git.sesse.net Git - ffmpeg/blob - libavcodec/x86/h264dsp_mmx.c

   1 /*
   2  * Copyright (c) 2004-2005 Michael Niedermayer, Loren Merritt
   3  *
   4  * This file is part of FFmpeg.
   5  *
   6  * FFmpeg is free software; you can redistribute it and/or
   7  * modify it under the terms of the GNU Lesser General Public
   8  * License as published by the Free Software Foundation; either
   9  * version 2.1 of the License, or (at your option) any later version.
  10  *
  11  * FFmpeg is distributed in the hope that it will be useful,
  12  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  13  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  14  * Lesser General Public License for more details.
  15  *
  16  * You should have received a copy of the GNU Lesser General Public
  17  * License along with FFmpeg; if not, write to the Free Software
  18  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
  19  */
  20
  21 #include "libavutil/cpu.h"
  22 #include "libavutil/x86_cpu.h"
  23 #include "libavcodec/h264dsp.h"
  24 #include "dsputil_mmx.h"
  25
  26 DECLARE_ALIGNED(8, static const uint64_t, ff_pb_3_1  ) = 0x0103010301030103ULL;
  27 DECLARE_ALIGNED(8, static const uint64_t, ff_pb_7_3  ) = 0x0307030703070307ULL;
  28
  29 /***********************************/
  30 /* IDCT */
  31
  32 void ff_h264_idct_add_mmx     (uint8_t *dst, int16_t *block, int stride);
  33 void ff_h264_idct8_add_mmx    (uint8_t *dst, int16_t *block, int stride);
  34 void ff_h264_idct8_add_sse2   (uint8_t *dst, int16_t *block, int stride);
  35 void ff_h264_idct_dc_add_mmx2 (uint8_t *dst, int16_t *block, int stride);
  36 void ff_h264_idct8_dc_add_mmx2(uint8_t *dst, int16_t *block, int stride);
  37
  38 void ff_h264_idct_add16_mmx      (uint8_t *dst, const int *block_offset,
  39                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  40 void ff_h264_idct8_add4_mmx      (uint8_t *dst, const int *block_offset,
  41                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  42 void ff_h264_idct_add16_mmx2     (uint8_t *dst, const int *block_offset,
  43                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  44 void ff_h264_idct_add16intra_mmx (uint8_t *dst, const int *block_offset,
  45                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  46 void ff_h264_idct_add16intra_mmx2(uint8_t *dst, const int *block_offset,
  47                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  48 void ff_h264_idct8_add4_mmx2     (uint8_t *dst, const int *block_offset,
  49                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  50 void ff_h264_idct8_add4_sse2     (uint8_t *dst, const int *block_offset,
  51                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  52 void ff_h264_idct_add8_mmx       (uint8_t **dest, const int *block_offset,
  53                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  54 void ff_h264_idct_add8_mmx2      (uint8_t **dest, const int *block_offset,
  55                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  56
  57 void ff_h264_idct_add16_sse2     (uint8_t *dst, const int *block_offset, DCTELEM *block,
  58                                   int stride, const uint8_t nnzc[6*8]);
  59 void ff_h264_idct_add16intra_sse2(uint8_t *dst, const int *block_offset, DCTELEM *block,
  60                                   int stride, const uint8_t nnzc[6*8]);
  61 void ff_h264_idct_add8_sse2      (uint8_t **dest, const int *block_offset, DCTELEM *block,
  62                                   int stride, const uint8_t nnzc[6*8]);
  63
  64 /***********************************/
  65 /* deblocking */
  66
  67 static void h264_loop_filter_strength_mmx2( int16_t bS[2][4][4], uint8_t nnz[40], int8_t ref[2][40], int16_t mv[2][40][2],
  68                                             int bidir, int edges, int step, int mask_mv0, int mask_mv1, int field ) {
  69     int dir;
  70     __asm__ volatile(
  71         "movq %0, %%mm7 \n"
  72         "movq %1, %%mm6 \n"
  73         ::"m"(ff_pb_1), "m"(ff_pb_3)
  74     );
  75     if(field)
  76         __asm__ volatile(
  77             "movq %0, %%mm6 \n"
  78             ::"m"(ff_pb_3_1)
  79         );
  80     __asm__ volatile(
  81         "movq  %%mm6, %%mm5 \n"
  82         "paddb %%mm5, %%mm5 \n"
  83     :);
  84
