31 #define LFC_FUNC(DIR, DEPTH, OPT) \ 32 void ff_hevc_ ## DIR ## _loop_filter_chroma_ ## DEPTH ## _ ## OPT(uint8_t *pix, ptrdiff_t stride, int *tc, uint8_t *no_p, uint8_t *no_q); 34 #define LFL_FUNC(DIR, DEPTH, OPT) \ 35 void ff_hevc_ ## DIR ## _loop_filter_luma_ ## DEPTH ## _ ## OPT(uint8_t *pix, ptrdiff_t stride, int beta, int *tc, uint8_t *no_p, uint8_t *no_q); 37 #define LFC_FUNCS(type, depth, opt) \ 38 LFC_FUNC(h, depth, opt) \ 39 LFC_FUNC(v, depth, opt) 41 #define LFL_FUNCS(type, depth, opt) \ 42 LFL_FUNC(h, depth, opt) \ 43 LFL_FUNC(v, depth, opt) 61 #define IDCT_DC_FUNCS(W, opt) \ 62 void ff_hevc_idct_ ## W ## _dc_8_ ## opt(int16_t *coeffs); \ 63 void ff_hevc_idct_ ## W ## _dc_10_ ## opt(int16_t *coeffs); \ 64 void ff_hevc_idct_ ## W ## _dc_12_ ## opt(int16_t *coeffs) 74 #define IDCT_FUNCS(opt) \ 75 void ff_hevc_idct_4x4_8_ ## opt(int16_t *coeffs, int col_limit); \ 76 void ff_hevc_idct_4x4_10_ ## opt(int16_t *coeffs, int col_limit); \ 77 void ff_hevc_idct_8x8_8_ ## opt(int16_t *coeffs, int col_limit); \ 78 void ff_hevc_idct_8x8_10_ ## opt(int16_t *coeffs, int col_limit); \ 79 void ff_hevc_idct_16x16_8_ ## opt(int16_t *coeffs, int col_limit); \ 80 void ff_hevc_idct_16x16_10_ ## opt(int16_t *coeffs, int col_limit); \ 81 void ff_hevc_idct_32x32_8_ ## opt(int16_t *coeffs, int col_limit); \ 82 void ff_hevc_idct_32x32_10_ ## opt(int16_t *coeffs, int col_limit); 87 #define mc_rep_func(name, bitd, step, W, opt) \ 88 void ff_hevc_put_hevc_##name##W##_##bitd##_##opt(int16_t *_dst, \ 89 uint8_t *_src, ptrdiff_t _srcstride, int height, \ 90 intptr_t mx, intptr_t my, int width) \ 95 for (i = 0; i < W; i += step) { \ 96 src = _src + (i * ((bitd + 7) / 8)); \ 98 ff_hevc_put_hevc_##name##step##_##bitd##_##opt(dst, src, _srcstride, height, mx, my, width); \ 101 #define mc_rep_uni_func(name, bitd, step, W, opt) \ 102 void ff_hevc_put_hevc_uni_##name##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t dststride, \ 103 uint8_t *_src, ptrdiff_t _srcstride, int height, \ 104 intptr_t mx, intptr_t my, int width) \ 109 for (i = 0; i < W; i += step) { \ 110 src = _src + (i * ((bitd + 7) / 8)); \ 111 dst = _dst + (i * ((bitd + 7) / 8)); \ 112 ff_hevc_put_hevc_uni_##name##step##_##bitd##_##opt(dst, dststride, src, _srcstride, \ 113 height, mx, my, width); \ 116 #define mc_rep_bi_func(name, bitd, step, W, opt) \ 117 void ff_hevc_put_hevc_bi_##name##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t dststride, uint8_t *_src, \ 118 ptrdiff_t _srcstride, int16_t* _src2, \ 119 int height, intptr_t mx, intptr_t my, int width) \ 125 for (i = 0; i < W ; i += step) { \ 126 src = _src + (i * ((bitd + 7) / 8)); \ 127 dst = _dst + (i * ((bitd + 7) / 8)); \ 129 ff_hevc_put_hevc_bi_##name##step##_##bitd##_##opt(dst, dststride, src, _srcstride, src2, \ 130 height, mx, my, width); \ 134 #define mc_rep_funcs(name, bitd, step, W, opt) \ 135 mc_rep_func(name, bitd, step, W, opt) \ 136 mc_rep_uni_func(name, bitd, step, W, opt) \ 137 mc_rep_bi_func(name, bitd, step, W, opt) 139 #define mc_rep_func2(name, bitd, step1, step2, W, opt) \ 140 void ff_hevc_put_hevc_##name##W##_##bitd##_##opt(int16_t *dst, \ 141 uint8_t *src, ptrdiff_t _srcstride, int height, \ 142 intptr_t mx, intptr_t my, int width) \ 144 ff_hevc_put_hevc_##name##step1##_##bitd##_##opt(dst, src, _srcstride, height, mx, my, width); \ 145 ff_hevc_put_hevc_##name##step2##_##bitd##_##opt(dst + step1, src + (step1 * ((bitd + 7) / 8)), \ 146 _srcstride, height, mx, my, width); \ 148 #define mc_rep_uni_func2(name, bitd, step1, step2, W, opt) \ 149 void ff_hevc_put_hevc_uni_##name##W##_##bitd##_##opt(uint8_t *dst, ptrdiff_t dststride, \ 150 uint8_t *src, ptrdiff_t _srcstride, int height, \ 151 intptr_t mx, intptr_t my, int width) \ 153 ff_hevc_put_hevc_uni_##name##step1##_##bitd##_##opt(dst, dststride, src, _srcstride, height, mx, my, width);\ 154 ff_hevc_put_hevc_uni_##name##step2##_##bitd##_##opt(dst + (step1 * ((bitd + 7) / 8)), dststride, \ 155 src + (step1 * ((bitd + 7) / 8)), _srcstride, \ 156 height, mx, my, width); \ 158 #define mc_rep_bi_func2(name, bitd, step1, step2, W, opt) \ 159 void ff_hevc_put_hevc_bi_##name##W##_##bitd##_##opt(uint8_t *dst, ptrdiff_t dststride, uint8_t *src, \ 160 ptrdiff_t _srcstride, int16_t* src2, \ 161 int height, intptr_t mx, intptr_t my, int width) \ 163 ff_hevc_put_hevc_bi_##name##step1##_##bitd##_##opt(dst, dststride, src, _srcstride, src2, height, mx, my, width);\ 164 ff_hevc_put_hevc_bi_##name##step2##_##bitd##_##opt(dst + (step1 * ((bitd + 7) / 8)), dststride, \ 165 src + (step1 * ((bitd + 7) / 8)), _srcstride, \ 166 src2 + step1, height, mx, my, width); \ 169 #define mc_rep_funcs2(name, bitd, step1, step2, W, opt) \ 170 mc_rep_func2(name, bitd, step1, step2, W, opt) \ 171 mc_rep_uni_func2(name, bitd, step1, step2, W, opt) \ 172 mc_rep_bi_func2(name, bitd, step1, step2, W, opt) 174 #if ARCH_X86_64 && HAVE_SSE4_EXTERNAL 176 #define mc_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4) \ 177 void ff_hevc_put_hevc_##name##width1##_10_##opt1(int16_t *dst, uint8_t *src, ptrdiff_t _srcstride, \ 178 int height, intptr_t mx, intptr_t my, int width) \ 181 ff_hevc_put_hevc_##name##width2##_10_##opt1(dst, src, _srcstride, height, mx, my, width); \ 182 ff_hevc_put_hevc_##name##width3##_10_##opt2(dst+ width2, src+ width4, _srcstride, height, mx, my, width); \ 185 #define mc_bi_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4) \ 186 void ff_hevc_put_hevc_bi_##name##width1##_10_##opt1(uint8_t *dst, ptrdiff_t dststride, uint8_t *src, \ 187 ptrdiff_t _srcstride, int16_t *src2, \ 188 int height, intptr_t mx, intptr_t my, int width) \ 190 ff_hevc_put_hevc_bi_##name##width2##_10_##opt1(dst, dststride, src, _srcstride, src2, \ 191 height, mx, my, width); \ 192 ff_hevc_put_hevc_bi_##name##width3##_10_##opt2(dst+width4, dststride, src+width4, _srcstride, src2+width2,\ 193 height, mx, my, width); \ 196 #define mc_uni_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4) \ 197 void ff_hevc_put_hevc_uni_##name##width1##_10_##opt1(uint8_t *dst, ptrdiff_t dststride, \ 198 uint8_t *src, ptrdiff_t _srcstride, int height, \ 199 intptr_t mx, intptr_t my, int width) \ 201 ff_hevc_put_hevc_uni_##name##width2##_10_##opt1(dst, dststride, src, _srcstride, \ 202 height, mx, my, width); \ 203 ff_hevc_put_hevc_uni_##name##width3##_10_##opt2(dst+width4, dststride, src+width4, _srcstride, \ 204 height, mx, my, width); \ 207 #define mc_rep_mixs_10(name, width1, width2, width3, opt1, opt2, width4) \ 208 mc_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4) \ 209 mc_bi_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4) \ 210 mc_uni_rep_mix_10(name, width1, width2, width3, opt1, opt2, width4) 212 #define mc_rep_mix_8(name, width1, width2, width3, opt1, opt2) \ 213 void ff_hevc_put_hevc_##name##width1##_8_##opt1(int16_t *dst, uint8_t *src, ptrdiff_t _srcstride, \ 214 int height, intptr_t mx, intptr_t my, int width) \ 217 ff_hevc_put_hevc_##name##width2##_8_##opt1(dst, src, _srcstride, height, mx, my, width); \ 218 ff_hevc_put_hevc_##name##width3##_8_##opt2(dst+ width2, src+ width2, _srcstride, height, mx, my, width); \ 221 #define mc_bi_rep_mix_8(name, width1, width2, width3, opt1, opt2) \ 222 void ff_hevc_put_hevc_bi_##name##width1##_8_##opt1(uint8_t *dst, ptrdiff_t dststride, uint8_t *src, \ 223 ptrdiff_t _srcstride, int16_t* src2, \ 224 int height, intptr_t mx, intptr_t my, int width) \ 226 ff_hevc_put_hevc_bi_##name##width2##_8_##opt1(dst, dststride, src, _srcstride, \ 227 src2, height, mx, my, width); \ 228 ff_hevc_put_hevc_bi_##name##width3##_8_##opt2(dst+width2, dststride, src+width2, _srcstride, \ 229 src2+width2, height, mx, my, width); \ 232 #define mc_uni_rep_mix_8(name, width1, width2, width3, opt1, opt2) \ 233 void ff_hevc_put_hevc_uni_##name##width1##_8_##opt1(uint8_t *dst, ptrdiff_t dststride, \ 234 uint8_t *src, ptrdiff_t _srcstride, int height, \ 235 intptr_t mx, intptr_t my, int width) \ 237 ff_hevc_put_hevc_uni_##name##width2##_8_##opt1(dst, dststride, src, _srcstride, \ 238 height, mx, my, width); \ 239 ff_hevc_put_hevc_uni_##name##width3##_8_##opt2(dst+width2, dststride, src+width2, _srcstride, \ 240 height, mx, my, width); \ 243 #define mc_rep_mixs_8(name, width1, width2, width3, opt1, opt2) \ 244 mc_rep_mix_8(name, width1, width2, width3, opt1, opt2) \ 245 mc_bi_rep_mix_8(name, width1, width2, width3, opt1, opt2) \ 246 mc_uni_rep_mix_8(name, width1, width2, width3, opt1, opt2) 248 #if HAVE_AVX2_EXTERNAL 250 mc_rep_mixs_8(pel_pixels, 48, 32, 16, avx2, sse4)
