00001
00002
00003
00004
00005
00006
00007
00008
00009
00010
00011
00012
00013
00014
00015
00016
00017
00018
00019
00020
00021 #include "dsputil_mmx.h"
00022
00023
00024
00025
00026 #define SUMSUB_BADC( a, b, c, d ) \
00027 "paddw "#b", "#a" \n\t"\
00028 "paddw "#d", "#c" \n\t"\
00029 "paddw "#b", "#b" \n\t"\
00030 "paddw "#d", "#d" \n\t"\
00031 "psubw "#a", "#b" \n\t"\
00032 "psubw "#c", "#d" \n\t"
00033
00034 #define SUMSUBD2_AB( a, b, t ) \
00035 "movq "#b", "#t" \n\t"\
00036 "psraw $1 , "#b" \n\t"\
00037 "paddw "#a", "#b" \n\t"\
00038 "psraw $1 , "#a" \n\t"\
00039 "psubw "#t", "#a" \n\t"
00040
00041 #define IDCT4_1D( s02, s13, d02, d13, t ) \
00042 SUMSUB_BA ( s02, d02 )\
00043 SUMSUBD2_AB( s13, d13, t )\
00044 SUMSUB_BADC( d13, s02, s13, d02 )
00045
00046 #define STORE_DIFF_4P( p, t, z ) \
00047 "psraw $6, "#p" \n\t"\
00048 "movd (%0), "#t" \n\t"\
00049 "punpcklbw "#z", "#t" \n\t"\
00050 "paddsw "#t", "#p" \n\t"\
00051 "packuswb "#z", "#p" \n\t"\
00052 "movd "#p", (%0) \n\t"
00053
00054 static void ff_h264_idct_add_mmx(uint8_t *dst, int16_t *block, int stride)
00055 {
00056
00057 asm volatile(
00058 "movq (%0), %%mm0 \n\t"
00059 "movq 8(%0), %%mm1 \n\t"
00060 "movq 16(%0), %%mm2 \n\t"
00061 "movq 24(%0), %%mm3 \n\t"
00062 :: "r"(block) );
00063
00064 asm volatile(
00065
00066 IDCT4_1D( %%mm2, %%mm1, %%mm0, %%mm3, %%mm4 )
00067
00068 "movq %0, %%mm6 \n\t"
00069
00070 TRANSPOSE4( %%mm3, %%mm1, %%mm0, %%mm2, %%mm4 )
00071
00072 "paddw %%mm6, %%mm3 \n\t"
00073
00074
00075 IDCT4_1D( %%mm4, %%mm2, %%mm3, %%mm0, %%mm1 )
00076
00077 "pxor %%mm7, %%mm7 \n\t"
00078 :: "m"(ff_pw_32));
00079
00080 asm volatile(
00081 STORE_DIFF_4P( %%mm0, %%mm1, %%mm7)
00082 "add %1, %0 \n\t"
00083 STORE_DIFF_4P( %%mm2, %%mm1, %%mm7)
00084 "add %1, %0 \n\t"
00085 STORE_DIFF_4P( %%mm3, %%mm1, %%mm7)
00086 "add %1, %0 \n\t"
00087 STORE_DIFF_4P( %%mm4, %%mm1, %%mm7)
00088 : "+r"(dst)
00089 : "r" ((long)stride)
00090 );
00091 }
00092
00093 static inline void h264_idct8_1d(int16_t *block)
00094 {
00095 asm volatile(
00096 "movq 112(%0), %%mm7 \n\t"
00097 "movq 80(%0), %%mm0 \n\t"
00098 "movq 48(%0), %%mm3 \n\t"
00099 "movq 16(%0), %%mm5 \n\t"
00100
00101 "movq %%mm0, %%mm4 \n\t"
00102 "movq %%mm5, %%mm1 \n\t"
00103 "psraw $1, %%mm4 \n\t"
00104 "psraw $1, %%mm1 \n\t"
00105 "paddw %%mm0, %%mm4 \n\t"
00106 "paddw %%mm5, %%mm1 \n\t"
00107 "paddw %%mm7, %%mm4 \n\t"
00108 "paddw %%mm0, %%mm1 \n\t"
00109 "psubw %%mm5, %%mm4 \n\t"
00110 "paddw %%mm3, %%mm1 \n\t"
00111
00112 "psubw %%mm3, %%mm5 \n\t"
00113 "psubw %%mm3, %%mm0 \n\t"
00114 "paddw %%mm7, %%mm5 \n\t"
00115 "psubw %%mm7, %%mm0 \n\t"
00116 "psraw $1, %%mm3 \n\t"
00117 "psraw $1, %%mm7 \n\t"
00118 "psubw %%mm3, %%mm5 \n\t"
00119 "psubw %%mm7, %%mm0 \n\t"
00120
00121 "movq %%mm4, %%mm3 \n\t"
00122 "movq %%mm1, %%mm7 \n\t"
00123 "psraw $2, %%mm1 \n\t"
00124 "psraw $2, %%mm3 \n\t"
00125 "paddw %%mm5, %%mm3 \n\t"
00126 "psraw $2, %%mm5 \n\t"
00127 "paddw %%mm0, %%mm1 \n\t"
00128 "psraw $2, %%mm0 \n\t"
00129 "psubw %%mm4, %%mm5 \n\t"
00130 "psubw %%mm0, %%mm7 \n\t"
00131
00132 "movq 32(%0), %%mm2 \n\t"
00133 "movq 96(%0), %%mm6 \n\t"
00134 "movq %%mm2, %%mm4 \n\t"
00135 "movq %%mm6, %%mm0 \n\t"
00136 "psraw $1, %%mm4 \n\t"
00137 "psraw $1, %%mm6 \n\t"
00138 "psubw %%mm0, %%mm4 \n\t"
00139 "paddw %%mm2, %%mm6 \n\t"
00140
00141 "movq (%0), %%mm2 \n\t"
00142 "movq 64(%0), %%mm0 \n\t"
00143 SUMSUB_BA( %%mm0, %%mm2 )
00144 SUMSUB_BA( %%mm6, %%mm0 )
00145 SUMSUB_BA( %%mm4, %%mm2 )
00146 SUMSUB_BA( %%mm7, %%mm6 )
00147 SUMSUB_BA( %%mm5, %%mm4 )
00148 SUMSUB_BA( %%mm3, %%mm2 )
00149 SUMSUB_BA( %%mm1, %%mm0 )
00150 :: "r"(block)
00151 );
00152 }
00153
00154 static void ff_h264_idct8_add_mmx(uint8_t *dst, int16_t *block, int stride)
00155 {
00156 int i;
00157 int16_t __attribute__ ((aligned(8))) b2[64];
00158
00159 block[0] += 32;
00160
00161 for(i=0; i<2; i++){
00162 DECLARE_ALIGNED_8(uint64_t, tmp);
00163
00164 h264_idct8_1d(block+4*i);
00165
00166 asm volatile(
00167 "movq %%mm7, %0 \n\t"
00168 TRANSPOSE4( %%mm0, %%mm2, %%mm4, %%mm6, %%mm7 )
00169 "movq %%mm0, 8(%1) \n\t"
00170 "movq %%mm6, 24(%1) \n\t"
00171 "movq %%mm7, 40(%1) \n\t"
00172 "movq %%mm4, 56(%1) \n\t"
00173 "movq %0, %%mm7 \n\t"
00174 TRANSPOSE4( %%mm7, %%mm5, %%mm3, %%mm1, %%mm0 )
00175 "movq %%mm7, (%1) \n\t"
00176 "movq %%mm1, 16(%1) \n\t"
00177 "movq %%mm0, 32(%1) \n\t"
00178 "movq %%mm3, 48(%1) \n\t"
00179 : "=m"(tmp)
00180 : "r"(b2+32*i)
00181 : "memory"
00182 );
00183 }
00184
00185 for(i=0; i<2; i++){
00186 h264_idct8_1d(b2+4*i);
00187
00188 asm volatile(
00189 "psraw $6, %%mm7 \n\t"
00190 "psraw $6, %%mm6 \n\t"
00191 "psraw $6, %%mm5 \n\t"
00192 "psraw $6, %%mm4 \n\t"
00193 "psraw $6, %%mm3 \n\t"
00194 "psraw $6, %%mm2 \n\t"
00195 "psraw $6, %%mm1 \n\t"
00196 "psraw $6, %%mm0 \n\t"
00197
00198 "movq %%mm7, (%0) \n\t"
00199 "movq %%mm5, 16(%0) \n\t"
00200 "movq %%mm3, 32(%0) \n\t"
00201 "movq %%mm1, 48(%0) \n\t"
00202 "movq %%mm0, 64(%0) \n\t"
00203 "movq %%mm2, 80(%0) \n\t"
00204 "movq %%mm4, 96(%0) \n\t"
00205 "movq %%mm6, 112(%0) \n\t"
00206 :: "r"(b2+4*i)
00207 : "memory"
00208 );
00209 }
00210
00211 add_pixels_clamped_mmx(b2, dst, stride);
00212 }
00213
00214 #define STORE_DIFF_8P( p, d, t, z )\
00215 "movq "#d", "#t" \n"\
00216 "psraw $6, "#p" \n"\
00217 "punpcklbw "#z", "#t" \n"\
00218 "paddsw "#t", "#p" \n"\
00219 "packuswb "#p", "#p" \n"\
00220 "movq "#p", "#d" \n"
00221
00222 #define H264_IDCT8_1D_SSE2(a,b,c,d,e,f,g,h)\
00223 "movdqa "#c", "#a" \n"\
00224 "movdqa "#g", "#e" \n"\
00225 "psraw $1, "#c" \n"\
00226 "psraw $1, "#g" \n"\
00227 "psubw "#e", "#c" \n"\
00228 "paddw "#a", "#g" \n"\
00229 "movdqa "#b", "#e" \n"\
00230 "psraw $1, "#e" \n"\
00231 "paddw "#b", "#e" \n"\
00232 "paddw "#d", "#e" \n"\
00233 "paddw "#f", "#e" \n"\
00234 "movdqa "#f", "#a" \n"\
00235 "psraw $1, "#a" \n"\
00236 "paddw "#f", "#a" \n"\
00237 "paddw "#h", "#a" \n"\
00238 "psubw "#b", "#a" \n"\
00239 "psubw "#d", "#b" \n"\
00240 "psubw "#d", "#f" \n"\
00241 "paddw "#h", "#b" \n"\
00242 "psubw "#h", "#f" \n"\
00243 "psraw $1, "#d" \n"\
00244 "psraw $1, "#h" \n"\
00245 "psubw "#d", "#b" \n"\
00246 "psubw "#h", "#f" \n"\
00247 "movdqa "#e", "#d" \n"\
00248 "movdqa "#a", "#h" \n"\
00249 "psraw $2, "#d" \n"\
00250 "psraw $2, "#h" \n"\
00251 "paddw "#f", "#d" \n"\
00252 "paddw "#b", "#h" \n"\
00253 "psraw $2, "#f" \n"\
00254 "psraw $2, "#b" \n"\
00255 "psubw "#f", "#e" \n"\
00256 "psubw "#a", "#b" \n"\
00257 "movdqa 0x00(%1), "#a" \n"\
00258 "movdqa 0x40(%1), "#f" \n"\
00259 SUMSUB_BA(f, a)\
00260 SUMSUB_BA(g, f)\
00261 SUMSUB_BA(c, a)\
00262 SUMSUB_BA(e, g)\
00263 SUMSUB_BA(b, c)\
00264 SUMSUB_BA(h, a)\
00265 SUMSUB_BA(d, f)
00266
00267 static void ff_h264_idct8_add_sse2(uint8_t *dst, int16_t *block, int stride)
00268 {
00269 asm volatile(
00270 "movdqa 0x10(%1), %%xmm1 \n"
00271 "movdqa 0x20(%1), %%xmm2 \n"
00272 "movdqa 0x30(%1), %%xmm3 \n"
00273 "movdqa 0x50(%1), %%xmm5 \n"
00274 "movdqa 0x60(%1), %%xmm6 \n"
00275 "movdqa 0x70(%1), %%xmm7 \n"
00276 H264_IDCT8_1D_SSE2(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm6, %%xmm7)
00277 TRANSPOSE8(%%xmm4, %%xmm1, %%xmm7, %%xmm3, %%xmm5, %%xmm0, %%xmm2, %%xmm6, (%1))
00278 "paddw %4, %%xmm4 \n"
00279 "movdqa %%xmm4, 0x00(%1) \n"
00280 "movdqa %%xmm2, 0x40(%1) \n"
00281 H264_IDCT8_1D_SSE2(%%xmm4, %%xmm0, %%xmm6, %%xmm3, %%xmm2, %%xmm5, %%xmm7, %%xmm1)
00282 "movdqa %%xmm6, 0x60(%1) \n"
00283 "movdqa %%xmm7, 0x70(%1) \n"
00284 "pxor %%xmm7, %%xmm7 \n"
00285 STORE_DIFF_8P(%%xmm2, (%0), %%xmm6, %%xmm7)
00286 STORE_DIFF_8P(%%xmm0, (%0,%2), %%xmm6, %%xmm7)
00287 STORE_DIFF_8P(%%xmm1, (%0,%2,2), %%xmm6, %%xmm7)
00288 STORE_DIFF_8P(%%xmm3, (%0,%3), %%xmm6, %%xmm7)
00289 "lea (%0,%2,4), %0 \n"
00290 STORE_DIFF_8P(%%xmm5, (%0), %%xmm6, %%xmm7)
00291 STORE_DIFF_8P(%%xmm4, (%0,%2), %%xmm6, %%xmm7)
00292 "movdqa 0x60(%1), %%xmm0 \n"
00293 "movdqa 0x70(%1), %%xmm1 \n"
00294 STORE_DIFF_8P(%%xmm0, (%0,%2,2), %%xmm6, %%xmm7)
00295 STORE_DIFF_8P(%%xmm1, (%0,%3), %%xmm6, %%xmm7)
00296 :"+r"(dst)
00297 :"r"(block), "r"((long)stride), "r"(3L*stride), "m"(ff_pw_32)
00298 );
00299 }
00300
00301 static void ff_h264_idct_dc_add_mmx2(uint8_t *dst, int16_t *block, int stride)
00302 {
00303 int dc = (block[0] + 32) >> 6;
00304 asm volatile(
00305 "movd %0, %%mm0 \n\t"
00306 "pshufw $0, %%mm0, %%mm0 \n\t"
00307 "pxor %%mm1, %%mm1 \n\t"
00308 "psubw %%mm0, %%mm1 \n\t"
00309 "packuswb %%mm0, %%mm0 \n\t"
00310 "packuswb %%mm1, %%mm1 \n\t"
00311 ::"r"(dc)
00312 );
00313 asm volatile(
00314 "movd %0, %%mm2 \n\t"
00315 "movd %1, %%mm3 \n\t"
00316 "movd %2, %%mm4 \n\t"
00317 "movd %3, %%mm5 \n\t"
00318 "paddusb %%mm0, %%mm2 \n\t"
00319 "paddusb %%mm0, %%mm3 \n\t"
00320 "paddusb %%mm0, %%mm4 \n\t"
00321 "paddusb %%mm0, %%mm5 \n\t"
