1 /* 2 * Copyright (c) 2010 The WebM 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 #include <emmintrin.h> // SSE2 12 13 #include "./vpx_config.h" 14 #include "./vpx_dsp_rtcd.h" 15 16 #include "vpx_ports/mem.h" 17 18 typedef void (*getNxMvar_fn_t)(const unsigned char *src, int src_stride, 19 const unsigned char *ref, int ref_stride, 20 unsigned int *sse, int *sum); 21 22 unsigned int vpx_get_mb_ss_sse2(const int16_t *src) { 23 __m128i vsum = _mm_setzero_si128(); 24 int i; 25 26 for (i = 0; i < 32; ++i) { 27 const __m128i v = _mm_loadu_si128((const __m128i *)src); 28 vsum = _mm_add_epi32(vsum, _mm_madd_epi16(v, v)); 29 src += 8; 30 } 31 32 vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 8)); 33 vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 4)); 34 return _mm_cvtsi128_si32(vsum); 35 } 36 37 #define READ64(p, stride, i) \ 38 _mm_unpacklo_epi8( \ 39 _mm_cvtsi32_si128(*(const uint32_t *)(p + i * stride)), \ 40 _mm_cvtsi32_si128(*(const uint32_t *)(p + (i + 1) * stride))) 41 42 static void get4x4var_sse2(const uint8_t *src, int src_stride, 43 const uint8_t *ref, int ref_stride, 44 unsigned int *sse, int *sum) { 45 const __m128i zero = _mm_setzero_si128(); 46 const __m128i src0 = _mm_unpacklo_epi8(READ64(src, src_stride, 0), zero); 47 const __m128i src1 = _mm_unpacklo_epi8(READ64(src, src_stride, 2), zero); 48 const __m128i ref0 = _mm_unpacklo_epi8(READ64(ref, ref_stride, 0), zero); 49 const __m128i ref1 = _mm_unpacklo_epi8(READ64(ref, ref_stride, 2), zero); 50 const __m128i diff0 = _mm_sub_epi16(src0, ref0); 51 const __m128i diff1 = _mm_sub_epi16(src1, ref1); 52 53 // sum 54 __m128i vsum = _mm_add_epi16(diff0, diff1); 55 vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8)); 56 vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4)); 57 vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 2)); 58 *sum = (int16_t)_mm_extract_epi16(vsum, 0); 59 60 // sse 61 vsum = 62 _mm_add_epi32(_mm_madd_epi16(diff0, diff0), _mm_madd_epi16(diff1, diff1)); 63 vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 8)); 64 vsum = _mm_add_epi32(vsum, _mm_srli_si128(vsum, 4)); 65 *sse = _mm_cvtsi128_si32(vsum); 66 } 67 68 void vpx_get8x8var_sse2(const uint8_t *src, int src_stride, const uint8_t *ref, 69 int ref_stride, unsigned int *sse, int *sum) { 70 const __m128i zero = _mm_setzero_si128(); 71 __m128i vsum = _mm_setzero_si128(); 72 __m128i vsse = _mm_setzero_si128(); 73 int i; 74 75 for (i = 0; i < 8; i += 2) { 76 const __m128i src0 = _mm_unpacklo_epi8( 77 _mm_loadl_epi64((const __m128i *)(src + i * src_stride)), zero); 78 const __m128i ref0 = _mm_unpacklo_epi8( 79 _mm_loadl_epi64((const __m128i *)(ref + i * ref_stride)), zero); 80 const __m128i diff0 = _mm_sub_epi16(src0, ref0); 81 82 const __m128i src1 = _mm_unpacklo_epi8( 83 _mm_loadl_epi64((const __m128i *)(src + (i + 1) * src_stride)), zero); 84 const __m128i ref1 = _mm_unpacklo_epi8( 85 _mm_loadl_epi64((const __m128i *)(ref + (i + 1) * ref_stride)), zero); 86 const __m128i diff1 = _mm_sub_epi16(src1, ref1); 87 88 vsum = _mm_add_epi16(vsum, diff0); 89 vsum = _mm_add_epi16(vsum, diff1); 90 vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff0, diff0)); 