251 mc_rep_mixs_8(epel_hv, 48, 32, 16, avx2, sse4)
252 mc_rep_mixs_8(epel_h , 48, 32, 16, avx2, sse4)
253 mc_rep_mixs_8(epel_v , 48, 32, 16, avx2, sse4)
255 mc_rep_mix_10(pel_pixels, 24, 16, 8, avx2, sse4, 32)
256 mc_bi_rep_mix_10(pel_pixels,24, 16, 8, avx2, sse4, 32)
257 mc_rep_mixs_10(epel_hv, 24, 16, 8, avx2, sse4, 32)
258 mc_rep_mixs_10(epel_h , 24, 16, 8, avx2, sse4, 32)
259 mc_rep_mixs_10(epel_v , 24, 16, 8, avx2, sse4, 32)
262 mc_rep_mixs_10(qpel_h , 24, 16, 8, avx2, sse4, 32)
263 mc_rep_mixs_10(qpel_v , 24, 16, 8, avx2, sse4, 32)
264 mc_rep_mixs_10(qpel_hv, 24, 16, 8, avx2, sse4, 32)
300 mc_rep_mixs_8(qpel_h , 48, 32, 16, avx2, sse4)
303 mc_rep_mixs_8(qpel_v, 48, 32, 16, avx2, sse4)
437 #define mc_rep_uni_w(bitd, step, W, opt) \ 438 void ff_hevc_put_hevc_uni_w##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t dststride, int16_t *_src, \ 439 int height, int denom, int _wx, int _ox) \ 444 for (i = 0; i < W; i += step) { \ 446 dst= _dst + (i * ((bitd + 7) / 8)); \ 447 ff_hevc_put_hevc_uni_w##step##_##bitd##_##opt(dst, dststride, src, \ 448 height, denom, _wx, _ox); \ 452 mc_rep_uni_w(8, 6, 12, sse4)
453 mc_rep_uni_w(8, 8, 16, sse4)
454 mc_rep_uni_w(8, 8, 24, sse4)
455 mc_rep_uni_w(8, 8, 32, sse4)
456 mc_rep_uni_w(8, 8, 48, sse4)
457 mc_rep_uni_w(8, 8, 64, sse4)
459 mc_rep_uni_w(10, 6, 12, sse4)
460 mc_rep_uni_w(10, 8, 16, sse4)
461 mc_rep_uni_w(10, 8, 24, sse4)
462 mc_rep_uni_w(10, 8, 32, sse4)
463 mc_rep_uni_w(10, 8, 48, sse4)
464 mc_rep_uni_w(10, 8, 64, sse4)
466 mc_rep_uni_w(12, 6, 12, sse4)
467 mc_rep_uni_w(12, 8, 16, sse4)
468 mc_rep_uni_w(12, 8, 24, sse4)
469 mc_rep_uni_w(12, 8, 32, sse4)
470 mc_rep_uni_w(12, 8, 48, sse4)
471 mc_rep_uni_w(12, 8, 64, sse4)
473 #define mc_rep_bi_w(bitd, step, W, opt) \ 474 void ff_hevc_put_hevc_bi_w##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t dststride, int16_t *_src, \ 475 int16_t *_src2, int height, \ 476 int denom, int _wx0, int _wx1, int _ox0, int _ox1) \ 482 for (i = 0; i < W; i += step) { \ 485 dst = _dst + (i * ((bitd + 7) / 8)); \ 486 ff_hevc_put_hevc_bi_w##step##_##bitd##_##opt(dst, dststride, src, src2, \ 487 height, denom, _wx0, _wx1, _ox0, _ox1); \ 491 mc_rep_bi_w(8, 6, 12, sse4)
492 mc_rep_bi_w(8, 8, 16, sse4)
493 mc_rep_bi_w(8, 8, 24, sse4)
494 mc_rep_bi_w(8, 8, 32, sse4)
495 mc_rep_bi_w(8, 8, 48, sse4)
496 mc_rep_bi_w(8, 8, 64, sse4)
498 mc_rep_bi_w(10, 6, 12, sse4)
499 mc_rep_bi_w(10, 8, 16, sse4)
500 mc_rep_bi_w(10, 8, 24, sse4)
501 mc_rep_bi_w(10, 8, 32, sse4)
502 mc_rep_bi_w(10, 8, 48, sse4)
503 mc_rep_bi_w(10, 8, 64, sse4)
505 mc_rep_bi_w(12, 6, 12, sse4)
506 mc_rep_bi_w(12, 8, 16, sse4)
507 mc_rep_bi_w(12, 8, 24, sse4)
508 mc_rep_bi_w(12, 8, 32, sse4)
509 mc_rep_bi_w(12, 8, 48, sse4)
510 mc_rep_bi_w(12, 8, 64, sse4)
512 #define mc_uni_w_func(name, bitd, W, opt) \ 513 void ff_hevc_put_hevc_uni_w_##name##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t _dststride, \ 514 uint8_t *_src, ptrdiff_t _srcstride, \ 515 int height, int denom, \ 517 intptr_t mx, intptr_t my, int width) \ 519 LOCAL_ALIGNED_16(int16_t, temp, [71 * MAX_PB_SIZE]); \ 520 ff_hevc_put_hevc_##name##W##_##bitd##_##opt(temp, _src, _srcstride, height, mx, my, width); \ 521 ff_hevc_put_hevc_uni_w##W##_##bitd##_##opt(_dst, _dststride, temp, height, denom, _wx, _ox);\ 524 #define mc_uni_w_funcs(name, bitd, opt) \ 525 mc_uni_w_func(name, bitd, 4, opt) \ 526 mc_uni_w_func(name, bitd, 8, opt) \ 527 mc_uni_w_func(name, bitd, 12, opt) \ 528 mc_uni_w_func(name, bitd, 16, opt) \ 529 mc_uni_w_func(name, bitd, 24, opt) \ 530 mc_uni_w_func(name, bitd, 32, opt) \ 531 mc_uni_w_func(name, bitd, 48, opt) \ 532 mc_uni_w_func(name, bitd, 64, opt) 534 mc_uni_w_funcs(pel_pixels, 8, sse4)
