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 av_cold int ff_init_hscaler_mmxext(int dstW, int xInc, uint8_t *filterCode,
  29                                        int16_t *filter, int32_t *filterPos,
  30                                        int numSplits)
  31 {
  32     uint8_t *fragmentA;
  33     x86_reg imm8OfPShufW1A;
  34     x86_reg imm8OfPShufW2A;
  35     x86_reg fragmentLengthA;
  36     uint8_t *fragmentB;
  37     x86_reg imm8OfPShufW1B;
  38     x86_reg imm8OfPShufW2B;
  39     x86_reg fragmentLengthB;
  40     int fragmentPos;
  41
  42     int xpos, i;
  43
  44     // create an optimized horizontal scaling routine
  45     /* This scaler is made of runtime-generated MMXEXT code using specially tuned
  46      * pshufw instructions. For every four output pixels, if four input pixels
  47      * are enough for the fast bilinear scaling, then a chunk of fragmentB is
  48      * used. If five input pixels are needed, then a chunk of fragmentA is used.
  49      */
  50
  51     // code fragment
  52
  53     __asm__ volatile (
  54         "jmp                         9f                 \n\t"
  55         // Begin
  56         "0:                                             \n\t"
  57         "movq    (%%"REG_d", %%"REG_a"), %%mm3          \n\t"
  58         "movd    (%%"REG_c", %%"REG_S"), %%mm0          \n\t"
  59         "movd   1(%%"REG_c", %%"REG_S"), %%mm1          \n\t"
  60         "punpcklbw                %%mm7, %%mm1          \n\t"
  61         "punpcklbw                %%mm7, %%mm0          \n\t"
  62         "pshufw                   $0xFF, %%mm1, %%mm1   \n\t"
  63         "1:                                             \n\t"
  64         "pshufw                   $0xFF, %%mm0, %%mm0   \n\t"
  65         "2:                                             \n\t"
  66         "psubw                    %%mm1, %%mm0          \n\t"
  67         "movl   8(%%"REG_b", %%"REG_a"), %%esi          \n\t"
  68         "pmullw                   %%mm3, %%mm0          \n\t"
  69         "psllw                       $7, %%mm1          \n\t"
  70         "paddw                    %%mm1, %%mm0          \n\t"
  71
  72         "movq                     %%mm0, (%%"REG_D", %%"REG_a") \n\t"
  73
  74         "add                         $8, %%"REG_a"      \n\t"
  75         // End
  76         "9:                                             \n\t"
  77         // "int $3                                         \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    (%%"REG_d", %%"REG_a"), %%mm3          \n\t"
  98         "movd    (%%"REG_c", %%"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(%%"REG_b", %%"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, (%%"REG_D", %%"REG_a") \n\t"
 111
 112         "add                         $8, %%"REG_a"      \n\t"
 113         // End
 114         "9:                                             \n\t"
 115         // "int                       $3                   \n\t"
 116         "lea       " LOCAL_MANGLE(0b) ", %0             \n\t"
 117         "lea       " LOCAL_MANGLE(1b) ", %1             \n\t"
 118         "lea       " LOCAL_MANGLE(2b) ", %2             \n\t"
 119         "dec                         %1                 \n\t"
 120         "dec                         %2                 \n\t"
 121         "sub                         %0, %1             \n\t"
 122         "sub                         %0, %2             \n\t"
 123         "lea       " LOCAL_MANGLE(9b) ", %3             \n\t"
 124         "sub                         %0, %3             \n\t"
 125
 126
 127         : "=r" (fragmentB), "=r" (imm8OfPShufW1B), "=r" (imm8OfPShufW2B),
 128           "=r" (fragmentLengthB)
 129         );
 130
 131     xpos        = 0; // lumXInc/2 - 0x8000; // difference between pixel centers
 132     fragmentPos = 0;
 133
 134     for (i = 0; i < dstW / numSplits; i++) {
 135         int xx = xpos >> 16;
 136
 137         if ((i & 3) == 0) {
 138             int a                  = 0;
 139             int b                  = ((xpos + xInc) >> 16) - xx;
 140             int c                  = ((xpos + xInc * 2) >> 16) - xx;
 141             int d                  = ((xpos + xInc * 3) >> 16) - xx;
 142             int inc                = (d + 1 < 4);
 143             uint8_t *fragment      = inc ? fragmentB : fragmentA;