  85     // could do a special case for dir==0 && edges==1, but it only reduces the
  86     // average filter time by 1.2%
  87     for( dir=1; dir>=0; dir-- ) {
  88         const x86_reg d_idx = dir ? -8 : -1;
  89         const int mask_mv = dir ? mask_mv1 : mask_mv0;
  90         DECLARE_ALIGNED(8, const uint64_t, mask_dir) = dir ? 0 : 0xffffffffffffffffULL;
  91         int b_idx, edge;
  92         for( b_idx=12, edge=0; edge<edges; edge+=step, b_idx+=8*step ) {
  93             __asm__ volatile(
  94                 "pand %0, %%mm0 \n\t"
  95                 ::"m"(mask_dir)
  96             );
  97             if(!(mask_mv & edge)) {
  98                 if(bidir) {
  99                     __asm__ volatile(
 100                         "movd         (%1,%0), %%mm2 \n"
 101                         "punpckldq  40(%1,%0), %%mm2 \n" // { ref0[bn], ref1[bn] }
 102                         "pshufw $0x44,   (%1), %%mm0 \n" // { ref0[b], ref0[b] }
 103                         "pshufw $0x44, 40(%1), %%mm1 \n" // { ref1[b], ref1[b] }
 104                         "pshufw $0x4E, %%mm2, %%mm3 \n"
 105                         "psubb         %%mm2, %%mm0 \n" // { ref0[b]!=ref0[bn], ref0[b]!=ref1[bn] }
 106                         "psubb         %%mm3, %%mm1 \n" // { ref1[b]!=ref1[bn], ref1[b]!=ref0[bn] }
 107                         "1: \n"
 108                         "por           %%mm1, %%mm0 \n"
 109                         "movq      (%2,%0,4), %%mm1 \n"
 110                         "movq     8(%2,%0,4), %%mm2 \n"
 111                         "movq          %%mm1, %%mm3 \n"
 112                         "movq          %%mm2, %%mm4 \n"
 113                         "psubw          (%2), %%mm1 \n"
 114                         "psubw         8(%2), %%mm2 \n"
 115                         "psubw       160(%2), %%mm3 \n"
 116                         "psubw       168(%2), %%mm4 \n"
 117                         "packsswb      %%mm2, %%mm1 \n"
 118                         "packsswb      %%mm4, %%mm3 \n"
 119                         "paddb         %%mm6, %%mm1 \n"
 120                         "paddb         %%mm6, %%mm3 \n"
 121                         "psubusb       %%mm5, %%mm1 \n" // abs(mv[b] - mv[bn]) >= limit
 122                         "psubusb       %%mm5, %%mm3 \n"
 123                         "packsswb      %%mm3, %%mm1 \n"
 124                         "add $40, %0 \n"
 125                         "cmp $40, %0 \n"
 126                         "jl 1b \n"
 127                         "sub $80, %0 \n"
 128                         "pshufw $0x4E, %%mm1, %%mm1 \n"
 129                         "por           %%mm1, %%mm0 \n"
 130                         "pshufw $0x4E, %%mm0, %%mm1 \n"
 131                         "pminub        %%mm1, %%mm0 \n"
 132                         ::"r"(d_idx),
 133                           "r"(ref[0]+b_idx),
 134                           "r"(mv[0]+b_idx)
 135                     );
 136                 } else {
 137                     __asm__ volatile(
 138                         "movd        (%1), %%mm0 \n"
 139                         "psubb    (%1,%0), %%mm0 \n" // ref[b] != ref[bn]
 140                         "movq        (%2), %%mm1 \n"
 141                         "movq       8(%2), %%mm2 \n"
 142                         "psubw  (%2,%0,4), %%mm1 \n"
 143                         "psubw 8(%2,%0,4), %%mm2 \n"
 144                         "packsswb   %%mm2, %%mm1 \n"
 145                         "paddb      %%mm6, %%mm1 \n"
 146                         "psubusb    %%mm5, %%mm1 \n" // abs(mv[b] - mv[bn]) >= limit
 147                         "packsswb   %%mm1, %%mm1 \n"
 148                         "por        %%mm1, %%mm0 \n"
 149                         ::"r"(d_idx),
 150                           "r"(ref[0]+b_idx),
 151                           "r"(mv[0]+b_idx)
 152                     );
 153                 }
 154             }
 155             __asm__ volatile(
 156                 "movd %0, %%mm1 \n"
 157                 "por  %1, %%mm1 \n" // nnz[b] || nnz[bn]
 158                 ::"m"(nnz[b_idx]),
 159                   "m"(nnz[b_idx+d_idx])
 160             );
 161             __asm__ volatile(
 162                 "pminub    %%mm7, %%mm1 \n"
 163                 "pminub    %%mm7, %%mm0 \n"
 164                 "psllw        $1, %%mm1 \n"
