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
  28 /***********************************/
  29 /* IDCT */
  30
  31 void ff_h264_idct_add_mmx     (uint8_t *dst, int16_t *block, int stride);
  32 void ff_h264_idct8_add_mmx    (uint8_t *dst, int16_t *block, int stride);
  33 void ff_h264_idct8_add_sse2   (uint8_t *dst, int16_t *block, int stride);
  34 void ff_h264_idct_dc_add_mmx2 (uint8_t *dst, int16_t *block, int stride);
  35 void ff_h264_idct8_dc_add_mmx2(uint8_t *dst, int16_t *block, int stride);
  36
  37 void ff_h264_idct_add16_mmx      (uint8_t *dst, const int *block_offset,
  38                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  39 void ff_h264_idct8_add4_mmx      (uint8_t *dst, const int *block_offset,
  40                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  41 void ff_h264_idct_add16_mmx2     (uint8_t *dst, const int *block_offset,
  42                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  43 void ff_h264_idct_add16intra_mmx (uint8_t *dst, const int *block_offset,
  44                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  45 void ff_h264_idct_add16intra_mmx2(uint8_t *dst, const int *block_offset,
  46                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  47 void ff_h264_idct8_add4_mmx2     (uint8_t *dst, const int *block_offset,
  48                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  49 void ff_h264_idct8_add4_sse2     (uint8_t *dst, const int *block_offset,
  50                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  51 void ff_h264_idct_add8_mmx       (uint8_t **dest, const int *block_offset,
  52                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  53 void ff_h264_idct_add8_mmx2      (uint8_t **dest, const int *block_offset,
  54                                   DCTELEM *block, int stride, const uint8_t nnzc[6*8]);
  55
  56 void ff_h264_idct_add16_sse2     (uint8_t *dst, const int *block_offset, DCTELEM *block,
  57                                   int stride, const uint8_t nnzc[6*8]);
  58 void ff_h264_idct_add16intra_sse2(uint8_t *dst, const int *block_offset, DCTELEM *block,
  59                                   int stride, const uint8_t nnzc[6*8]);
  60 void ff_h264_idct_add8_sse2      (uint8_t **dest, const int *block_offset, DCTELEM *block,
  61                                   int stride, const uint8_t nnzc[6*8]);
  62
  63 /***********************************/
  64 /* deblocking */
  65
  66 static av_always_inline void h264_loop_filter_strength_iteration_mmx2(int16_t bS[2][4][4], uint8_t nnz[40],
  67                                                                       int8_t ref[2][40],   int16_t mv[2][40][2],
  68                                                                       int bidir,   int edges, int step,
  69                                                                       int mask_mv, int dir)
  70 {
  71         const x86_reg d_idx = dir ? -8 : -1;
  72         DECLARE_ALIGNED(8, const uint64_t, mask_dir) = dir ? 0 : 0xffffffffffffffffULL;
  73         int b_idx, edge;
  74         for( b_idx=12, edge=0; edge<edges; edge+=step, b_idx+=8*step ) {
  75             __asm__ volatile(
  76                 "pand %0, %%mm0 \n\t"
  77                 ::"m"(mask_dir)
  78             );
  79             if(!(mask_mv & edge)) {
  80                 if(bidir) {
  81                     __asm__ volatile(
  82                         "movd         (%1,%0), %%mm2 \n"
  83                         "punpckldq  40(%1,%0), %%mm2 \n" // { ref0[bn], ref1[bn] }
  84                         "pshufw $0x44,   (%1), %%mm0 \n" // { ref0[b], ref0[b] }
  85                         "pshufw $0x44, 40(%1), %%mm1 \n" // { ref1[b], ref1[b] }
  86                         "pshufw $0x4E, %%mm2, %%mm3 \n"
  87                         "psubb         %%mm2, %%mm0 \n" // { ref0[b]!=ref0[bn], ref0[b]!=ref1[bn] }
  88                         "psubb         %%mm3, %%mm1 \n" // { ref1[b]!=ref1[bn], ref1[b]!=ref0[bn] }
  89                         "1: \n"
  90                         "por           %%mm1, %%mm0 \n"
  91                         "movq      (%2,%0,4), %%mm1 \n"
