git.sesse.net Git - ffmpeg/blob - libswscale/x86/hscale_fast_bilinear_simd.c

   1 /*
   2  * Copyright (C) 2001-2003 Michael Niedermayer <michaelni@gmx.at>
   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 "../swscale_internal.h"
  22 #include "libavutil/x86/asm.h"
  23 #include "libavutil/x86/cpu.h"
  24
  25 #define RET 0xC3 // near return opcode for x86
  26 #define PREFETCH "prefetchnta"
  27
  28 #if HAVE_INLINE_ASM
  29 av_cold int ff_init_hscaler_mmxext(int dstW, int xInc, uint8_t *filterCode,
  30                                        int16_t *filter, int32_t *filterPos,
  31                                        int numSplits)
  32 {
  33     uint8_t *fragmentA;
  34     x86_reg imm8OfPShufW1A;
  35     x86_reg imm8OfPShufW2A;
  36     x86_reg fragmentLengthA;
  37     uint8_t *fragmentB;
  38     x86_reg imm8OfPShufW1B;
  39     x86_reg imm8OfPShufW2B;
  40     x86_reg fragmentLengthB;
  41     int fragmentPos;
  42
  43     int xpos, i;
  44
  45     // create an optimized horizontal scaling routine
  46     /* This scaler is made of runtime-generated MMXEXT code using specially tuned
  47      * pshufw instructions. For every four output pixels, if four input pixels
  48      * are enough for the fast bilinear scaling, then a chunk of fragmentB is
  49      * used. If five input pixels are needed, then a chunk of fragmentA is used.
  50      */
  51
  52     // code fragment
  53
  54     __asm__ volatile (
  55         "jmp                         9f                 \n\t"
  56         // Begin
  57         "0:                                             \n\t"
  58         "movq    (%%"FF_REG_d", %%"FF_REG_a"), %%mm3    \n\t"
  59         "movd    (%%"FF_REG_c", %%"FF_REG_S"), %%mm0    \n\t"
  60         "movd   1(%%"FF_REG_c", %%"FF_REG_S"), %%mm1    \n\t"
  61         "punpcklbw                %%mm7, %%mm1          \n\t"
  62         "punpcklbw                %%mm7, %%mm0          \n\t"
  63         "pshufw                   $0xFF, %%mm1, %%mm1   \n\t"
  64         "1:                                             \n\t"
  65         "pshufw                   $0xFF, %%mm0, %%mm0   \n\t"
  66         "2:                                             \n\t"
  67         "psubw                    %%mm1, %%mm0          \n\t"
  68         "movl   8(%%"FF_REG_b", %%"FF_REG_a"), %%esi    \n\t"
  69         "pmullw                   %%mm3, %%mm0          \n\t"
  70         "psllw                       $7, %%mm1          \n\t"
  71         "paddw                    %%mm1, %%mm0          \n\t"
  72
  73         "movq                     %%mm0, (%%"FF_REG_D", %%"FF_REG_a") \n\t"
  74
  75         "add                         $8, %%"FF_REG_a"   \n\t"
  76         // End
  77         "9:                                             \n\t"
  78         "lea       " LOCAL_MANGLE(0b) ", %0             \n\t"
  79         "lea       " LOCAL_MANGLE(1b) ", %1             \n\t"
  80         "lea       " LOCAL_MANGLE(2b) ", %2             \n\t"
  81         "dec                         %1                 \n\t"
  82         "dec                         %2                 \n\t"
  83         "sub                         %0, %1             \n\t"
  84         "sub                         %0, %2             \n\t"
  85         "lea       " LOCAL_MANGLE(9b) ", %3             \n\t"
  86         "sub                         %0, %3             \n\t"
  87
  88
  89         : "=r" (fragmentA), "=r" (imm8OfPShufW1A), "=r" (imm8OfPShufW2A),
  90           "=r" (fragmentLengthA)
  91         );
  92
  93     __asm__ volatile (
  94         "jmp                         9f                 \n\t"
  95         // Begin
  96         "0:                                             \n\t"
