_t pred = ClampedAddSubtractHalf(left, top[0], top[-1]); 153 return pred; 154} 155 156//------------------------------------------------------------------------------ 157// Subtract-Green Transform 158 159static void SubtractGreenFromBlueAndRed(uint32_t* argb_data, int num_pixels) { 160 const __m128i mask = _mm_set1_epi32(0x0000ff00); 161 int i; 162 for (i = 0; i + 4 <= num_pixels; i += 4) { 163 const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); 164 const __m128i in_00g0 = _mm_and_si128(in, mask); // 00g0|00g0|... 165 const __m128i in_0g00 = _mm_slli_epi32(in_00g0, 8); // 0g00|0g00|... 166 const __m128i in_000g = _mm_srli_epi32(in_00g0, 8); // 000g|000g|... 167 const __m128i in_0g0g = _mm_or_si128(in_0g00, in_000g); 168 const __m128i out = _mm_sub_epi8(in, in_0g0g); 169 _mm_storeu_si128((__m128i*)&argb_data[i], out); 170 } 171 // fallthrough and finish off with plain-C 172 VP8LSubtractGreenFromBlueAndRed_C(argb_data + i, num_pixels - i); 173} 174 175static void AddGreenToBlueAndRed(uint32_t* argb_data, int num_pixels) { 176 const __m128i mask = _mm_set1_epi32(0x0000ff00); 177 int i; 178 for (i = 0; i + 4 <= num_pixels; i += 4) { 179 const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); 180 const __m128i in_00g0 = _mm_and_si128(in, mask); // 00g0|00g0|... 181 const __m128i in_0g00 = _mm_slli_epi32(in_00g0, 8); // 0g00|0g00|... 182 const __m128i in_000g = _mm_srli_epi32(in_00g0, 8); // 000g|000g|... 183 const __m128i in_0g0g = _mm_or_si128(in_0g00, in_000g); 184 const __m128i out = _mm_add_epi8(in, in_0g0g); 185 _mm_storeu_si128((__m128i*)&argb_data[i], out); 186 } 187 // fallthrough and finish off with plain-C 188 VP8LAddGreenToBlueAndRed_C(argb_data + i, num_pixels - i); 189} 190 191//------------------------------------------------------------------------------ 192// Color Transform 193 194static WEBP_INLINE __m128i ColorTransformDelta(__m128i color_pred, 195 __m128i color) { 196 // We simulate signed 8-bit multiplication as: 197 // * Left shift the two (8-bit) numbers by 8 bits, 198 // * Perform a 16-bit signed multiplication and retain the higher 16-bits. 199 const __m128i color_pred_shifted = _mm_slli_epi32(color_pred, 8); 200 const __m128i color_shifted = _mm_slli_epi32(color, 8); 201 // Note: This performs multiplication on 8 packed 16-bit numbers, 4 of which 202 // happen to be zeroes. 203 const __m128i signed_mult = 204 _mm_mulhi_epi16(color_pred_shifted, color_shifted); 205 return _mm_srli_epi32(signed_mult, 5); 206} 207 208static WEBP_INLINE void TransformColor(const VP8LMultipliers* const m, 209 uint32_t* argb_data, 210 int num_pixels) { 211 const __m128i g_to_r = _mm_set1_epi32(m->green_to_red_); // multipliers 212 const __m128i g_to_b = _mm_set1_epi32(m->green_to_blue_); 213 const __m128i r_to_b = _mm_set1_epi32(m->red_to_blue_); 214 215 int i; 216 217 for (i = 0; i + 4 <= num_pixels; i += 4) { 218 const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); 219 const __m128i alpha_green_mask = _mm_set1_epi32(0xff00ff00); // masks 220 const __m128i red_mask = _mm_set1_epi32(0x00ff0000); 221 const __m128i green_mask = _mm_set1_epi32(0x0000ff00); 222 const __m128i lower_8bit_mask = _mm_set1_epi32(0x000000ff); 223 const __m128i ag = _mm_and_si128(in, alpha_green_mask); // alpha, green 224 const __m128i r = _mm_srli_epi32(_mm_and_si128(in, red_mask), 16); 225 const __m128i g = _mm_srli_epi32(_mm_and_si128(in, green_mask), 8); 226 const __m128i b = in; 227 228 const __m128i