Lines 127-140
Link Here
|
127 |
"psrlq $8, %%mm3 \n\t" \ |
127 |
"psrlq $8, %%mm3 \n\t" \ |
128 |
"psrlq $8, %%mm6 \n\t" \ |
128 |
"psrlq $8, %%mm6 \n\t" \ |
129 |
"psrlq $8, %%mm7 \n\t" \ |
129 |
"psrlq $8, %%mm7 \n\t" \ |
130 |
"pand "MANGLE(mask24l)", %%mm0\n\t" \ |
130 |
"pand %2, %%mm0\n\t" \ |
131 |
"pand "MANGLE(mask24l)", %%mm1\n\t" \ |
131 |
"pand %2, %%mm1\n\t" \ |
132 |
"pand "MANGLE(mask24l)", %%mm4\n\t" \ |
132 |
"pand %2, %%mm4\n\t" \ |
133 |
"pand "MANGLE(mask24l)", %%mm5\n\t" \ |
133 |
"pand %2, %%mm5\n\t" \ |
134 |
"pand "MANGLE(mask24h)", %%mm2\n\t" \ |
134 |
"pand %3, %%mm2\n\t" \ |
135 |
"pand "MANGLE(mask24h)", %%mm3\n\t" \ |
135 |
"pand %3, %%mm3\n\t" \ |
136 |
"pand "MANGLE(mask24h)", %%mm6\n\t" \ |
136 |
"pand %3, %%mm6\n\t" \ |
137 |
"pand "MANGLE(mask24h)", %%mm7\n\t" \ |
137 |
"pand %3, %%mm7\n\t" \ |
138 |
"por %%mm2, %%mm0 \n\t" \ |
138 |
"por %%mm2, %%mm0 \n\t" \ |
139 |
"por %%mm3, %%mm1 \n\t" \ |
139 |
"por %%mm3, %%mm1 \n\t" \ |
140 |
"por %%mm6, %%mm4 \n\t" \ |
140 |
"por %%mm6, %%mm4 \n\t" \ |
Lines 144-157
Link Here
|
144 |
"movq %%mm4, %%mm3 \n\t" \ |
144 |
"movq %%mm4, %%mm3 \n\t" \ |
145 |
"psllq $48, %%mm2 \n\t" \ |
145 |
"psllq $48, %%mm2 \n\t" \ |
146 |
"psllq $32, %%mm3 \n\t" \ |
146 |
"psllq $32, %%mm3 \n\t" \ |
147 |
"pand "MANGLE(mask24hh)", %%mm2\n\t" \ |
147 |
"pand %4, %%mm2\n\t" \ |
148 |
"pand "MANGLE(mask24hhh)", %%mm3\n\t" \ |
148 |
"pand %5, %%mm3\n\t" \ |
149 |
"por %%mm2, %%mm0 \n\t" \ |
149 |
"por %%mm2, %%mm0 \n\t" \ |
150 |
"psrlq $16, %%mm1 \n\t" \ |
150 |
"psrlq $16, %%mm1 \n\t" \ |
151 |
"psrlq $32, %%mm4 \n\t" \ |
151 |
"psrlq $32, %%mm4 \n\t" \ |
152 |
"psllq $16, %%mm5 \n\t" \ |
152 |
"psllq $16, %%mm5 \n\t" \ |
153 |
"por %%mm3, %%mm1 \n\t" \ |
153 |
"por %%mm3, %%mm1 \n\t" \ |
154 |
"pand "MANGLE(mask24hhhh)", %%mm5\n\t" \ |
154 |
"pand %6, %%mm5\n\t" \ |
155 |
"por %%mm5, %%mm4 \n\t" \ |
155 |
"por %%mm5, %%mm4 \n\t" \ |
156 |
\ |
156 |
\ |
157 |
MOVNTQ" %%mm0, %0 \n\t" \ |
157 |
MOVNTQ" %%mm0, %0 \n\t" \ |
Lines 184-190
Link Here
|
184 |
"movq %%mm5, %%mm7 \n\t" |
184 |
"movq %%mm5, %%mm7 \n\t" |
185 |
STORE_BGR24_MMX |
185 |
STORE_BGR24_MMX |
186 |
:"=m"(*dest) |