00322 "psubusb %%mm1, %%mm2 \n\t"
00323 "psubusb %%mm1, %%mm3 \n\t"
00324 "psubusb %%mm1, %%mm4 \n\t"
00325 "psubusb %%mm1, %%mm5 \n\t"
00326 "movd %%mm2, %0 \n\t"
00327 "movd %%mm3, %1 \n\t"
00328 "movd %%mm4, %2 \n\t"
00329 "movd %%mm5, %3 \n\t"
00330 :"+m"(*(uint32_t*)(dst+0*stride)),
00331 "+m"(*(uint32_t*)(dst+1*stride)),
00332 "+m"(*(uint32_t*)(dst+2*stride)),
00333 "+m"(*(uint32_t*)(dst+3*stride))
00334 );
00335 }
00336
00337 static void ff_h264_idct8_dc_add_mmx2(uint8_t *dst, int16_t *block, int stride)
00338 {
00339 int dc = (block[0] + 32) >> 6;
00340 int y;
00341 asm volatile(
00342 "movd %0, %%mm0 \n\t"
00343 "pshufw $0, %%mm0, %%mm0 \n\t"
00344 "pxor %%mm1, %%mm1 \n\t"
00345 "psubw %%mm0, %%mm1 \n\t"
00346 "packuswb %%mm0, %%mm0 \n\t"
00347 "packuswb %%mm1, %%mm1 \n\t"
00348 ::"r"(dc)
00349 );
00350 for(y=2; y--; dst += 4*stride){
00351 asm volatile(
00352 "movq %0, %%mm2 \n\t"
00353 "movq %1, %%mm3 \n\t"
00354 "movq %2, %%mm4 \n\t"
00355 "movq %3, %%mm5 \n\t"
00356 "paddusb %%mm0, %%mm2 \n\t"
00357 "paddusb %%mm0, %%mm3 \n\t"
00358 "paddusb %%mm0, %%mm4 \n\t"
00359 "paddusb %%mm0, %%mm5 \n\t"
00360 "psubusb %%mm1, %%mm2 \n\t"
00361 "psubusb %%mm1, %%mm3 \n\t"
00362 "psubusb %%mm1, %%mm4 \n\t"
00363 "psubusb %%mm1, %%mm5 \n\t"
00364 "movq %%mm2, %0 \n\t"
00365 "movq %%mm3, %1 \n\t"
00366 "movq %%mm4, %2 \n\t"
00367 "movq %%mm5, %3 \n\t"
00368 :"+m"(*(uint64_t*)(dst+0*stride)),
00369 "+m"(*(uint64_t*)(dst+1*stride)),
00370 "+m"(*(uint64_t*)(dst+2*stride)),
00371 "+m"(*(uint64_t*)(dst+3*stride))
00372 );
00373 }
00374 }
00375
00376
00377
00378
00379
00380
00381
00382 #define DIFF_GT_MMX(x,y,a,o,t)\
00383 "movq "#y", "#t" \n\t"\
00384 "movq "#x", "#o" \n\t"\
00385 "psubusb "#x", "#t" \n\t"\
00386 "psubusb "#y", "#o" \n\t"\
00387 "por "#t", "#o" \n\t"\
00388 "psubusb "#a", "#o" \n\t"
00389
00390
00391
00392 #define DIFF_GT2_MMX(x,y,a,o,t)\
00393 "movq "#y", "#t" \n\t"\
00394 "movq "#x", "#o" \n\t"\
00395 "psubusb "#x", "#t" \n\t"\
00396 "psubusb "#y", "#o" \n\t"\
00397 "psubusb "#a", "#t" \n\t"\
00398 "psubusb "#a", "#o" \n\t"\
00399 "pcmpeqb "#t", "#o" \n\t"\
00400
00401
00402
00403
00404 #define H264_DEBLOCK_MASK(alpha1, beta1) \
00405 "pshufw $0, "#alpha1", %%mm4 \n\t"\
00406 "pshufw $0, "#beta1 ", %%mm5 \n\t"\
00407 "packuswb %%mm4, %%mm4 \n\t"\
00408 "packuswb %%mm5, %%mm5 \n\t"\
00409 DIFF_GT_MMX(%%mm1, %%mm2, %%mm4, %%mm7, %%mm6) \
00410 DIFF_GT_MMX(%%mm0, %%mm1, %%mm5, %%mm4, %%mm6) \
00411 "por %%mm4, %%mm7 \n\t"\
00412 DIFF_GT_MMX(%%mm3, %%mm2, %%mm5, %%mm4, %%mm6) \
00413 "por %%mm4, %%mm7 \n\t"\
00414 "pxor %%mm6, %%mm6 \n\t"\
00415 "pcmpeqb %%mm6, %%mm7 \n\t"
00416
00417
00418
00419
00420 #define H264_DEBLOCK_P0_Q0(pb_01, pb_3f)\
00421 "movq %%mm1 , %%mm5 \n\t"\
00422 "pxor %%mm2 , %%mm5 \n\t" \
00423 "pand "#pb_01" , %%mm5 \n\t" \
00424 "pcmpeqb %%mm4 , %%mm4 \n\t"\
00425 "pxor %%mm4 , %%mm3 \n\t"\
00426 "pavgb %%mm0 , %%mm3 \n\t" \
00427 "pavgb "MANGLE(ff_pb_3)" , %%mm3 \n\t" \
00428 "pxor %%mm1 , %%mm4 \n\t"\
00429 "pavgb %%mm2 , %%mm4 \n\t" \
00430 "pavgb %%mm5 , %%mm3 \n\t"\
00431 "paddusb %%mm4 , %%mm3 \n\t" \
00432 "movq "MANGLE(ff_pb_A1)" , %%mm6 \n\t"\
00433 "psubusb %%mm3 , %%mm6 \n\t"\
00434 "psubusb "MANGLE(ff_pb_A1)" , %%mm3 \n\t"\
00435 "pminub %%mm7 , %%mm6 \n\t"\
00436 "pminub %%mm7 , %%mm3 \n\t"\
00437 "psubusb %%mm6 , %%mm1 \n\t"\
00438 "psubusb %%mm3 , %%mm2 \n\t"\
00439 "paddusb %%mm3 , %%mm1 \n\t"\
00440 "paddusb %%mm6 , %%mm2 \n\t"
00441
00442
00443
00444
00445 #define H264_DEBLOCK_Q1(p1, q2, q2addr, q1addr, tc0, tmp)\
00446 "movq %%mm1, "#tmp" \n\t"\
00447 "pavgb %%mm2, "#tmp" \n\t"\
00448 "pavgb "#tmp", "#q2" \n\t" \
00449 "pxor "q2addr", "#tmp" \n\t"\
00450 "pand %8, "#tmp" \n\t" \
00451 "psubusb "#tmp", "#q2" \n\t" \
00452 "movq "#p1", "#tmp" \n\t"\
00453 "psubusb "#tc0", "#tmp" \n\t"\
00454 "paddusb "#p1", "#tc0" \n\t"\
00455 "pmaxub "#tmp", "#q2" \n\t"\
00456 "pminub "#tc0", "#q2" \n\t"\
00457 "movq "#q2", "q1addr" \n\t"
00458
00459 static inline void h264_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha1, int beta1, int8_t *tc0)
00460 {
00461 DECLARE_ALIGNED_8(uint64_t, tmp0[2]);
00462
00463 asm volatile(
00464 "movq (%1,%3), %%mm0 \n\t"
00465 "movq (%1,%3,2), %%mm1 \n\t"
00466 "movq (%2), %%mm2 \n\t"
00467 "movq (%2,%3), %%mm3 \n\t"
00468 H264_DEBLOCK_MASK(%6, %7)
00469
00470 "movd %5, %%mm4 \n\t"
00471 "punpcklbw %%mm4, %%mm4 \n\t"
00472 "punpcklwd %%mm4, %%mm4 \n\t"
00473 "pcmpeqb %%mm3, %%mm3 \n\t"
00474 "movq %%mm4, %%mm6 \n\t"
00475 "pcmpgtb %%mm3, %%mm4 \n\t"
00476 "movq %%mm6, 8+%0 \n\t"
00477 "pand %%mm4, %%mm7 \n\t"
00478 "movq %%mm7, %0 \n\t"
00479
00480
00481 "movq (%1), %%mm3 \n\t"
00482 DIFF_GT2_MMX(%%mm1, %%mm3, %%mm5, %%mm6, %%mm4)
00483 "pand %%mm7, %%mm6 \n\t"
00484 "pand 8+%0, %%mm7 \n\t"
00485 "movq %%mm7, %%mm4 \n\t"
00486 "psubb %%mm6, %%mm7 \n\t"
00487 "pand %%mm4, %%mm6 \n\t"
00488 H264_DEBLOCK_Q1(%%mm0, %%mm3, "(%1)", "(%1,%3)", %%mm6, %%mm4)
00489
00490
00491 "movq (%2,%3,2), %%mm4 \n\t"
00492 DIFF_GT2_MMX(%%mm2, %%mm4, %%mm5, %%mm6, %%mm3)
00493 "pand %0, %%mm6 \n\t"
00494 "movq 8+%0, %%mm5 \n\t"
00495 "pand %%mm6, %%mm5 \n\t"
00496 "psubb %%mm6, %%mm7 \n\t"
00497 "movq (%2,%3), %%mm3 \n\t"
00498 H264_DEBLOCK_Q1(%%mm3, %%mm4, "(%2,%3,2)", "(%2,%3)", %%mm5, %%mm6)
00499
00500
00501 H264_DEBLOCK_P0_Q0(%8, unused)
00502 "movq %%mm1, (%1,%3,2) \n\t"
00503 "movq %%mm2, (%2) \n\t"
00504
00505 : "=m"(*tmp0)
00506 : "r"(pix-3*stride), "r"(pix), "r"((long)stride),
00507 "m"(*tmp0), "m"(*(uint32_t*)tc0), "m"(alpha1), "m"(beta1),
00508 "m"(ff_bone)
00509 );
00510 }
00511
00512 static void h264_v_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
00513 {
00514 if((tc0[0] & tc0[1]) >= 0)
00515 h264_loop_filter_luma_mmx2(pix, stride, alpha-1, beta-1, tc0);
00516 if((tc0[2] & tc0[3]) >= 0)
00517 h264_loop_filter_luma_mmx2(pix+8, stride, alpha-1, beta-1, tc0+2);
00518 }
00519 static void h264_h_loop_filter_luma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
00520 {
00521
00522
00523 DECLARE_ALIGNED_8(uint8_t, trans[8*8]);
00524 int i;
00525 for(i=0; i<2; i++, pix+=8*stride, tc0+=2) {
00526 if((tc0[0] & tc0[1]) < 0)
00527 continue;
00528 transpose4x4(trans, pix-4, 8, stride);
00529 transpose4x4(trans +4*8, pix, 8, stride);
00530 transpose4x4(trans+4, pix-4+4*stride, 8, stride);
00531 transpose4x4(trans+4+4*8, pix +4*stride, 8, stride);
00532 h264_loop_filter_luma_mmx2(trans+4*8, 8, alpha-1, beta-1, tc0);
00533 transpose4x4(pix-2, trans +2*8, stride, 8);
00534 transpose4x4(pix-2+4*stride, trans+4+2*8, stride, 8);
00535 }
00536 }
00537
00538 static inline void h264_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha1, int beta1, int8_t *tc0)
00539 {
00540 asm volatile(
00541 "movq (%0), %%mm0 \n\t"
00542 "movq (%0,%2), %%mm1 \n\t"
00543 "movq (%1), %%mm2 \n\t"
00544 "movq (%1,%2), %%mm3 \n\t"
00545 H264_DEBLOCK_MASK(%4, %5)
00546 "movd %3, %%mm6 \n\t"
00547 "punpcklbw %%mm6, %%mm6 \n\t"
00548 "pand %%mm6, %%mm7 \n\t"
00549 H264_DEBLOCK_P0_Q0(%6, %7)
00550 "movq %%mm1, (%0,%2) \n\t"
00551 "movq %%mm2, (%1) \n\t"
00552
00553 :: "r"(pix-2*stride), "r"(pix), "r"((long)stride),
00554 "r"(*(uint32_t*)tc0),
00555 "m"(alpha1), "m"(beta1), "m"(ff_bone), "m"(ff_pb_3F)
00556 );
00557 }
00558
00559 static void h264_v_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
00560 {
00561 h264_loop_filter_chroma_mmx2(pix, stride, alpha-1, beta-1, tc0);
00562 }
00563
00564 static void h264_h_loop_filter_chroma_mmx2(uint8_t *pix, int stride, int alpha, int beta, int8_t *tc0)
00565 {
00566
00567 DECLARE_ALIGNED_8(uint8_t, trans[8*4]);
00568 transpose4x4(trans, pix-2, 8, stride);
00569 transpose4x4(trans+4, pix-2+4*stride, 8, stride);
00570 h264_loop_filter_chroma_mmx2(trans+2*8, 8, alpha-1, beta-1, tc0);
00571 transpose4x4(pix-2, trans, stride, 8);
00572 transpose4x4(pix-2+4*stride, trans+4, stride, 8);
00573 }
00574
00575
00576 #define H264_FILTER_CHROMA4(p0, p1, q1, one) \
00577 "movq "#p0", %%mm4 \n\t"\
00578 "pxor "#q1", %%mm4 \n\t"\
00579 "pand "#one", %%mm4 \n\t" \
00580 "pavgb "#q1", "#p0" \n\t"\
00581 "psubusb %%mm4, "#p0" \n\t"\
00582 "pavgb "#p1", "#p0" \n\t" \
00583
00584 static inline void h264_loop_filter_chroma_intra_mmx2(uint8_t *pix, int stride, int alpha1, int beta1)
00585 {
00586 asm volatile(
00587 "movq (%0), %%mm0 \n\t"
00588 "movq (%0,%2), %%mm1 \n\t"
00589 "movq (%1), %%mm2 \n\t"
00590 "movq (%1,%2), %%mm3 \n\t"
00591 H264_DEBLOCK_MASK(%3, %4)
00592 "movq %%mm1, %%mm5 \n\t"
00593 "movq %%mm2, %%mm6 \n\t"
00594 H264_FILTER_CHROMA4(%%mm1, %%mm0, %%mm3, %5)
00595 H264_FILTER_CHROMA4(%%mm2, %%mm3, %%mm0, %5)
00596 "psubb %%mm5, %%mm1 \n\t"
00597 "psubb %%mm6, %%mm2 \n\t"
00598 "pand %%mm7, %%mm1 \n\t"
00599 "pand %%mm7, %%mm2 \n\t"