r_delta = ColorTransformDelta(g_to_r, g); // red 229 const __m128i r_new = 230 _mm_and_si128(_mm_sub_epi32(r, r_delta), lower_8bit_mask); 231 const __m128i r_new_shifted = _mm_slli_epi32(r_new, 16); 232 233 const __m128i b_delta_1 = ColorTransformDelta(g_to_b, g); // blue 234 const __m128i b_delta_2 = ColorTransformDelta(r_to_b, r); 235 const __m128i b_delta = _mm_add_epi32(b_delta_1, b_delta_2); 236 const __m128i b_new = 237 _mm_and_si128(_mm_sub_epi32(b, b_delta), lower_8bit_mask); 238 239 const __m128i out = _mm_or_si128(_mm_or_si128(ag, r_new_shifted), b_new); 240 _mm_storeu_si128((__m128i*)&argb_data[i], out); 241 } 242 243 // Fall-back to C-version for left-overs. 244 VP8LTransformColor_C(m, argb_data + i, num_pixels - i); 245} 246 247static WEBP_INLINE void TransformColorInverse(const VP8LMultipliers* const m, 248 uint32_t* argb_data, 249 int num_pixels) { 250 const __m128i g_to_r = _mm_set1_epi32(m->green_to_red_); // multipliers 251 const __m128i g_to_b = _mm_set1_epi32(m->green_to_blue_); 252 const __m128i r_to_b = _mm_set1_epi32(m->red_to_blue_); 253 254 int i; 255 256 for (i = 0; i + 4 <= num_pixels; i += 4) { 257 const __m128i in = _mm_loadu_si128((__m128i*)&argb_data[i]); 258 const __m128i alpha_green_mask = _mm_set1_epi32(0xff00ff00); // masks 259 const __m128i red_mask = _mm_set1_epi32(0x00ff0000); 260 const __m128i green_mask = _mm_set1_epi32(0x0000ff00); 261 const __m128i lower_8bit_mask = _mm_set1_epi32(0x000000ff); 262 const __m128i ag = _mm_and_si128(in, alpha_green_mask); // alpha, green 263 const __m128i r = _mm_srli_epi32(_mm_and_si128(in, red_mask), 16); 264 const __m128i g = _mm_srli_epi32(_mm_and_si128(in, green_mask), 8); 265 const __m128i b = in; 266 267 const __m128i r_delta = ColorTransformDelta(g_to_r, g); // red 268 const __m128i r_new = 269 _mm_and_si128(_mm_add_epi32(r, r_delta), lower_8bit_mask); 270 const __m128i r_new_shifted = _mm_slli_epi32(r_new, 16); 271 272 const __m128i b_delta_1 = ColorTransformDelta(g_to_b, g); // blue 273 const __m128i b_delta_2 = ColorTransformDelta(r_to_b, r_new); 274 const __m128i b_delta = _mm_add_epi32(b_delta_1, b_delta_2); 275 const __m128i b_new = 276 _mm_and_si128(_mm_add_epi32(b, b_delta), lower_8bit_mask); 277 278 const __m128i out = _mm_or_si128(_mm_or_si128(ag, r_new_shifted), b_new); 279 _mm_storeu_si128((__m128i*)&argb_data[i], out); 280 } 281 282 // Fall-back to C-version for left-overs. 283 VP8LTransformColorInverse_C(m, argb_data + i, num_pixels - i); 284} 285 286//------------------------------------------------------------------------------ 287// Color-space conversion functions 288 289static void ConvertBGRAToRGBA(const uint32_t* src, 290 int num_pixels, uint8_t* dst) { 291 const __m128i* in = (const __m128i*)src; 292 __m128i* out = (__m128i*)dst; 293 while (num_pixels >= 8) { 294 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 295 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 296 const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... 297 const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... 298 const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... 299 const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... 300 const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 301 const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 302 const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 303 const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 304 const __m128i rg0 = _mm_unpacklo_epi8(rb0, ga0); // r0g0r1g1 ... r6g6r7g7 305 const __m128i ba0 = _mm_unpackhi_epi8(rb0, ga0); // b0a0b1a1 ... b6a6b7a7 306 const __m128i rgba0 = _mm_unpacklo_epi16(rg0, ba0); // rgba0|rgba1... 