91 vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff1, diff1)); 92 } 93 94 // sum 95 vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8)); 96 vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4)); 97 vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 2)); 98 *sum = (int16_t)_mm_extract_epi16(vsum, 0); 99 100 // sse 101 vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 8)); 102 vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 4)); 103 *sse = _mm_cvtsi128_si32(vsse); 104 } 105 106 void vpx_get16x16var_sse2(const uint8_t *src, int src_stride, 107 const uint8_t *ref, int ref_stride, unsigned int *sse, 108 int *sum) { 109 const __m128i zero = _mm_setzero_si128(); 110 __m128i vsum = _mm_setzero_si128(); 111 __m128i vsse = _mm_setzero_si128(); 112 int i; 113 114 for (i = 0; i < 16; ++i) { 115 const __m128i s = _mm_loadu_si128((const __m128i *)src); 116 const __m128i r = _mm_loadu_si128((const __m128i *)ref); 117 118 const __m128i src0 = _mm_unpacklo_epi8(s, zero); 119 const __m128i ref0 = _mm_unpacklo_epi8(r, zero); 120 const __m128i diff0 = _mm_sub_epi16(src0, ref0); 121 122 const __m128i src1 = _mm_unpackhi_epi8(s, zero); 123 const __m128i ref1 = _mm_unpackhi_epi8(r, zero); 124 const __m128i diff1 = _mm_sub_epi16(src1, ref1); 125 126 vsum = _mm_add_epi16(vsum, diff0); 127 vsum = _mm_add_epi16(vsum, diff1); 128 vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff0, diff0)); 129 vsse = _mm_add_epi32(vsse, _mm_madd_epi16(diff1, diff1)); 130 131 src += src_stride; 132 ref += ref_stride; 133 } 134 135 // sum 136 vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 8)); 137 vsum = _mm_add_epi16(vsum, _mm_srli_si128(vsum, 4)); 138 *sum = 139 (int16_t)_mm_extract_epi16(vsum, 0) + (int16_t)_mm_extract_epi16(vsum, 1); 140 141 // sse 142 vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 8)); 143 vsse = _mm_add_epi32(vsse, _mm_srli_si128(vsse, 4)); 144 *sse = _mm_cvtsi128_si32(vsse); 145 } 146 147 static void variance_sse2(const unsigned char *src, int src_stride, 148 const unsigned char *ref, int ref_stride, int w, 149 int h, unsigned int *sse, int *sum, 150 getNxMvar_fn_t var_fn, int block_size) { 151 int i, j; 152 153 *sse = 0; 154 *sum = 0; 155 156 for (i = 0; i < h; i += block_size) { 157 for (j = 0; j < w; j += block_size) { 158 unsigned int sse0; 159 int sum0; 160 var_fn(src + src_stride * i + j, src_stride, ref + ref_stride * i + j, 161 ref_stride, &sse0, &sum0); 162 *sse += sse0; 163 *sum += sum0; 164 } 165 } 166 } 167 168 unsigned int vpx_variance4x4_sse2(const unsigned char *src, int src_stride, 169 const unsigned char *ref, int ref_stride, 170 unsigned int *sse) { 171 int sum; 172 get4x4var_sse2(src, src_stride, ref, ref_stride, sse, &sum); 173 return *sse - ((sum * sum) >> 4); 174 } 175 176 unsigned int vpx_variance8x4_sse2(const uint8_t *src, int src_stride, 177 const uint8_t *ref, int ref_stride, 178 unsigned int *sse) { 179 int sum; 180 variance_sse2(src, src_stride, ref, ref_stride, 8, 4, sse, &sum, 181 get4x4var_sse2, 4); 182 return *sse - ((sum * sum) >> 5); 183 } 184 185 unsigned int vpx_variance4x8_sse2(const uint8_t *src, int src_stride, 186 const uint8_t *ref, int