libavcodec/x86/motion_est_mmx.c
Go to the documentation of this file.
00001 /*
00002  * MMX optimized motion estimation
00003  * Copyright (c) 2001 Fabrice Bellard
00004  * Copyright (c) 2002-2004 Michael Niedermayer
00005  *
00006  * mostly by Michael Niedermayer <michaelni@gmx.at>
00007  *
00008  * This file is part of FFmpeg.
00009  *
00010  * FFmpeg is free software; you can redistribute it and/or
00011  * modify it under the terms of the GNU Lesser General Public
00012  * License as published by the Free Software Foundation; either
00013  * version 2.1 of the License, or (at your option) any later version.
00014  *
00015  * FFmpeg is distributed in the hope that it will be useful,
00016  * but WITHOUT ANY WARRANTY; without even the implied warranty of
00017  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
00018  * Lesser General Public License for more details.
00019  *
00020  * You should have received a copy of the GNU Lesser General Public
00021  * License along with FFmpeg; if not, write to the Free Software
00022  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
00023  */
00024 
00025 #include "libavutil/x86_cpu.h"
00026 #include "libavcodec/dsputil.h"
00027 #include "dsputil_mmx.h"
00028 
00029 DECLARE_ASM_CONST(8, uint64_t, round_tab)[3]={
00030 0x0000000000000000ULL,
00031 0x0001000100010001ULL,
00032 0x0002000200020002ULL,
00033 };
00034 
00035 DECLARE_ASM_CONST(8, uint64_t, bone)= 0x0101010101010101LL;
00036 
00037 static inline void sad8_1_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
00038 {
00039     x86_reg len= -(stride*h);
00040     __asm__ volatile(
00041         ".p2align 4                     \n\t"
00042         "1:                             \n\t"
00043         "movq (%1, %%"REG_a"), %%mm0    \n\t"
00044         "movq (%2, %%"REG_a"), %%mm2    \n\t"
00045         "movq (%2, %%"REG_a"), %%mm4    \n\t"
00046         "add %3, %%"REG_a"              \n\t"
00047         "psubusb %%mm0, %%mm2           \n\t"
00048         "psubusb %%mm4, %%mm0           \n\t"
00049         "movq (%1, %%"REG_a"), %%mm1    \n\t"
00050         "movq (%2, %%"REG_a"), %%mm3    \n\t"
00051         "movq (%2, %%"REG_a"), %%mm5    \n\t"
00052         "psubusb %%mm1, %%mm3           \n\t"
00053         "psubusb %%mm5, %%mm1           \n\t"
00054         "por %%mm2, %%mm0               \n\t"
00055         "por %%mm1, %%mm3               \n\t"
00056         "movq %%mm0, %%mm1              \n\t"
00057         "movq %%mm3, %%mm2              \n\t"
00058         "punpcklbw %%mm7, %%mm0         \n\t"
00059         "punpckhbw %%mm7, %%mm1         \n\t"
00060         "punpcklbw %%mm7, %%mm3         \n\t"
00061         "punpckhbw %%mm7, %%mm2         \n\t"
00062         "paddw %%mm1, %%mm0             \n\t"
00063         "paddw %%mm3, %%mm2             \n\t"
00064         "paddw %%mm2, %%mm0             \n\t"
00065         "paddw %%mm0, %%mm6             \n\t"
00066         "add %3, %%"REG_a"              \n\t"
00067         " js 1b                         \n\t"
00068         : "+a" (len)
00069         : "r" (blk1 - len), "r" (blk2 - len), "r" ((x86_reg)stride)
00070     );
00071 }
00072 
00073 static inline void sad8_1_mmx2(uint8_t *blk1, uint8_t *blk2, int stride, int h)
00074 {
00075     __asm__ volatile(
00076         ".p2align 4                     \n\t"
00077         "1:                             \n\t"
00078         "movq (%1), %%mm0               \n\t"
00079         "movq (%1, %3), %%mm1           \n\t"
00080         "psadbw (%2), %%mm0             \n\t"
00081         "psadbw (%2, %3), %%mm1         \n\t"
00082         "paddw %%mm0, %%mm6             \n\t"
00083         "paddw %%mm1, %%mm6             \n\t"
00084         "lea (%1,%3,2), %1              \n\t"
00085         "lea (%2,%3,2), %2              \n\t"