00600 "paddb %%mm5, %%mm1 \n\t"
00601 "paddb %%mm6, %%mm2 \n\t"
00602 "movq %%mm1, (%0,%2) \n\t"
00603 "movq %%mm2, (%1) \n\t"
00604 :: "r"(pix-2*stride), "r"(pix), "r"((long)stride),
00605 "m"(alpha1), "m"(beta1), "m"(ff_bone)
00606 );
00607 }
00608
00609 static void h264_v_loop_filter_chroma_intra_mmx2(uint8_t *pix, int stride, int alpha, int beta)
00610 {
00611 h264_loop_filter_chroma_intra_mmx2(pix, stride, alpha-1, beta-1);
00612 }
00613
00614 static void h264_h_loop_filter_chroma_intra_mmx2(uint8_t *pix, int stride, int alpha, int beta)
00615 {
00616
00617 DECLARE_ALIGNED_8(uint8_t, trans[8*4]);
00618 transpose4x4(trans, pix-2, 8, stride);
00619 transpose4x4(trans+4, pix-2+4*stride, 8, stride);
00620 h264_loop_filter_chroma_intra_mmx2(trans+2*8, 8, alpha-1, beta-1);
00621 transpose4x4(pix-2, trans, stride, 8);
00622 transpose4x4(pix-2+4*stride, trans+4, stride, 8);
00623 }
00624
00625 static void h264_loop_filter_strength_mmx2( int16_t bS[2][4][4], uint8_t nnz[40], int8_t ref[2][40], int16_t mv[2][40][2],
00626 int bidir, int edges, int step, int mask_mv0, int mask_mv1 ) {
00627 int dir;
00628 asm volatile(
00629 "pxor %%mm7, %%mm7 \n\t"
00630 "movq %0, %%mm6 \n\t"
00631 "movq %1, %%mm5 \n\t"
00632 "movq %2, %%mm4 \n\t"
00633 ::"m"(ff_pb_1), "m"(ff_pb_3), "m"(ff_pb_7)
00634 );
00635
00636
00637 for( dir=1; dir>=0; dir-- ) {
00638 const int d_idx = dir ? -8 : -1;
00639 const int mask_mv = dir ? mask_mv1 : mask_mv0;
00640 DECLARE_ALIGNED_8(const uint64_t, mask_dir) = dir ? 0 : 0xffffffffffffffffULL;
00641 int b_idx, edge, l;
00642 for( b_idx=12, edge=0; edge<edges; edge+=step, b_idx+=8*step ) {
00643 asm volatile(
00644 "pand %0, %%mm0 \n\t"
00645 ::"m"(mask_dir)
00646 );
00647 if(!(mask_mv & edge)) {
00648 asm volatile("pxor %%mm0, %%mm0 \n\t":);
00649 for( l = bidir; l >= 0; l-- ) {
00650 asm volatile(
00651 "movd %0, %%mm1 \n\t"
00652 "punpckldq %1, %%mm1 \n\t"
00653 "movq %%mm1, %%mm2 \n\t"
00654 "psrlw $7, %%mm2 \n\t"
00655 "pand %%mm6, %%mm2 \n\t"
00656 "por %%mm2, %%mm1 \n\t"
00657 "punpckldq %%mm1, %%mm2 \n\t"
00658 "pcmpeqb %%mm2, %%mm1 \n\t"
00659 "paddb %%mm6, %%mm1 \n\t"
00660 "punpckhbw %%mm7, %%mm1 \n\t"
00661 "por %%mm1, %%mm0 \n\t"
00662
00663 "movq %2, %%mm1 \n\t"
00664 "movq %3, %%mm2 \n\t"
00665 "psubw %4, %%mm1 \n\t"
00666 "psubw %5, %%mm2 \n\t"
00667 "packsswb %%mm2, %%mm1 \n\t"
00668 "paddb %%mm5, %%mm1 \n\t"
00669 "pminub %%mm4, %%mm1 \n\t"
00670 "pcmpeqb %%mm4, %%mm1 \n\t"
00671 "por %%mm1, %%mm0 \n\t"
00672 ::"m"(ref[l][b_idx]),
00673 "m"(ref[l][b_idx+d_idx]),
00674 "m"(mv[l][b_idx][0]),
00675 "m"(mv[l][b_idx+2][0]),
00676 "m"(mv[l][b_idx+d_idx][0]),
00677 "m"(mv[l][b_idx+d_idx+2][0])
00678 );
00679 }
00680 }
00681 asm volatile(
00682 "movd %0, %%mm1 \n\t"
00683 "por %1, %%mm1 \n\t"
00684 "punpcklbw %%mm7, %%mm1 \n\t"
00685 "pcmpgtw %%mm7, %%mm1 \n\t"
00686 ::"m"(nnz[b_idx]),
00687 "m"(nnz[b_idx+d_idx])
00688 );
00689 asm volatile(
00690 "pcmpeqw %%mm7, %%mm0 \n\t"
00691 "pcmpeqw %%mm7, %%mm0 \n\t"
00692 "psrlw $15, %%mm0 \n\t"
00693 "psrlw $14, %%mm1 \n\t"
00694 "movq %%mm0, %%mm2 \n\t"
00695 "por %%mm1, %%mm2 \n\t"
00696 "psrlw $1, %%mm1 \n\t"
00697 "pandn %%mm2, %%mm1 \n\t"
00698 "movq %%mm1, %0 \n\t"
00699 :"=m"(*bS[dir][edge])
00700 ::"memory"
00701 );
00702 }
00703 edges = 4;
00704 step = 1;
00705 }
00706 asm volatile(
00707 "movq (%0), %%mm0 \n\t"
00708 "movq 8(%0), %%mm1 \n\t"
00709 "movq 16(%0), %%mm2 \n\t"
00710 "movq 24(%0), %%mm3 \n\t"
00711 TRANSPOSE4(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4)
00712 "movq %%mm0, (%0) \n\t"
00713 "movq %%mm3, 8(%0) \n\t"
00714 "movq %%mm4, 16(%0) \n\t"
00715 "movq %%mm2, 24(%0) \n\t"
00716 ::"r"(bS[0])
00717 :"memory"
00718 );
00719 }
00720
00721
00722
00723
00724 #define QPEL_H264V_MM(A,B,C,D,E,F,OP,T,Z,d,q)\
00725 "mov"#q" "#C", "#T" \n\t"\
00726 "mov"#d" (%0), "#F" \n\t"\
00727 "paddw "#D", "#T" \n\t"\
00728 "psllw $2, "#T" \n\t"\
00729 "psubw "#B", "#T" \n\t"\
00730 "psubw "#E", "#T" \n\t"\
00731 "punpcklbw "#Z", "#F" \n\t"\
00732 "pmullw %4, "#T" \n\t"\
00733 "paddw %5, "#A" \n\t"\
00734 "add %2, %0 \n\t"\
00735 "paddw "#F", "#A" \n\t"\
00736 "paddw "#A", "#T" \n\t"\
00737 "psraw $5, "#T" \n\t"\
00738 "packuswb "#T", "#T" \n\t"\
00739 OP(T, (%1), A, d)\
00740 "add %3, %1 \n\t"
00741
00742 #define QPEL_H264HV_MM(A,B,C,D,E,F,OF,T,Z,d,q)\
00743 "mov"#q" "#C", "#T" \n\t"\
00744 "mov"#d" (%0), "#F" \n\t"\
00745 "paddw "#D", "#T" \n\t"\
00746 "psllw $2, "#T" \n\t"\
00747 "paddw %4, "#A" \n\t"\
00748 "psubw "#B", "#T" \n\t"\
00749 "psubw "#E", "#T" \n\t"\
00750 "punpcklbw "#Z", "#F" \n\t"\
00751 "pmullw %3, "#T" \n\t"\
00752 "paddw "#F", "#A" \n\t"\
00753 "add %2, %0 \n\t"\
00754 "paddw "#A", "#T" \n\t"\
00755 "mov"#q" "#T", "#OF"(%1) \n\t"
00756
00757 #define QPEL_H264V(A,B,C,D,E,F,OP) QPEL_H264V_MM(A,B,C,D,E,F,OP,%%mm6,%%mm7,d,q)
00758 #define QPEL_H264HV(A,B,C,D,E,F,OF) QPEL_H264HV_MM(A,B,C,D,E,F,OF,%%mm6,%%mm7,d,q)
00759 #define QPEL_H264V_XMM(A,B,C,D,E,F,OP) QPEL_H264V_MM(A,B,C,D,E,F,OP,%%xmm6,%%xmm7,q,dqa)
00760 #define QPEL_H264HV_XMM(A,B,C,D,E,F,OF) QPEL_H264HV_MM(A,B,C,D,E,F,OF,%%xmm6,%%xmm7,q,dqa)
00761
00762
00763 #define QPEL_H264(OPNAME, OP, MMX)\
00764 static av_noinline void OPNAME ## h264_qpel4_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
00765 int h=4;\
00766 \
00767 asm volatile(\
00768 "pxor %%mm7, %%mm7 \n\t"\
00769 "movq %5, %%mm4 \n\t"\
00770 "movq %6, %%mm5 \n\t"\
00771 "1: \n\t"\
00772 "movd -1(%0), %%mm1 \n\t"\
00773 "movd (%0), %%mm2 \n\t"\
00774 "movd 1(%0), %%mm3 \n\t"\
00775 "movd 2(%0), %%mm0 \n\t"\
00776 "punpcklbw %%mm7, %%mm1 \n\t"\
00777 "punpcklbw %%mm7, %%mm2 \n\t"\
00778 "punpcklbw %%mm7, %%mm3 \n\t"\
00779 "punpcklbw %%mm7, %%mm0 \n\t"\
00780 "paddw %%mm0, %%mm1 \n\t"\
00781 "paddw %%mm3, %%mm2 \n\t"\
00782 "movd -2(%0), %%mm0 \n\t"\
00783 "movd 3(%0), %%mm3 \n\t"\
00784 "punpcklbw %%mm7, %%mm0 \n\t"\
00785 "punpcklbw %%mm7, %%mm3 \n\t"\
00786 "paddw %%mm3, %%mm0 \n\t"\
00787 "psllw $2, %%mm2 \n\t"\
00788 "psubw %%mm1, %%mm2 \n\t"\
00789 "pmullw %%mm4, %%mm2 \n\t"\
00790 "paddw %%mm5, %%mm0 \n\t"\
00791 "paddw %%mm2, %%mm0 \n\t"\
00792 "psraw $5, %%mm0 \n\t"\
00793 "packuswb %%mm0, %%mm0 \n\t"\
00794 OP(%%mm0, (%1),%%mm6, d)\
00795 "add %3, %0 \n\t"\
00796 "add %4, %1 \n\t"\
00797 "decl %2 \n\t"\
00798 " jnz 1b \n\t"\
00799 : "+a"(src), "+c"(dst), "+m"(h)\
00800 : "d"((long)srcStride), "S"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
00801 : "memory"\
00802 );\
00803 }\
00804 static av_noinline void OPNAME ## h264_qpel4_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
00805 int h=4;\
00806 asm volatile(\
00807 "pxor %%mm7, %%mm7 \n\t"\
00808 "movq %0, %%mm4 \n\t"\
00809 "movq %1, %%mm5 \n\t"\
00810 :: "m"(ff_pw_5), "m"(ff_pw_16)\
00811 );\
00812 do{\
00813 asm volatile(\
00814 "movd -1(%0), %%mm1 \n\t"\
00815 "movd (%0), %%mm2 \n\t"\
00816 "movd 1(%0), %%mm3 \n\t"\
00817 "movd 2(%0), %%mm0 \n\t"\
00818 "punpcklbw %%mm7, %%mm1 \n\t"\
00819 "punpcklbw %%mm7, %%mm2 \n\t"\
00820 "punpcklbw %%mm7, %%mm3 \n\t"\
00821 "punpcklbw %%mm7, %%mm0 \n\t"\
00822 "paddw %%mm0, %%mm1 \n\t"\
00823 "paddw %%mm3, %%mm2 \n\t"\
00824 "movd -2(%0), %%mm0 \n\t"\
00825 "movd 3(%0), %%mm3 \n\t"\
00826 "punpcklbw %%mm7, %%mm0 \n\t"\
00827 "punpcklbw %%mm7, %%mm3 \n\t"\
00828 "paddw %%mm3, %%mm0 \n\t"\
00829 "psllw $2, %%mm2 \n\t"\
00830 "psubw %%mm1, %%mm2 \n\t"\
00831 "pmullw %%mm4, %%mm2 \n\t"\
00832 "paddw %%mm5, %%mm0 \n\t"\
00833 "paddw %%mm2, %%mm0 \n\t"\
00834 "movd (%2), %%mm3 \n\t"\
00835 "psraw $5, %%mm0 \n\t"\
00836 "packuswb %%mm0, %%mm0 \n\t"\
00837 PAVGB" %%mm3, %%mm0 \n\t"\
00838 OP(%%mm0, (%1),%%mm6, d)\
00839 "add %4, %0 \n\t"\
00840 "add %4, %1 \n\t"\
00841 "add %3, %2 \n\t"\
00842 : "+a"(src), "+c"(dst), "+d"(src2)\
00843 : "D"((long)src2Stride), "S"((long)dstStride)\
00844 : "memory"\
00845 );\
00846 }while(--h);\
00847 }\
00848 static av_noinline void OPNAME ## h264_qpel4_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
00849 src -= 2*srcStride;\
00850 asm volatile(\
00851 "pxor %%mm7, %%mm7 \n\t"\
00852 "movd (%0), %%mm0 \n\t"\
00853 "add %2, %0 \n\t"\
00854 "movd (%0), %%mm1 \n\t"\
00855 "add %2, %0 \n\t"\
00856 "movd (%0), %%mm2 \n\t"\
00857 "add %2, %0 \n\t"\
00858 "movd (%0), %%mm3 \n\t"\
00859 "add %2, %0 \n\t"\
00860 "movd (%0), %%mm4 \n\t"\
00861 "add %2, %0 \n\t"\
00862 "punpcklbw %%mm7, %%mm0 \n\t"\
00863 "punpcklbw %%mm7, %%mm1 \n\t"\
00864 "punpcklbw %%mm7, %%mm2 \n\t"\
00865 "punpcklbw %%mm7, %%mm3 \n\t"\
00866 "punpcklbw %%mm7, %%mm4 \n\t"\
00867 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\
00868 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\
00869 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\
