Cinelerra: hvirtual/quicktime/ffmpeg/libavcodec/liba52/resample

00001 
00002 // MMX optimizations from Michael Niedermayer (michaelni@gmx.at) (under GPL)
00003 
00004 /* optimization TODO / NOTES 
00005     movntq is slightly faster (0.5% with the current test.c benchmark) 
00006         (but thats just test.c so that needs to be testd in reallity)
00007         and it would mean (C / MMX2 / MMX / 3DNOW) versions 
00008 */
00009 
00010 static uint64_t __attribute__((aligned(8))) attribute_used magicF2W= 0x43c0000043c00000LL;
00011 static uint64_t __attribute__((aligned(8))) attribute_used wm1010= 0xFFFF0000FFFF0000LL;
00012 static uint64_t __attribute__((aligned(8))) attribute_used wm0101= 0x0000FFFF0000FFFFLL;
00013 static uint64_t __attribute__((aligned(8))) attribute_used wm1100= 0xFFFFFFFF00000000LL;
00014 
00015 static int a52_resample_MONO_to_5_MMX(float * _f, int16_t * s16){
00016     int32_t * f = (int32_t *) _f;
00017         asm volatile(
00018                 "movl $-512, %%esi              \n\t"
00019                 "movq "MANGLE(magicF2W)", %%mm7 \n\t"
00020                 "movq "MANGLE(wm1100)", %%mm3   \n\t"
00021                 "movq "MANGLE(wm0101)", %%mm4   \n\t"
00022                 "movq "MANGLE(wm1010)", %%mm5   \n\t"
00023                 "pxor %%mm6, %%mm6              \n\t"
00024                 "1:                             \n\t"
00025                 "movq (%1, %%esi, 2), %%mm0     \n\t"
00026                 "movq 8(%1, %%esi, 2), %%mm1    \n\t"
00027                 "leal (%%esi, %%esi, 4), %%edi  \n\t"
00028                 "psubd %%mm7, %%mm0             \n\t"
00029                 "psubd %%mm7, %%mm1             \n\t"
00030                 "packssdw %%mm1, %%mm0          \n\t"
00031                 "movq %%mm0, %%mm1              \n\t"
00032                 "pand %%mm4, %%mm0              \n\t"
00033                 "pand %%mm5, %%mm1              \n\t"
00034                 "movq %%mm6, (%0, %%edi)        \n\t" // 0 0 0 0
00035                 "movd %%mm0, 8(%0, %%edi)       \n\t" // A 0
00036                 "pand %%mm3, %%mm0              \n\t"
00037                 "movd %%mm6, 12(%0, %%edi)      \n\t" // 0 0
00038                 "movd %%mm1, 16(%0, %%edi)      \n\t" // 0 B
00039                 "pand %%mm3, %%mm1              \n\t"
00040                 "movd %%mm6, 20(%0, %%edi)      \n\t" // 0 0
00041                 "movq %%mm0, 24(%0, %%edi)      \n\t" // 0 0 C 0
00042                 "movq %%mm1, 32(%0, %%edi)      \n\t" // 0 0 0 B
00043                 "addl $8, %%esi                 \n\t"
00044                 " jnz 1b                        \n\t"
00045                 "emms                           \n\t"
00046                 :: "r" (s16+1280), "r" (f+256)
00047                 :"%esi", "%edi", "memory"
00048         );
00049     return 5*256;
00050 }
00051 
00052 static int a52_resample_STEREO_to_2_MMX(float * _f, int16_t * s16){
00053     int32_t * f = (int32_t *) _f;
00054 /* benchmark scores are 0.3% better with SSE but we would need to set bias=0 and premultiply it
00055 #ifdef HAVE_SSE
00056         asm volatile(
00057                 "movl $-1024, %%esi             \n\t"
00058                 "1:                             \n\t"
00059                 "cvtps2pi (%1, %%esi), %%mm0    \n\t"
00060                 "cvtps2pi 1024(%1, %%esi), %%mm2\n\t"
00061                 "movq %%mm0, %%mm1              \n\t"