307 const __m128i rgba4 = _mm_unpackhi_epi16(rg0, ba0); // rgba4|rgba5... 308 _mm_storeu_si128(out++, rgba0); 309 _mm_storeu_si128(out++, rgba4); 310 num_pixels -= 8; 311 } 312 // left-overs 313 VP8LConvertBGRAToRGBA_C((const uint32_t*)in, num_pixels, (uint8_t*)out); 314} 315 316static void ConvertBGRAToRGBA4444(const uint32_t* src, 317 int num_pixels, uint8_t* dst) { 318 const __m128i mask_0x0f = _mm_set1_epi8(0x0f); 319 const __m128i mask_0xf0 = _mm_set1_epi8(0xf0); 320 const __m128i* in = (const __m128i*)src; 321 __m128i* out = (__m128i*)dst; 322 while (num_pixels >= 8) { 323 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 324 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 325 const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... 326 const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... 327 const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... 328 const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... 329 const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 330 const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 331 const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 332 const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 333 const __m128i ga1 = _mm_srli_epi16(ga0, 4); // g0-|g1-|...|a6-|a7- 334 const __m128i rb1 = _mm_and_si128(rb0, mask_0xf0); // -r0|-r1|...|-b6|-a7 335 const __m128i ga2 = _mm_and_si128(ga1, mask_0x0f); // g0-|g1-|...|a6-|a7- 336 const __m128i rgba0 = _mm_or_si128(ga2, rb1); // rg0..rg7 | ba0..ba7 337 const __m128i rgba1 = _mm_srli_si128(rgba0, 8); // ba0..ba7 | 0 338#ifdef WEBP_SWAP_16BIT_CSP 339 const __m128i rgba = _mm_unpacklo_epi8(rgba1, rgba0); // barg0...barg7 340#else 341 const __m128i rgba = _mm_unpacklo_epi8(rgba0, rgba1); // rgba0...rgba7 342#endif 343 _mm_storeu_si128(out++, rgba); 344 num_pixels -= 8; 345 } 346 // left-overs 347 VP8LConvertBGRAToRGBA4444_C((const uint32_t*)in, num_pixels, (uint8_t*)out); 348} 349 350static void ConvertBGRAToRGB565(const uint32_t* src, 351 int num_pixels, uint8_t* dst) { 352 const __m128i mask_0xe0 = _mm_set1_epi8(0xe0); 353 const __m128i mask_0xf8 = _mm_set1_epi8(0xf8); 354 const __m128i mask_0x07 = _mm_set1_epi8(0x07); 355 const __m128i* in = (const __m128i*)src; 356 __m128i* out = (__m128i*)dst; 357 while (num_pixels >= 8) { 358 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 359 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 360 const __m128i v0l = _mm_unpacklo_epi8(bgra0, bgra4); // b0b4g0g4r0r4a0a4... 361 const __m128i v0h = _mm_unpackhi_epi8(bgra0, bgra4); // b2b6g2g6r2r6a2a6... 362 const __m128i v1l = _mm_unpacklo_epi8(v0l, v0h); // b0b2b4b6g0g2g4g6... 363 const __m128i v1h = _mm_unpackhi_epi8(v0l, v0h); // b1b3b5b7g1g3g5g7... 364 const __m128i v2l = _mm_unpacklo_epi8(v1l, v1h); // b0...b7 | g0...g7 365 const __m128i v2h = _mm_unpackhi_epi8(v1l, v1h); // r0...r7 | a0...a7 366 const __m128i ga0 = _mm_unpackhi_epi64(v2l, v2h); // g0...g7 | a0...a7 367 const __m128i rb0 = _mm_unpacklo_epi64(v2h, v2l); // r0...r7 | b0...b7 368 const __m128i rb1 = _mm_and_si128(rb0, mask_0xf8); // -r0..-r7|-b0..-b7 369 const __m128i g_lo1 = _mm_srli_epi16(ga0, 5); 370 const __m128i g_lo2 = _mm_and_si128(g_lo1, mask_0x07); // g0-...g7-|xx (3b) 371 const __m128i g_hi1 = _mm_slli_epi16(ga0, 3); 372 const __m128i g_hi2 = _mm_and_si128(g_hi1, mask_0xe0); // -g0...