535 mc_uni_w_func(pel_pixels, 8, 6, sse4)
536 mc_uni_w_funcs(epel_h, 8, sse4)
537 mc_uni_w_func(epel_h, 8, 6, sse4)
538 mc_uni_w_funcs(epel_v, 8, sse4)
539 mc_uni_w_func(epel_v, 8, 6, sse4)
540 mc_uni_w_funcs(epel_hv, 8, sse4)
541 mc_uni_w_func(epel_hv, 8, 6, sse4)
542 mc_uni_w_funcs(qpel_h, 8, sse4)
543 mc_uni_w_funcs(qpel_v, 8, sse4)
544 mc_uni_w_funcs(qpel_hv, 8, sse4)
546 mc_uni_w_funcs(pel_pixels, 10, sse4)
547 mc_uni_w_func(pel_pixels, 10, 6, sse4)
548 mc_uni_w_funcs(epel_h, 10, sse4)
549 mc_uni_w_func(epel_h, 10, 6, sse4)
550 mc_uni_w_funcs(epel_v, 10, sse4)
551 mc_uni_w_func(epel_v, 10, 6, sse4)
552 mc_uni_w_funcs(epel_hv, 10, sse4)
553 mc_uni_w_func(epel_hv, 10, 6, sse4)
554 mc_uni_w_funcs(qpel_h, 10, sse4)
555 mc_uni_w_funcs(qpel_v, 10, sse4)
556 mc_uni_w_funcs(qpel_hv, 10, sse4)
558 mc_uni_w_funcs(pel_pixels, 12, sse4)
559 mc_uni_w_func(pel_pixels, 12, 6, sse4)
560 mc_uni_w_funcs(epel_h, 12, sse4)
561 mc_uni_w_func(epel_h, 12, 6, sse4)
562 mc_uni_w_funcs(epel_v, 12, sse4)
563 mc_uni_w_func(epel_v, 12, 6, sse4)
564 mc_uni_w_funcs(epel_hv, 12, sse4)
565 mc_uni_w_func(epel_hv, 12, 6, sse4)
566 mc_uni_w_funcs(qpel_h, 12, sse4)
567 mc_uni_w_funcs(qpel_v, 12, sse4)
568 mc_uni_w_funcs(qpel_hv, 12, sse4)
570 #define mc_bi_w_func(name, bitd, W, opt) \ 571 void ff_hevc_put_hevc_bi_w_##name##W##_##bitd##_##opt(uint8_t *_dst, ptrdiff_t _dststride, \ 572 uint8_t *_src, ptrdiff_t _srcstride, \ 574 int height, int denom, \ 575 int _wx0, int _wx1, int _ox0, int _ox1, \ 576 intptr_t mx, intptr_t my, int width) \ 578 LOCAL_ALIGNED_16(int16_t, temp, [71 * MAX_PB_SIZE]); \ 579 ff_hevc_put_hevc_##name##W##_##bitd##_##opt(temp, _src, _srcstride, height, mx, my, width); \ 580 ff_hevc_put_hevc_bi_w##W##_##bitd##_##opt(_dst, _dststride, temp, _src2, \ 581 height, denom, _wx0, _wx1, _ox0, _ox1); \ 584 #define mc_bi_w_funcs(name, bitd, opt) \ 585 mc_bi_w_func(name, bitd, 4, opt) \ 586 mc_bi_w_func(name, bitd, 8, opt) \ 587 mc_bi_w_func(name, bitd, 12, opt) \ 588 mc_bi_w_func(name, bitd, 16, opt) \ 589 mc_bi_w_func(name, bitd, 24, opt) \ 590 mc_bi_w_func(name, bitd, 32, opt) \ 591 mc_bi_w_func(name, bitd, 48, opt) \ 592 mc_bi_w_func(name, bitd, 64, opt) 594 mc_bi_w_funcs(pel_pixels, 8, sse4)
595 mc_bi_w_func(pel_pixels, 8, 6, sse4)
596 mc_bi_w_funcs(epel_h, 8, sse4)
597 mc_bi_w_func(epel_h, 8, 6, sse4)
598 mc_bi_w_funcs(epel_v, 8, sse4)
599 mc_bi_w_func(epel_v, 8, 6, sse4)
600 mc_bi_w_funcs(epel_hv, 8, sse4)
601 mc_bi_w_func(epel_hv, 8, 6, sse4)
602 mc_bi_w_funcs(qpel_h, 8, sse4)
603 mc_bi_w_funcs(qpel_v, 8, sse4)
604 mc_bi_w_funcs(qpel_hv, 8, sse4)
606 mc_bi_w_funcs(pel_pixels, 10, sse4)
607 mc_bi_w_func(pel_pixels, 10, 6, sse4)
608 mc_bi_w_funcs(epel_h, 10, sse4)
609 mc_bi_w_func(epel_h, 10, 6, sse4)
610 mc_bi_w_funcs(epel_v, 10, sse4)
611 mc_bi_w_func(epel_v, 10, 6, sse4)
612 mc_bi_w_funcs(epel_hv, 10, sse4)
613 mc_bi_w_func(epel_hv, 10, 6, sse4)
614 mc_bi_w_funcs(qpel_h, 10, sse4)
615 mc_bi_w_funcs(qpel_v, 10, sse4)
616 mc_bi_w_funcs(qpel_hv, 10, sse4)
618 mc_bi_w_funcs(pel_pixels, 12, sse4)
619 mc_bi_w_func(pel_pixels, 12, 6, sse4)
620 mc_bi_w_funcs(epel_h, 12, sse4)
621 mc_bi_w_func(epel_h, 12, 6, sse4)
622 mc_bi_w_funcs(epel_v, 12, sse4)
623 mc_bi_w_func(epel_v, 12, 6, sse4)
624 mc_bi_w_funcs(epel_hv, 12, sse4)
625 mc_bi_w_func(epel_hv, 12, 6, sse4)
626 mc_bi_w_funcs(qpel_h, 12, sse4)
627 mc_bi_w_funcs(qpel_v, 12, sse4)
628 mc_bi_w_funcs(qpel_hv, 12, sse4)