00062                 "punpcklwd %%mm2, %%mm0         \n\t"
00063                 "punpckhwd %%mm2, %%mm1         \n\t"
00064                 "movq %%mm0, (%0, %%esi)        \n\t"
00065                 "movq %%mm1, 8(%0, %%esi)       \n\t"
00066                 "addl $16, %%esi                \n\t"
00067                 " jnz 1b                        \n\t"
00068                 "emms                           \n\t"
00069                 :: "r" (s16+512), "r" (f+256)
00070                 :"%esi", "memory"
00071         );*/
00072         asm volatile(
00073                 "movl $-1024, %%esi             \n\t"
00074                 "movq "MANGLE(magicF2W)", %%mm7 \n\t"
00075                 "1:                             \n\t"
00076                 "movq (%1, %%esi), %%mm0        \n\t"
00077                 "movq 8(%1, %%esi), %%mm1       \n\t"
00078                 "movq 1024(%1, %%esi), %%mm2    \n\t"
00079                 "movq 1032(%1, %%esi), %%mm3    \n\t"
00080                 "psubd %%mm7, %%mm0             \n\t"
00081                 "psubd %%mm7, %%mm1             \n\t"
00082                 "psubd %%mm7, %%mm2             \n\t"
00083                 "psubd %%mm7, %%mm3             \n\t"
00084                 "packssdw %%mm1, %%mm0          \n\t"
00085                 "packssdw %%mm3, %%mm2          \n\t"
00086                 "movq %%mm0, %%mm1              \n\t"
00087                 "punpcklwd %%mm2, %%mm0         \n\t"
00088                 "punpckhwd %%mm2, %%mm1         \n\t"
00089                 "movq %%mm0, (%0, %%esi)        \n\t"
00090                 "movq %%mm1, 8(%0, %%esi)       \n\t"
00091                 "addl $16, %%esi                \n\t"
00092                 " jnz 1b                        \n\t"
00093                 "emms                           \n\t"
00094                 :: "r" (s16+512), "r" (f+256)
00095                 :"%esi", "memory"
00096         );
00097     return 2*256;
00098 }
00099 
00100 static int a52_resample_3F_to_5_MMX(float * _f, int16_t * s16){
00101     int32_t * f = (int32_t *) _f;
00102         asm volatile(
00103                 "movl $-1024, %%esi             \n\t"
00104                 "movq "MANGLE(magicF2W)", %%mm7 \n\t"
00105                 "pxor %%mm6, %%mm6              \n\t"
00106                 "movq %%mm7, %%mm5              \n\t"
00107                 "punpckldq %%mm6, %%mm5         \n\t"
00108                 "1:                             \n\t"
00109                 "movd (%1, %%esi), %%mm0        \n\t"
00110                 "punpckldq 2048(%1, %%esi), %%mm0\n\t"
00111                 "movd 1024(%1, %%esi), %%mm1    \n\t"
00112                 "punpckldq 4(%1, %%esi), %%mm1  \n\t"
00113                 "movd 2052(%1, %%esi), %%mm2    \n\t"
00114                 "movq %%mm7, %%mm3              \n\t"
00115                 "punpckldq 1028(%1, %%esi), %%mm3\n\t"
00116                 "movd 8(%1, %%esi), %%mm4       \n\t"
00117                 "punpckldq 2056(%1, %%esi), %%mm4\n\t"
00118                 "leal (%%esi, %%esi, 4), %%edi  \n\t"
00119                 "sarl $1, %%edi                 \n\t"
00120                 "psubd %%mm7, %%mm0             \n\t"
00121                 "psubd %%mm7, %%mm1             \n\t"
00122                 "psubd %%mm5, %%mm2             \n\t"
00123                 "psubd %%mm7, %%mm3             \n\t"
00124                 "psubd %%mm7, %%mm4             \n\t"