 144             x86_reg imm8OfPShufW1  = inc ? imm8OfPShufW1B : imm8OfPShufW1A;
 145             x86_reg imm8OfPShufW2  = inc ? imm8OfPShufW2B : imm8OfPShufW2A;
 146             x86_reg fragmentLength = inc ? fragmentLengthB : fragmentLengthA;
 147             int maxShift           = 3 - (d + inc);
 148             int shift              = 0;
 149
 150             if (filterCode) {
 151                 filter[i]        = ((xpos              & 0xFFFF) ^ 0xFFFF) >> 9;
 152                 filter[i + 1]    = (((xpos + xInc)     & 0xFFFF) ^ 0xFFFF) >> 9;
 153                 filter[i + 2]    = (((xpos + xInc * 2) & 0xFFFF) ^ 0xFFFF) >> 9;
 154                 filter[i + 3]    = (((xpos + xInc * 3) & 0xFFFF) ^ 0xFFFF) >> 9;
 155                 filterPos[i / 2] = xx;
 156
 157                 memcpy(filterCode + fragmentPos, fragment, fragmentLength);
 158
 159                 filterCode[fragmentPos + imm8OfPShufW1] =  (a + inc)       |
 160                                                           ((b + inc) << 2) |
 161                                                           ((c + inc) << 4) |
 162                                                           ((d + inc) << 6);
 163                 filterCode[fragmentPos + imm8OfPShufW2] =  a | (b << 2) |
 164                                                                (c << 4) |
 165                                                                (d << 6);
 166
 167                 if (i + 4 - inc >= dstW)
 168                     shift = maxShift;               // avoid overread
 169                 else if ((filterPos[i / 2] & 3) <= maxShift)
 170                     shift = filterPos[i / 2] & 3;   // align
 171
 172                 if (shift && i >= shift) {
 173                     filterCode[fragmentPos + imm8OfPShufW1] += 0x55 * shift;
 174                     filterCode[fragmentPos + imm8OfPShufW2] += 0x55 * shift;
 175                     filterPos[i / 2]                        -= shift;
 176                 }
 177             }
 178
 179             fragmentPos += fragmentLength;
 180
 181             if (filterCode)
 182                 filterCode[fragmentPos] = RET;
 183         }
 184         xpos += xInc;
 185     }
 186     if (filterCode)
 187         filterPos[((i / 2) + 1) & (~1)] = xpos >> 16;  // needed to jump to the next part
 188
 189     return fragmentPos + 1;
 190 }
 191
 192 void ff_hyscale_fast_mmxext(SwsContext *c, int16_t *dst,
 193                                  int dstWidth, const uint8_t *src,
 194                                  int srcW, int xInc)
 195 {
 196     int32_t *filterPos = c->hLumFilterPos;
 197     int16_t *filter    = c->hLumFilter;
 198     void    *mmxextFilterCode = c->lumMmxextFilterCode;
 199     int i;
 200 #if defined(PIC)
 201     uint64_t ebxsave;
 202 #endif
 203 #if ARCH_X86_64
 204     uint64_t retsave;
 205 #endif
 206
 207     __asm__ volatile(
 208 #if defined(PIC)
 209         "mov               %%"REG_b", %5        \n\t"
 210 #if ARCH_X86_64
 211         "mov               -8(%%rsp), %%"REG_a" \n\t"
 212         "mov               %%"REG_a", %6        \n\t"
 213 #endif
 214 #else
 215 #if ARCH_X86_64
 216         "mov               -8(%%rsp), %%"REG_a" \n\t"
 217         "mov               %%"REG_a", %5        \n\t"
 218 #endif
 219 #endif
 220         "pxor                  %%mm7, %%mm7     \n\t"
 221         "mov                      %0, %%"REG_c" \n\t"
 222         "mov                      %1, %%"REG_D" \n\t"
 223         "mov                      %2, %%"REG_d" \n\t"
 224         "mov                      %3, %%"REG_b" \n\t"
 225         "xor               %%"REG_a", %%"REG_a" \n\t" // i
 226         PREFETCH"        (%%"REG_c")            \n\t"
 227         PREFETCH"      32(%%"REG_c")            \n\t"
 228         PREFETCH"      64(%%"REG_c")            \n\t"
 229
 230 #if ARCH_X86_64
 231 #define CALL_MMXEXT_FILTER_CODE \
 232         "movl            (%%"REG_b"), %%esi     \n\t"\
 233         "call                    *%4            \n\t"\
 234         "movl (%%"REG_b", %%"REG_a"), %%esi     \n\t"\
 235         "add               %%"REG_S", %%"REG_c" \n\t"\
 236         "add               %%"REG_a", %%"REG_D" \n\t"\
 237         "xor               %%"REG_a", %%"REG_a" \n\t"\
 238
 239 #else
 240 #define CALL_MMXEXT_FILTER_CODE \
 241         "movl (%%"REG_b"), %%esi        \n\t"\
 242         "call         *%4                       \n\t"\