  97         "movq    (%%"FF_REG_d", %%"FF_REG_a"), %%mm3    \n\t"
  98         "movd    (%%"FF_REG_c", %%"FF_REG_S"), %%mm0    \n\t"
  99         "punpcklbw                %%mm7, %%mm0          \n\t"
 100         "pshufw                   $0xFF, %%mm0, %%mm1   \n\t"
 101         "1:                                             \n\t"
 102         "pshufw                   $0xFF, %%mm0, %%mm0   \n\t"
 103         "2:                                             \n\t"
 104         "psubw                    %%mm1, %%mm0          \n\t"
 105         "movl   8(%%"FF_REG_b", %%"FF_REG_a"), %%esi    \n\t"
 106         "pmullw                   %%mm3, %%mm0          \n\t"
 107         "psllw                       $7, %%mm1          \n\t"
 108         "paddw                    %%mm1, %%mm0          \n\t"
 109
 110         "movq                     %%mm0, (%%"FF_REG_D", %%"FF_REG_a") \n\t"
 111
 112         "add                         $8, %%"FF_REG_a"   \n\t"
 113         // End
 114         "9:                                             \n\t"
 115         "lea       " LOCAL_MANGLE(0b) ", %0             \n\t"
 116         "lea       " LOCAL_MANGLE(1b) ", %1             \n\t"
 117         "lea       " LOCAL_MANGLE(2b) ", %2             \n\t"
 118         "dec                         %1                 \n\t"
 119         "dec                         %2                 \n\t"
 120         "sub                         %0, %1             \n\t"
 121         "sub                         %0, %2             \n\t"
 122         "lea       " LOCAL_MANGLE(9b) ", %3             \n\t"
 123         "sub                         %0, %3             \n\t"
 124
 125
 126         : "=r" (fragmentB), "=r" (imm8OfPShufW1B), "=r" (imm8OfPShufW2B),
 127           "=r" (fragmentLengthB)
 128         );
 129
 130     xpos        = 0; // lumXInc/2 - 0x8000; // difference between pixel centers
 131     fragmentPos = 0;
 132
 133     for (i = 0; i < dstW / numSplits; i++) {
 134         int xx = xpos >> 16;
 135
 136         if ((i & 3) == 0) {
 137             int a                  = 0;
 138             int b                  = ((xpos + xInc) >> 16) - xx;
 139             int c                  = ((xpos + xInc * 2) >> 16) - xx;
 140             int d                  = ((xpos + xInc * 3) >> 16) - xx;
 141             int inc                = (d + 1 < 4);
 142             uint8_t *fragment      = inc ? fragmentB : fragmentA;
 143             x86_reg imm8OfPShufW1  = inc ? imm8OfPShufW1B : imm8OfPShufW1A;
 144             x86_reg imm8OfPShufW2  = inc ? imm8OfPShufW2B : imm8OfPShufW2A;
 145             x86_reg fragmentLength = inc ? fragmentLengthB : fragmentLengthA;
 146             int maxShift           = 3 - (d + inc);
 147             int shift              = 0;
 148
 149             if (filterCode) {
 150                 filter[i]        = ((xpos              & 0xFFFF) ^ 0xFFFF) >> 9;
 151                 filter[i + 1]    = (((xpos + xInc)     & 0xFFFF) ^ 0xFFFF) >> 9;
 152                 filter[i + 2]    = (((xpos + xInc * 2) & 0xFFFF) ^ 0xFFFF) >> 9;
 153                 filter[i + 3]    = (((xpos + xInc * 3) & 0xFFFF) ^ 0xFFFF) >> 9;
 154                 filterPos[i / 2] = xx;
 155
 156                 memcpy(filterCode + fragmentPos, fragment, fragmentLength);
 157
 158                 filterCode[fragmentPos + imm8OfPShufW1] =  (a + inc)       |
 159                                                           ((b + inc) << 2) |
 160                                                           ((c + inc) << 4) |
 161                                                           ((d + inc) << 6);
 162                 filterCode[fragmentPos + imm8OfPShufW2] =  a | (b << 2) |
 163                                                                (c << 4) |