ref_stride, 187 unsigned int *sse) { 188 int sum; 189 variance_sse2(src, src_stride, ref, ref_stride, 4, 8, sse, &sum, 190 get4x4var_sse2, 4); 191 return *sse - ((sum * sum) >> 5); 192 } 193 194 unsigned int vpx_variance8x8_sse2(const unsigned char *src, int src_stride, 195 const unsigned char *ref, int ref_stride, 196 unsigned int *sse) { 197 int sum; 198 vpx_get8x8var_sse2(src, src_stride, ref, ref_stride, sse, &sum); 199 return *sse - ((sum * sum) >> 6); 200 } 201 202 unsigned int vpx_variance16x8_sse2(const unsigned char *src, int src_stride, 203 const unsigned char *ref, int ref_stride, 204 unsigned int *sse) { 205 int sum; 206 variance_sse2(src, src_stride, ref, ref_stride, 16, 8, sse, &sum, 207 vpx_get8x8var_sse2, 8); 208 return *sse - ((sum * sum) >> 7); 209 } 210 211 unsigned int vpx_variance8x16_sse2(const unsigned char *src, int src_stride, 212 const unsigned char *ref, int ref_stride, 213 unsigned int *sse) { 214 int sum; 215 variance_sse2(src, src_stride, ref, ref_stride, 8, 16, sse, &sum, 216 vpx_get8x8var_sse2, 8); 217 return *sse - ((sum * sum) >> 7); 218 } 219 220 unsigned int vpx_variance16x16_sse2(const unsigned char *src, int src_stride, 221 const unsigned char *ref, int ref_stride, 222 unsigned int *sse) { 223 int sum; 224 vpx_get16x16var_sse2(src, src_stride, ref, ref_stride, sse, &sum); 225 return *sse - (((uint32_t)((int64_t)sum * sum)) >> 8); 226 } 227 228 unsigned int vpx_variance32x32_sse2(const uint8_t *src, int src_stride, 229 const uint8_t *ref, int ref_stride, 230 unsigned int *sse) { 231 int sum; 232 variance_sse2(src, src_stride, ref, ref_stride, 32, 32, sse, &sum, 233 vpx_get16x16var_sse2, 16); 234 return *sse - (unsigned int)(((int64_t)sum * sum) >> 10); 235 } 236 237 unsigned int vpx_variance32x16_sse2(const uint8_t *src, int src_stride, 238 const uint8_t *ref, int ref_stride, 239 unsigned int *sse) { 240 int sum; 241 variance_sse2(src, src_stride, ref, ref_stride, 32, 16, sse, &sum, 242 vpx_get16x16var_sse2, 16); 243 return *sse - (unsigned int)(((int64_t)sum * sum) >> 9); 244 } 245 246 unsigned int vpx_variance16x32_sse2(const uint8_t *src, int src_stride, 247 const uint8_t *ref, int ref_stride, 248 unsigned int *sse) { 249 int sum; 250 variance_sse2(src, src_stride, ref, ref_stride, 16, 32, sse, &sum, 251 vpx_get16x16var_sse2, 16); 252 return *sse - (unsigned int)(((int64_t)sum * sum) >> 9); 253 } 254 255 unsigned int vpx_variance64x64_sse2(const uint8_t *src, int src_stride, 256 const uint8_t *ref, int ref_stride, 257 unsigned int *sse) { 258 int sum; 259 variance_sse2(src, src_stride, ref, ref_stride, 64, 64, sse, &sum, 260 vpx_get16x16var_sse2, 16); 261 return *sse - (unsigned int)(((int64_t)sum * sum) >> 12); 262 } 263 264 unsigned int vpx_variance64x32_sse2(const uint8_t *src, int src_stride, 265 const uint8_t *ref, int ref_stride, 266 unsigned int *sse) { 267 int sum; 268 variance_sse2(src, src_stride, ref, ref_stride, 64, 32, sse, &sum, 269 vpx_get16x16var_sse2, 16); 270 return *sse - (unsigned int)(((int64_t)sum * sum) >> 11); 271 } 272 273 unsigned int vpx_variance32x64_sse2(const uint8_t *src, int src_stride, 274 const uint8_t *ref, int ref_stride, 275 unsigned int *sse) { 