00870 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\
00871 \
00872 : "+a"(src), "+c"(dst)\
00873 : "S"((long)srcStride), "D"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
00874 : "memory"\
00875 );\
00876 }\
00877 static av_noinline void OPNAME ## h264_qpel4_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
00878 int h=4;\
00879 int w=3;\
00880 src -= 2*srcStride+2;\
00881 while(w--){\
00882 asm volatile(\
00883 "pxor %%mm7, %%mm7 \n\t"\
00884 "movd (%0), %%mm0 \n\t"\
00885 "add %2, %0 \n\t"\
00886 "movd (%0), %%mm1 \n\t"\
00887 "add %2, %0 \n\t"\
00888 "movd (%0), %%mm2 \n\t"\
00889 "add %2, %0 \n\t"\
00890 "movd (%0), %%mm3 \n\t"\
00891 "add %2, %0 \n\t"\
00892 "movd (%0), %%mm4 \n\t"\
00893 "add %2, %0 \n\t"\
00894 "punpcklbw %%mm7, %%mm0 \n\t"\
00895 "punpcklbw %%mm7, %%mm1 \n\t"\
00896 "punpcklbw %%mm7, %%mm2 \n\t"\
00897 "punpcklbw %%mm7, %%mm3 \n\t"\
00898 "punpcklbw %%mm7, %%mm4 \n\t"\
00899 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 0*8*3)\
00900 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 1*8*3)\
00901 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 2*8*3)\
00902 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 3*8*3)\
00903 \
00904 : "+a"(src)\
00905 : "c"(tmp), "S"((long)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)\
00906 : "memory"\
00907 );\
00908 tmp += 4;\
00909 src += 4 - 9*srcStride;\
00910 }\
00911 tmp -= 3*4;\
00912 asm volatile(\
00913 "1: \n\t"\
00914 "movq (%0), %%mm0 \n\t"\
00915 "paddw 10(%0), %%mm0 \n\t"\
00916 "movq 2(%0), %%mm1 \n\t"\
00917 "paddw 8(%0), %%mm1 \n\t"\
00918 "movq 4(%0), %%mm2 \n\t"\
00919 "paddw 6(%0), %%mm2 \n\t"\
00920 "psubw %%mm1, %%mm0 \n\t"\
00921 "psraw $2, %%mm0 \n\t"\
00922 "psubw %%mm1, %%mm0 \n\t"\
00923 "paddsw %%mm2, %%mm0 \n\t"\
00924 "psraw $2, %%mm0 \n\t"\
00925 "paddw %%mm2, %%mm0 \n\t"\
00926 "psraw $6, %%mm0 \n\t"\
00927 "packuswb %%mm0, %%mm0 \n\t"\
00928 OP(%%mm0, (%1),%%mm7, d)\
00929 "add $24, %0 \n\t"\
00930 "add %3, %1 \n\t"\
00931 "decl %2 \n\t"\
00932 " jnz 1b \n\t"\
00933 : "+a"(tmp), "+c"(dst), "+m"(h)\
00934 : "S"((long)dstStride)\
00935 : "memory"\
00936 );\
00937 }\
00938 \
00939 static av_noinline void OPNAME ## h264_qpel8_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
00940 int h=8;\
00941 asm volatile(\
00942 "pxor %%mm7, %%mm7 \n\t"\
00943 "movq %5, %%mm6 \n\t"\
00944 "1: \n\t"\
00945 "movq (%0), %%mm0 \n\t"\
00946 "movq 1(%0), %%mm2 \n\t"\
00947 "movq %%mm0, %%mm1 \n\t"\
00948 "movq %%mm2, %%mm3 \n\t"\
00949 "punpcklbw %%mm7, %%mm0 \n\t"\
00950 "punpckhbw %%mm7, %%mm1 \n\t"\
00951 "punpcklbw %%mm7, %%mm2 \n\t"\
00952 "punpckhbw %%mm7, %%mm3 \n\t"\
00953 "paddw %%mm2, %%mm0 \n\t"\
00954 "paddw %%mm3, %%mm1 \n\t"\
00955 "psllw $2, %%mm0 \n\t"\
00956 "psllw $2, %%mm1 \n\t"\
00957 "movq -1(%0), %%mm2 \n\t"\
00958 "movq 2(%0), %%mm4 \n\t"\
00959 "movq %%mm2, %%mm3 \n\t"\
00960 "movq %%mm4, %%mm5 \n\t"\
00961 "punpcklbw %%mm7, %%mm2 \n\t"\
00962 "punpckhbw %%mm7, %%mm3 \n\t"\
00963 "punpcklbw %%mm7, %%mm4 \n\t"\
00964 "punpckhbw %%mm7, %%mm5 \n\t"\
00965 "paddw %%mm4, %%mm2 \n\t"\
00966 "paddw %%mm3, %%mm5 \n\t"\
00967 "psubw %%mm2, %%mm0 \n\t"\
00968 "psubw %%mm5, %%mm1 \n\t"\
00969 "pmullw %%mm6, %%mm0 \n\t"\
00970 "pmullw %%mm6, %%mm1 \n\t"\
00971 "movd -2(%0), %%mm2 \n\t"\
00972 "movd 7(%0), %%mm5 \n\t"\
00973 "punpcklbw %%mm7, %%mm2 \n\t"\
00974 "punpcklbw %%mm7, %%mm5 \n\t"\
00975 "paddw %%mm3, %%mm2 \n\t"\
00976 "paddw %%mm5, %%mm4 \n\t"\
00977 "movq %6, %%mm5 \n\t"\
00978 "paddw %%mm5, %%mm2 \n\t"\
00979 "paddw %%mm5, %%mm4 \n\t"\
00980 "paddw %%mm2, %%mm0 \n\t"\
00981 "paddw %%mm4, %%mm1 \n\t"\
00982 "psraw $5, %%mm0 \n\t"\
00983 "psraw $5, %%mm1 \n\t"\
00984 "packuswb %%mm1, %%mm0 \n\t"\
00985 OP(%%mm0, (%1),%%mm5, q)\
00986 "add %3, %0 \n\t"\
00987 "add %4, %1 \n\t"\
00988 "decl %2 \n\t"\
00989 " jnz 1b \n\t"\
00990 : "+a"(src), "+c"(dst), "+m"(h)\
00991 : "d"((long)srcStride), "S"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
00992 : "memory"\
00993 );\
00994 }\
00995 \
00996 static av_noinline void OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
00997 int h=8;\
00998 asm volatile(\
00999 "pxor %%mm7, %%mm7 \n\t"\
01000 "movq %0, %%mm6 \n\t"\
01001 :: "m"(ff_pw_5)\
01002 );\
01003 do{\
01004 asm volatile(\
01005 "movq (%0), %%mm0 \n\t"\
01006 "movq 1(%0), %%mm2 \n\t"\
01007 "movq %%mm0, %%mm1 \n\t"\
01008 "movq %%mm2, %%mm3 \n\t"\
01009 "punpcklbw %%mm7, %%mm0 \n\t"\
01010 "punpckhbw %%mm7, %%mm1 \n\t"\
01011 "punpcklbw %%mm7, %%mm2 \n\t"\
01012 "punpckhbw %%mm7, %%mm3 \n\t"\
01013 "paddw %%mm2, %%mm0 \n\t"\
01014 "paddw %%mm3, %%mm1 \n\t"\
01015 "psllw $2, %%mm0 \n\t"\
01016 "psllw $2, %%mm1 \n\t"\
01017 "movq -1(%0), %%mm2 \n\t"\
01018 "movq 2(%0), %%mm4 \n\t"\
01019 "movq %%mm2, %%mm3 \n\t"\
01020 "movq %%mm4, %%mm5 \n\t"\
01021 "punpcklbw %%mm7, %%mm2 \n\t"\
01022 "punpckhbw %%mm7, %%mm3 \n\t"\
01023 "punpcklbw %%mm7, %%mm4 \n\t"\
01024 "punpckhbw %%mm7, %%mm5 \n\t"\
01025 "paddw %%mm4, %%mm2 \n\t"\
01026 "paddw %%mm3, %%mm5 \n\t"\
01027 "psubw %%mm2, %%mm0 \n\t"\
01028 "psubw %%mm5, %%mm1 \n\t"\
01029 "pmullw %%mm6, %%mm0 \n\t"\
01030 "pmullw %%mm6, %%mm1 \n\t"\
01031 "movd -2(%0), %%mm2 \n\t"\
01032 "movd 7(%0), %%mm5 \n\t"\
01033 "punpcklbw %%mm7, %%mm2 \n\t"\
01034 "punpcklbw %%mm7, %%mm5 \n\t"\
01035 "paddw %%mm3, %%mm2 \n\t"\
01036 "paddw %%mm5, %%mm4 \n\t"\
01037 "movq %5, %%mm5 \n\t"\
01038 "paddw %%mm5, %%mm2 \n\t"\
01039 "paddw %%mm5, %%mm4 \n\t"\
01040 "paddw %%mm2, %%mm0 \n\t"\
01041 "paddw %%mm4, %%mm1 \n\t"\
01042 "psraw $5, %%mm0 \n\t"\
01043 "psraw $5, %%mm1 \n\t"\
01044 "movq (%2), %%mm4 \n\t"\
01045 "packuswb %%mm1, %%mm0 \n\t"\
01046 PAVGB" %%mm4, %%mm0 \n\t"\
01047 OP(%%mm0, (%1),%%mm5, q)\
01048 "add %4, %0 \n\t"\
01049 "add %4, %1 \n\t"\
01050 "add %3, %2 \n\t"\
01051 : "+a"(src), "+c"(dst), "+d"(src2)\
01052 : "D"((long)src2Stride), "S"((long)dstStride),\
01053 "m"(ff_pw_16)\
01054 : "memory"\
01055 );\
01056 }while(--h);\
01057 }\
01058 \
01059 static av_noinline void OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
01060 int w= 2;\
01061 src -= 2*srcStride;\
01062 \
01063 while(w--){\
01064 asm volatile(\
01065 "pxor %%mm7, %%mm7 \n\t"\
01066 "movd (%0), %%mm0 \n\t"\
01067 "add %2, %0 \n\t"\
01068 "movd (%0), %%mm1 \n\t"\
01069 "add %2, %0 \n\t"\
01070 "movd (%0), %%mm2 \n\t"\
01071 "add %2, %0 \n\t"\
01072 "movd (%0), %%mm3 \n\t"\
01073 "add %2, %0 \n\t"\
01074 "movd (%0), %%mm4 \n\t"\
01075 "add %2, %0 \n\t"\
01076 "punpcklbw %%mm7, %%mm0 \n\t"\
01077 "punpcklbw %%mm7, %%mm1 \n\t"\
01078 "punpcklbw %%mm7, %%mm2 \n\t"\
01079 "punpcklbw %%mm7, %%mm3 \n\t"\
01080 "punpcklbw %%mm7, %%mm4 \n\t"\
01081 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\
01082 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\
01083 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\
01084 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\
01085 QPEL_H264V(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, OP)\
01086 QPEL_H264V(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, OP)\
01087 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\
01088 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\
01089 \
01090 : "+a"(src), "+c"(dst)\
01091 : "S"((long)srcStride), "D"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
01092 : "memory"\
01093 );\
01094 if(h==16){\
01095 asm volatile(\
01096 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\
01097 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\
01098 QPEL_H264V(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, OP)\
01099 QPEL_H264V(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, OP)\
01100 QPEL_H264V(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, OP)\
01101 QPEL_H264V(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, OP)\
01102 QPEL_H264V(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, OP)\
01103 QPEL_H264V(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, OP)\
01104 \
01105 : "+a"(src), "+c"(dst)\
01106 : "S"((long)srcStride), "D"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
01107 : "memory"\
01108 );\
01109 }\
01110 src += 4-(h+5)*srcStride;\
01111 dst += 4-h*dstStride;\
01112 }\
01113 }\
01114 static av_noinline void OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride, int size){\
01115 int h = size;\
01116 int w = (size+8)>>2;\
01117 src -= 2*srcStride+2;\
01118 while(w--){\
01119 asm volatile(\
01120 "pxor %%mm7, %%mm7 \n\t"\
01121 "movd (%0), %%mm0 \n\t"\
01122 "add %2, %0 \n\t"\
01123 "movd (%0), %%mm1 \n\t"\
01124 "add %2, %0 \n\t"\
01125 "movd (%0), %%mm2 \n\t"\
01126 "add %2, %0 \n\t"\
01127 "movd (%0), %%mm3 \n\t"\
01128 "add %2, %0 \n\t"\
01129 "movd (%0), %%mm4 \n\t"\
01130 "add %2, %0 \n\t"\
