git.sesse.net Git - vlc/blob - modules/codec/avcodec/copy.c

   1 /*****************************************************************************
   2  * copy.c: Fast YV12/NV12 copy
   3  *****************************************************************************
   4  * Copyright (C) 2010 Laurent Aimar
   5  * $Id$
   6  *
   7  * Authors: Laurent Aimar <fenrir _AT_ videolan _DOT_ org>
   8  *
   9  * This program is free software; you can redistribute it and/or modify
  10  * it under the terms of the GNU General Public License as published by
  11  * the Free Software Foundation; either version 2 of the License, or
  12  * (at your option) any later version.
  13  *
  14  * This program is distributed in the hope that it will be useful,
  15  * but WITHOUT ANY WARRANTY; without even the implied warranty of
  16  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
  17  * GNU General Public License for more details.
  18  *
  19  * You should have received a copy of the GNU General Public License
  20  * along with this program; if not, write to the Free Software
  21  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston MA 02110-1301, USA.
  22  *****************************************************************************/
  23
  24 #ifdef HAVE_CONFIG_H
  25 # include "config.h"
  26 #endif
  27
  28 #include <vlc_common.h>
  29 #include <vlc_picture.h>
  30 #include <vlc_cpu.h>
  31 #include <assert.h>
  32
  33 #include "copy.h"
  34
  35 int CopyInitCache(copy_cache_t *cache, unsigned width)
  36 {
  37 #ifdef CAN_COMPILE_SSE2
  38     cache->size = __MAX((width + 0x0f) & ~ 0x0f, 4096);
  39     cache->buffer = vlc_memalign(16, cache->size);
  40     if (!cache->buffer)
  41         return VLC_EGENERIC;
  42 #else
  43     (void) cache; (void) width;
  44 #endif
  45     return VLC_SUCCESS;
  46 }
  47
  48 void CopyCleanCache(copy_cache_t *cache)
  49 {
  50 #ifdef CAN_COMPILE_SSE2
  51     vlc_free(cache->buffer);
  52     cache->buffer = NULL;
  53     cache->size   = 0;
  54 #else
  55     (void) cache;
  56 #endif
  57 }
  58
  59 #ifdef CAN_COMPILE_SSE2
  60 /* Copy 64 bytes from srcp to dstp loading data with the SSE>=2 instruction
  61  * load and storing data with the SSE>=2 instruction store.
  62  */
  63 #define COPY64(dstp, srcp, load, store) \
  64     asm volatile (                      \
  65         load "  0(%[src]), %%xmm1\n"    \
  66         load " 16(%[src]), %%xmm2\n"    \
  67         load " 32(%[src]), %%xmm3\n"    \
  68         load " 48(%[src]), %%xmm4\n"    \
  69         store " %%xmm1,    0(%[dst])\n" \
  70         store " %%xmm2,   16(%[dst])\n" \
  71         store " %%xmm3,   32(%[dst])\n" \
  72         store " %%xmm4,   48(%[dst])\n" \
  73         : : [dst]"r"(dstp), [src]"r"(srcp) : "memory")
  74
  75 #ifndef __SSE4_1__
  76 # undef vlc_CPU_SSE4_1
  77 # define vlc_CPU_SSE4_1() ((cpu & VLC_CPU_SSE4_1) != 0)
  78 #endif
  79
  80 #ifndef __SSSE3__
  81 # undef vlc_CPU_SSSE3
  82 # define vlc_CPU_SSSE3() ((cpu & VLC_CPU_SSSE3) != 0)
  83 #endif
  84
  85 #ifndef __SSE2__
  86 # undef vlc_CPU_SSE2
  87 # define vlc_CPU_SSE2() ((cpu & VLC_CPU_SSE2) != 0)
  88 #endif
  89
  90 /* Optimized copy from "Uncacheable Speculative Write Combining" memory
  91  * as used by some video surface.
  92  * XXX It is really efficient only when SSE4.1 is available.