276 int sum; 277 variance_sse2(src, src_stride, ref, ref_stride, 32, 64, sse, &sum, 278 vpx_get16x16var_sse2, 16); 279 return *sse - (unsigned int)(((int64_t)sum * sum) >> 11); 280 } 281 282 unsigned int vpx_mse8x8_sse2(const uint8_t *src, int src_stride, 283 const uint8_t *ref, int ref_stride, 284 unsigned int *sse) { 285 vpx_variance8x8_sse2(src, src_stride, ref, ref_stride, sse); 286 return *sse; 287 } 288 289 unsigned int vpx_mse8x16_sse2(const uint8_t *src, int src_stride, 290 const uint8_t *ref, int ref_stride, 291 unsigned int *sse) { 292 vpx_variance8x16_sse2(src, src_stride, ref, ref_stride, sse); 293 return *sse; 294 } 295 296 unsigned int vpx_mse16x8_sse2(const uint8_t *src, int src_stride, 297 const uint8_t *ref, int ref_stride, 298 unsigned int *sse) { 299 vpx_variance16x8_sse2(src, src_stride, ref, ref_stride, sse); 300 return *sse; 301 } 302 303 unsigned int vpx_mse16x16_sse2(const uint8_t *src, int src_stride, 304 const uint8_t *ref, int ref_stride, 305 unsigned int *sse) { 306 vpx_variance16x16_sse2(src, src_stride, ref, ref_stride, sse); 307 return *sse; 308 } 309 310 // The 2 unused parameters are place holders for PIC enabled build. 311 // These definitions are for functions defined in subpel_variance.asm 312 #define DECL(w, opt) \ 313 int vpx_sub_pixel_variance##w##xh_##opt( \ 314 const uint8_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \ 315 const uint8_t *dst, ptrdiff_t dst_stride, int height, unsigned int *sse, \ 316 void *unused0, void *unused) 317 #define DECLS(opt1, opt2) \ 318 DECL(4, opt1); \ 319 DECL(8, opt1); \ 320 DECL(16, opt1) 321 322 DECLS(sse2, sse2); 323 DECLS(ssse3, ssse3); 324 #undef DECLS 325 #undef DECL 326 327 #define FN(w, h, wf, wlog2, hlog2, opt, cast_prod, cast) \ 328 unsigned int vpx_sub_pixel_variance##w##x##h##_##opt( \ 329 const uint8_t *src, int src_stride, int x_offset, int y_offset, \ 330 const uint8_t *dst, int dst_stride, unsigned int *sse_ptr) { \ 331 unsigned int sse; \ 332 int se = vpx_sub_pixel_variance##wf##xh_##opt(src, src_stride, x_offset, \ 333 y_offset, dst, dst_stride, \ 334 h, &sse, NULL, NULL); \ 335 if (w > wf) { \ 336 unsigned int sse2; \ 337 int se2 = vpx_sub_pixel_variance##wf##xh_##opt( \ 338 src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride, h, \ 339 &sse2, NULL, NULL); \ 340 se += se2; \ 341 sse += sse2; \ 342 if (w > wf * 2) { \ 343 se2 = vpx_sub_pixel_variance##wf##xh_##opt( \ 344 src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, h, \ 345 &sse2, NULL, NULL); \ 346 se += se2; \ 347 sse += sse2; \ 348 se2 = vpx_sub_pixel_variance##wf##xh_##opt( \ 349 src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride, h, \ 350 &sse2, NULL, NULL); \ 351 se += se2; \ 352 sse += sse2; \ 353 } \ 354 } \ 355 *sse_ptr = sse; \ 356 return sse - (unsigned int)(cast_prod(cast se * se) >> (wlog2 + hlog2)); \ 357 } 358 359 #define FNS(opt1, opt2) \ 360 FN(64, 64, 16, 6, 6, opt1, (int64_t), (int64_t)); \ 361 FN(64, 32, 16, 6, 5, opt1, (int64_t), (int64_t)); \ 362 FN(32, 64, 16, 5, 6, opt1, (int64_t), (int64_t)); \ 363 FN(32, 32, 16, 5, 5, opt1, (int64_t), (int64_t)); \ 364 FN(32, 16, 16, 5, 4, opt1, (int64_t), (int64_t)); \ 365 FN(16, 