00086         "sub $2, %0                     \n\t"
00087         " jg 1b                         \n\t"
00088         : "+r" (h), "+r" (blk1), "+r" (blk2)
00089         : "r" ((x86_reg)stride)
00090     );
00091 }
00092 
00093 static int sad16_sse2(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)
00094 {
00095     int ret;
00096     __asm__ volatile(
00097         "pxor %%xmm2, %%xmm2            \n\t"
00098         ".p2align 4                     \n\t"
00099         "1:                             \n\t"
00100         "movdqu (%1), %%xmm0            \n\t"
00101         "movdqu (%1, %4), %%xmm1        \n\t"
00102         "psadbw (%2), %%xmm0            \n\t"
00103         "psadbw (%2, %4), %%xmm1        \n\t"
00104         "paddw %%xmm0, %%xmm2           \n\t"
00105         "paddw %%xmm1, %%xmm2           \n\t"
00106         "lea (%1,%4,2), %1              \n\t"
00107         "lea (%2,%4,2), %2              \n\t"
00108         "sub $2, %0                     \n\t"
00109         " jg 1b                         \n\t"
00110         "movhlps %%xmm2, %%xmm0         \n\t"
00111         "paddw   %%xmm0, %%xmm2         \n\t"
00112         "movd    %%xmm2, %3             \n\t"
00113         : "+r" (h), "+r" (blk1), "+r" (blk2), "=r"(ret)
00114         : "r" ((x86_reg)stride)
00115     );
00116     return ret;
00117 }
00118 
00119 static inline void sad8_x2a_mmx2(uint8_t *blk1, uint8_t *blk2, int stride, int h)
00120 {
00121     __asm__ volatile(
00122         ".p2align 4                     \n\t"
00123         "1:                             \n\t"
00124         "movq (%1), %%mm0               \n\t"
00125         "movq (%1, %3), %%mm1           \n\t"
00126         "pavgb 1(%1), %%mm0             \n\t"
00127         "pavgb 1(%1, %3), %%mm1         \n\t"
00128         "psadbw (%2), %%mm0             \n\t"
00129         "psadbw (%2, %3), %%mm1         \n\t"
00130         "paddw %%mm0, %%mm6             \n\t"
00131         "paddw %%mm1, %%mm6             \n\t"
00132         "lea (%1,%3,2), %1              \n\t"
00133         "lea (%2,%3,2), %2              \n\t"
00134         "sub $2, %0                     \n\t"
00135         " jg 1b                         \n\t"
00136         : "+r" (h), "+r" (blk1), "+r" (blk2)
00137         : "r" ((x86_reg)stride)
00138     );
00139 }
00140 
00141 static inline void sad8_y2a_mmx2(uint8_t *blk1, uint8_t *blk2, int stride, int h)
00142 {
00143     __asm__ volatile(
00144         "movq (%1), %%mm0               \n\t"
00145         "add %3, %1                     \n\t"
00146         ".p2align 4                     \n\t"
00147         "1:                             \n\t"
00148         "movq (%1), %%mm1               \n\t"
00149         "movq (%1, %3), %%mm2           \n\t"
00150         "pavgb %%mm1, %%mm0             \n\t"
00151         "pavgb %%mm2, %%mm1             \n\t"
00152         "psadbw (%2), %%mm0             \n\t"
00153         "psadbw (%2, %3), %%mm1         \n\t"
00154         "paddw %%mm0, %%mm6             \n\t"
00155         "paddw %%mm1, %%mm6             \n\t"
00156         "movq %%mm2, %%mm0              \n\t"
00157         "lea (%1,%3,2), %1              \n\t"
00158         "lea (%2,%3,2), %2              \n\t"
00159         "sub $2, %0                     \n\t"
00160         " jg 1b                         \n\t"
00161         : "+r" (h), "+r" (blk1), "+r" (blk2)
00162         : "r" ((x86_reg)stride)
00163     );
00164 }
00165 
00166 static inline void sad8_4_mmx2(uint8_t *blk1, uint8_t *blk2, int stride, int h)
00167 {
00168     __asm__ volatile(
00169         "movq "MANGLE(bone)", %%mm5     \n\t"
00170         "movq (%1), %%mm0               \n\t"
00171         "pavgb 1(%1), %%mm0             \n\t"
00172         "add %3, %1                     \n\t"
00173         ".p2align 4                     \n\t"