186 |
:"=m"(*dest) |
187 |
:"m"(*s) |
187 |
:"m"(*s), "m"(mask24l), "m"(mask24h), "m"(mask24hh), "m"(mask24hhh), "m"(mask24hhhh) |
188 |
:"memory"); |
188 |
:"memory"); |
189 |
dest += 24; |
189 |
dest += 24; |
190 |
s += 32; |
190 |
s += 32; |
Lines 977-983
Link Here
|
977 |
STORE_BGR24_MMX |
977 |
STORE_BGR24_MMX |
978 |
|
978 |
|
979 |
:"=m"(*d) |
979 |
:"=m"(*d) |
980 |
:"m"(*s) |
980 |
:"m"(*s), "m"(mask24l), "m"(mask24h), "m"(mask24hh), "m"(mask24hhh), "m"(mask24hhhh) |
981 |
:"memory"); |
981 |
:"memory"); |
982 |
d += 24; |
982 |
d += 24; |
983 |
s += 8; |
983 |
s += 8; |
Lines 1083-1089
Link Here
|
1083 |
STORE_BGR24_MMX |
1083 |
STORE_BGR24_MMX |
1084 |
|
1084 |
|
1085 |
:"=m"(*d) |
1085 |
:"=m"(*d) |
1086 |
:"m"(*s) |
1086 |
:"m"(*s), "m"(mask24l), "m"(mask24h), "m"(mask24hh), "m"(mask24hhh), "m"(mask24hhhh) |
1087 |
:"memory"); |
1087 |
:"memory"); |
1088 |
d += 24; |
1088 |
d += 24; |
1089 |
s += 8; |
1089 |
s += 8; |
Lines 1297-1305
Link Here
|
1297 |
__asm__ volatile ( |
1297 |
__asm__ volatile ( |
1298 |
"test %%"REG_a", %%"REG_a" \n\t" |
1298 |
"test %%"REG_a", %%"REG_a" \n\t" |
1299 |
"jns 2f \n\t" |
1299 |
"jns 2f \n\t" |
1300 |
"movq "MANGLE(mask24r)", %%mm5 \n\t" |
1300 |
"movq %3, %%mm5 \n\t" |
1301 |
"movq "MANGLE(mask24g)", %%mm6 \n\t" |
1301 |
"movq %4, %%mm6 \n\t" |
1302 |
"movq "MANGLE(mask24b)", %%mm7 \n\t" |
1302 |
"movq %5, %%mm7 \n\t" |
1303 |
ASMALIGN(4) |
1303 |
ASMALIGN(4) |
1304 |
"1: \n\t" |
1304 |
"1: \n\t" |
1305 |
PREFETCH" 32(%1, %%"REG_a") \n\t" |
1305 |
PREFETCH" 32(%1, %%"REG_a") \n\t" |
Lines 1335-1341
Link Here
|
1335 |
" js 1b \n\t" |
1335 |
" js 1b \n\t" |
1336 |
"2: \n\t" |
1336 |
"2: \n\t" |
1337 |
: "+a" (mmx_size) |
1337 |
: "+a" (mmx_size) |
1338 |
: "r" (src-mmx_size), "r"(dst-mmx_size) |
1338 |
: "r" (src-mmx_size), "r"(dst-mmx_size), "m"(mask24r), "m"(mask24g), "m"(mask24b) |
1339 |
); |
1339 |
); |
1340 |
|
1340 |
|
1341 |
__asm__ volatile(SFENCE:::"memory"); |
1341 |
__asm__ volatile(SFENCE:::"memory"); |
Lines 2007-2014
Link Here
|
2007 |
for (i=0; i<2; i++) { |
2007 |
for (i=0; i<2; i++) { |
2008 |
__asm__ volatile( |
2008 |
__asm__ volatile( |
2009 |