629 #endif //ARCH_X86_64 && HAVE_SSE4_EXTERNAL 631 #define SAO_BAND_FILTER_FUNCS(bitd, opt) \ 632 void ff_hevc_sao_band_filter_8_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \ 633 int16_t *sao_offset_val, int sao_left_class, int width, int height); \ 634 void ff_hevc_sao_band_filter_16_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \ 635 int16_t *sao_offset_val, int sao_left_class, int width, int height); \ 636 void ff_hevc_sao_band_filter_32_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \ 637 int16_t *sao_offset_val, int sao_left_class, int width, int height); \ 638 void ff_hevc_sao_band_filter_48_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \ 639 int16_t *sao_offset_val, int sao_left_class, int width, int height); \ 640 void ff_hevc_sao_band_filter_64_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, \ 641 int16_t *sao_offset_val, int sao_left_class, int width, int height); 653 #define SAO_BAND_INIT(bitd, opt) do { \ 654 c->sao_band_filter[0] = ff_hevc_sao_band_filter_8_##bitd##_##opt; \ 655 c->sao_band_filter[1] = ff_hevc_sao_band_filter_16_##bitd##_##opt; \ 656 c->sao_band_filter[2] = ff_hevc_sao_band_filter_32_##bitd##_##opt; \ 657 c->sao_band_filter[3] = ff_hevc_sao_band_filter_48_##bitd##_##opt; \ 658 c->sao_band_filter[4] = ff_hevc_sao_band_filter_64_##bitd##_##opt; \ 661 #define SAO_EDGE_FILTER_FUNCS(bitd, opt) \ 662 void ff_hevc_sao_edge_filter_8_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val, \ 663 int eo, int width, int height); \ 664 void ff_hevc_sao_edge_filter_16_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val, \ 665 int eo, int width, int height); \ 666 void ff_hevc_sao_edge_filter_32_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val, \ 667 int eo, int width, int height); \ 668 void ff_hevc_sao_edge_filter_48_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val, \ 669 int eo, int width, int height); \ 670 void ff_hevc_sao_edge_filter_64_##bitd##_##opt(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val, \ 671 int eo, int width, int height); \ 680 #define SAO_EDGE_INIT(bitd, opt) do { \ 681 c->sao_edge_filter[0] = ff_hevc_sao_edge_filter_8_##bitd##_##opt; \ 682 c->sao_edge_filter[1] = ff_hevc_sao_edge_filter_16_##bitd##_##opt; \ 683 c->sao_edge_filter[2] = ff_hevc_sao_edge_filter_32_##bitd##_##opt; \ 684 c->sao_edge_filter[3] = ff_hevc_sao_edge_filter_48_##bitd##_##opt; \ 685 c->sao_edge_filter[4] = ff_hevc_sao_edge_filter_64_##bitd##_##opt; \ 688 #define EPEL_LINKS(pointer, my, mx, fname, bitd, opt ) \ 689 PEL_LINK(pointer, 1, my , mx , fname##4 , bitd, opt ); \ 690 PEL_LINK(pointer, 2, my , mx , fname##6 , bitd, opt ); \ 691 PEL_LINK(pointer, 3, my , mx , fname##8 , bitd, opt ); \ 692 PEL_LINK(pointer, 4, my , mx , fname##12, bitd, opt ); \ 693 PEL_LINK(pointer, 5, my , mx , fname##16, bitd, opt ); \ 694 PEL_LINK(pointer, 6, my , mx , fname##24, bitd, opt ); \ 695 PEL_LINK(pointer, 7, my , mx , fname##32, bitd, opt ); \ 696 PEL_LINK(pointer, 8, my , mx , fname##48, bitd, opt ); \ 697 PEL_LINK(pointer, 9, my , mx , fname##64, bitd, opt ) 698 #define QPEL_LINKS(pointer, my, mx, fname, bitd, opt) \ 699 PEL_LINK(pointer, 1, my , mx , fname##4 , bitd, opt ); \ 700 PEL_LINK(pointer, 3, my , mx , fname##8 , bitd, opt ); \ 701 PEL_LINK(pointer, 4, my , mx , fname##12, bitd, opt ); \ 702 PEL_LINK(pointer, 5, my , mx , fname##16, bitd, opt ); \ 703 PEL_LINK(pointer, 6, my , mx , fname##24, bitd, opt ); \ 704 PEL_LINK(pointer, 7, my , mx , fname##32, bitd, opt ); \ 705 PEL_LINK(pointer, 8, my , mx , fname##48, bitd, opt ); \ 706 PEL_LINK(pointer, 9, my , mx , fname##64, bitd, opt ) 712 if (bit_depth == 8) {
714 c->
idct_dc[0] = ff_hevc_idct_4x4_dc_8_mmxext;
715 c->
idct_dc[1] = ff_hevc_idct_8x8_dc_8_mmxext;
726 c->
idct[2] = ff_hevc_idct_16x16_8_sse2;
727 c->
idct[3] = ff_hevc_idct_32x32_8_sse2;
731 c->
idct_dc[1] = ff_hevc_idct_8x8_dc_8_sse2;
732 c->
idct_dc[2] = ff_hevc_idct_16x16_dc_8_sse2;
733 c->
idct_dc[3] = ff_hevc_idct_32x32_dc_8_sse2;
735 c->
idct[0] = ff_hevc_idct_4x4_8_sse2;
736 c->
idct[1] = ff_hevc_idct_8x8_8_sse2;
768 c->
idct[2] = ff_hevc_idct_16x16_8_avx;
769 c->
idct[3] = ff_hevc_idct_32x32_8_avx;
773 c->
idct[0] = ff_hevc_idct_4x4_8_avx;
774 c->
idct[1] = ff_hevc_idct_8x8_8_avx;