  93  */
  94 static void CopyFromUswc(uint8_t *dst, size_t dst_pitch,
  95                          const uint8_t *src, size_t src_pitch,
  96                          unsigned width, unsigned height,
  97                          unsigned cpu)
  98 {
  99     assert(((intptr_t)dst & 0x0f) == 0 && (dst_pitch & 0x0f) == 0);
 100
 101     asm volatile ("mfence");
 102
 103     for (unsigned y = 0; y < height; y++) {
 104         const unsigned unaligned = (-(uintptr_t)src) & 0x0f;
 105         unsigned x = 0;
 106
 107         for (; x < unaligned; x++)
 108             dst[x] = src[x];
 109
 110 #ifdef CAN_COMPILE_SSE4_1
 111         if (vlc_CPU_SSE4_1()) {
 112             if (!unaligned) {
 113                 for (; x+63 < width; x += 64)
 114                     COPY64(&dst[x], &src[x], "movntdqa", "movdqa");
 115             } else {
 116                 for (; x+63 < width; x += 64)
 117                     COPY64(&dst[x], &src[x], "movntdqa", "movdqu");
 118             }
 119         } else
 120 #endif
 121         {
 122             if (!unaligned) {
 123                 for (; x+63 < width; x += 64)
 124                     COPY64(&dst[x], &src[x], "movdqa", "movdqa");
 125             } else {
 126                 for (; x+63 < width; x += 64)
 127                     COPY64(&dst[x], &src[x], "movdqa", "movdqu");
 128             }
 129         }
 130
 131         for (; x < width; x++)
 132             dst[x] = src[x];
 133
 134         src += src_pitch;
 135         dst += dst_pitch;
 136     }
 137 }
 138
 139 static void Copy2d(uint8_t *dst, size_t dst_pitch,
 140                    const uint8_t *src, size_t src_pitch,
 141                    unsigned width, unsigned height)
 142 {
 143     assert(((intptr_t)src & 0x0f) == 0 && (src_pitch & 0x0f) == 0);
 144
 145     asm volatile ("mfence");
 146
 147     for (unsigned y = 0; y < height; y++) {
 148         unsigned x = 0;
 149
 150         bool unaligned = ((intptr_t)dst & 0x0f) != 0;
 151         if (!unaligned) {
 152             for (; x+63 < width; x += 64)
 153                 COPY64(&dst[x], &src[x], "movdqa", "movntdq");
 154         } else {
 155             for (; x+63 < width; x += 64)
 156                 COPY64(&dst[x], &src[x], "movdqa", "movdqu");
 157         }
 158
 159         for (; x < width; x++)
 160             dst[x] = src[x];
 161
 162         src += src_pitch;
 163         dst += dst_pitch;
 164     }
 165 }
 166
 167 static void SSE_SplitUV(uint8_t *dstu, size_t dstu_pitch,
 168                         uint8_t *dstv, size_t dstv_pitch,
 169                         const uint8_t *src, size_t src_pitch,
 170                         unsigned width, unsigned height, unsigned cpu)
 171 {
 172     const uint8_t shuffle[] = { 0, 2, 4, 6, 8, 10, 12, 14,
 173                                 1, 3, 5, 7, 9, 11, 13, 15 };
 174     const uint8_t mask[] = { 0xff, 0x00, 0xff, 0x00, 0xff, 0x00, 0xff, 0x00,
 175                              0xff, 0x00, 0xff, 0x00, 0xff, 0x00, 0xff, 0x00 };
 176
 177     assert(((intptr_t)src & 0x0f) == 0 && (src_pitch & 0x0f) == 0);
 178
 179     asm volatile ("mfence");
 180
 181     for (unsigned y = 0; y < height; y++) {
 182         unsigned x = 0;
 183
 184 #define LOAD64 \
 185     "movdqa  0(%[src]), %%xmm0\n" \
 186     "movdqa 16(%[src]), %%xmm1\n" \
 187     "movdqa 32(%[src]), %%xmm2\n" \
 188     "movdqa 48(%[src]), %%xmm3\n"
 189
 190 #define STORE2X32 \
 191     "movq   %%xmm0,   0(%[dst1])\n" \
 192     "movq   %%xmm1,   8(%[dst1])\n" \
 193     "movhpd %%xmm0,   0(%[dst2])\n" \
 194     "movhpd %%xmm1,   8(%[dst2])\n" \
 195     "movq   %%xmm2,  16(%[dst1])\n" \
 196     "movq   %%xmm3,  24(%[dst1])\n" \
 197     "movhpd %%xmm2,  16(%[dst2])\n" \
 198     "movhpd %%xmm3,  24(%[dst2])\n"
 199
 200 #ifdef CAN_COMPILE_SSSE3
 201         if (vlc_CPU_SSSE3())
 202         {
 203             for (x = 0; x < (width & ~31); x += 32) {
 204                 asm volatile (
 205                     "movdqu (%[shuffle]), %%xmm7\n"
 206                     LOAD64