  92                         "movq     8(%2,%0,4), %%mm2 \n"
  93                         "movq          %%mm1, %%mm3 \n"
  94                         "movq          %%mm2, %%mm4 \n"
  95                         "psubw          (%2), %%mm1 \n"
  96                         "psubw         8(%2), %%mm2 \n"
  97                         "psubw       160(%2), %%mm3 \n"
  98                         "psubw       168(%2), %%mm4 \n"
  99                         "packsswb      %%mm2, %%mm1 \n"
 100                         "packsswb      %%mm4, %%mm3 \n"
 101                         "paddb         %%mm6, %%mm1 \n"
 102                         "paddb         %%mm6, %%mm3 \n"
 103                         "psubusb       %%mm5, %%mm1 \n" // abs(mv[b] - mv[bn]) >= limit
 104                         "psubusb       %%mm5, %%mm3 \n"
 105                         "packsswb      %%mm3, %%mm1 \n"
 106                         "add $40, %0 \n"
 107                         "cmp $40, %0 \n"
 108                         "jl 1b \n"
 109                         "sub $80, %0 \n"
 110                         "pshufw $0x4E, %%mm1, %%mm1 \n"
 111                         "por           %%mm1, %%mm0 \n"
 112                         "pshufw $0x4E, %%mm0, %%mm1 \n"
 113                         "pminub        %%mm1, %%mm0 \n"
 114                         ::"r"(d_idx),
 115                           "r"(ref[0]+b_idx),
 116                           "r"(mv[0]+b_idx)
 117                     );
 118                 } else {
 119                     __asm__ volatile(
 120                         "movd        (%1), %%mm0 \n"
 121                         "psubb    (%1,%0), %%mm0 \n" // ref[b] != ref[bn]
 122                         "movq        (%2), %%mm1 \n"
 123                         "movq       8(%2), %%mm2 \n"
 124                         "psubw  (%2,%0,4), %%mm1 \n"
 125                         "psubw 8(%2,%0,4), %%mm2 \n"
 126                         "packsswb   %%mm2, %%mm1 \n"
 127                         "paddb      %%mm6, %%mm1 \n"
 128                         "psubusb    %%mm5, %%mm1 \n" // abs(mv[b] - mv[bn]) >= limit
 129                         "packsswb   %%mm1, %%mm1 \n"
 130                         "por        %%mm1, %%mm0 \n"
 131                         ::"r"(d_idx),
 132                           "r"(ref[0]+b_idx),
 133                           "r"(mv[0]+b_idx)
 134                     );
 135                 }
 136             }
 137             __asm__ volatile(
 138                 "movd %0, %%mm1 \n"
 139                 "por  %1, %%mm1 \n" // nnz[b] || nnz[bn]
 140                 ::"m"(nnz[b_idx]),
 141                   "m"(nnz[b_idx+d_idx])
 142             );
 143             __asm__ volatile(
 144                 "pminub    %%mm7, %%mm1 \n"
 145                 "pminub    %%mm7, %%mm0 \n"
 146                 "psllw        $1, %%mm1 \n"
 147                 "pxor      %%mm2, %%mm2 \n"
 148                 "pmaxub    %%mm0, %%mm1 \n"
 149                 "punpcklbw %%mm2, %%mm1 \n"
 150                 "movq      %%mm1, %0    \n"
 151                 :"=m"(*bS[dir][edge])
 152                 ::"memory"
 153             );
 154         }
 155 }
 156
 157 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],
 158                                             int bidir, int edges, int step, int mask_mv0, int mask_mv1, int field ) {
 159     __asm__ volatile(
 160         "movq %0, %%mm7 \n"
 161         "movq %1, %%mm6 \n"
 162         ::"m"(ff_pb_1), "m"(ff_pb_3)
 163     );
 164     if(field)
 165         __asm__ volatile(
 166             "movq %0, %%mm6 \n"
 167             ::"m"(ff_pb_3_1)
 168         );
 169     __asm__ volatile(
 170         "movq  %%mm6, %%mm5 \n"
 171         "paddb %%mm5, %%mm5 \n"
 172     :);
 173
 174     // could do a special case for dir==0 && edges==1, but it only reduces the
 175     // average filter time by 1.2%
 176     h264_loop_filter_strength_iteration_mmx2(bS, nnz, ref, mv, bidir, edges, step, mask_mv1, 1);
 177     h264_loop_filter_strength_iteration_mmx2(bS, nnz, ref, mv, bidir,     4,    1, mask_mv0, 0);
 178
 179     __asm__ volatile(