 165                 "pxor      %%mm2, %%mm2 \n"
 166                 "pmaxub    %%mm0, %%mm1 \n"
 167                 "punpcklbw %%mm2, %%mm1 \n"
 168                 "movq      %%mm1, %0    \n"
 169                 :"=m"(*bS[dir][edge])
 170                 ::"memory"
 171             );
 172         }
 173         edges = 4;
 174         step = 1;
 175     }
 176     __asm__ volatile(
 177         "movq   (%0), %%mm0 \n\t"
 178         "movq  8(%0), %%mm1 \n\t"
 179         "movq 16(%0), %%mm2 \n\t"
 180         "movq 24(%0), %%mm3 \n\t"
 181         TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4)
 182         "movq %%mm0,   (%0) \n\t"
 183         "movq %%mm3,  8(%0) \n\t"
 184         "movq %%mm4, 16(%0) \n\t"
 185         "movq %%mm2, 24(%0) \n\t"
 186         ::"r"(bS[0])
 187         :"memory"
 188     );
 189 }
 190
 191 #define LF_FUNC(DIR, TYPE, OPT) \
 192 void ff_x264_deblock_ ## DIR ## _ ## TYPE ## _ ## OPT (uint8_t *pix, int stride, \
 193                                                int alpha, int beta, int8_t *tc0);
 194 #define LF_IFUNC(DIR, TYPE, OPT) \
 195 void ff_x264_deblock_ ## DIR ## _ ## TYPE ## _ ## OPT (uint8_t *pix, int stride, \
 196                                                int alpha, int beta);
 197
 198 LF_FUNC (h,  chroma,       mmxext)
 199 LF_IFUNC(h,  chroma_intra, mmxext)
 200 LF_FUNC (v,  chroma,       mmxext)
 201 LF_IFUNC(v,  chroma_intra, mmxext)
 202
 203 LF_FUNC (h,  luma,         mmxext)
 204 LF_IFUNC(h,  luma_intra,   mmxext)
 205 #if HAVE_YASM && ARCH_X86_32
 206 LF_FUNC (v8, luma,         mmxext)
 207 static void ff_x264_deblock_v_luma_mmxext(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
 208 {
 209     if((tc0[0] & tc0[1]) >= 0)
 210         ff_x264_deblock_v8_luma_mmxext(pix+0, stride, alpha, beta, tc0);
 211     if((tc0[2] & tc0[3]) >= 0)
 212         ff_x264_deblock_v8_luma_mmxext(pix+8, stride, alpha, beta, tc0+2);
 213 }
 214 LF_IFUNC(v8, luma_intra,   mmxext)
 215 static void ff_x264_deblock_v_luma_intra_mmxext(uint8_t *pix, int stride, int alpha, int beta)
 216 {
 217     ff_x264_deblock_v8_luma_intra_mmxext(pix+0, stride, alpha, beta);
 218     ff_x264_deblock_v8_luma_intra_mmxext(pix+8, stride, alpha, beta);
 219 }
 220 #endif
 221
 222 LF_FUNC (h,  luma,         sse2)
 223 LF_IFUNC(h,  luma_intra,   sse2)
 224 LF_FUNC (v,  luma,         sse2)
 225 LF_IFUNC(v,  luma_intra,   sse2)
 226
 227 /***********************************/
 228 /* weighted prediction */
 229
 230 #define H264_WEIGHT(W, H, OPT) \
 231 void ff_h264_weight_ ## W ## x ## H ## _ ## OPT(uint8_t *dst, \
 232     int stride, int log2_denom, int weight, int offset);
 233
 234 #define H264_BIWEIGHT(W, H, OPT) \
 235 void ff_h264_biweight_ ## W ## x ## H ## _ ## OPT(uint8_t *dst, \
 236     uint8_t *src, int stride, int log2_denom, int weightd, \
 237     int weights, int offset);
 238
 239 #define H264_BIWEIGHT_MMX(W,H) \
 240 H264_WEIGHT  (W, H, mmx2) \
 241 H264_BIWEIGHT(W, H, mmx2)
 242
 243 #define H264_BIWEIGHT_MMX_SSE(W,H) \
 244 H264_BIWEIGHT_MMX(W, H) \
 245 H264_WEIGHT      (W, H, sse2) \
 246 H264_BIWEIGHT    (W, H, sse2) \
 247 H264_BIWEIGHT    (W, H, ssse3)
 248
 249 H264_BIWEIGHT_MMX_SSE(16, 16)
 250 H264_BIWEIGHT_MMX_SSE(16,  8)
 251 H264_BIWEIGHT_MMX_SSE( 8, 16)
 252 H264_BIWEIGHT_MMX_SSE( 8,  8)
 253 H264_BIWEIGHT_MMX_SSE( 8,  4)
 254 H264_BIWEIGHT_MMX    ( 4,  8)
 255 H264_BIWEIGHT_MMX    ( 4,  4)
 256 H264_BIWEIGHT_MMX    ( 4,  2)
 257
 258 void ff_h264dsp_init_x86(H264DSPContext *c)
 259 {
 260     int mm_flags = av_get_cpu_flags();
 261
 262     if (mm_flags & AV_CPU_FLAG_MMX2) {
 263         c->h264_loop_filter_strength= h264_loop_filter_strength_mmx2;
 264     }
 265 #if HAVE_YASM
 266     if (mm_flags & AV_CPU_FLAG_MMX) {
 267         c->h264_idct_dc_add=
 268         c->h264_idct_add= ff_h264_idct_add_mmx;
 269         c->h264_idct8_dc_add=
 270         c->h264_idct8_add= ff_h264_idct8_add_mmx;
 271
 272         c->h264_idct_add16     = ff_h264_idct_add16_mmx;
 273         c->h264_idct8_add4     = ff_h264_idct8_add4_mmx;
 274         c->h264_idct_add8      = ff_h264_idct_add8_mmx;
 275         c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx;
 276
 277         if (mm_flags & AV_CPU_FLAG_MMX2) {
 278             c->h264_idct_dc_add= ff_h264_idct_dc_add_mmx2;
 279             c->h264_idct8_dc_add= ff_h264_idct8_dc_add_mmx2;