785 c->
idct_dc[2] = ff_hevc_idct_16x16_dc_8_avx2;
786 c->
idct_dc[3] = ff_hevc_idct_32x32_dc_8_avx2;
812 c->
put_hevc_epel[7][0][1] = ff_hevc_put_hevc_epel_h32_8_avx2;
813 c->
put_hevc_epel[8][0][1] = ff_hevc_put_hevc_epel_h48_8_avx2;
814 c->
put_hevc_epel[9][0][1] = ff_hevc_put_hevc_epel_h64_8_avx2;
824 c->
put_hevc_epel[7][1][0] = ff_hevc_put_hevc_epel_v32_8_avx2;
825 c->
put_hevc_epel[8][1][0] = ff_hevc_put_hevc_epel_v48_8_avx2;
826 c->
put_hevc_epel[9][1][0] = ff_hevc_put_hevc_epel_v64_8_avx2;
836 c->
put_hevc_epel[7][1][1] = ff_hevc_put_hevc_epel_hv32_8_avx2;
837 c->
put_hevc_epel[8][1][1] = ff_hevc_put_hevc_epel_hv48_8_avx2;
838 c->
put_hevc_epel[9][1][1] = ff_hevc_put_hevc_epel_hv64_8_avx2;
848 c->
put_hevc_qpel[7][0][1] = ff_hevc_put_hevc_qpel_h32_8_avx2;
849 c->
put_hevc_qpel[8][0][1] = ff_hevc_put_hevc_qpel_h48_8_avx2;
850 c->
put_hevc_qpel[9][0][1] = ff_hevc_put_hevc_qpel_h64_8_avx2;
852 c->
put_hevc_qpel[7][1][0] = ff_hevc_put_hevc_qpel_v32_8_avx2;
853 c->
put_hevc_qpel[8][1][0] = ff_hevc_put_hevc_qpel_v48_8_avx2;
854 c->
put_hevc_qpel[9][1][0] = ff_hevc_put_hevc_qpel_v64_8_avx2;
880 }
else if (bit_depth == 10) {
883 c->
idct_dc[0] = ff_hevc_idct_4x4_dc_10_mmxext;
884 c->
idct_dc[1] = ff_hevc_idct_8x8_dc_10_mmxext;
893 c->
idct[2] = ff_hevc_idct_16x16_10_sse2;
894 c->
idct[3] = ff_hevc_idct_32x32_10_sse2;
899 c->
idct_dc[1] = ff_hevc_idct_8x8_dc_10_sse2;
900 c->
idct_dc[2] = ff_hevc_idct_16x16_dc_10_sse2;
901 c->
idct_dc[3] = ff_hevc_idct_32x32_dc_10_sse2;
903 c->
idct[0] = ff_hevc_idct_4x4_10_sse2;
904 c->
idct[1] = ff_hevc_idct_8x8_10_sse2;
932 c->
idct[2] = ff_hevc_idct_16x16_10_avx;
933 c->
idct[3] = ff_hevc_idct_32x32_10_avx;
936 c->
idct[0] = ff_hevc_idct_4x4_10_avx;
937 c->
idct[1] = ff_hevc_idct_8x8_10_avx;
945 c->
idct_dc[2] = ff_hevc_idct_16x16_dc_10_avx2;
946 c->
idct_dc[3] = ff_hevc_idct_32x32_dc_10_avx2;
983 c->
put_hevc_epel[5][0][1] = ff_hevc_put_hevc_epel_h16_10_avx2;
984 c->
put_hevc_epel[6][0][1] = ff_hevc_put_hevc_epel_h24_10_avx2;
985 c->
put_hevc_epel[7][0][1] = ff_hevc_put_hevc_epel_h32_10_avx2;
986 c->
put_hevc_epel[8][0][1] = ff_hevc_put_hevc_epel_h48_10_avx2;
987 c->
put_hevc_epel[9][0][1] = ff_hevc_put_hevc_epel_h64_10_avx2;
1001 c->
put_hevc_epel[5][1][0] = ff_hevc_put_hevc_epel_v16_10_avx2;
1002 c->
put_hevc_epel[6][1][0] = ff_hevc_put_hevc_epel_v24_10_avx2;
1003 c->
put_hevc_epel[7][1][0] = ff_hevc_put_hevc_epel_v32_10_avx2;
1004 c->
put_hevc_epel[8][1][0] = ff_hevc_put_hevc_epel_v48_10_avx2;
1005 c->
put_hevc_epel[9][1][0] = ff_hevc_put_hevc_epel_v64_10_avx2;
1019 c->
put_hevc_epel[5][1][1] = ff_hevc_put_hevc_epel_hv16_10_avx2;
1020 c->
put_hevc_epel[6][1][1] = ff_hevc_put_hevc_epel_hv24_10_avx2;
1021 c->
put_hevc_epel[7][1][1] = ff_hevc_put_hevc_epel_hv32_10_avx2;
1022 c->
put_hevc_epel[8][1][1] = ff_hevc_put_hevc_epel_hv48_10_avx2;
1023 c->
put_hevc_epel[9][1][1] = ff_hevc_put_hevc_epel_hv64_10_avx2;
1037 c->
put_hevc_qpel[5][0][1] = ff_hevc_put_hevc_qpel_h16_10_avx2;
1038 c->
put_hevc_qpel[6][0][1] = ff_hevc_put_hevc_qpel_h24_10_avx2;
1039 c->
put_hevc_qpel[7][0][1] = ff_hevc_put_hevc_qpel_h32_10_avx2;
1040 c->
put_hevc_qpel[8][0][1] = ff_hevc_put_hevc_qpel_h48_10_avx2;
1041 c->
put_hevc_qpel[9][0][1] = ff_hevc_put_hevc_qpel_h64_10_avx2;
1055 c->
put_hevc_qpel[5][1][0] = ff_hevc_put_hevc_qpel_v16_10_avx2;
1056 c->
put_hevc_qpel[6][1][0] = ff_hevc_put_hevc_qpel_v24_10_avx2;
1057 c->
put_hevc_qpel[7][1][0] = ff_hevc_put_hevc_qpel_v32_10_avx2;
1058 c->
put_hevc_qpel[8][1][0] = ff_hevc_put_hevc_qpel_v48_10_avx2;
1059 c->
put_hevc_qpel[9][1][0] = ff_hevc_put_hevc_qpel_v64_10_avx2;
1073 c->
put_hevc_qpel[5][1][1] = ff_hevc_put_hevc_qpel_hv16_10_avx2;
1074 c->
put_hevc_qpel[6][1][1] = ff_hevc_put_hevc_qpel_hv24_10_avx2;
1075 c->
put_hevc_qpel[7][1][1] = ff_hevc_put_hevc_qpel_hv32_10_avx2;
1076 c->
put_hevc_qpel[8][1][1] = ff_hevc_put_hevc_qpel_hv48_10_avx2;
1077 c->
put_hevc_qpel[9][1][1] = ff_hevc_put_hevc_qpel_hv64_10_avx2;
1097 }
else if (bit_depth == 12) {