01131 "punpcklbw %%mm7, %%mm0 \n\t"\
01132 "punpcklbw %%mm7, %%mm1 \n\t"\
01133 "punpcklbw %%mm7, %%mm2 \n\t"\
01134 "punpcklbw %%mm7, %%mm3 \n\t"\
01135 "punpcklbw %%mm7, %%mm4 \n\t"\
01136 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 0*48)\
01137 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 1*48)\
01138 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 2*48)\
01139 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 3*48)\
01140 QPEL_H264HV(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, 4*48)\
01141 QPEL_H264HV(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, 5*48)\
01142 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 6*48)\
01143 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 7*48)\
01144 : "+a"(src)\
01145 : "c"(tmp), "S"((long)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)\
01146 : "memory"\
01147 );\
01148 if(size==16){\
01149 asm volatile(\
01150 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 8*48)\
01151 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 9*48)\
01152 QPEL_H264HV(%%mm4, %%mm5, %%mm0, %%mm1, %%mm2, %%mm3, 10*48)\
01153 QPEL_H264HV(%%mm5, %%mm0, %%mm1, %%mm2, %%mm3, %%mm4, 11*48)\
01154 QPEL_H264HV(%%mm0, %%mm1, %%mm2, %%mm3, %%mm4, %%mm5, 12*48)\
01155 QPEL_H264HV(%%mm1, %%mm2, %%mm3, %%mm4, %%mm5, %%mm0, 13*48)\
01156 QPEL_H264HV(%%mm2, %%mm3, %%mm4, %%mm5, %%mm0, %%mm1, 14*48)\
01157 QPEL_H264HV(%%mm3, %%mm4, %%mm5, %%mm0, %%mm1, %%mm2, 15*48)\
01158 : "+a"(src)\
01159 : "c"(tmp), "S"((long)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)\
01160 : "memory"\
01161 );\
01162 }\
01163 tmp += 4;\
01164 src += 4 - (size+5)*srcStride;\
01165 }\
01166 tmp -= size+8;\
01167 w = size>>4;\
01168 do{\
01169 h = size;\
01170 asm volatile(\
01171 "1: \n\t"\
01172 "movq (%0), %%mm0 \n\t"\
01173 "movq 8(%0), %%mm3 \n\t"\
01174 "movq 2(%0), %%mm1 \n\t"\
01175 "movq 10(%0), %%mm4 \n\t"\
01176 "paddw %%mm4, %%mm0 \n\t"\
01177 "paddw %%mm3, %%mm1 \n\t"\
01178 "paddw 18(%0), %%mm3 \n\t"\
01179 "paddw 16(%0), %%mm4 \n\t"\
01180 "movq 4(%0), %%mm2 \n\t"\
01181 "movq 12(%0), %%mm5 \n\t"\
01182 "paddw 6(%0), %%mm2 \n\t"\
01183 "paddw 14(%0), %%mm5 \n\t"\
01184 "psubw %%mm1, %%mm0 \n\t"\
01185 "psubw %%mm4, %%mm3 \n\t"\
01186 "psraw $2, %%mm0 \n\t"\
01187 "psraw $2, %%mm3 \n\t"\
01188 "psubw %%mm1, %%mm0 \n\t"\
01189 "psubw %%mm4, %%mm3 \n\t"\
01190 "paddsw %%mm2, %%mm0 \n\t"\
01191 "paddsw %%mm5, %%mm3 \n\t"\
01192 "psraw $2, %%mm0 \n\t"\
01193 "psraw $2, %%mm3 \n\t"\
01194 "paddw %%mm2, %%mm0 \n\t"\
01195 "paddw %%mm5, %%mm3 \n\t"\
01196 "psraw $6, %%mm0 \n\t"\
01197 "psraw $6, %%mm3 \n\t"\
01198 "packuswb %%mm3, %%mm0 \n\t"\
01199 OP(%%mm0, (%1),%%mm7, q)\
01200 "add $48, %0 \n\t"\
01201 "add %3, %1 \n\t"\
01202 "decl %2 \n\t"\
01203 " jnz 1b \n\t"\
01204 : "+a"(tmp), "+c"(dst), "+m"(h)\
01205 : "S"((long)dstStride)\
01206 : "memory"\
01207 );\
01208 tmp += 8 - size*24;\
01209 dst += 8 - size*dstStride;\
01210 }while(w--);\
01211 }\
01212 \
01213 static void OPNAME ## h264_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
01214 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst , src , dstStride, srcStride, 8);\
01215 }\
01216 static av_noinline void OPNAME ## h264_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
01217 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst , src , dstStride, srcStride, 16);\
01218 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride, 16);\
01219 }\
01220 \
01221 static av_noinline void OPNAME ## h264_qpel16_h_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
01222 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst , src , dstStride, srcStride);\
01223 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\
01224 src += 8*srcStride;\
01225 dst += 8*dstStride;\
01226 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst , src , dstStride, srcStride);\
01227 OPNAME ## h264_qpel8_h_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride);\
01228 }\
01229 \
01230 static av_noinline void OPNAME ## h264_qpel16_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
01231 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst , src , src2 , dstStride, src2Stride);\
01232 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\
01233 src += 8*dstStride;\
01234 dst += 8*dstStride;\
01235 src2 += 8*src2Stride;\
01236 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst , src , src2 , dstStride, src2Stride);\
01237 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\
01238 }\
01239 \
01240 static void OPNAME ## h264_qpel8_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
01241 OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst , tmp , src , dstStride, tmpStride, srcStride, 8);\
01242 }\
01243 \
01244 static void OPNAME ## h264_qpel16_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
01245 OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst , tmp , src , dstStride, tmpStride, srcStride, 16);\
01246 }\
01247 \
01248 static av_noinline void OPNAME ## pixels4_l2_shift5_ ## MMX(uint8_t *dst, int16_t *src16, uint8_t *src8, int dstStride, int src8Stride, int h)\
01249 {\
01250 asm volatile(\
01251 "movq (%1), %%mm0 \n\t"\
01252 "movq 24(%1), %%mm1 \n\t"\
01253 "psraw $5, %%mm0 \n\t"\
01254 "psraw $5, %%mm1 \n\t"\
01255 "packuswb %%mm0, %%mm0 \n\t"\
01256 "packuswb %%mm1, %%mm1 \n\t"\
01257 PAVGB" (%0), %%mm0 \n\t"\
01258 PAVGB" (%0,%3), %%mm1 \n\t"\
01259 OP(%%mm0, (%2), %%mm4, d)\
01260 OP(%%mm1, (%2,%4), %%mm5, d)\
01261 "lea (%0,%3,2), %0 \n\t"\
01262 "lea (%2,%4,2), %2 \n\t"\
01263 "movq 48(%1), %%mm0 \n\t"\
01264 "movq 72(%1), %%mm1 \n\t"\
01265 "psraw $5, %%mm0 \n\t"\
01266 "psraw $5, %%mm1 \n\t"\
01267 "packuswb %%mm0, %%mm0 \n\t"\
01268 "packuswb %%mm1, %%mm1 \n\t"\
01269 PAVGB" (%0), %%mm0 \n\t"\
01270 PAVGB" (%0,%3), %%mm1 \n\t"\
01271 OP(%%mm0, (%2), %%mm4, d)\
01272 OP(%%mm1, (%2,%4), %%mm5, d)\
01273 :"+a"(src8), "+c"(src16), "+d"(dst)\
01274 :"S"((long)src8Stride), "D"((long)dstStride)\
01275 :"memory");\
01276 }\
01277 static av_noinline void OPNAME ## pixels8_l2_shift5_ ## MMX(uint8_t *dst, int16_t *src16, uint8_t *src8, int dstStride, int src8Stride, int h)\
01278 {\
01279 do{\
01280 asm volatile(\
01281 "movq (%1), %%mm0 \n\t"\
01282 "movq 8(%1), %%mm1 \n\t"\
01283 "movq 48(%1), %%mm2 \n\t"\
01284 "movq 8+48(%1), %%mm3 \n\t"\
01285 "psraw $5, %%mm0 \n\t"\
01286 "psraw $5, %%mm1 \n\t"\
01287 "psraw $5, %%mm2 \n\t"\
01288 "psraw $5, %%mm3 \n\t"\
01289 "packuswb %%mm1, %%mm0 \n\t"\
01290 "packuswb %%mm3, %%mm2 \n\t"\
01291 PAVGB" (%0), %%mm0 \n\t"\
01292 PAVGB" (%0,%3), %%mm2 \n\t"\
01293 OP(%%mm0, (%2), %%mm5, q)\
01294 OP(%%mm2, (%2,%4), %%mm5, q)\
01295 ::"a"(src8), "c"(src16), "d"(dst),\
01296 "r"((long)src8Stride), "r"((long)dstStride)\
01297 :"memory");\
01298 src8 += 2L*src8Stride;\
01299 src16 += 48;\
01300 dst += 2L*dstStride;\
01301 }while(h-=2);\
01302 }\
01303 static void OPNAME ## pixels16_l2_shift5_ ## MMX(uint8_t *dst, int16_t *src16, uint8_t *src8, int dstStride, int src8Stride, int h)\
01304 {\
01305 OPNAME ## pixels8_l2_shift5_ ## MMX(dst , src16 , src8 , dstStride, src8Stride, h);\
01306 OPNAME ## pixels8_l2_shift5_ ## MMX(dst+8, src16+8, src8+8, dstStride, src8Stride, h);\
01307 }\
01308
01309 #ifdef ARCH_X86_64
01310 #define QPEL_H264_HL2_XMM(OPNAME, OP, MMX)\
01311 static av_noinline void OPNAME ## h264_qpel16_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
01312 int h=16;\
01313 asm volatile(\
01314 "pxor %%xmm15, %%xmm15 \n\t"\
01315 "movdqa %6, %%xmm14 \n\t"\
01316 "movdqa %7, %%xmm13 \n\t"\
01317 "1: \n\t"\
01318 "lddqu 3(%0), %%xmm1 \n\t"\
01319 "lddqu -5(%0), %%xmm7 \n\t"\
01320 "movdqa %%xmm1, %%xmm0 \n\t"\
01321 "punpckhbw %%xmm15, %%xmm1 \n\t"\
01322 "punpcklbw %%xmm15, %%xmm0 \n\t"\
01323 "punpcklbw %%xmm15, %%xmm7 \n\t"\
01324 "movdqa %%xmm1, %%xmm2 \n\t"\
01325 "movdqa %%xmm0, %%xmm6 \n\t"\
01326 "movdqa %%xmm1, %%xmm3 \n\t"\
01327 "movdqa %%xmm0, %%xmm8 \n\t"\
01328 "movdqa %%xmm1, %%xmm4 \n\t"\
01329 "movdqa %%xmm0, %%xmm9 \n\t"\
01330 "movdqa %%xmm1, %%xmm5 \n\t"\
01331 "movdqa %%xmm0, %%xmm10 \n\t"\
01332 "palignr $6, %%xmm0, %%xmm5 \n\t"\
01333 "palignr $6, %%xmm7, %%xmm10\n\t"\
01334 "palignr $8, %%xmm0, %%xmm4 \n\t"\
01335 "palignr $8, %%xmm7, %%xmm9 \n\t"\
01336 "palignr $10,%%xmm0, %%xmm3 \n\t"\
01337 "palignr $10,%%xmm7, %%xmm8 \n\t"\
01338 "paddw %%xmm1, %%xmm5 \n\t"\
01339 "paddw %%xmm0, %%xmm10 \n\t"\
01340 "palignr $12,%%xmm0, %%xmm2 \n\t"\
01341 "palignr $12,%%xmm7, %%xmm6 \n\t"\
01342 "palignr $14,%%xmm0, %%xmm1 \n\t"\
01343 "palignr $14,%%xmm7, %%xmm0 \n\t"\
01344 "paddw %%xmm3, %%xmm2 \n\t"\
01345 "paddw %%xmm8, %%xmm6 \n\t"\
01346 "paddw %%xmm4, %%xmm1 \n\t"\
01347 "paddw %%xmm9, %%xmm0 \n\t"\