 207                     "pshufb  %%xmm7, %%xmm0\n"
 208                     "pshufb  %%xmm7, %%xmm1\n"
 209                     "pshufb  %%xmm7, %%xmm2\n"
 210                     "pshufb  %%xmm7, %%xmm3\n"
 211                     STORE2X32
 212                     : : [dst1]"r"(&dstu[x]), [dst2]"r"(&dstv[x]), [src]"r"(&src[2*x]), [shuffle]"r"(shuffle) : "memory");
 213             }
 214         } else
 215 #endif
 216         {
 217             for (x = 0; x < (width & ~31); x += 32) {
 218                 asm volatile (
 219                     "movdqu (%[mask]), %%xmm7\n"
 220                     LOAD64
 221                     "movdqa   %%xmm0, %%xmm4\n"
 222                     "movdqa   %%xmm1, %%xmm5\n"
 223                     "movdqa   %%xmm2, %%xmm6\n"
 224                     "psrlw    $8,     %%xmm0\n"
 225                     "psrlw    $8,     %%xmm1\n"
 226                     "pand     %%xmm7, %%xmm4\n"
 227                     "pand     %%xmm7, %%xmm5\n"
 228                     "pand     %%xmm7, %%xmm6\n"
 229                     "packuswb %%xmm4, %%xmm0\n"
 230                     "packuswb %%xmm5, %%xmm1\n"
 231                     "pand     %%xmm3, %%xmm7\n"
 232                     "psrlw    $8,     %%xmm2\n"
 233                     "psrlw    $8,     %%xmm3\n"
 234                     "packuswb %%xmm6, %%xmm2\n"
 235                     "packuswb %%xmm7, %%xmm3\n"
 236                     STORE2X32
 237                     : : [dst2]"r"(&dstu[x]), [dst1]"r"(&dstv[x]), [src]"r"(&src[2*x]), [mask]"r"(mask) : "memory");
 238             }
 239         }
 240 #undef STORE2X32
 241 #undef LOAD64
 242
 243         for (; x < width; x++) {
 244             dstu[x] = src[2*x+0];
 245             dstv[x] = src[2*x+1];
 246         }
 247         src  += src_pitch;
 248         dstu += dstu_pitch;
 249         dstv += dstv_pitch;
 250     }
 251 }
 252
 253 static void SSE_CopyPlane(uint8_t *dst, size_t dst_pitch,
 254                           const uint8_t *src, size_t src_pitch,
 255                           uint8_t *cache, size_t cache_size,
 256                           unsigned width, unsigned height, unsigned cpu)
 257 {
 258     const unsigned w16 = (width+15) & ~15;
 259     const unsigned hstep = cache_size / w16;
 260     assert(hstep > 0);
 261
 262     for (unsigned y = 0; y < height; y += hstep) {
 263         const unsigned hblock =  __MIN(hstep, height - y);
 264
 265         /* Copy a bunch of line into our cache */
 266         CopyFromUswc(cache, w16,
 267                      src, src_pitch,
 268                      width, hblock, cpu);
 269
 270         /* Copy from our cache to the destination */
 271         Copy2d(dst, dst_pitch,
 272                cache, w16,
 273                width, hblock);
 274
 275         /* */
 276         src += src_pitch * hblock;
 277         dst += dst_pitch * hblock;
 278     }
 279     asm volatile ("mfence");
 280 }
 281
 282 static void SSE_SplitPlanes(uint8_t *dstu, size_t dstu_pitch,
 283                             uint8_t *dstv, size_t dstv_pitch,
 284                             const uint8_t *src, size_t src_pitch,
 285                             uint8_t *cache, size_t cache_size,
 286                             unsigned width, unsigned height, unsigned cpu)
 287 {
 288     const unsigned w2_16 = (2*width+15) & ~15;
 289     const unsigned hstep = cache_size / w2_16;
 290     assert(hstep > 0);
 291
 292     for (unsigned y = 0; y < height; y += hstep) {
 293         const unsigned hblock =  __MIN(hstep, height - y);
 294
 295         /* Copy a bunch of line into our cache */
 296         CopyFromUswc(cache, w2_16, src, src_pitch,
 297                      2*width, hblock, cpu);
 298
 299         /* Copy from our cache to the destination */
 300         SSE_SplitUV(dstu, dstu_pitch, dstv, dstv_pitch,
 301                     cache, w2_16, width, hblock, cpu);
 302
 303         /* */
 304         src  += src_pitch  * hblock;
 305         dstu += dstu_pitch * hblock;
 306         dstv += dstv_pitch * hblock;
 307     }
 308     asm volatile ("mfence");