-g7|xx (3b) 373 const __m128i b0 = _mm_srli_si128(rb1, 8); // -b0...-b7|0 374 const __m128i rg1 = _mm_or_si128(rb1, g_lo2); // gr0...gr7|xx 375 const __m128i b1 = _mm_srli_epi16(b0, 3); 376 const __m128i gb1 = _mm_or_si128(b1, g_hi2); // bg0...bg7|xx 377#ifdef WEBP_SWAP_16BIT_CSP 378 const __m128i rgba = _mm_unpacklo_epi8(gb1, rg1); // rggb0...rggb7 379#else 380 const __m128i rgba = _mm_unpacklo_epi8(rg1, gb1); // bgrb0...bgrb7 381#endif 382 _mm_storeu_si128(out++, rgba); 383 num_pixels -= 8; 384 } 385 // left-overs 386 VP8LConvertBGRAToRGB565_C((const uint32_t*)in, num_pixels, (uint8_t*)out); 387} 388 389static void ConvertBGRAToBGR(const uint32_t* src, 390 int num_pixels, uint8_t* dst) { 391 const __m128i mask_l = _mm_set_epi32(0, 0x00ffffff, 0, 0x00ffffff); 392 const __m128i mask_h = _mm_set_epi32(0x00ffffff, 0, 0x00ffffff, 0); 393 const __m128i* in = (const __m128i*)src; 394 const uint8_t* const end = dst + num_pixels * 3; 395 // the last storel_epi64 below writes 8 bytes starting at offset 18 396 while (dst + 26 <= end) { 397 const __m128i bgra0 = _mm_loadu_si128(in++); // bgra0|bgra1|bgra2|bgra3 398 const __m128i bgra4 = _mm_loadu_si128(in++); // bgra4|bgra5|bgra6|bgra7 399 const __m128i a0l = _mm_and_si128(bgra0, mask_l); // bgr0|0|bgr0|0 400 const __m128i a4l = _mm_and_si128(bgra4, mask_l); // bgr0|0|bgr0|0 401 const __m128i a0h = _mm_and_si128(bgra0, mask_h); // 0|bgr0|0|bgr0 402 const __m128i a4h = _mm_and_si128(bgra4, mask_h); // 0|bgr0|0|bgr0 403 const __m128i b0h = _mm_srli_epi64(a0h, 8); // 000b|gr00|000b|gr00 404 const __m128i b4h = _mm_srli_epi64(a4h, 8); // 000b|gr00|000b|gr00 405 const __m128i c0 = _mm_or_si128(a0l, b0h); // rgbrgb00|rgbrgb00 406 const __m128i c4 = _mm_or_si128(a4l, b4h); // rgbrgb00|rgbrgb00 407 const __m128i c2 = _mm_srli_si128(c0, 8); 408 const __m128i c6 = _mm_srli_si128(c4, 8); 409 _mm_storel_epi64((__m128i*)(dst + 0), c0); 410 _mm_storel_epi64((__m128i*)(dst + 6), c2); 411 _mm_storel_epi64((__m128i*)(dst + 12), c4); 412 _mm_storel_epi64((__m128i*)(dst + 18), c6); 413 dst += 24; 414 num_pixels -= 8; 415 } 416 // left-overs 417 VP8LConvertBGRAToBGR_C((const uint32_t*)in, num_pixels, dst); 418} 419 420//------------------------------------------------------------------------------ 421 422#define LINE_SIZE 16 // 8 or 16 423static void AddVector(const uint32_t* a, const uint32_t* b, uint32_t* out, 424 int size) { 425 int i; 426 assert(size % LINE_SIZE == 0); 427 for (i = 0; i < size; i += LINE_SIZE) { 428 const __m128i a0 = _mm_loadu_si128((__m128i*)&a[i + 0]); 429 const __m128i a1 = _mm_loadu_si128((__m128i*)&a[i + 4]); 430#if (LINE_SIZE == 16) 431 const __m128i a2 = _mm_loadu_si128((__m128i*)&a[i + 8]); 432 const __m128i a3 = _mm_loadu_si128((__m128i*)&a[i + 12]); 433#endif 434 const __m128i b0 = _mm_loadu_si128((__m128i*)&b[i + 0]); 435 const __m128i b1 = _mm_loadu_si128((__m128i*)&b[i + 4]); 436#if (LINE_SIZE == 16) 437 const __m128i b2 = _mm_loadu_si128((__m128i*)&b[i + 8]); 438 const __m128i b3 = _mm_loadu_si128((__m128i*)&b[i + 12]); 439#endif 440 _mm_storeu_si128((__m128i*)&out[i + 0], _mm_add_epi32(a0, b0)); 441 _mm_storeu_si128((__m128i*)&out[i + 4], _mm_add_epi32(a1, b1)); 442#if (LINE_SIZE == 16) 443 _mm_storeu_si128((__m128i*)&out[i + 8], _mm_add_epi32(a2, b2)); 444 _mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3)); 445#endif 