00125                 "packssdw %%mm6, %%mm0          \n\t"
00126                 "packssdw %%mm2, %%mm1          \n\t"
00127                 "packssdw %%mm4, %%mm3          \n\t"
00128                 "movq %%mm0, (%0, %%edi)        \n\t"
00129                 "movq %%mm1, 8(%0, %%edi)       \n\t"
00130                 "movq %%mm3, 16(%0, %%edi)      \n\t"
00131                 
00132                 "movd 1032(%1, %%esi), %%mm1    \n\t"
00133                 "punpckldq 12(%1, %%esi), %%mm1\n\t"
00134                 "movd 2060(%1, %%esi), %%mm2    \n\t"
00135                 "movq %%mm7, %%mm3              \n\t"
00136                 "punpckldq 1036(%1, %%esi), %%mm3\n\t"
00137                 "pxor %%mm0, %%mm0              \n\t"
00138                 "psubd %%mm7, %%mm1             \n\t"
00139                 "psubd %%mm5, %%mm2             \n\t"
00140                 "psubd %%mm7, %%mm3             \n\t"
00141                 "packssdw %%mm1, %%mm0          \n\t"
00142                 "packssdw %%mm3, %%mm2          \n\t"
00143                 "movq %%mm0, 24(%0, %%edi)      \n\t"
00144                 "movq %%mm2, 32(%0, %%edi)      \n\t"
00145                                 
00146                 "addl $16, %%esi                \n\t"
00147                 " jnz 1b                        \n\t"
00148                 "emms                           \n\t"
00149                 :: "r" (s16+1280), "r" (f+256)
00150                 :"%esi", "%edi", "memory"
00151         );
00152     return 5*256;
00153 }
00154 
00155 static int a52_resample_2F_2R_to_4_MMX(float * _f, int16_t * s16){
00156     int32_t * f = (int32_t *) _f;
00157         asm volatile(
00158                 "movl $-1024, %%esi             \n\t"
00159                 "movq "MANGLE(magicF2W)", %%mm7 \n\t"
00160                 "1:                             \n\t"
00161                 "movq (%1, %%esi), %%mm0        \n\t"
00162                 "movq 8(%1, %%esi), %%mm1       \n\t"
00163                 "movq 1024(%1, %%esi), %%mm2    \n\t"
00164                 "movq 1032(%1, %%esi), %%mm3    \n\t"
00165                 "psubd %%mm7, %%mm0             \n\t"
00166                 "psubd %%mm7, %%mm1             \n\t"
00167                 "psubd %%mm7, %%mm2             \n\t"
00168                 "psubd %%mm7, %%mm3             \n\t"
00169                 "packssdw %%mm1, %%mm0          \n\t"
00170                 "packssdw %%mm3, %%mm2          \n\t"
00171                 "movq 2048(%1, %%esi), %%mm3    \n\t"
00172                 "movq 2056(%1, %%esi), %%mm4    \n\t"
00173                 "movq 3072(%1, %%esi), %%mm5    \n\t"
00174                 "movq 3080(%1, %%esi), %%mm6    \n\t"
00175                 "psubd %%mm7, %%mm3             \n\t"
00176                 "psubd %%mm7, %%mm4             \n\t"
00177                 "psubd %%mm7, %%mm5             \n\t"
00178                 "psubd %%mm7, %%mm6             \n\t"
00179                 "packssdw %%mm4, %%mm3          \n\t"
00180                 "packssdw %%mm6, %%mm5          \n\t"
00181                 "movq %%mm0, %%mm1              \n\t"
00182                 "movq %%mm3, %%mm4              \n\t"
00183                 "punpcklwd %%mm2, %%mm0         \n\t"
00184                 "punpckhwd %%mm2, %%mm1         \n\t"
00185                 "punpcklwd %%mm5, %%mm3         \n\t"
00186                 "punpckhwd %%mm5, %%mm4         \n\t"
00187                 "movq %%mm0, %%mm2              \n\t"