 280             c->h264_idct_add16     = ff_h264_idct_add16_mmx2;
 281             c->h264_idct8_add4     = ff_h264_idct8_add4_mmx2;
 282             c->h264_idct_add8      = ff_h264_idct_add8_mmx2;
 283             c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx2;
 284
 285             c->h264_v_loop_filter_chroma= ff_x264_deblock_v_chroma_mmxext;
 286             c->h264_h_loop_filter_chroma= ff_x264_deblock_h_chroma_mmxext;
 287             c->h264_v_loop_filter_chroma_intra= ff_x264_deblock_v_chroma_intra_mmxext;
 288             c->h264_h_loop_filter_chroma_intra= ff_x264_deblock_h_chroma_intra_mmxext;
 289 #if ARCH_X86_32
 290             c->h264_v_loop_filter_luma= ff_x264_deblock_v_luma_mmxext;
 291             c->h264_h_loop_filter_luma= ff_x264_deblock_h_luma_mmxext;
 292             c->h264_v_loop_filter_luma_intra = ff_x264_deblock_v_luma_intra_mmxext;
 293             c->h264_h_loop_filter_luma_intra = ff_x264_deblock_h_luma_intra_mmxext;
 294 #endif
 295             c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_mmx2;
 296             c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_mmx2;
 297             c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_mmx2;
 298             c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_mmx2;
 299             c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_mmx2;
 300             c->weight_h264_pixels_tab[5]= ff_h264_weight_4x8_mmx2;
 301             c->weight_h264_pixels_tab[6]= ff_h264_weight_4x4_mmx2;
 302             c->weight_h264_pixels_tab[7]= ff_h264_weight_4x2_mmx2;
 303
 304             c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_mmx2;
 305             c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_mmx2;
 306             c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_mmx2;
 307             c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_mmx2;
 308             c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_mmx2;
 309             c->biweight_h264_pixels_tab[5]= ff_h264_biweight_4x8_mmx2;
 310             c->biweight_h264_pixels_tab[6]= ff_h264_biweight_4x4_mmx2;
 311             c->biweight_h264_pixels_tab[7]= ff_h264_biweight_4x2_mmx2;
 312
 313             if (mm_flags&AV_CPU_FLAG_SSE2) {
 314                 c->h264_idct8_add = ff_h264_idct8_add_sse2;
 315                 c->h264_idct8_add4= ff_h264_idct8_add4_sse2;
 316
 317                 c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_sse2;
 318                 c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_sse2;
 319                 c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_sse2;
 320                 c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_sse2;
 321                 c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_sse2;
 322
 323                 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_sse2;
 324                 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_sse2;
 325                 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_sse2;
 326                 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_sse2;
 327                 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_sse2;
 328
 329                 c->h264_v_loop_filter_luma = ff_x264_deblock_v_luma_sse2;
 330                 c->h264_h_loop_filter_luma = ff_x264_deblock_h_luma_sse2;
 331                 c->h264_v_loop_filter_luma_intra = ff_x264_deblock_v_luma_intra_sse2;
 332                 c->h264_h_loop_filter_luma_intra = ff_x264_deblock_h_luma_intra_sse2;
 333
 334                 c->h264_idct_add16 = ff_h264_idct_add16_sse2;
 335                 c->h264_idct_add8  = ff_h264_idct_add8_sse2;
 336                 c->h264_idct_add16intra = ff_h264_idct_add16intra_sse2;
 337             }
 338             if (mm_flags&AV_CPU_FLAG_SSSE3) {
 339                 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_ssse3;
 340                 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_ssse3;
 341                 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_ssse3;
 342                 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_ssse3;
 343                 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_ssse3;
 344             }
 345         }
 346     }
 347 #endif
 348 }