01348 "psllw $2, %%xmm2 \n\t"\
01349 "psllw $2, %%xmm6 \n\t"\
01350 "psubw %%xmm1, %%xmm2 \n\t"\
01351 "psubw %%xmm0, %%xmm6 \n\t"\
01352 "paddw %%xmm13,%%xmm5 \n\t"\
01353 "paddw %%xmm13,%%xmm10 \n\t"\
01354 "pmullw %%xmm14,%%xmm2 \n\t"\
01355 "pmullw %%xmm14,%%xmm6 \n\t"\
01356 "lddqu (%2), %%xmm3 \n\t"\
01357 "paddw %%xmm5, %%xmm2 \n\t"\
01358 "paddw %%xmm10,%%xmm6 \n\t"\
01359 "psraw $5, %%xmm2 \n\t"\
01360 "psraw $5, %%xmm6 \n\t"\
01361 "packuswb %%xmm2,%%xmm6 \n\t"\
01362 "pavgb %%xmm3, %%xmm6 \n\t"\
01363 OP(%%xmm6, (%1), %%xmm4, dqa)\
01364 "add %5, %0 \n\t"\
01365 "add %5, %1 \n\t"\
01366 "add %4, %2 \n\t"\
01367 "dec %3 \n\t"\
01368 "jg 1b \n\t"\
01369 : "+a"(src), "+c"(dst), "+d"(src2), "+g"(h)\
01370 : "D"((long)src2Stride), "S"((long)dstStride),\
01371 "m"(ff_pw_5), "m"(ff_pw_16)\
01372 : "memory"\
01373 );\
01374 }
01375 #else // ARCH_X86_64
01376 #define QPEL_H264_HL2_XMM(OPNAME, OP, MMX)\
01377 static av_noinline void OPNAME ## h264_qpel16_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
01378 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst , src , src2 , dstStride, src2Stride);\
01379 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\
01380 src += 8*dstStride;\
01381 dst += 8*dstStride;\
01382 src2 += 8*src2Stride;\
01383 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst , src , src2 , dstStride, src2Stride);\
01384 OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(dst+8, src+8, src2+8, dstStride, src2Stride);\
01385 }
01386 #endif // ARCH_X86_64
01387
01388 #define QPEL_H264_XMM(OPNAME, OP, MMX)\
01389 static av_noinline void OPNAME ## h264_qpel8_h_lowpass_l2_ ## MMX(uint8_t *dst, uint8_t *src, uint8_t *src2, int dstStride, int src2Stride){\
01390 int h=8;\
01391 asm volatile(\
01392 "pxor %%xmm7, %%xmm7 \n\t"\
01393 "movdqa %0, %%xmm6 \n\t"\
01394 :: "m"(ff_pw_5)\
01395 );\
01396 do{\
01397 asm volatile(\
01398 "lddqu -5(%0), %%xmm1 \n\t"\
01399 "movdqa %%xmm1, %%xmm0 \n\t"\
01400 "punpckhbw %%xmm7, %%xmm1 \n\t"\
01401 "punpcklbw %%xmm7, %%xmm0 \n\t"\
01402 "movdqa %%xmm1, %%xmm2 \n\t"\
01403 "movdqa %%xmm1, %%xmm3 \n\t"\
01404 "movdqa %%xmm1, %%xmm4 \n\t"\
01405 "movdqa %%xmm1, %%xmm5 \n\t"\
01406 "palignr $6, %%xmm0, %%xmm5 \n\t"\
01407 "palignr $8, %%xmm0, %%xmm4 \n\t"\
01408 "palignr $10,%%xmm0, %%xmm3 \n\t"\
01409 "paddw %%xmm1, %%xmm5 \n\t"\
01410 "palignr $12,%%xmm0, %%xmm2 \n\t"\
01411 "palignr $14,%%xmm0, %%xmm1 \n\t"\
01412 "paddw %%xmm3, %%xmm2 \n\t"\
01413 "paddw %%xmm4, %%xmm1 \n\t"\
01414 "psllw $2, %%xmm2 \n\t"\
01415 "movq (%2), %%xmm3 \n\t"\
01416 "psubw %%xmm1, %%xmm2 \n\t"\
01417 "paddw %5, %%xmm5 \n\t"\
01418 "pmullw %%xmm6, %%xmm2 \n\t"\
01419 "paddw %%xmm5, %%xmm2 \n\t"\
01420 "psraw $5, %%xmm2 \n\t"\
01421 "packuswb %%xmm2, %%xmm2 \n\t"\
01422 "pavgb %%xmm3, %%xmm2 \n\t"\
01423 OP(%%xmm2, (%1), %%xmm4, q)\
01424 "add %4, %0 \n\t"\
01425 "add %4, %1 \n\t"\
01426 "add %3, %2 \n\t"\
01427 : "+a"(src), "+c"(dst), "+d"(src2)\
01428 : "D"((long)src2Stride), "S"((long)dstStride),\
01429 "m"(ff_pw_16)\
01430 : "memory"\
01431 );\
01432 }while(--h);\
01433 }\
01434 QPEL_H264_HL2_XMM(OPNAME, OP, MMX)\
01435 \
01436 static av_noinline void OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride, int h){\
01437 src -= 2*srcStride;\
01438 \
01439 asm volatile(\
01440 "pxor %%xmm7, %%xmm7 \n\t"\
01441 "movq (%0), %%xmm0 \n\t"\
01442 "add %2, %0 \n\t"\
01443 "movq (%0), %%xmm1 \n\t"\
01444 "add %2, %0 \n\t"\
01445 "movq (%0), %%xmm2 \n\t"\
01446 "add %2, %0 \n\t"\
01447 "movq (%0), %%xmm3 \n\t"\
01448 "add %2, %0 \n\t"\
01449 "movq (%0), %%xmm4 \n\t"\
01450 "add %2, %0 \n\t"\
01451 "punpcklbw %%xmm7, %%xmm0 \n\t"\
01452 "punpcklbw %%xmm7, %%xmm1 \n\t"\
01453 "punpcklbw %%xmm7, %%xmm2 \n\t"\
01454 "punpcklbw %%xmm7, %%xmm3 \n\t"\
01455 "punpcklbw %%xmm7, %%xmm4 \n\t"\
01456 QPEL_H264V_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, OP)\
01457 QPEL_H264V_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, OP)\
01458 QPEL_H264V_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, OP)\
01459 QPEL_H264V_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, OP)\
01460 QPEL_H264V_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, OP)\
01461 QPEL_H264V_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, OP)\
01462 QPEL_H264V_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, OP)\
01463 QPEL_H264V_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, OP)\
01464 \
01465 : "+a"(src), "+c"(dst)\
01466 : "S"((long)srcStride), "D"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
01467 : "memory"\
01468 );\
01469 if(h==16){\
01470 asm volatile(\
01471 QPEL_H264V_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, OP)\
01472 QPEL_H264V_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, OP)\
01473 QPEL_H264V_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, OP)\
01474 QPEL_H264V_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, OP)\
01475 QPEL_H264V_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, OP)\
01476 QPEL_H264V_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, OP)\
01477 QPEL_H264V_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, OP)\
01478 QPEL_H264V_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, OP)\
01479 \
01480 : "+a"(src), "+c"(dst)\
01481 : "S"((long)srcStride), "D"((long)dstStride), "m"(ff_pw_5), "m"(ff_pw_16)\
01482 : "memory"\
01483 );\
01484 }\
01485 }\
01486 static av_noinline void OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride, int size){\
01487 int h = size;\
01488 int w = (size+8)>>3;\
01489 src -= 2*srcStride+2;\
01490 while(w--){\
01491 asm volatile(\
01492 "pxor %%xmm7, %%xmm7 \n\t"\
01493 "movq (%0), %%xmm0 \n\t"\
01494 "add %2, %0 \n\t"\
01495 "movq (%0), %%xmm1 \n\t"\
01496 "add %2, %0 \n\t"\
01497 "movq (%0), %%xmm2 \n\t"\
01498 "add %2, %0 \n\t"\
01499 "movq (%0), %%xmm3 \n\t"\
01500 "add %2, %0 \n\t"\
01501 "movq (%0), %%xmm4 \n\t"\
01502 "add %2, %0 \n\t"\
01503 "punpcklbw %%xmm7, %%xmm0 \n\t"\
01504 "punpcklbw %%xmm7, %%xmm1 \n\t"\
01505 "punpcklbw %%xmm7, %%xmm2 \n\t"\
01506 "punpcklbw %%xmm7, %%xmm3 \n\t"\
01507 "punpcklbw %%xmm7, %%xmm4 \n\t"\
01508 QPEL_H264HV_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, 0*48)\
01509 QPEL_H264HV_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, 1*48)\
01510 QPEL_H264HV_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, 2*48)\
01511 QPEL_H264HV_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, 3*48)\
01512 QPEL_H264HV_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, 4*48)\
01513 QPEL_H264HV_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, 5*48)\
01514 QPEL_H264HV_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, 6*48)\
01515 QPEL_H264HV_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, 7*48)\
01516 : "+a"(src)\
01517 : "c"(tmp), "S"((long)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)\
01518 : "memory"\
01519 );\
01520 if(size==16){\
01521 asm volatile(\
01522 QPEL_H264HV_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, 8*48)\
01523 QPEL_H264HV_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, 9*48)\
01524 QPEL_H264HV_XMM(%%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, 10*48)\
01525 QPEL_H264HV_XMM(%%xmm5, %%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, 11*48)\
01526 QPEL_H264HV_XMM(%%xmm0, %%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, 12*48)\
01527 QPEL_H264HV_XMM(%%xmm1, %%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, 13*48)\
01528 QPEL_H264HV_XMM(%%xmm2, %%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, 14*48)\
01529 QPEL_H264HV_XMM(%%xmm3, %%xmm4, %%xmm5, %%xmm0, %%xmm1, %%xmm2, 15*48)\
01530 : "+a"(src)\
01531 : "c"(tmp), "S"((long)srcStride), "m"(ff_pw_5), "m"(ff_pw_16)\
01532 : "memory"\
01533 );\
01534 }\
01535 tmp += 8;\
01536 src += 8 - (size+5)*srcStride;\
01537 }\
01538 tmp -= size+8;\
01539 h = size;\
01540 if(size==16){\
01541 asm volatile(\
01542 "1: \n\t"\
01543 "movdqa 32(%0), %%xmm4 \n\t"\
01544 "movdqa 16(%0), %%xmm5 \n\t"\
01545 "movdqa (%0), %%xmm7 \n\t"\
01546 "movdqa %%xmm4, %%xmm3 \n\t"\
01547 "movdqa %%xmm4, %%xmm2 \n\t"\
01548 "movdqa %%xmm4, %%xmm1 \n\t"\
01549 "movdqa %%xmm4, %%xmm0 \n\t"\
01550 "palignr $10, %%xmm5, %%xmm0 \n\t"\
01551 "palignr $8, %%xmm5, %%xmm1 \n\t"\
01552 "palignr $6, %%xmm5, %%xmm2 \n\t"\
01553 "palignr $4, %%xmm5, %%xmm3 \n\t"\
01554 "palignr $2, %%xmm5, %%xmm4 \n\t"\
01555 "paddw %%xmm5, %%xmm0 \n\t"\
01556 "paddw %%xmm4, %%xmm1 \n\t"\
01557 "paddw %%xmm3, %%xmm2 \n\t"\
01558 "movdqa %%xmm5, %%xmm6 \n\t"\
01559 "movdqa %%xmm5, %%xmm4 \n\t"\
01560 "movdqa %%xmm5, %%xmm3 \n\t"\
01561 "palignr $8, %%xmm7, %%xmm4 \n\t"\
01562 "palignr $2, %%xmm7, %%xmm6 \n\t"\