 164                                                                (d << 6);
 165
 166                 if (i + 4 - inc >= dstW)
 167                     shift = maxShift;               // avoid overread
 168                 else if ((filterPos[i / 2] & 3) <= maxShift)
 169                     shift = filterPos[i / 2] & 3;   // align
 170
 171                 if (shift && i >= shift) {
 172                     filterCode[fragmentPos + imm8OfPShufW1] += 0x55 * shift;
 173                     filterCode[fragmentPos + imm8OfPShufW2] += 0x55 * shift;
 174                     filterPos[i / 2]                        -= shift;
 175                 }
 176             }
 177
 178             fragmentPos += fragmentLength;
 179
 180             if (filterCode)
 181                 filterCode[fragmentPos] = RET;
 182         }
 183         xpos += xInc;
 184     }
 185     if (filterCode)
 186         filterPos[((i / 2) + 1) & (~1)] = xpos >> 16;  // needed to jump to the next part
 187
 188     return fragmentPos + 1;
 189 }
 190
 191 void ff_hyscale_fast_mmxext(SwsContext *c, int16_t *dst,
 192                                  int dstWidth, const uint8_t *src,
 193                                  int srcW, int xInc)
 194 {
 195     int32_t *filterPos = c->hLumFilterPos;
 196     int16_t *filter    = c->hLumFilter;
 197     void    *mmxextFilterCode = c->lumMmxextFilterCode;
 198     int i;
 199 #if ARCH_X86_64
 200     uint64_t retsave;
 201 #else
 202 #if !HAVE_EBX_AVAILABLE
 203     uint64_t ebxsave;
 204 #endif
 205 #endif
 206
 207     __asm__ volatile(
 208 #if ARCH_X86_64
 209         "mov               -8(%%rsp), %%"FF_REG_a"    \n\t"
 210         "mov            %%"FF_REG_a", %5              \n\t"  // retsave
 211 #else
 212 #if !HAVE_EBX_AVAILABLE
 213         "mov            %%"FF_REG_b", %5              \n\t"  // ebxsave
 214 #endif
 215 #endif
 216         "pxor                  %%mm7, %%mm7           \n\t"
 217         "mov                      %0, %%"FF_REG_c"    \n\t"
 218         "mov                      %1, %%"FF_REG_D"    \n\t"
 219         "mov                      %2, %%"FF_REG_d"    \n\t"
 220         "mov                      %3, %%"FF_REG_b"    \n\t"
 221         "xor            %%"FF_REG_a", %%"FF_REG_a"    \n\t" // i
 222         PREFETCH"      (%%"FF_REG_c")                 \n\t"
 223         PREFETCH"    32(%%"FF_REG_c")                 \n\t"
 224         PREFETCH"    64(%%"FF_REG_c")                 \n\t"
 225
 226 #if ARCH_X86_64
 227 #define CALL_MMXEXT_FILTER_CODE \
 228         "movl               (%%"FF_REG_b"), %%esi        \n\t"\
 229         "call                          *%4               \n\t"\
 230         "movl (%%"FF_REG_b", %%"FF_REG_a"), %%esi        \n\t"\
 231         "add                  %%"FF_REG_S", %%"FF_REG_c" \n\t"\
 232         "add                  %%"FF_REG_a", %%"FF_REG_D" \n\t"\
 233         "xor                  %%"FF_REG_a", %%"FF_REG_a" \n\t"\
 234
 235 #else
 236 #define CALL_MMXEXT_FILTER_CODE \
 237         "movl               (%%"FF_REG_b"), %%esi        \n\t"\
 238         "call                          *%4               \n\t"\
 239         "addl (%%"FF_REG_b", %%"FF_REG_a"), %%"FF_REG_c" \n\t"\
 240         "add                  %%"FF_REG_a", %%"FF_REG_D" \n\t"\
 241         "xor                  %%"FF_REG_a", %%"FF_REG_a" \n\t"\
 242
 243 #endif /* ARCH_X86_64 */
 244
 245         CALL_MMXEXT_FILTER_CODE
 246         CALL_MMXEXT_FILTER_CODE
 247         CALL_MMXEXT_FILTER_CODE
 248         CALL_MMXEXT_FILTER_CODE
 249         CALL_MMXEXT_FILTER_CODE
 250         CALL_MMXEXT_FILTER_CODE
 251         CALL_MMXEXT_FILTER_CODE
 252         CALL_MMXEXT_FILTER_CODE
 253
 254 #if ARCH_X86_64
 255         "mov                      %5, %%"FF_REG_a" \n\t"