00174         "1:                             \n\t"
00175         "movq (%1), %%mm1               \n\t"
00176         "movq (%1,%3), %%mm2            \n\t"
00177         "pavgb 1(%1), %%mm1             \n\t"
00178         "pavgb 1(%1,%3), %%mm2          \n\t"
00179         "psubusb %%mm5, %%mm1           \n\t"
00180         "pavgb %%mm1, %%mm0             \n\t"
00181         "pavgb %%mm2, %%mm1             \n\t"
00182         "psadbw (%2), %%mm0             \n\t"
00183         "psadbw (%2,%3), %%mm1          \n\t"
00184         "paddw %%mm0, %%mm6             \n\t"
00185         "paddw %%mm1, %%mm6             \n\t"
00186         "movq %%mm2, %%mm0              \n\t"
00187         "lea (%1,%3,2), %1              \n\t"
00188         "lea (%2,%3,2), %2              \n\t"
00189         "sub $2, %0                     \n\t"
00190         " jg 1b                         \n\t"
00191         : "+r" (h), "+r" (blk1), "+r" (blk2)
00192         : "r" ((x86_reg)stride)
00193     );
00194 }
00195 
00196 static inline void sad8_2_mmx(uint8_t *blk1a, uint8_t *blk1b, uint8_t *blk2, int stride, int h)
00197 {
00198     x86_reg len= -(stride*h);
00199     __asm__ volatile(
00200         ".p2align 4                     \n\t"
00201         "1:                             \n\t"
00202         "movq (%1, %%"REG_a"), %%mm0    \n\t"
00203         "movq (%2, %%"REG_a"), %%mm1    \n\t"
00204         "movq (%1, %%"REG_a"), %%mm2    \n\t"
00205         "movq (%2, %%"REG_a"), %%mm3    \n\t"
00206         "punpcklbw %%mm7, %%mm0         \n\t"
00207         "punpcklbw %%mm7, %%mm1         \n\t"
00208         "punpckhbw %%mm7, %%mm2         \n\t"
00209         "punpckhbw %%mm7, %%mm3         \n\t"
00210         "paddw %%mm0, %%mm1             \n\t"
00211         "paddw %%mm2, %%mm3             \n\t"
00212         "movq (%3, %%"REG_a"), %%mm4    \n\t"
00213         "movq (%3, %%"REG_a"), %%mm2    \n\t"
00214         "paddw %%mm5, %%mm1             \n\t"
00215         "paddw %%mm5, %%mm3             \n\t"
00216         "psrlw $1, %%mm1                \n\t"
00217         "psrlw $1, %%mm3                \n\t"
00218         "packuswb %%mm3, %%mm1          \n\t"
00219         "psubusb %%mm1, %%mm4           \n\t"
00220         "psubusb %%mm2, %%mm1           \n\t"
00221         "por %%mm4, %%mm1               \n\t"
00222         "movq %%mm1, %%mm0              \n\t"
00223         "punpcklbw %%mm7, %%mm0         \n\t"
00224         "punpckhbw %%mm7, %%mm1         \n\t"
00225         "paddw %%mm1, %%mm0             \n\t"
00226         "paddw %%mm0, %%mm6             \n\t"
00227         "add %4, %%"REG_a"              \n\t"
00228         " js 1b                         \n\t"
00229         : "+a" (len)
00230         : "r" (blk1a - len), "r" (blk1b -len), "r" (blk2 - len), "r" ((x86_reg)stride)
00231     );
00232 }
00233 
00234 static inline void sad8_4_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
00235 {
00236     x86_reg len= -(stride*h);
00237     __asm__ volatile(
00238         "movq (%1, %%"REG_a"), %%mm0    \n\t"
00239         "movq 1(%1, %%"REG_a"), %%mm2   \n\t"
00240         "movq %%mm0, %%mm1              \n\t"
00241         "movq %%mm2, %%mm3              \n\t"
00242         "punpcklbw %%mm7, %%mm0         \n\t"
00243         "punpckhbw %%mm7, %%mm1         \n\t"
00244         "punpcklbw %%mm7, %%mm2         \n\t"
00245         "punpckhbw %%mm7, %%mm3         \n\t"
00246         "paddw %%mm2, %%mm0             \n\t"
00247         "paddw %%mm3, %%mm1             \n\t"
00248         ".p2align 4                     \n\t"
00249         "1:                             \n\t"
00250         "movq (%2, %%"REG_a"), %%mm2    \n\t"
00251         "movq 1(%2, %%"REG_a"), %%mm4   \n\t"
00252         "movq %%mm2, %%mm3              \n\t"
00253         "movq %%mm4, %%mm5              \n\t"