446 } 447} 448 449static void AddVectorEq(const uint32_t* a, uint32_t* out, int size) { 450 int i; 451 assert(size % LINE_SIZE == 0); 452 for (i = 0; i < size; i += LINE_SIZE) { 453 const __m128i a0 = _mm_loadu_si128((__m128i*)&a[i + 0]); 454 const __m128i a1 = _mm_loadu_si128((__m128i*)&a[i + 4]); 455#if (LINE_SIZE == 16) 456 const __m128i a2 = _mm_loadu_si128((__m128i*)&a[i + 8]); 457 const __m128i a3 = _mm_loadu_si128((__m128i*)&a[i + 12]); 458#endif 459 const __m128i b0 = _mm_loadu_si128((__m128i*)&out[i + 0]); 460 const __m128i b1 = _mm_loadu_si128((__m128i*)&out[i + 4]); 461#if (LINE_SIZE == 16) 462 const __m128i b2 = _mm_loadu_si128((__m128i*)&out[i + 8]); 463 const __m128i b3 = _mm_loadu_si128((__m128i*)&out[i + 12]); 464#endif 465 _mm_storeu_si128((__m128i*)&out[i + 0], _mm_add_epi32(a0, b0)); 466 _mm_storeu_si128((__m128i*)&out[i + 4], _mm_add_epi32(a1, b1)); 467#if (LINE_SIZE == 16) 468 _mm_storeu_si128((__m128i*)&out[i + 8], _mm_add_epi32(a2, b2)); 469 _mm_storeu_si128((__m128i*)&out[i + 12], _mm_add_epi32(a3, b3)); 470#endif 471 } 472} 473#undef LINE_SIZE 474 475// Note we are adding uint32_t's as *signed* int32's (using _mm_add_epi32). But 476// that's ok since the histogram values are less than 1<<28 (max picture size). 477static void HistogramAdd(const VP8LHistogram* const a, 478 const VP8LHistogram* const b, 479 VP8LHistogram* const out) { 480 int i; 481 const int literal_size = VP8LHistogramNumCodes(a->palette_code_bits_); 482 assert(a->palette_code_bits_ == b->palette_code_bits_); 483 if (b != out) { 484 AddVector(a->literal_, b->literal_, out->literal_, NUM_LITERAL_CODES); 485 AddVector(a->red_, b->red_, out->red_, NUM_LITERAL_CODES); 486 AddVector(a->blue_, b->blue_, out->blue_, NUM_LITERAL_CODES); 487 AddVector(a->alpha_, b->alpha_, out->alpha_, NUM_LITERAL_CODES); 488 } else { 489 AddVectorEq(a->literal_, out->literal_, NUM_LITERAL_CODES); 490 AddVectorEq(a->red_, out->red_, NUM_LITERAL_CODES); 491 AddVectorEq(a->blue_, out->blue_, NUM_LITERAL_CODES); 492 AddVectorEq(a->alpha_, out->alpha_, NUM_LITERAL_CODES); 493 } 494 for (i = NUM_LITERAL_CODES; i < literal_size; ++i) { 495 out->literal_[i] = a->literal_[i] + b->literal_[i]; 496 } 497 for (i = 0; i < NUM_DISTANCE_CODES; ++i) { 498 out->distance_[i] = a->distance_[i] + b->distance_[i]; 499 } 500} 501 502#endif // WEBP_USE_SSE2 503 504//------------------------------------------------------------------------------ 505 506extern void VP8LDspInitSSE2(void); 507 508void VP8LDspInitSSE2(void) { 509#if defined(WEBP_USE_SSE2) 510 VP8LPredictors[5] = Predictor5; 511 VP8LPredictors[6] = Predictor6; 512 VP8LPredictors[7] = Predictor7; 513 VP8LPredictors[8] = Predictor8; 514 VP8LPredictors[9] = Predictor9; 515 VP8LPredictors[10] = Predictor10; 516 VP8LPredictors[11] = Predictor11; 517 VP8LPredictors[12] = Predictor12; 518 VP8LPredictors[13] = Predictor13; 519 520 VP8LSubtractGreenFromBlueAndRed = SubtractGreenFromBlueAndRed; 521 VP8LAddGreenToBlueAndRed = AddGreenToBlueAndRed; 522 523 VP8LTransformColor = TransformColor; 524 VP8LTransformColorInverse = TransformColorInverse; 525 526 VP8LConvertBGRAToRGBA = ConvertBGRAToRGBA; 527 VP8LConvertBGRAToRGBA4444 = ConvertBGRAToRGBA4444; 528 VP8LConvertBGRAToRGB565 = ConvertBGRAToRGB565; 529 VP8LConvertBGRAToBGR = ConvertBGRAToBGR; 530 531 VP8LHistogramAdd = HistogramAdd; 532#endif // WEBP_USE_SSE2 533} 534 535//------------------------------------------------------------------------------ 536