00188                 "movq %%mm1, %%mm5              \n\t"
00189                 "punpckldq %%mm3, %%mm0         \n\t"
00190                 "punpckhdq %%mm3, %%mm2         \n\t"
00191                 "punpckldq %%mm4, %%mm1         \n\t"
00192                 "punpckhdq %%mm4, %%mm5         \n\t"
00193                 "movq %%mm0, (%0, %%esi,2)      \n\t"
00194                 "movq %%mm2, 8(%0, %%esi,2)     \n\t"
00195                 "movq %%mm1, 16(%0, %%esi,2)    \n\t"
00196                 "movq %%mm5, 24(%0, %%esi,2)    \n\t"
00197                 "addl $16, %%esi                \n\t"
00198                 " jnz 1b                        \n\t"
00199                 "emms                           \n\t"
00200                 :: "r" (s16+1024), "r" (f+256)
00201                 :"%esi", "memory"
00202         );
00203     return 4*256;
00204 }
00205 
00206 static int a52_resample_3F_2R_to_5_MMX(float * _f, int16_t * s16){
00207     int32_t * f = (int32_t *) _f;
00208         asm volatile(
00209                 "movl $-1024, %%esi             \n\t"
00210                 "movq "MANGLE(magicF2W)", %%mm7 \n\t"
00211                 "1:                             \n\t"
00212                 "movd (%1, %%esi), %%mm0        \n\t"
00213                 "punpckldq 2048(%1, %%esi), %%mm0\n\t"
00214                 "movd 3072(%1, %%esi), %%mm1    \n\t"
00215                 "punpckldq 4096(%1, %%esi), %%mm1\n\t"
00216                 "movd 1024(%1, %%esi), %%mm2    \n\t"
00217                 "punpckldq 4(%1, %%esi), %%mm2  \n\t"
00218                 "movd 2052(%1, %%esi), %%mm3    \n\t"
00219                 "punpckldq 3076(%1, %%esi), %%mm3\n\t"
00220                 "movd 4100(%1, %%esi), %%mm4    \n\t"
00221                 "punpckldq 1028(%1, %%esi), %%mm4\n\t"
00222                 "movd 8(%1, %%esi), %%mm5       \n\t"
00223                 "punpckldq 2056(%1, %%esi), %%mm5\n\t"
00224                 "leal (%%esi, %%esi, 4), %%edi  \n\t"
00225                 "sarl $1, %%edi                 \n\t"
00226                 "psubd %%mm7, %%mm0             \n\t"
00227                 "psubd %%mm7, %%mm1             \n\t"
00228                 "psubd %%mm7, %%mm2             \n\t"
00229                 "psubd %%mm7, %%mm3             \n\t"
00230                 "psubd %%mm7, %%mm4             \n\t"
00231                 "psubd %%mm7, %%mm5             \n\t"
00232                 "packssdw %%mm1, %%mm0          \n\t"
00233                 "packssdw %%mm3, %%mm2          \n\t"
00234                 "packssdw %%mm5, %%mm4          \n\t"
00235                 "movq %%mm0, (%0, %%edi)        \n\t"
00236                 "movq %%mm2, 8(%0, %%edi)       \n\t"
00237                 "movq %%mm4, 16(%0, %%edi)      \n\t"
00238                 
00239                 "movd 3080(%1, %%esi), %%mm0    \n\t"
00240                 "punpckldq 4104(%1, %%esi), %%mm0\n\t"
00241                 "movd 1032(%1, %%esi), %%mm1    \n\t"
00242                 "punpckldq 12(%1, %%esi), %%mm1\n\t"
00243                 "movd 2060(%1, %%esi), %%mm2    \n\t"
00244                 "punpckldq 3084(%1, %%esi), %%mm2\n\t"
00245                 "movd 4108(%1, %%esi), %%mm3    \n\t"
00246                 "punpckldq 1036(%1, %%esi), %%mm3\n\t"
00247                 "psubd %%mm7, %%mm0             \n\t"
00248                 "psubd %%mm7, %%mm1             \n\t"
00249                 "psubd %%mm7, %%mm2             \n\t"