32, 16, 4, 5, opt1, (int64_t), (int64_t)); \ 366 FN(16, 16, 16, 4, 4, opt1, (uint32_t), (int64_t)); \ 367 FN(16, 8, 16, 4, 3, opt1, (int32_t), (int32_t)); \ 368 FN(8, 16, 8, 3, 4, opt1, (int32_t), (int32_t)); \ 369 FN(8, 8, 8, 3, 3, opt1, (int32_t), (int32_t)); \ 370 FN(8, 4, 8, 3, 2, opt1, (int32_t), (int32_t)); \ 371 FN(4, 8, 4, 2, 3, opt1, (int32_t), (int32_t)); \ 372 FN(4, 4, 4, 2, 2, opt1, (int32_t), (int32_t)) 373 374 FNS(sse2, sse2); 375 FNS(ssse3, ssse3); 376 377 #undef FNS 378 #undef FN 379 380 // The 2 unused parameters are place holders for PIC enabled build. 381 #define DECL(w, opt) \ 382 int vpx_sub_pixel_avg_variance##w##xh_##opt( \ 383 const uint8_t *src, ptrdiff_t src_stride, int x_offset, int y_offset, \ 384 const uint8_t *dst, ptrdiff_t dst_stride, const uint8_t *sec, \ 385 ptrdiff_t sec_stride, int height, unsigned int *sse, void *unused0, \ 386 void *unused) 387 #define DECLS(opt1, opt2) \ 388 DECL(4, opt1); \ 389 DECL(8, opt1); \ 390 DECL(16, opt1) 391 392 DECLS(sse2, sse2); 393 DECLS(ssse3, ssse3); 394 #undef DECL 395 #undef DECLS 396 397 #define FN(w, h, wf, wlog2, hlog2, opt, cast_prod, cast) \ 398 unsigned int vpx_sub_pixel_avg_variance##w##x##h##_##opt( \ 399 const uint8_t *src, int src_stride, int x_offset, int y_offset, \ 400 const uint8_t *dst, int dst_stride, unsigned int *sseptr, \ 401 const uint8_t *sec) { \ 402 unsigned int sse; \ 403 int se = vpx_sub_pixel_avg_variance##wf##xh_##opt( \ 404 src, src_stride, x_offset, y_offset, dst, dst_stride, sec, w, h, &sse, \ 405 NULL, NULL); \ 406 if (w > wf) { \ 407 unsigned int sse2; \ 408 int se2 = vpx_sub_pixel_avg_variance##wf##xh_##opt( \ 409 src + 16, src_stride, x_offset, y_offset, dst + 16, dst_stride, \ 410 sec + 16, w, h, &sse2, NULL, NULL); \ 411 se += se2; \ 412 sse += sse2; \ 413 if (w > wf * 2) { \ 414 se2 = vpx_sub_pixel_avg_variance##wf##xh_##opt( \ 415 src + 32, src_stride, x_offset, y_offset, dst + 32, dst_stride, \ 416 sec + 32, w, h, &sse2, NULL, NULL); \ 417 se += se2; \ 418 sse += sse2; \ 419 se2 = vpx_sub_pixel_avg_variance##wf##xh_##opt( \ 420 src + 48, src_stride, x_offset, y_offset, dst + 48, dst_stride, \ 421 sec + 48, w, h, &sse2, NULL, NULL); \ 422 se += se2; \ 423 sse += sse2; \ 424 } \ 425 } \ 426 *sseptr = sse; \ 427 return sse - (unsigned int)(cast_prod(cast se * se) >> (wlog2 + hlog2)); \ 428 } 429 430 #define FNS(opt1, opt2) \ 431 FN(64, 64, 16, 6, 6, opt1, (int64_t), (int64_t)); \ 432 FN(64, 32, 16, 6, 5, opt1, (int64_t), (int64_t)); \ 433 FN(32, 64, 16, 5, 6, opt1, (int64_t), (int64_t)); \ 434 FN(32, 32, 16, 5, 5, opt1, (int64_t), (int64_t)); \ 435 FN(32, 16, 16, 5, 4, opt1, (int64_t), (int64_t)); \ 436 FN(16, 32, 16, 4, 5, opt1, (int64_t), (int64_t)); \ 437 FN(16, 16, 16, 4, 4, opt1, (uint32_t), (int64_t)); \ 438 FN(16, 8, 16, 4, 3, opt1, (uint32_t), (int32_t)); \ 439 FN(8, 16, 8, 3, 4, opt1, (uint32_t), (int32_t)); \ 440 FN(8, 8, 8, 3, 3, opt1, (uint32_t), (int32_t)); \ 441 FN(8, 4, 8, 3, 2, opt1, (uint32_t), (int32_t)); \ 442 FN(4, 8, 4, 2, 3, opt1, (uint32_t), (int32_t)); \ 443 FN(4, 4, 4, 2, 2, opt1, (uint32_t), (int32_t)) 444 445 FNS(sse2, sse); 446 FNS(ssse3, ssse3); 447 448 #undef FNS 449 #undef FN 450