00254         "punpcklbw %%mm7, %%mm2         \n\t"
00255         "punpckhbw %%mm7, %%mm3         \n\t"
00256         "punpcklbw %%mm7, %%mm4         \n\t"
00257         "punpckhbw %%mm7, %%mm5         \n\t"
00258         "paddw %%mm4, %%mm2             \n\t"
00259         "paddw %%mm5, %%mm3             \n\t"
00260         "movq 16+"MANGLE(round_tab)", %%mm5 \n\t"
00261         "paddw %%mm2, %%mm0             \n\t"
00262         "paddw %%mm3, %%mm1             \n\t"
00263         "paddw %%mm5, %%mm0             \n\t"
00264         "paddw %%mm5, %%mm1             \n\t"
00265         "movq (%3, %%"REG_a"), %%mm4    \n\t"
00266         "movq (%3, %%"REG_a"), %%mm5    \n\t"
00267         "psrlw $2, %%mm0                \n\t"
00268         "psrlw $2, %%mm1                \n\t"
00269         "packuswb %%mm1, %%mm0          \n\t"
00270         "psubusb %%mm0, %%mm4           \n\t"
00271         "psubusb %%mm5, %%mm0           \n\t"
00272         "por %%mm4, %%mm0               \n\t"
00273         "movq %%mm0, %%mm4              \n\t"
00274         "punpcklbw %%mm7, %%mm0         \n\t"
00275         "punpckhbw %%mm7, %%mm4         \n\t"
00276         "paddw %%mm0, %%mm6             \n\t"
00277         "paddw %%mm4, %%mm6             \n\t"
00278         "movq  %%mm2, %%mm0             \n\t"
00279         "movq  %%mm3, %%mm1             \n\t"
00280         "add %4, %%"REG_a"              \n\t"
00281         " js 1b                         \n\t"
00282         : "+a" (len)
00283         : "r" (blk1 - len), "r" (blk1 -len + stride), "r" (blk2 - len), "r" ((x86_reg)stride)
00284     );
00285 }
00286 
00287 static inline int sum_mmx(void)
00288 {
00289     int ret;
00290     __asm__ volatile(
00291         "movq %%mm6, %%mm0              \n\t"
00292         "psrlq $32, %%mm6               \n\t"
00293         "paddw %%mm0, %%mm6             \n\t"
00294         "movq %%mm6, %%mm0              \n\t"
00295         "psrlq $16, %%mm6               \n\t"
00296         "paddw %%mm0, %%mm6             \n\t"
00297         "movd %%mm6, %0                 \n\t"
00298         : "=r" (ret)
00299     );
00300     return ret&0xFFFF;
00301 }
00302 
00303 static inline int sum_mmx2(void)
00304 {
00305     int ret;
00306     __asm__ volatile(
00307         "movd %%mm6, %0                 \n\t"
00308         : "=r" (ret)
00309     );
00310     return ret;
00311 }
00312 
00313 static inline void sad8_x2a_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
00314 {
00315     sad8_2_mmx(blk1, blk1+1, blk2, stride, h);
00316 }
00317 static inline void sad8_y2a_mmx(uint8_t *blk1, uint8_t *blk2, int stride, int h)
00318 {
00319     sad8_2_mmx(blk1, blk1+stride, blk2, stride, h);
00320 }
00321 
00322 
00323 #define PIX_SAD(suf)\
00324 static int sad8_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
00325 {\
00326     assert(h==8);\
00327     __asm__ volatile("pxor %%mm7, %%mm7     \n\t"\
00328                  "pxor %%mm6, %%mm6     \n\t":);\
00329 \
00330     sad8_1_ ## suf(blk1, blk2, stride, 8);\
00331 \
00332     return sum_ ## suf();\
00333 }\
00334 static int sad8_x2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
00335 {\
00336     assert(h==8);\
00337     __asm__ volatile("pxor %%mm7, %%mm7     \n\t"\
00338                  "pxor %%mm6, %%mm6     \n\t"\
00339                  "movq %0, %%mm5        \n\t"\
00340                  :: "m"(round_tab[1]) \
00341                  );\
00342 \
00343     sad8_x2a_ ## suf(blk1, blk2, stride, 8);\
00344 \
00345     return sum_ ## suf();\
00346 }\
00347 \
00348 static int sad8_y2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
00349 {\
00350     assert(h==8);\
00351     __asm__ volatile("pxor %%mm7, %%mm7     \n\t"\
00352                  "pxor %%mm6, %%mm6     \n\t"\
00353                  "movq %0, %%mm5        \n\t"\