 180         "movq   (%0), %%mm0 \n\t"
 181         "movq  8(%0), %%mm1 \n\t"
 182         "movq 16(%0), %%mm2 \n\t"
 183         "movq 24(%0), %%mm3 \n\t"
 184         TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4)
 185         "movq %%mm0,   (%0) \n\t"
 186         "movq %%mm3,  8(%0) \n\t"
 187         "movq %%mm4, 16(%0) \n\t"
 188         "movq %%mm2, 24(%0) \n\t"
 189         ::"r"(bS[0])
 190         :"memory"
 191     );
 192 }
 193
 194 #define LF_FUNC(DIR, TYPE, OPT) \
 195 void ff_x264_deblock_ ## DIR ## _ ## TYPE ## _ ## OPT (uint8_t *pix, int stride, \
 196                                                int alpha, int beta, int8_t *tc0);
 197 #define LF_IFUNC(DIR, TYPE, OPT) \
 198 void ff_x264_deblock_ ## DIR ## _ ## TYPE ## _ ## OPT (uint8_t *pix, int stride, \
 199                                                int alpha, int beta);
 200
 201 LF_FUNC (h,  chroma,       mmxext)
 202 LF_IFUNC(h,  chroma_intra, mmxext)
 203 LF_FUNC (v,  chroma,       mmxext)
 204 LF_IFUNC(v,  chroma_intra, mmxext)
 205
 206 LF_FUNC (h,  luma,         mmxext)
 207 LF_IFUNC(h,  luma_intra,   mmxext)
 208 #if HAVE_YASM && ARCH_X86_32
 209 LF_FUNC (v8, luma,         mmxext)
 210 static void ff_x264_deblock_v_luma_mmxext(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
 211 {
 212     if((tc0[0] & tc0[1]) >= 0)
 213         ff_x264_deblock_v8_luma_mmxext(pix+0, stride, alpha, beta, tc0);
 214     if((tc0[2] & tc0[3]) >= 0)
 215         ff_x264_deblock_v8_luma_mmxext(pix+8, stride, alpha, beta, tc0+2);
 216 }
 217 LF_IFUNC(v8, luma_intra,   mmxext)
 218 static void ff_x264_deblock_v_luma_intra_mmxext(uint8_t *pix, int stride, int alpha, int beta)
 219 {
 220     ff_x264_deblock_v8_luma_intra_mmxext(pix+0, stride, alpha, beta);
 221     ff_x264_deblock_v8_luma_intra_mmxext(pix+8, stride, alpha, beta);
 222 }
 223 #endif
 224
 225 LF_FUNC (h,  luma,         sse2)
 226 LF_IFUNC(h,  luma_intra,   sse2)
 227 LF_FUNC (v,  luma,         sse2)
 228 LF_IFUNC(v,  luma_intra,   sse2)
 229
 230 /***********************************/
 231 /* weighted prediction */
 232
 233 #define H264_WEIGHT(W, H, OPT) \
 234 void ff_h264_weight_ ## W ## x ## H ## _ ## OPT(uint8_t *dst, \
 235     int stride, int log2_denom, int weight, int offset);
 236
 237 #define H264_BIWEIGHT(W, H, OPT) \
 238 void ff_h264_biweight_ ## W ## x ## H ## _ ## OPT(uint8_t *dst, \
 239     uint8_t *src, int stride, int log2_denom, int weightd, \
 240     int weights, int offset);
 241
 242 #define H264_BIWEIGHT_MMX(W,H) \
 243 H264_WEIGHT  (W, H, mmx2) \
 244 H264_BIWEIGHT(W, H, mmx2)
 245
 246 #define H264_BIWEIGHT_MMX_SSE(W,H) \
 247 H264_BIWEIGHT_MMX(W, H) \
 248 H264_WEIGHT      (W, H, sse2) \
 249 H264_BIWEIGHT    (W, H, sse2) \
 250 H264_BIWEIGHT    (W, H, ssse3)
 251
 252 H264_BIWEIGHT_MMX_SSE(16, 16)
 253 H264_BIWEIGHT_MMX_SSE(16,  8)
 254 H264_BIWEIGHT_MMX_SSE( 8, 16)
 255 H264_BIWEIGHT_MMX_SSE( 8,  8)
 256 H264_BIWEIGHT_MMX_SSE( 8,  4)
 257 H264_BIWEIGHT_MMX    ( 4,  8)
 258 H264_BIWEIGHT_MMX    ( 4,  4)
 259 H264_BIWEIGHT_MMX    ( 4,  2)
 260
 261 void ff_h264dsp_init_x86(H264DSPContext *c)
 262 {
 263     int mm_flags = av_get_cpu_flags();
 264
 265     if (mm_flags & AV_CPU_FLAG_MMX2) {
 266         c->h264_loop_filter_strength= h264_loop_filter_strength_mmx2;
 267     }
 268 #if HAVE_YASM
 269     if (mm_flags & AV_CPU_FLAG_MMX) {
 270         c->h264_idct_dc_add=
 271         c->h264_idct_add= ff_h264_idct_add_mmx;
 272         c->h264_idct8_dc_add=
 273         c->h264_idct8_add= ff_h264_idct8_add_mmx;
 274
 275         c->h264_idct_add16     = ff_h264_idct_add16_mmx;
 276         c->h264_idct8_add4     = ff_h264_idct8_add4_mmx;
 277         c->h264_idct_add8      = ff_h264_idct_add8_mmx;
 278         c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx;
 279
 280         if (mm_flags & AV_CPU_FLAG_MMX2) {
 281             c->h264_idct_dc_add= ff_h264_idct_dc_add_mmx2;
 282             c->h264_idct8_dc_add= ff_h264_idct8_dc_add_mmx2;
 283             c->h264_idct_add16     = ff_h264_idct_add16_mmx2;
 284             c->h264_idct8_add4     = ff_h264_idct8_add4_mmx2;
 285             c->h264_idct_add8      = ff_h264_idct_add8_mmx2;