1099 c->
idct_dc[0] = ff_hevc_idct_4x4_dc_12_mmxext;
1100 c->
idct_dc[1] = ff_hevc_idct_8x8_dc_12_mmxext;
1112 c->
idct_dc[1] = ff_hevc_idct_8x8_dc_12_sse2;
1113 c->
idct_dc[2] = ff_hevc_idct_16x16_dc_12_sse2;
1114 c->
idct_dc[3] = ff_hevc_idct_32x32_dc_12_sse2;
1144 c->
idct_dc[2] = ff_hevc_idct_16x16_dc_12_avx2;
1145 c->
idct_dc[3] = ff_hevc_idct_32x32_dc_12_avx2;
void(* hevc_v_loop_filter_chroma)(uint8_t *pix, ptrdiff_t stride, int32_t *tc, uint8_t *no_p, uint8_t *no_q)
void(* hevc_h_loop_filter_luma)(uint8_t *pix, ptrdiff_t stride, int beta, int32_t *tc, uint8_t *no_p, uint8_t *no_q)
void ff_hevc_add_residual_16_8_avx(uint8_t *dst, int16_t *res, ptrdiff_t stride)
void ff_hevc_put_hevc_pel_pixels16_10_avx2(int16_t *dst, uint8_t *_src, ptrdiff_t _srcstride, int height, intptr_t mx, intptr_t my, int width)
void ff_hevc_put_hevc_bi_pel_pixels48_10_avx2(uint8_t *_dst, ptrdiff_t _dststride, uint8_t *_src, ptrdiff_t _srcstride, int16_t *src2, int height, intptr_t mx, intptr_t my, int width)
void ff_hevc_add_residual_32_10_avx2(uint8_t *dst, int16_t *res, ptrdiff_t stride)
#define SAO_EDGE_INIT(bitd, opt)
void(* put_hevc_epel_uni[10][2][2])(uint8_t *dst, ptrdiff_t dststride, uint8_t *_src, ptrdiff_t _srcstride, int height, intptr_t mx, intptr_t my, int width)
static atomic_int cpu_flags
#define LFL_FUNCS(type, depth, opt)
void ff_hevc_dsp_init_x86(HEVCDSPContext *c, const int bit_depth)
#define SAO_EDGE_FILTER_FUNCS(bitd, opt)
void(* idct[4])(int16_t *coeffs, int col_limit)
void ff_hevc_put_hevc_uni_pel_pixels128_8_avx2(uint8_t *dst, ptrdiff_t dststride, uint8_t *_src, ptrdiff_t _srcstride, int height, intptr_t mx, intptr_t my, int width)
void ff_hevc_put_hevc_pel_pixels64_8_avx2(int16_t *dst, uint8_t *_src, ptrdiff_t _srcstride, int height, intptr_t mx, intptr_t my, int width)
#define EXTERNAL_SSE4(flags)
void(* put_hevc_epel[10][2][2])(int16_t *dst, uint8_t *src, ptrdiff_t srcstride, int height, intptr_t mx, intptr_t my, int width)
void ff_hevc_put_hevc_bi_pel_pixels32_10_avx2(uint8_t *_dst, ptrdiff_t _dststride, uint8_t *_src, ptrdiff_t _srcstride, int16_t *src2, int height, intptr_t mx, intptr_t my, int width)
#define mc_rep_func(name, bitd, step, W, opt)
#define mc_rep_funcs2(name, bitd, step1, step2, W, opt)
bitstream reader API header.
void ff_hevc_put_hevc_pel_pixels32_8_avx2(int16_t *dst, uint8_t *_src, ptrdiff_t _srcstride, int height, intptr_t mx, intptr_t my, int width)
void ff_hevc_add_residual_8_8_sse2(uint8_t *dst, int16_t *res, ptrdiff_t stride)
void ff_hevc_put_hevc_uni_pel_pixels48_8_avx2(uint8_t *dst, ptrdiff_t dststride, uint8_t *_src, ptrdiff_t _srcstride, int height, intptr_t mx, intptr_t my, int width)
void ff_hevc_put_hevc_pel_pixels64_10_avx2(int16_t *dst, uint8_t *_src, ptrdiff_t _srcstride, int height, intptr_t mx, intptr_t my, int width)
#define EXTERNAL_SSE2(flags)
void(* put_hevc_qpel[10][2][2])(int16_t *dst, uint8_t *src, ptrdiff_t srcstride, int height, intptr_t mx, intptr_t my, int width)
void ff_hevc_add_residual_32_8_avx(uint8_t *dst, int16_t *res, ptrdiff_t stride)
void ff_hevc_put_hevc_pel_pixels48_10_avx2(int16_t *dst, uint8_t *_src, ptrdiff_t _srcstride, int height, intptr_t mx, intptr_t my, int width)
#define EXTERNAL_AVX2_FAST(flags)
void ff_hevc_add_residual_8_10_sse2(uint8_t *dst, int16_t *res, ptrdiff_t stride)
void ff_hevc_add_residual_4_8_mmxext(uint8_t *dst, int16_t *res, ptrdiff_t stride)
#define mc_rep_funcs(name, bitd, step, W, opt)
void(* sao_edge_filter[5])(uint8_t *_dst, uint8_t *_src, ptrdiff_t stride_dst, int16_t *sao_offset_val, int sao_eo_class, int width, int height)
#define EPEL_LINKS(pointer, my, mx, fname, bitd, opt)
void ff_hevc_put_hevc_bi_pel_pixels32_8_avx2(uint8_t *_dst, ptrdiff_t _dststride, uint8_t *_src, ptrdiff_t _srcstride, int16_t *src2, int height, intptr_t mx, intptr_t my, int width)
void ff_hevc_put_hevc_uni_pel_pixels32_8_avx2(uint8_t *dst, ptrdiff_t dststride, uint8_t *_src, ptrdiff_t _srcstride, int height, intptr_t mx, intptr_t my, int width)