00354                  :: "m"(round_tab[1]) \
00355                  );\
00356 \
00357     sad8_y2a_ ## suf(blk1, blk2, stride, 8);\
00358 \
00359     return sum_ ## suf();\
00360 }\
00361 \
00362 static int sad8_xy2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
00363 {\
00364     assert(h==8);\
00365     __asm__ volatile("pxor %%mm7, %%mm7     \n\t"\
00366                  "pxor %%mm6, %%mm6     \n\t"\
00367                  ::);\
00368 \
00369     sad8_4_ ## suf(blk1, blk2, stride, 8);\
00370 \
00371     return sum_ ## suf();\
00372 }\
00373 \
00374 static int sad16_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
00375 {\
00376     __asm__ volatile("pxor %%mm7, %%mm7     \n\t"\
00377                  "pxor %%mm6, %%mm6     \n\t":);\
00378 \
00379     sad8_1_ ## suf(blk1  , blk2  , stride, h);\
00380     sad8_1_ ## suf(blk1+8, blk2+8, stride, h);\
00381 \
00382     return sum_ ## suf();\
00383 }\
00384 static int sad16_x2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
00385 {\
00386     __asm__ volatile("pxor %%mm7, %%mm7     \n\t"\
00387                  "pxor %%mm6, %%mm6     \n\t"\
00388                  "movq %0, %%mm5        \n\t"\
00389                  :: "m"(round_tab[1]) \
00390                  );\
00391 \
00392     sad8_x2a_ ## suf(blk1  , blk2  , stride, h);\
00393     sad8_x2a_ ## suf(blk1+8, blk2+8, stride, h);\
00394 \
00395     return sum_ ## suf();\
00396 }\
00397 static int sad16_y2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
00398 {\
00399     __asm__ volatile("pxor %%mm7, %%mm7     \n\t"\
00400                  "pxor %%mm6, %%mm6     \n\t"\
00401                  "movq %0, %%mm5        \n\t"\
00402                  :: "m"(round_tab[1]) \
00403                  );\
00404 \
00405     sad8_y2a_ ## suf(blk1  , blk2  , stride, h);\
00406     sad8_y2a_ ## suf(blk1+8, blk2+8, stride, h);\
00407 \
00408     return sum_ ## suf();\
00409 }\
00410 static int sad16_xy2_ ## suf(void *v, uint8_t *blk2, uint8_t *blk1, int stride, int h)\
00411 {\
00412     __asm__ volatile("pxor %%mm7, %%mm7     \n\t"\
00413                  "pxor %%mm6, %%mm6     \n\t"\
00414                  ::);\
00415 \
00416     sad8_4_ ## suf(blk1  , blk2  , stride, h);\
00417     sad8_4_ ## suf(blk1+8, blk2+8, stride, h);\
00418 \
00419     return sum_ ## suf();\
00420 }\
00421 
00422 PIX_SAD(mmx)
00423 PIX_SAD(mmx2)
00424 
00425 void dsputil_init_pix_mmx(DSPContext* c, AVCodecContext *avctx)
00426 {
00427     int mm_flags = av_get_cpu_flags();
00428 
00429     if (mm_flags & AV_CPU_FLAG_MMX) {
00430         c->pix_abs[0][0] = sad16_mmx;
00431         c->pix_abs[0][1] = sad16_x2_mmx;
00432         c->pix_abs[0][2] = sad16_y2_mmx;
00433         c->pix_abs[0][3] = sad16_xy2_mmx;
00434         c->pix_abs[1][0] = sad8_mmx;
00435         c->pix_abs[1][1] = sad8_x2_mmx;
00436         c->pix_abs[1][2] = sad8_y2_mmx;
00437         c->pix_abs[1][3] = sad8_xy2_mmx;
00438 
00439         c->sad[0]= sad16_mmx;
00440         c->sad[1]= sad8_mmx;
00441     }
00442     if (mm_flags & AV_CPU_FLAG_MMX2) {
00443         c->pix_abs[0][0] = sad16_mmx2;
00444         c->pix_abs[1][0] = sad8_mmx2;
00445 
00446         c->sad[0]= sad16_mmx2;
00447         c->sad[1]= sad8_mmx2;
00448 
00449         if(!(avctx->flags & CODEC_FLAG_BITEXACT)){
00450             c->pix_abs[0][1] = sad16_x2_mmx2;
00451             c->pix_abs[0][2] = sad16_y2_mmx2;
00452             c->pix_abs[0][3] = sad16_xy2_mmx2;
00453             c->pix_abs[1][1] = sad8_x2_mmx2;
00454             c->pix_abs[1][2] = sad8_y2_mmx2;
00455             c->pix_abs[1][3] = sad8_xy2_mmx2;
00456         }
00457     }
00458     if ((mm_flags & AV_CPU_FLAG_SSE2) && !(mm_flags & AV_CPU_FLAG_3DNOW) && avctx->codec_id != CODEC_ID_SNOW) {
00459         c->sad[0]= sad16_sse2;
00460     }
00461 }