00250                 "psubd %%mm7, %%mm3             \n\t"
00251                 "packssdw %%mm1, %%mm0          \n\t"
00252                 "packssdw %%mm3, %%mm2          \n\t"
00253                 "movq %%mm0, 24(%0, %%edi)      \n\t"
00254                 "movq %%mm2, 32(%0, %%edi)      \n\t"
00255                                 
00256                 "addl $16, %%esi                \n\t"
00257                 " jnz 1b                        \n\t"
00258                 "emms                           \n\t"
00259                 :: "r" (s16+1280), "r" (f+256)
00260                 :"%esi", "%edi", "memory"
00261         );
00262     return 5*256;
00263 }
00264 
00265 static int a52_resample_MONO_LFE_to_6_MMX(float * _f, int16_t * s16){
00266     int32_t * f = (int32_t *) _f;
00267         asm volatile(
00268                 "movl $-1024, %%esi             \n\t"
00269                 "movq "MANGLE(magicF2W)", %%mm7 \n\t"
00270                 "pxor %%mm6, %%mm6              \n\t"
00271                 "1:                             \n\t"
00272                 "movq 1024(%1, %%esi), %%mm0    \n\t"
00273                 "movq 1032(%1, %%esi), %%mm1    \n\t"
00274                 "movq (%1, %%esi), %%mm2        \n\t"
00275                 "movq 8(%1, %%esi), %%mm3       \n\t"
00276                 "psubd %%mm7, %%mm0             \n\t"
00277                 "psubd %%mm7, %%mm1             \n\t"
00278                 "psubd %%mm7, %%mm2             \n\t"
00279                 "psubd %%mm7, %%mm3             \n\t"
00280                 "packssdw %%mm1, %%mm0          \n\t"
00281                 "packssdw %%mm3, %%mm2          \n\t"
00282                 "movq %%mm0, %%mm1              \n\t"
00283                 "punpcklwd %%mm2, %%mm0         \n\t"
00284                 "punpckhwd %%mm2, %%mm1         \n\t"
00285                 "leal (%%esi, %%esi, 2), %%edi  \n\t"
00286                 "movq %%mm6, (%0, %%edi)        \n\t"
00287                 "movd %%mm0, 8(%0, %%edi)       \n\t"
00288                 "punpckhdq %%mm0, %%mm0         \n\t"
00289                 "movq %%mm6, 12(%0, %%edi)      \n\t"
00290                 "movd %%mm0, 20(%0, %%edi)      \n\t"
00291                 "movq %%mm6, 24(%0, %%edi)      \n\t"
00292                 "movd %%mm1, 32(%0, %%edi)      \n\t"
00293                 "punpckhdq %%mm1, %%mm1         \n\t"
00294                 "movq %%mm6, 36(%0, %%edi)      \n\t"
00295                 "movd %%mm1, 44(%0, %%edi)      \n\t"
00296                 "addl $16, %%esi                \n\t"
00297                 " jnz 1b                        \n\t"
00298                 "emms                           \n\t"
00299                 :: "r" (s16+1536), "r" (f+256)
00300                 :"%esi", "%edi", "memory"
00301         );
00302     return 6*256;
00303 }
00304 
00305 static int a52_resample_STEREO_LFE_to_6_MMX(float * _f, int16_t * s16){
00306     int32_t * f = (int32_t *) _f;
00307         asm volatile(
00308                 "movl $-1024, %%esi             \n\t"
00309                 "movq "MANGLE(magicF2W)", %%mm7 \n\t"
00310                 "pxor %%mm6, %%mm6              \n\t"
00311                 "1:                             \n\t"
00312                 "movq 1024(%1, %%esi), %%mm0    \n\t"
00313                 "movq 2048(%1, %%esi), %%mm1    \n\t"
00314                 "movq (%1, %%esi), %%mm5        \n\t" 
00315                 "psubd %%mm7, %%mm0             \n\t"