01563 "palignr $10, %%xmm7, %%xmm3 \n\t"\
01564 "paddw %%xmm6, %%xmm4 \n\t"\
01565 "movdqa %%xmm5, %%xmm6 \n\t"\
01566 "palignr $6, %%xmm7, %%xmm5 \n\t"\
01567 "palignr $4, %%xmm7, %%xmm6 \n\t"\
01568 "paddw %%xmm7, %%xmm3 \n\t"\
01569 "paddw %%xmm6, %%xmm5 \n\t"\
01570 \
01571 "psubw %%xmm1, %%xmm0 \n\t"\
01572 "psubw %%xmm4, %%xmm3 \n\t"\
01573 "psraw $2, %%xmm0 \n\t"\
01574 "psraw $2, %%xmm3 \n\t"\
01575 "psubw %%xmm1, %%xmm0 \n\t"\
01576 "psubw %%xmm4, %%xmm3 \n\t"\
01577 "paddw %%xmm2, %%xmm0 \n\t"\
01578 "paddw %%xmm5, %%xmm3 \n\t"\
01579 "psraw $2, %%xmm0 \n\t"\
01580 "psraw $2, %%xmm3 \n\t"\
01581 "paddw %%xmm2, %%xmm0 \n\t"\
01582 "paddw %%xmm5, %%xmm3 \n\t"\
01583 "psraw $6, %%xmm0 \n\t"\
01584 "psraw $6, %%xmm3 \n\t"\
01585 "packuswb %%xmm0, %%xmm3 \n\t"\
01586 OP(%%xmm3, (%1), %%xmm7, dqa)\
01587 "add $48, %0 \n\t"\
01588 "add %3, %1 \n\t"\
01589 "decl %2 \n\t"\
01590 " jnz 1b \n\t"\
01591 : "+a"(tmp), "+c"(dst), "+g"(h)\
01592 : "S"((long)dstStride)\
01593 : "memory"\
01594 );\
01595 }else{\
01596 asm volatile(\
01597 "1: \n\t"\
01598 "movdqa 16(%0), %%xmm1 \n\t"\
01599 "movdqa (%0), %%xmm0 \n\t"\
01600 "movdqa %%xmm1, %%xmm2 \n\t"\
01601 "movdqa %%xmm1, %%xmm3 \n\t"\
01602 "movdqa %%xmm1, %%xmm4 \n\t"\
01603 "movdqa %%xmm1, %%xmm5 \n\t"\
01604 "palignr $10, %%xmm0, %%xmm5 \n\t"\
01605 "palignr $8, %%xmm0, %%xmm4 \n\t"\
01606 "palignr $6, %%xmm0, %%xmm3 \n\t"\
01607 "palignr $4, %%xmm0, %%xmm2 \n\t"\
01608 "palignr $2, %%xmm0, %%xmm1 \n\t"\
01609 "paddw %%xmm5, %%xmm0 \n\t"\
01610 "paddw %%xmm4, %%xmm1 \n\t"\
01611 "paddw %%xmm3, %%xmm2 \n\t"\
01612 "psubw %%xmm1, %%xmm0 \n\t"\
01613 "psraw $2, %%xmm0 \n\t"\
01614 "psubw %%xmm1, %%xmm0 \n\t"\
01615 "paddw %%xmm2, %%xmm0 \n\t"\
01616 "psraw $2, %%xmm0 \n\t"\
01617 "paddw %%xmm2, %%xmm0 \n\t"\
01618 "psraw $6, %%xmm0 \n\t"\
01619 "packuswb %%xmm0, %%xmm0 \n\t"\
01620 OP(%%xmm0, (%1), %%xmm7, q)\
01621 "add $48, %0 \n\t"\
01622 "add %3, %1 \n\t"\
01623 "decl %2 \n\t"\
01624 " jnz 1b \n\t"\
01625 : "+a"(tmp), "+c"(dst), "+g"(h)\
01626 : "S"((long)dstStride)\
01627 : "memory"\
01628 );\
01629 }\
01630 }\
01631 \
01632 static void OPNAME ## h264_qpel8_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
01633 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst , src , dstStride, srcStride, 8);\
01634 }\
01635 static av_noinline void OPNAME ## h264_qpel16_v_lowpass_ ## MMX(uint8_t *dst, uint8_t *src, int dstStride, int srcStride){\
01636 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst , src , dstStride, srcStride, 16);\
01637 OPNAME ## h264_qpel8or16_v_lowpass_ ## MMX(dst+8, src+8, dstStride, srcStride, 16);\
01638 }\
01639 static void OPNAME ## h264_qpel8_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
01640 OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst , tmp , src , dstStride, tmpStride, srcStride, 8);\
01641 }\
01642 static void OPNAME ## h264_qpel16_hv_lowpass_ ## MMX(uint8_t *dst, int16_t *tmp, uint8_t *src, int dstStride, int tmpStride, int srcStride){\
01643 OPNAME ## h264_qpel8or16_hv_lowpass_ ## MMX(dst , tmp , src , dstStride, tmpStride, srcStride, 16);\
01644 }\
01645
01646 #define put_pixels8_ssse3 put_pixels8_mmx2
01647 #define put_pixels16_ssse3 put_pixels16_sse2
01648 #define avg_pixels8_ssse3 avg_pixels8_mmx2
01649 #define avg_pixels16_ssse3 avg_pixels16_sse2
01650
01651 #define put_pixels8_l2_ssse3 put_pixels8_l2_mmx2
01652 #define put_pixels16_l2_ssse3 put_pixels16_l2_mmx2
01653 #define avg_pixels8_l2_ssse3 avg_pixels8_l2_mmx2
01654 #define avg_pixels16_l2_ssse3 avg_pixels16_l2_mmx2
01655
01656 #define put_pixels8_l2_shift5_ssse3 put_pixels8_l2_shift5_mmx2
01657 #define put_pixels16_l2_shift5_ssse3 put_pixels16_l2_shift5_mmx2
01658 #define avg_pixels8_l2_shift5_ssse3 avg_pixels8_l2_shift5_mmx2
01659 #define avg_pixels16_l2_shift5_ssse3 avg_pixels16_l2_shift5_mmx2
01660
01661 #define put_h264_qpel8_h_lowpass_ssse3 put_h264_qpel8_h_lowpass_mmx2
01662 #define put_h264_qpel16_h_lowpass_ssse3 put_h264_qpel16_h_lowpass_mmx2
01663 #define avg_h264_qpel8_h_lowpass_ssse3 avg_h264_qpel8_h_lowpass_mmx2
01664 #define avg_h264_qpel16_h_lowpass_ssse3 avg_h264_qpel16_h_lowpass_mmx2
01665
01666 #define H264_MC(OPNAME, SIZE, MMX, ALIGN) \
01667 static void OPNAME ## h264_qpel ## SIZE ## _mc00_ ## MMX (uint8_t *dst, uint8_t *src, int stride){\
01668 OPNAME ## pixels ## SIZE ## _ ## MMX(dst, src, stride, SIZE);\
01669 }\
01670 \
01671 static void OPNAME ## h264_qpel ## SIZE ## _mc10_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
01672 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, src, stride, stride);\
01673 }\
01674 \
01675 static void OPNAME ## h264_qpel ## SIZE ## _mc20_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
01676 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_ ## MMX(dst, src, stride, stride);\
01677 }\
01678 \
01679 static void OPNAME ## h264_qpel ## SIZE ## _mc30_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
01680 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, src+1, stride, stride);\
01681 }\
01682 \
01683 static void OPNAME ## h264_qpel ## SIZE ## _mc01_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
01684 DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*SIZE/8]);\
01685 uint8_t * const half= (uint8_t*)temp;\
01686 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(half, src, SIZE, stride);\
01687 OPNAME ## pixels ## SIZE ## _l2_ ## MMX(dst, src, half, stride, stride, SIZE);\
01688 }\
01689 \
01690 static void OPNAME ## h264_qpel ## SIZE ## _mc02_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
01691 OPNAME ## h264_qpel ## SIZE ## _v_lowpass_ ## MMX(dst, src, stride, stride);\
01692 }\
01693 \
01694 static void OPNAME ## h264_qpel ## SIZE ## _mc03_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
01695 DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*SIZE/8]);\
01696 uint8_t * const half= (uint8_t*)temp;\
01697 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(half, src, SIZE, stride);\
01698 OPNAME ## pixels ## SIZE ## _l2_ ## MMX(dst, src+stride, half, stride, stride, SIZE);\
01699 }\
01700 \
01701 static void OPNAME ## h264_qpel ## SIZE ## _mc11_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
01702 DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*SIZE/8]);\
01703 uint8_t * const halfV= (uint8_t*)temp;\
01704 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(halfV, src, SIZE, stride);\
01705 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, halfV, stride, SIZE);\
01706 }\
01707 \
01708 static void OPNAME ## h264_qpel ## SIZE ## _mc31_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
01709 DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*SIZE/8]);\
01710 uint8_t * const halfV= (uint8_t*)temp;\
01711 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(halfV, src+1, SIZE, stride);\
01712 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, halfV, stride, SIZE);\
01713 }\
01714 \
01715 static void OPNAME ## h264_qpel ## SIZE ## _mc13_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
01716 DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*SIZE/8]);\
01717 uint8_t * const halfV= (uint8_t*)temp;\
01718 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(halfV, src, SIZE, stride);\
01719 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, halfV, stride, SIZE);\
01720 }\
01721 \
01722 static void OPNAME ## h264_qpel ## SIZE ## _mc33_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
01723 DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*SIZE/8]);\
01724 uint8_t * const halfV= (uint8_t*)temp;\
01725 put_h264_qpel ## SIZE ## _v_lowpass_ ## MMX(halfV, src+1, SIZE, stride);\
01726 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, halfV, stride, SIZE);\
01727 }\
01728 \
01729 static void OPNAME ## h264_qpel ## SIZE ## _mc22_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
01730 DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*(SIZE<8?12:24)/4]);\
01731 int16_t * const tmp= (int16_t*)temp;\
01732 OPNAME ## h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(dst, tmp, src, stride, SIZE, stride);\
01733 }\
01734 \
01735 static void OPNAME ## h264_qpel ## SIZE ## _mc21_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
01736 DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8]);\
01737 uint8_t * const halfHV= (uint8_t*)temp;\
01738 int16_t * const tmp= ((int16_t*)temp) + SIZE*SIZE/2;\
01739 assert(((int)temp & 7) == 0);\
01740 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, tmp, src, SIZE, SIZE, stride);\
01741 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src, halfHV, stride, SIZE);\
01742 }\
01743 \