 286             c->h264_idct_add16intra= ff_h264_idct_add16intra_mmx2;
 287
 288             c->h264_v_loop_filter_chroma= ff_x264_deblock_v_chroma_mmxext;
 289             c->h264_h_loop_filter_chroma= ff_x264_deblock_h_chroma_mmxext;
 290             c->h264_v_loop_filter_chroma_intra= ff_x264_deblock_v_chroma_intra_mmxext;
 291             c->h264_h_loop_filter_chroma_intra= ff_x264_deblock_h_chroma_intra_mmxext;
 292 #if ARCH_X86_32
 293             c->h264_v_loop_filter_luma= ff_x264_deblock_v_luma_mmxext;
 294             c->h264_h_loop_filter_luma= ff_x264_deblock_h_luma_mmxext;
 295             c->h264_v_loop_filter_luma_intra = ff_x264_deblock_v_luma_intra_mmxext;
 296             c->h264_h_loop_filter_luma_intra = ff_x264_deblock_h_luma_intra_mmxext;
 297 #endif
 298             c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_mmx2;
 299             c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_mmx2;
 300             c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_mmx2;
 301             c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_mmx2;
 302             c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_mmx2;
 303             c->weight_h264_pixels_tab[5]= ff_h264_weight_4x8_mmx2;
 304             c->weight_h264_pixels_tab[6]= ff_h264_weight_4x4_mmx2;
 305             c->weight_h264_pixels_tab[7]= ff_h264_weight_4x2_mmx2;
 306
 307             c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_mmx2;
 308             c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_mmx2;
 309             c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_mmx2;
 310             c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_mmx2;
 311             c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_mmx2;
 312             c->biweight_h264_pixels_tab[5]= ff_h264_biweight_4x8_mmx2;
 313             c->biweight_h264_pixels_tab[6]= ff_h264_biweight_4x4_mmx2;
 314             c->biweight_h264_pixels_tab[7]= ff_h264_biweight_4x2_mmx2;
 315
 316             if (mm_flags&AV_CPU_FLAG_SSE2) {
 317                 c->h264_idct8_add = ff_h264_idct8_add_sse2;
 318                 c->h264_idct8_add4= ff_h264_idct8_add4_sse2;
 319
 320                 c->weight_h264_pixels_tab[0]= ff_h264_weight_16x16_sse2;
 321                 c->weight_h264_pixels_tab[1]= ff_h264_weight_16x8_sse2;
 322                 c->weight_h264_pixels_tab[2]= ff_h264_weight_8x16_sse2;
 323                 c->weight_h264_pixels_tab[3]= ff_h264_weight_8x8_sse2;
 324                 c->weight_h264_pixels_tab[4]= ff_h264_weight_8x4_sse2;
 325
 326                 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_sse2;
 327                 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_sse2;
 328                 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_sse2;
 329                 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_sse2;
 330                 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_sse2;
 331
 332 #if HAVE_ALIGNED_STACK
 333                 c->h264_v_loop_filter_luma = ff_x264_deblock_v_luma_sse2;
 334                 c->h264_h_loop_filter_luma = ff_x264_deblock_h_luma_sse2;
 335                 c->h264_v_loop_filter_luma_intra = ff_x264_deblock_v_luma_intra_sse2;
 336                 c->h264_h_loop_filter_luma_intra = ff_x264_deblock_h_luma_intra_sse2;
 337 #endif
 338
 339                 c->h264_idct_add16 = ff_h264_idct_add16_sse2;
 340                 c->h264_idct_add8  = ff_h264_idct_add8_sse2;
 341                 c->h264_idct_add16intra = ff_h264_idct_add16intra_sse2;
 342             }
 343             if (mm_flags&AV_CPU_FLAG_SSSE3) {
 344                 c->biweight_h264_pixels_tab[0]= ff_h264_biweight_16x16_ssse3;
 345                 c->biweight_h264_pixels_tab[1]= ff_h264_biweight_16x8_ssse3;
 346                 c->biweight_h264_pixels_tab[2]= ff_h264_biweight_8x16_ssse3;
 347                 c->biweight_h264_pixels_tab[3]= ff_h264_biweight_8x8_ssse3;
 348                 c->biweight_h264_pixels_tab[4]= ff_h264_biweight_8x4_ssse3;
 349             }
 350         }
 351     }
 352 #endif
 353 }