00316                 "psubd %%mm7, %%mm1             \n\t"
00317                 "psubd %%mm7, %%mm5             \n\t"
00318                 "leal (%%esi, %%esi, 2), %%edi  \n\t"
00319                 
00320                 "pxor %%mm4, %%mm4              \n\t"
00321                 "packssdw %%mm5, %%mm0          \n\t" // FfAa
00322                 "packssdw %%mm4, %%mm1          \n\t" // 00Bb
00323                 "punpckhwd %%mm0, %%mm4         \n\t" // F0f0
00324                 "punpcklwd %%mm1, %%mm0         \n\t" // BAba
00325                 "movq %%mm0, %%mm1              \n\t" // BAba
00326                 "punpckldq %%mm4, %%mm3         \n\t" // f0XX
00327                 "punpckldq %%mm6, %%mm0         \n\t" // 00ba
00328                 "punpckhdq %%mm1, %%mm3         \n\t" // BAf0
00329                 
00330                 "movq %%mm0, (%0, %%edi)        \n\t" // 00ba
00331                 "punpckhdq %%mm4, %%mm0         \n\t" // F000
00332                 "movq %%mm3, 8(%0, %%edi)       \n\t" // BAf0
00333                 "movq %%mm0, 16(%0, %%edi)      \n\t" // F000
00334                 "addl $8, %%esi                 \n\t"
00335                 " jnz 1b                        \n\t"
00336                 "emms                           \n\t"
00337                 :: "r" (s16+1536), "r" (f+256)
00338                 :"%esi", "%edi", "memory"
00339         );
00340     return 6*256;
00341 }
00342 
00343 static int a52_resample_3F_LFE_to_6_MMX(float * _f, int16_t * s16){
00344     int32_t * f = (int32_t *) _f;
00345         asm volatile(
00346                 "movl $-1024, %%esi             \n\t"
00347                 "movq "MANGLE(magicF2W)", %%mm7 \n\t"
00348                 "pxor %%mm6, %%mm6              \n\t"
00349                 "1:                             \n\t"
00350                 "movq 1024(%1, %%esi), %%mm0    \n\t"
00351                 "movq 3072(%1, %%esi), %%mm1    \n\t"
00352                 "movq 2048(%1, %%esi), %%mm4    \n\t"
00353                 "movq (%1, %%esi), %%mm5        \n\t" 
00354                 "psubd %%mm7, %%mm0             \n\t"
00355                 "psubd %%mm7, %%mm1             \n\t"
00356                 "psubd %%mm7, %%mm4             \n\t"
00357                 "psubd %%mm7, %%mm5             \n\t"
00358                 "leal (%%esi, %%esi, 2), %%edi  \n\t"
00359                 
00360                 "packssdw %%mm4, %%mm0          \n\t" // EeAa
00361                 "packssdw %%mm5, %%mm1          \n\t" // FfBb
00362                 "movq %%mm0, %%mm2              \n\t" // EeAa
00363                 "punpcklwd %%mm1, %%mm0         \n\t" // BAba
00364                 "punpckhwd %%mm1, %%mm2         \n\t" // FEfe
00365                 "movq %%mm0, %%mm1              \n\t" // BAba
00366                 "punpckldq %%mm6, %%mm0         \n\t" // 00ba
00367                 "punpckhdq %%mm1, %%mm1         \n\t" // BABA
00368                 
00369                 "movq %%mm0, (%0, %%edi)        \n\t"
00370                 "punpckhdq %%mm2, %%mm0         \n\t" // FE00
00371                 "punpckldq %%mm1, %%mm2         \n\t" // BAfe
00372                 "movq %%mm2, 8(%0, %%edi)       \n\t"
00373                 "movq %%mm0, 16(%0, %%edi)      \n\t"
00374                 "addl $8, %%esi                 \n\t"
00375                 " jnz 1b                        \n\t"