01744 static void OPNAME ## h264_qpel ## SIZE ## _mc23_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
01745 DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8]);\
01746 uint8_t * const halfHV= (uint8_t*)temp;\
01747 int16_t * const tmp= ((int16_t*)temp) + SIZE*SIZE/2;\
01748 assert(((int)temp & 7) == 0);\
01749 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, tmp, src, SIZE, SIZE, stride);\
01750 OPNAME ## h264_qpel ## SIZE ## _h_lowpass_l2_ ## MMX(dst, src+stride, halfHV, stride, SIZE);\
01751 }\
01752 \
01753 static void OPNAME ## h264_qpel ## SIZE ## _mc12_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
01754 DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8]);\
01755 int16_t * const halfV= ((int16_t*)temp) + SIZE*SIZE/2;\
01756 uint8_t * const halfHV= ((uint8_t*)temp);\
01757 assert(((int)temp & 7) == 0);\
01758 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, halfV, src, SIZE, SIZE, stride);\
01759 OPNAME ## pixels ## SIZE ## _l2_shift5_ ## MMX(dst, halfV+2, halfHV, stride, SIZE, SIZE);\
01760 }\
01761 \
01762 static void OPNAME ## h264_qpel ## SIZE ## _mc32_ ## MMX(uint8_t *dst, uint8_t *src, int stride){\
01763 DECLARE_ALIGNED(ALIGN, uint64_t, temp[SIZE*(SIZE<8?12:24)/4 + SIZE*SIZE/8]);\
01764 int16_t * const halfV= ((int16_t*)temp) + SIZE*SIZE/2;\
01765 uint8_t * const halfHV= ((uint8_t*)temp);\
01766 assert(((int)temp & 7) == 0);\
01767 put_h264_qpel ## SIZE ## _hv_lowpass_ ## MMX(halfHV, halfV, src, SIZE, SIZE, stride);\
01768 OPNAME ## pixels ## SIZE ## _l2_shift5_ ## MMX(dst, halfV+3, halfHV, stride, SIZE, SIZE);\
01769 }\
01770
01771
01772 #define AVG_3DNOW_OP(a,b,temp, size) \
01773 "mov" #size " " #b ", " #temp " \n\t"\
01774 "pavgusb " #temp ", " #a " \n\t"\
01775 "mov" #size " " #a ", " #b " \n\t"
01776 #define AVG_MMX2_OP(a,b,temp, size) \
01777 "mov" #size " " #b ", " #temp " \n\t"\
01778 "pavgb " #temp ", " #a " \n\t"\
01779 "mov" #size " " #a ", " #b " \n\t"
01780
01781 #define PAVGB "pavgusb"
01782 QPEL_H264(put_, PUT_OP, 3dnow)
01783 QPEL_H264(avg_, AVG_3DNOW_OP, 3dnow)
01784 #undef PAVGB
01785 #define PAVGB "pavgb"
01786 QPEL_H264(put_, PUT_OP, mmx2)
01787 QPEL_H264(avg_, AVG_MMX2_OP, mmx2)
01788 #ifdef HAVE_SSSE3
01789 QPEL_H264_XMM(put_, PUT_OP, ssse3)
01790 QPEL_H264_XMM(avg_, AVG_MMX2_OP, ssse3)
01791 #endif
01792 #undef PAVGB
01793
01794 H264_MC(put_, 4, 3dnow, 8)
01795 H264_MC(put_, 8, 3dnow, 8)
01796 H264_MC(put_, 16,3dnow, 8)
01797 H264_MC(avg_, 4, 3dnow, 8)
01798 H264_MC(avg_, 8, 3dnow, 8)
01799 H264_MC(avg_, 16,3dnow, 8)
01800 H264_MC(put_, 4, mmx2, 8)
01801 H264_MC(put_, 8, mmx2, 8)
01802 H264_MC(put_, 16,mmx2, 8)
01803 H264_MC(avg_, 4, mmx2, 8)
01804 H264_MC(avg_, 8, mmx2, 8)
01805 H264_MC(avg_, 16,mmx2, 8)
01806 #ifdef HAVE_SSSE3
01807
01808 H264_MC(put_, 8, ssse3, 16)
01809 H264_MC(put_, 16,ssse3, 16)
01810 H264_MC(avg_, 8, ssse3, 16)
01811 H264_MC(avg_, 16,ssse3, 16)
01812 #endif
01813
01814
01815 #define put_h264_qpel8_mc20_ssse3 put_h264_qpel8_mc20_mmx2
01816 #define put_h264_qpel16_mc20_ssse3 put_h264_qpel16_mc20_mmx2
01817 #define avg_h264_qpel8_mc20_ssse3 avg_h264_qpel8_mc20_mmx2
01818 #define avg_h264_qpel16_mc20_ssse3 avg_h264_qpel16_mc20_mmx2
01819
01820
01821 #define H264_CHROMA_OP(S,D)
01822 #define H264_CHROMA_OP4(S,D,T)
01823 #define H264_CHROMA_MC8_TMPL put_h264_chroma_mc8_mmx
01824 #define H264_CHROMA_MC4_TMPL put_h264_chroma_mc4_mmx
01825 #define H264_CHROMA_MC2_TMPL put_h264_chroma_mc2_mmx2
01826 #define H264_CHROMA_MC8_MV0 put_pixels8_mmx
01827 #include "dsputil_h264_template_mmx.c"
01828
01829 static void put_h264_chroma_mc8_mmx_rnd(uint8_t *dst, uint8_t *src, int stride, int h, int x, int y)
01830 {
01831 put_h264_chroma_mc8_mmx(dst, src, stride, h, x, y, 1);
01832 }
01833 static void put_h264_chroma_mc8_mmx_nornd(uint8_t *dst, uint8_t *src, int stride, int h, int x, int y)
01834 {
01835 put_h264_chroma_mc8_mmx(dst, src, stride, h, x, y, 0);
01836 }
01837
01838 #undef H264_CHROMA_OP
01839 #undef H264_CHROMA_OP4
01840 #undef H264_CHROMA_MC8_TMPL
01841 #undef H264_CHROMA_MC4_TMPL
01842 #undef H264_CHROMA_MC2_TMPL
01843 #undef H264_CHROMA_MC8_MV0
01844
01845 #define H264_CHROMA_OP(S,D) "pavgb " #S ", " #D " \n\t"
01846 #define H264_CHROMA_OP4(S,D,T) "movd " #S ", " #T " \n\t"\
01847 "pavgb " #T ", " #D " \n\t"
01848 #define H264_CHROMA_MC8_TMPL avg_h264_chroma_mc8_mmx2
01849 #define H264_CHROMA_MC4_TMPL avg_h264_chroma_mc4_mmx2
01850 #define H264_CHROMA_MC2_TMPL avg_h264_chroma_mc2_mmx2
01851 #define H264_CHROMA_MC8_MV0 avg_pixels8_mmx2
01852 #include "dsputil_h264_template_mmx.c"
01853 static void avg_h264_chroma_mc8_mmx2_rnd(uint8_t *dst, uint8_t *src, int stride, int h, int x, int y)
01854 {
01855 avg_h264_chroma_mc8_mmx2(dst, src, stride, h, x, y, 1);
01856 }
01857 #undef H264_CHROMA_OP
01858 #undef H264_CHROMA_OP4
01859 #undef H264_CHROMA_MC8_TMPL
01860 #undef H264_CHROMA_MC4_TMPL
01861 #undef H264_CHROMA_MC2_TMPL
01862 #undef H264_CHROMA_MC8_MV0
01863
01864 #define H264_CHROMA_OP(S,D) "pavgusb " #S ", " #D " \n\t"
01865 #define H264_CHROMA_OP4(S,D,T) "movd " #S ", " #T " \n\t"\
01866 "pavgusb " #T ", " #D " \n\t"
01867 #define H264_CHROMA_MC8_TMPL avg_h264_chroma_mc8_3dnow
01868 #define H264_CHROMA_MC4_TMPL avg_h264_chroma_mc4_3dnow
01869 #define H264_CHROMA_MC8_MV0 avg_pixels8_3dnow
01870 #include "dsputil_h264_template_mmx.c"
01871 static void avg_h264_chroma_mc8_3dnow_rnd(uint8_t *dst, uint8_t *src, int stride, int h, int x, int y)
01872 {
01873 avg_h264_chroma_mc8_3dnow(dst, src, stride, h, x, y, 1);
01874 }
01875 #undef H264_CHROMA_OP
01876 #undef H264_CHROMA_OP4
01877 #undef H264_CHROMA_MC8_TMPL
01878 #undef H264_CHROMA_MC4_TMPL
01879 #undef H264_CHROMA_MC8_MV0
01880
01881
01882
01883
01884 static inline void ff_h264_weight_WxH_mmx2(uint8_t *dst, int stride, int log2_denom, int weight, int offset, int w, int h)
01885 {
01886 int x, y;
01887 offset <<= log2_denom;
01888 offset += (1 << log2_denom) >> 1;
01889 asm volatile(
01890 "movd %0, %%mm4 \n\t"
01891 "movd %1, %%mm5 \n\t"
01892 "movd %2, %%mm6 \n\t"
01893 "pshufw $0, %%mm4, %%mm4 \n\t"
01894 "pshufw $0, %%mm5, %%mm5 \n\t"
01895 "pxor %%mm7, %%mm7 \n\t"
01896 :: "g"(weight), "g"(offset), "g"(log2_denom)
01897 );
01898 for(y=0; y<h; y+=2){
01899 for(x=0; x<w; x+=4){
01900 asm volatile(
01901 "movd %0, %%mm0 \n\t"
01902 "movd %1, %%mm1 \n\t"
01903 "punpcklbw %%mm7, %%mm0 \n\t"
01904 "punpcklbw %%mm7, %%mm1 \n\t"
01905 "pmullw %%mm4, %%mm0 \n\t"
01906 "pmullw %%mm4, %%mm1 \n\t"
01907 "paddsw %%mm5, %%mm0 \n\t"
01908 "paddsw %%mm5, %%mm1 \n\t"
01909 "psraw %%mm6, %%mm0 \n\t"
01910 "psraw %%mm6, %%mm1 \n\t"
01911 "packuswb %%mm7, %%mm0 \n\t"
01912 "packuswb %%mm7, %%mm1 \n\t"
01913 "movd %%mm0, %0 \n\t"
01914 "movd %%mm1, %1 \n\t"
01915 : "+m"(*(uint32_t*)(dst+x)),
01916 "+m"(*(uint32_t*)(dst+x+stride))
01917 );
01918 }
01919 dst += 2*stride;
01920 }
01921 }
01922
01923 static inline void ff_h264_biweight_WxH_mmx2(uint8_t *dst, uint8_t *src, int stride, int log2_denom, int weightd, int weights, int offset, int w, int h)
01924 {
01925 int x, y;
01926 offset = ((offset + 1) | 1) << log2_denom;
01927 asm volatile(
01928 "movd %0, %%mm3 \n\t"
01929 "movd %1, %%mm4 \n\t"
01930 "movd %2, %%mm5 \n\t"
01931 "movd %3, %%mm6 \n\t"
01932 "pshufw $0, %%mm3, %%mm3 \n\t"
01933 "pshufw $0, %%mm4, %%mm4 \n\t"
01934 "pshufw $0, %%mm5, %%mm5 \n\t"
01935 "pxor %%mm7, %%mm7 \n\t"
01936 :: "g"(weightd), "g"(weights), "g"(offset), "g"(log2_denom+1)
01937 );
01938 for(y=0; y<h; y++){
01939 for(x=0; x<w; x+=4){
01940 asm volatile(
01941 "movd %0, %%mm0 \n\t"
01942 "movd %1, %%mm1 \n\t"
01943 "punpcklbw %%mm7, %%mm0 \n\t"
01944 "punpcklbw %%mm7, %%mm1 \n\t"
01945 "pmullw %%mm3, %%mm0 \n\t"
01946 "pmullw %%mm4, %%mm1 \n\t"
01947 "paddsw %%mm1, %%mm0 \n\t"
01948 "paddsw %%mm5, %%mm0 \n\t"
01949 "psraw %%mm6, %%mm0 \n\t"
01950 "packuswb %%mm0, %%mm0 \n\t"
01951 "movd %%mm0, %0 \n\t"
01952 : "+m"(*(uint32_t*)(dst+x))
01953 : "m"(*(uint32_t*)(src+x))
01954 );
01955 }
01956 src += stride;
01957 dst += stride;
01958 }
01959 }
01960
01961 #define H264_WEIGHT(W,H) \
01962 static void ff_h264_biweight_ ## W ## x ## H ## _mmx2(uint8_t *dst, uint8_t *src, int stride, int log2_denom, int weightd, int weights, int offset){ \
01963 ff_h264_biweight_WxH_mmx2(dst, src, stride, log2_denom, weightd, weights, offset, W, H); \
01964 } \
01965 static void ff_h264_weight_ ## W ## x ## H ## _mmx2(uint8_t *dst, int stride, int log2_denom, int weight, int offset){ \
01966 ff_h264_weight_WxH_mmx2(dst, stride, log2_denom, weight, offset, W, H); \
01967 }
01968
01969 H264_WEIGHT(16,16)
01970 H264_WEIGHT(16, 8)
01971 H264_WEIGHT( 8,16)
01972 H264_WEIGHT( 8, 8)
01973 H264_WEIGHT( 8, 4)
01974 H264_WEIGHT( 4, 8)
01975 H264_WEIGHT( 4, 4)
01976 H264_WEIGHT( 4, 2)
01977