void ff_hevc_put_hevc_bi_pel_pixels24_10_avx2(uint8_t *_dst, ptrdiff_t _dststride, uint8_t *_src, ptrdiff_t _srcstride, int16_t *src2, int height, intptr_t mx, intptr_t my, int width)
static void bit_depth(AudioStatsContext *s, uint64_t mask, uint64_t imask, AVRational *depth)
void ff_hevc_add_residual_16_10_avx2(uint8_t *dst, int16_t *res, ptrdiff_t stride)
#define SAO_BAND_FILTER_FUNCS(bitd, opt)
void(* idct_dc[4])(int16_t *coeffs)
#define LFC_FUNCS(type, depth, opt)
void(* put_hevc_qpel_bi[10][2][2])(uint8_t *dst, ptrdiff_t dststride, uint8_t *_src, ptrdiff_t _srcstride, int16_t *src2, int height, intptr_t mx, intptr_t my, int width)
#define mc_rep_uni_func(name, bitd, step, W, opt)
void ff_hevc_add_residual_32_10_sse2(uint8_t *dst, int16_t *res, ptrdiff_t stride)
void(* hevc_v_loop_filter_luma)(uint8_t *pix, ptrdiff_t stride, int beta, int32_t *tc, uint8_t *no_p, uint8_t *no_q)
void(* sao_band_filter[5])(uint8_t *_dst, uint8_t *_src, ptrdiff_t _stride_dst, ptrdiff_t _stride_src, int16_t *sao_offset_val, int sao_left_class, int width, int height)
void ff_hevc_add_residual_8_8_avx(uint8_t *dst, int16_t *res, ptrdiff_t stride)
#define EXTERNAL_SSSE3(flags)
void ff_hevc_add_residual_4_10_mmxext(uint8_t *dst, int16_t *res, ptrdiff_t stride)
void(* add_residual[4])(uint8_t *dst, int16_t *res, ptrdiff_t stride)
void ff_hevc_add_residual_32_8_sse2(uint8_t *dst, int16_t *res, ptrdiff_t stride)
void ff_hevc_put_hevc_uni_pel_pixels96_8_avx2(uint8_t *dst, ptrdiff_t dststride, uint8_t *_src, ptrdiff_t _srcstride, int height, intptr_t mx, intptr_t my, int width)
void ff_hevc_put_hevc_pel_pixels24_10_avx2(int16_t *dst, uint8_t *_src, ptrdiff_t _srcstride, int height, intptr_t mx, intptr_t my, int width)
void ff_hevc_put_hevc_bi_pel_pixels16_10_avx2(uint8_t *_dst, ptrdiff_t _dststride, uint8_t *_src, ptrdiff_t _srcstride, int16_t *src2, int height, intptr_t mx, intptr_t my, int width)
void(* put_hevc_qpel_uni[10][2][2])(uint8_t *dst, ptrdiff_t dststride, uint8_t *src, ptrdiff_t srcstride, int height, intptr_t mx, intptr_t my, int width)
int av_get_cpu_flags(void)
Return the flags which specify extensions supported by the CPU.
#define EXTERNAL_AVX2(flags)
#define EXTERNAL_MMXEXT(flags)
void(* hevc_h_loop_filter_chroma)(uint8_t *pix, ptrdiff_t stride, int32_t *tc, uint8_t *no_p, uint8_t *no_q)
void ff_hevc_put_hevc_pel_pixels32_10_avx2(int16_t *dst, uint8_t *_src, ptrdiff_t _srcstride, int height, intptr_t mx, intptr_t my, int width)
void ff_hevc_add_residual_16_8_sse2(uint8_t *dst, int16_t *res, ptrdiff_t stride)
void ff_hevc_put_hevc_bi_pel_pixels64_8_avx2(uint8_t *_dst, ptrdiff_t _dststride, uint8_t *_src, ptrdiff_t _srcstride, int16_t *src2, int height, intptr_t mx, intptr_t my, int width)
#define SAO_BAND_INIT(bitd, opt)
#define mc_rep_bi_func(name, bitd, step, W, opt)
void ff_hevc_put_hevc_pel_pixels48_8_avx2(int16_t *dst, uint8_t *_src, ptrdiff_t _srcstride, int height, intptr_t mx, intptr_t my, int width)
void ff_hevc_put_hevc_bi_pel_pixels64_10_avx2(uint8_t *_dst, ptrdiff_t _dststride, uint8_t *_src, ptrdiff_t _srcstride, int16_t *src2, int height, intptr_t mx, intptr_t my, int width)
void(* put_hevc_epel_bi[10][2][2])(uint8_t *dst, ptrdiff_t dststride, uint8_t *_src, ptrdiff_t _srcstride, int16_t *src2, int height, intptr_t mx, intptr_t my, int width)
#define IDCT_DC_FUNCS(W, opt)
void ff_hevc_add_residual_32_8_avx2(uint8_t *dst, int16_t *res, ptrdiff_t stride)
#define QPEL_LINKS(pointer, my, mx, fname, bitd, opt)
#define EXTERNAL_AVX(flags)
void ff_hevc_add_residual_16_10_sse2(uint8_t *dst, int16_t *res, ptrdiff_t stride)
void ff_hevc_put_hevc_bi_pel_pixels48_8_avx2(uint8_t *_dst, ptrdiff_t _dststride, uint8_t *_src, ptrdiff_t _srcstride, int16_t *src2, int height, intptr_t mx, intptr_t my, int width)
void ff_hevc_put_hevc_uni_pel_pixels64_8_avx2(uint8_t *dst, ptrdiff_t dststride, uint8_t *_src, ptrdiff_t _srcstride, int height, intptr_t mx, intptr_t my, int width)