 309 }
 310
 311 static void SSE_CopyFromNv12(picture_t *dst,
 312                              uint8_t *src[2], size_t src_pitch[2],
 313                              unsigned width, unsigned height,
 314                              copy_cache_t *cache)
 315 {
 316     const unsigned cpu = vlc_CPU();
 317
 318     /* */
 319     SSE_CopyPlane(dst->p[0].p_pixels, dst->p[0].i_pitch,
 320                   src[0], src_pitch[0],
 321                   cache->buffer, cache->size,
 322                   width, height, cpu);
 323     SSE_SplitPlanes(dst->p[2].p_pixels, dst->p[2].i_pitch,
 324                     dst->p[1].p_pixels, dst->p[1].i_pitch,
 325                     src[1], src_pitch[1],
 326                     cache->buffer, cache->size,
 327                     width/2, height/2, cpu);
 328     asm volatile ("emms");
 329 }
 330
 331 static void SSE_CopyFromYv12(picture_t *dst,
 332                              uint8_t *src[3], size_t src_pitch[3],
 333                              unsigned width, unsigned height,
 334                              copy_cache_t *cache)
 335 {
 336     const unsigned cpu = vlc_CPU();
 337
 338     /* */
 339     for (unsigned n = 0; n < 3; n++) {
 340         const unsigned d = n > 0 ? 2 : 1;
 341         SSE_CopyPlane(dst->p[n].p_pixels, dst->p[n].i_pitch,
 342                       src[n], src_pitch[n],
 343                       cache->buffer, cache->size,
 344                       width/d, height/d, cpu);
 345     }
 346     asm volatile ("emms");
 347 }
 348 #undef COPY64
 349 #endif /* CAN_COMPILE_SSE2 */
 350
 351 static void CopyPlane(uint8_t *dst, size_t dst_pitch,
 352                       const uint8_t *src, size_t src_pitch,
 353                       unsigned width, unsigned height)
 354 {
 355     for (unsigned y = 0; y < height; y++) {
 356         memcpy(dst, src, width);
 357         src += src_pitch;
 358         dst += dst_pitch;
 359     }
 360 }
 361
 362 static void SplitPlanes(uint8_t *dstu, size_t dstu_pitch,
 363                         uint8_t *dstv, size_t dstv_pitch,
 364                         const uint8_t *src, size_t src_pitch,
 365                         unsigned width, unsigned height)
 366 {
 367     for (unsigned y = 0; y < height; y++) {
 368         for (unsigned x = 0; x < width; x++) {
 369             dstu[x] = src[2*x+0];
 370             dstv[x] = src[2*x+1];
 371         }
 372         src  += src_pitch;
 373         dstu += dstu_pitch;
 374         dstv += dstv_pitch;
 375     }
 376 }
 377
 378 void CopyFromNv12(picture_t *dst, uint8_t *src[2], size_t src_pitch[2],
 379                   unsigned width, unsigned height,
 380                   copy_cache_t *cache)
 381 {
 382 #ifdef CAN_COMPILE_SSE2
 383     unsigned cpu = vlc_CPU();
 384     if (vlc_CPU_SSE2())
 385         return SSE_CopyFromNv12(dst, src, src_pitch, width, height, cache);
 386 #else
 387     (void) cache;
 388 #endif
 389
 390     CopyPlane(dst->p[0].p_pixels, dst->p[0].i_pitch,
 391               src[0], src_pitch[0],
 392               width, height);
 393     SplitPlanes(dst->p[2].p_pixels, dst->p[2].i_pitch,
 394                 dst->p[1].p_pixels, dst->p[1].i_pitch,
 395                 src[1], src_pitch[1],
 396                 width/2, height/2);
 397 }
 398
 399 void CopyFromYv12(picture_t *dst, uint8_t *src[3], size_t src_pitch[3],
 400                   unsigned width, unsigned height,
 401                   copy_cache_t *cache)
 402 {
 403 #ifdef CAN_COMPILE_SSE2
 404     unsigned cpu = vlc_CPU();
 405     if (vlc_CPU_SSE2())
 406         return SSE_CopyFromYv12(dst, src, src_pitch, width, height, cache);
 407 #else
 408     (void) cache;
 409 #endif
 410
 411      CopyPlane(dst->p[0].p_pixels, dst->p[0].i_pitch,
 412                src[0], src_pitch[0], width, height);
 413      CopyPlane(dst->p[1].p_pixels, dst->p[1].i_pitch,
 414                src[1], src_pitch[1], width / 2, height / 2);
 415      CopyPlane(dst->p[2].p_pixels, dst->p[2].i_pitch,
 416                src[1], src_pitch[2], width / 2, height / 2);
 417 }