00376                 "emms                           \n\t"
00377                 :: "r" (s16+1536), "r" (f+256)
00378                 :"%esi", "%edi", "memory"
00379         );
00380     return 6*256;
00381 }
00382 
00383 static int a52_resample_2F_2R_LFE_to_6_MMX(float * _f, int16_t * s16){
00384     int32_t * f = (int32_t *) _f;
00385         asm volatile(
00386                 "movl $-1024, %%esi             \n\t"
00387                 "movq "MANGLE(magicF2W)", %%mm7 \n\t"
00388 //              "pxor %%mm6, %%mm6              \n\t"
00389                 "1:                             \n\t"
00390                 "movq 1024(%1, %%esi), %%mm0    \n\t"
00391                 "movq 2048(%1, %%esi), %%mm1    \n\t"
00392                 "movq 3072(%1, %%esi), %%mm2    \n\t"
00393                 "movq 4096(%1, %%esi), %%mm3    \n\t"
00394                 "movq (%1, %%esi), %%mm5        \n\t" 
00395                 "psubd %%mm7, %%mm0             \n\t"
00396                 "psubd %%mm7, %%mm1             \n\t"
00397                 "psubd %%mm7, %%mm2             \n\t"
00398                 "psubd %%mm7, %%mm3             \n\t"
00399                 "psubd %%mm7, %%mm5             \n\t"
00400                 "leal (%%esi, %%esi, 2), %%edi  \n\t"
00401                 
00402                 "packssdw %%mm2, %%mm0          \n\t" // CcAa
00403                 "packssdw %%mm3, %%mm1          \n\t" // DdBb
00404                 "packssdw %%mm5, %%mm5          \n\t" // FfFf
00405                 "movq %%mm0, %%mm2              \n\t" // CcAa
00406                 "punpcklwd %%mm1, %%mm0         \n\t" // BAba
00407                 "punpckhwd %%mm1, %%mm2         \n\t" // DCdc
00408                 "pxor %%mm4, %%mm4              \n\t" // 0000
00409                 "punpcklwd %%mm5, %%mm4         \n\t" // F0f0
00410                 "movq %%mm0, %%mm1              \n\t" // BAba
00411                 "movq %%mm4, %%mm3              \n\t" // F0f0
00412                 "punpckldq %%mm2, %%mm0         \n\t" // dcba
00413                 "punpckhdq %%mm1, %%mm1         \n\t" // BABA
00414                 "punpckldq %%mm1, %%mm4         \n\t" // BAf0
00415                 "punpckhdq %%mm3, %%mm2         \n\t" // F0DC
00416                 
00417                 "movq %%mm0, (%0, %%edi)        \n\t"
00418                 "movq %%mm4, 8(%0, %%edi)       \n\t"
00419                 "movq %%mm2, 16(%0, %%edi)      \n\t"
00420                 "addl $8, %%esi                 \n\t"
00421                 " jnz 1b                        \n\t"
00422                 "emms                           \n\t"
00423                 :: "r" (s16+1536), "r" (f+256)
00424                 :"%esi", "%edi", "memory"
00425         );
00426     return 6*256;
00427 }
00428 
00429 static int a52_resample_3F_2R_LFE_to_6_MMX(float * _f, int16_t * s16){
00430     int32_t * f = (int32_t *) _f;
00431         asm volatile(
00432                 "movl $-1024, %%esi             \n\t"
00433                 "movq "MANGLE(magicF2W)", %%mm7 \n\t"
00434 //              "pxor %%mm6, %%mm6              \n\t"
00435                 "1:                             \n\t"
00436                 "movq 1024(%1, %%esi), %%mm0    \n\t"
00437                 "movq 3072(%1, %%esi), %%mm1    \n\t"
00438                 "movq 4096(%1, %%esi), %%mm2    \n\t"
00439                 "movq 5120(%1, %%esi), %%mm3    \n\t"
00440                 "movq 2048(%1, %%esi), %%mm4    \n\t"