 256         "mov            %%"FF_REG_a", -8(%%rsp)    \n\t"
 257 #else
 258 #if !HAVE_EBX_AVAILABLE
 259         "mov                      %5, %%"FF_REG_b" \n\t"
 260 #endif
 261 #endif
 262         :: "m" (src), "m" (dst), "m" (filter), "m" (filterPos),
 263            "m" (mmxextFilterCode)
 264 #if ARCH_X86_64
 265           ,"m"(retsave)
 266 #else
 267 #if !HAVE_EBX_AVAILABLE
 268           ,"m" (ebxsave)
 269 #endif
 270 #endif
 271         : "%"FF_REG_a, "%"FF_REG_c, "%"FF_REG_d, "%"FF_REG_S, "%"FF_REG_D
 272 #if ARCH_X86_64 || HAVE_EBX_AVAILABLE
 273          ,"%"FF_REG_b
 274 #endif
 275     );
 276
 277     for (i=dstWidth-1; (i*xInc)>>16 >=srcW-1; i--)
 278         dst[i] = src[srcW-1]*128;
 279 }
 280
 281 void ff_hcscale_fast_mmxext(SwsContext *c, int16_t *dst1, int16_t *dst2,
 282                                  int dstWidth, const uint8_t *src1,
 283                                  const uint8_t *src2, int srcW, int xInc)
 284 {
 285     int32_t *filterPos = c->hChrFilterPos;
 286     int16_t *filter    = c->hChrFilter;
 287     void    *mmxextFilterCode = c->chrMmxextFilterCode;
 288     int i;
 289 #if ARCH_X86_64
 290     DECLARE_ALIGNED(8, uint64_t, retsave);
 291 #else
 292 #if !HAVE_EBX_AVAILABLE
 293     DECLARE_ALIGNED(8, uint64_t, ebxsave);
 294 #endif
 295 #endif
 296     __asm__ volatile(
 297 #if ARCH_X86_64
 298         "mov          -8(%%rsp), %%"FF_REG_a"    \n\t"
 299         "mov       %%"FF_REG_a", %7              \n\t"  // retsave
 300 #else
 301 #if !HAVE_EBX_AVAILABLE
 302         "mov       %%"FF_REG_b", %7              \n\t"  // ebxsave
 303 #endif
 304 #endif
 305         "pxor             %%mm7, %%mm7           \n\t"
 306         "mov                 %0, %%"FF_REG_c"    \n\t"
 307         "mov                 %1, %%"FF_REG_D"    \n\t"
 308         "mov                 %2, %%"FF_REG_d"    \n\t"
 309         "mov                 %3, %%"FF_REG_b"    \n\t"
 310         "xor          %%"FF_REG_a", %%"FF_REG_a" \n\t" // i
 311         PREFETCH"   (%%"FF_REG_c")               \n\t"
 312         PREFETCH" 32(%%"FF_REG_c")               \n\t"
 313         PREFETCH" 64(%%"FF_REG_c")               \n\t"
 314
 315         CALL_MMXEXT_FILTER_CODE
 316         CALL_MMXEXT_FILTER_CODE
 317         CALL_MMXEXT_FILTER_CODE
 318         CALL_MMXEXT_FILTER_CODE
 319         "xor          %%"FF_REG_a", %%"FF_REG_a" \n\t" // i
 320         "mov                    %5, %%"FF_REG_c" \n\t" // src2
 321         "mov                    %6, %%"FF_REG_D" \n\t" // dst2
 322         PREFETCH"   (%%"FF_REG_c")               \n\t"
 323         PREFETCH" 32(%%"FF_REG_c")               \n\t"
 324         PREFETCH" 64(%%"FF_REG_c")               \n\t"
 325
 326         CALL_MMXEXT_FILTER_CODE
 327         CALL_MMXEXT_FILTER_CODE
 328         CALL_MMXEXT_FILTER_CODE
 329         CALL_MMXEXT_FILTER_CODE
 330
 331 #if ARCH_X86_64
 332         "mov                    %7, %%"FF_REG_a" \n\t"
 333         "mov          %%"FF_REG_a", -8(%%rsp)    \n\t"
 334 #else
 335 #if !HAVE_EBX_AVAILABLE
 336         "mov %7, %%"FF_REG_b"    \n\t"
 337 #endif
 338 #endif
 339         :: "m" (src1), "m" (dst1), "m" (filter), "m" (filterPos),
 340            "m" (mmxextFilterCode), "m" (src2), "m"(dst2)
 341 #if ARCH_X86_64
 342           ,"m"(retsave)
 343 #else
 344 #if !HAVE_EBX_AVAILABLE
 345           ,"m" (ebxsave)
 346 #endif
 347 #endif
 348         : "%"FF_REG_a, "%"FF_REG_c, "%"FF_REG_d, "%"FF_REG_S, "%"FF_REG_D
 349 #if ARCH_X86_64 || HAVE_EBX_AVAILABLE
 350          ,"%"FF_REG_b
 351 #endif
 352     );
 353
 354     for (i=dstWidth-1; (i*xInc)>>16 >=srcW-1; i--) {
 355         dst1[i] = src1[srcW-1]*128;
 356         dst2[i] = src2[srcW-1]*128;
 357     }
 358 }
 359 #endif //HAVE_INLINE_ASM