 243         "addl (%%"REG_b", %%"REG_a"), %%"REG_c" \n\t"\
 244         "add               %%"REG_a", %%"REG_D" \n\t"\
 245         "xor               %%"REG_a", %%"REG_a" \n\t"\
 246
 247 #endif /* ARCH_X86_64 */
 248
 249         CALL_MMXEXT_FILTER_CODE
 250         CALL_MMXEXT_FILTER_CODE
 251         CALL_MMXEXT_FILTER_CODE
 252         CALL_MMXEXT_FILTER_CODE
 253         CALL_MMXEXT_FILTER_CODE
 254         CALL_MMXEXT_FILTER_CODE
 255         CALL_MMXEXT_FILTER_CODE
 256         CALL_MMXEXT_FILTER_CODE
 257
 258 #if defined(PIC)
 259         "mov                      %5, %%"REG_b" \n\t"
 260 #if ARCH_X86_64
 261         "mov                      %6, %%"REG_a" \n\t"
 262         "mov               %%"REG_a", -8(%%rsp) \n\t"
 263 #endif
 264 #else
 265 #if ARCH_X86_64
 266         "mov                      %5, %%"REG_a" \n\t"
 267         "mov               %%"REG_a", -8(%%rsp) \n\t"
 268 #endif
 269 #endif
 270         :: "m" (src), "m" (dst), "m" (filter), "m" (filterPos),
 271            "m" (mmxextFilterCode)
 272 #if defined(PIC)
 273           ,"m" (ebxsave)
 274 #endif
 275 #if ARCH_X86_64
 276           ,"m"(retsave)
 277 #endif
 278         : "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_S, "%"REG_D
 279 #if !defined(PIC)
 280          ,"%"REG_b
 281 #endif
 282     );
 283
 284     for (i=dstWidth-1; (i*xInc)>>16 >=srcW-1; i--)
 285         dst[i] = src[srcW-1]*128;
 286 }
 287
 288 void ff_hcscale_fast_mmxext(SwsContext *c, int16_t *dst1, int16_t *dst2,
 289                                  int dstWidth, const uint8_t *src1,
 290                                  const uint8_t *src2, int srcW, int xInc)
 291 {
 292     int32_t *filterPos = c->hChrFilterPos;
 293     int16_t *filter    = c->hChrFilter;
 294     void    *mmxextFilterCode = c->chrMmxextFilterCode;
 295     int i;
 296 #if defined(PIC)
 297     DECLARE_ALIGNED(8, uint64_t, ebxsave);
 298 #endif
 299 #if ARCH_X86_64
 300     DECLARE_ALIGNED(8, uint64_t, retsave);
 301 #endif
 302
 303     __asm__ volatile(
 304 #if defined(PIC)
 305         "mov          %%"REG_b", %7         \n\t"
 306 #if ARCH_X86_64
 307         "mov          -8(%%rsp), %%"REG_a"  \n\t"
 308         "mov          %%"REG_a", %8         \n\t"
 309 #endif
 310 #else
 311 #if ARCH_X86_64
 312         "mov          -8(%%rsp), %%"REG_a"  \n\t"
 313         "mov          %%"REG_a", %7         \n\t"
 314 #endif
 315 #endif
 316         "pxor             %%mm7, %%mm7      \n\t"
 317         "mov                 %0, %%"REG_c"  \n\t"
 318         "mov                 %1, %%"REG_D"  \n\t"
 319         "mov                 %2, %%"REG_d"  \n\t"
 320         "mov                 %3, %%"REG_b"  \n\t"
 321         "xor          %%"REG_a", %%"REG_a"  \n\t" // i
 322         PREFETCH"   (%%"REG_c")             \n\t"
 323         PREFETCH" 32(%%"REG_c")             \n\t"
 324         PREFETCH" 64(%%"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         "xor          %%"REG_a", %%"REG_a"  \n\t" // i
 331         "mov                 %5, %%"REG_c"  \n\t" // src
 332         "mov                 %6, %%"REG_D"  \n\t" // buf2
 333         PREFETCH"   (%%"REG_c")             \n\t"
 334         PREFETCH" 32(%%"REG_c")             \n\t"
 335         PREFETCH" 64(%%"REG_c")             \n\t"
 336
 337         CALL_MMXEXT_FILTER_CODE
 338         CALL_MMXEXT_FILTER_CODE
 339         CALL_MMXEXT_FILTER_CODE
 340         CALL_MMXEXT_FILTER_CODE
 341
 342 #if defined(PIC)
 343         "mov %7, %%"REG_b"    \n\t"
 344 #if ARCH_X86_64
 345         "mov                 %8, %%"REG_a"  \n\t"
 346         "mov          %%"REG_a", -8(%%rsp)  \n\t"
 347 #endif
 348 #else
 349 #if ARCH_X86_64
 350         "mov                 %7, %%"REG_a"  \n\t"
 351         "mov          %%"REG_a", -8(%%rsp)  \n\t"
 352 #endif
 353 #endif
 354         :: "m" (src1), "m" (dst1), "m" (filter), "m" (filterPos),
 355            "m" (mmxextFilterCode), "m" (src2), "m"(dst2)
 356 #if defined(PIC)
 357           ,"m" (ebxsave)
 358 #endif
 359 #if ARCH_X86_64
 360           ,"m"(retsave)
 361 #endif
 362         : "%"REG_a, "%"REG_c, "%"REG_d, "%"REG_S, "%"REG_D
 363 #if !defined(PIC)
 364          ,"%"REG_b
 365 #endif
 366     );
 367
 368     for (i=dstWidth-1; (i*xInc)>>16 >=srcW-1; i--) {
 369         dst1[i] = src1[srcW-1]*128;
 370         dst2[i] = src2[srcW-1]*128;
 371     }
 372 }