00441                 "movq (%1, %%esi), %%mm5        \n\t" 
00442                 "psubd %%mm7, %%mm0             \n\t"
00443                 "psubd %%mm7, %%mm1             \n\t"
00444                 "psubd %%mm7, %%mm2             \n\t"
00445                 "psubd %%mm7, %%mm3             \n\t"
00446                 "psubd %%mm7, %%mm4             \n\t"
00447                 "psubd %%mm7, %%mm5             \n\t"
00448                 "leal (%%esi, %%esi, 2), %%edi  \n\t"
00449                 
00450                 "packssdw %%mm2, %%mm0          \n\t" // CcAa
00451                 "packssdw %%mm3, %%mm1          \n\t" // DdBb
00452                 "packssdw %%mm4, %%mm4          \n\t" // EeEe
00453                 "packssdw %%mm5, %%mm5          \n\t" // FfFf
00454                 "movq %%mm0, %%mm2              \n\t" // CcAa
00455                 "punpcklwd %%mm1, %%mm0         \n\t" // BAba
00456                 "punpckhwd %%mm1, %%mm2         \n\t" // DCdc
00457                 "punpcklwd %%mm5, %%mm4         \n\t" // FEfe
00458                 "movq %%mm0, %%mm1              \n\t" // BAba
00459                 "movq %%mm4, %%mm3              \n\t" // FEfe
00460                 "punpckldq %%mm2, %%mm0         \n\t" // dcba
00461                 "punpckhdq %%mm1, %%mm1         \n\t" // BABA
00462                 "punpckldq %%mm1, %%mm4         \n\t" // BAfe
00463                 "punpckhdq %%mm3, %%mm2         \n\t" // FEDC
00464                 
00465                 "movq %%mm0, (%0, %%edi)        \n\t"
00466                 "movq %%mm4, 8(%0, %%edi)       \n\t"
00467                 "movq %%mm2, 16(%0, %%edi)      \n\t"
00468                 "addl $8, %%esi                 \n\t"
00469                 " jnz 1b                        \n\t"
00470                 "emms                           \n\t"
00471                 :: "r" (s16+1536), "r" (f+256)
00472                 :"%esi", "%edi", "memory"
00473         );
00474     return 6*256;
00475 }
00476 
00477 
00478 static void* a52_resample_MMX(int flags, int ch){
00479     switch (flags) {
00480     case A52_MONO:
00481         if(ch==5) return a52_resample_MONO_to_5_MMX;
00482         break;
00483     case A52_CHANNEL:
00484     case A52_STEREO:
00485     case A52_DOLBY:
00486         if(ch==2) return a52_resample_STEREO_to_2_MMX;
00487         break;
00488     case A52_3F:
00489         if(ch==5) return a52_resample_3F_to_5_MMX;
00490         break;
00491     case A52_2F2R:
00492         if(ch==4) return a52_resample_2F_2R_to_4_MMX;
00493         break;
00494     case A52_3F2R:
00495         if(ch==5) return a52_resample_3F_2R_to_5_MMX;
00496         break;
00497     case A52_MONO | A52_LFE:
00498         if(ch==6) return a52_resample_MONO_LFE_to_6_MMX;
00499         break;
00500     case A52_CHANNEL | A52_LFE:
00501     case A52_STEREO | A52_LFE:
00502     case A52_DOLBY | A52_LFE:
00503         if(ch==6) return a52_resample_STEREO_LFE_to_6_MMX;
00504         break;
00505     case A52_3F | A52_LFE:
00506         if(ch==6) return a52_resample_3F_LFE_to_6_MMX;
00507         break;
00508     case A52_2F2R | A52_LFE:
00509         if(ch==6) return a52_resample_2F_2R_LFE_to_6_MMX;
00510         break;
00511     case A52_3F2R | A52_LFE:
00512         if(ch==6) return a52_resample_3F_2R_LFE_to_6_MMX;
00513         break;
00514     }
00515     return NULL;
00516 }
00517 
00518
hvirtual/quicktime/ffmpeg/libavcodec/liba52/resample_mmx.c