"mov %2, %%"REG_a" \n\t" |
2009 |
"mov %2, %%"REG_a" \n\t" |
2010 |
"movq "MANGLE(ff_bgr2YCoeff)", %%mm6 \n\t" |
2010 |
"movq %3, %%mm6 \n\t" |
2011 |
"movq "MANGLE(ff_w1111)", %%mm5 \n\t" |
2011 |
"movq %4, %%mm5 \n\t" |
2012 |
"pxor %%mm7, %%mm7 \n\t" |
2012 |
"pxor %%mm7, %%mm7 \n\t" |
2013 |
"lea (%%"REG_a", %%"REG_a", 2), %%"REG_d" \n\t" |
2013 |
"lea (%%"REG_a", %%"REG_a", 2), %%"REG_d" \n\t" |
2014 |
ASMALIGN(4) |
2014 |
ASMALIGN(4) |
Lines 2066-2077
Link Here
|
2066 |
"psraw $7, %%mm4 \n\t" |
2066 |
"psraw $7, %%mm4 \n\t" |
2067 |
|
2067 |
|
2068 |
"packuswb %%mm4, %%mm0 \n\t" |
2068 |
"packuswb %%mm4, %%mm0 \n\t" |
2069 |
"paddusb "MANGLE(ff_bgr2YOffset)", %%mm0 \n\t" |
2069 |
"paddusb %5, %%mm0 \n\t" |
2070 |
|
2070 |
|
2071 |
MOVNTQ" %%mm0, (%1, %%"REG_a") \n\t" |
2071 |
MOVNTQ" %%mm0, (%1, %%"REG_a") \n\t" |
2072 |
"add $8, %%"REG_a" \n\t" |
2072 |
"add $8, %%"REG_a" \n\t" |
2073 |
" js 1b \n\t" |
2073 |
" js 1b \n\t" |
2074 |
: : "r" (src+width*3), "r" (ydst+width), "g" ((x86_reg)-width) |
2074 |
: : "r" (src+width*3), "r" (ydst+width), "g" ((x86_reg)-width), |
|
|
2075 |
"m"(ff_bgr2YCoeff), "m"(ff_w1111), "m"(ff_bgr2YOffset) |
2075 |
: "%"REG_a, "%"REG_d |
2076 |
: "%"REG_a, "%"REG_d |
2076 |
); |
2077 |
); |
2077 |
ydst += lumStride; |
2078 |
ydst += lumStride; |
Lines 2080-2087
Link Here
|
2080 |
src -= srcStride*2; |
2081 |
src -= srcStride*2; |
2081 |
__asm__ volatile( |
2082 |
__asm__ volatile( |
2082 |
"mov %4, %%"REG_a" \n\t" |
2083 |
"mov %4, %%"REG_a" \n\t" |
2083 |
"movq "MANGLE(ff_w1111)", %%mm5 \n\t" |
2084 |
"movq %5, %%mm5 \n\t" |
2084 |
"movq "MANGLE(ff_bgr2UCoeff)", %%mm6 \n\t" |
2085 |
"movq %6, %%mm6 \n\t" |
2085 |
"pxor %%mm7, %%mm7 \n\t" |
2086 |
"pxor %%mm7, %%mm7 \n\t" |
2086 |
"lea (%%"REG_a", %%"REG_a", 2), %%"REG_d" \n\t" |
2087 |
"lea (%%"REG_a", %%"REG_a", 2), %%"REG_d" \n\t" |
2087 |
"add %%"REG_d", %%"REG_d" \n\t" |
2088 |
"add %%"REG_d", %%"REG_d" \n\t" |
Lines 2130-2137
Link Here
|
2130 |
"psrlw $2, %%mm0 \n\t" |
2131 |
"psrlw $2, %%mm0 \n\t" |
2131 |
"psrlw $2, %%mm2 \n\t" |
2132 |
"psrlw $2, %%mm2 \n\t" |
2132 |
#endif |
2133 |
#endif |
2133 |
"movq "MANGLE(ff_bgr2VCoeff)", %%mm1 \n\t" |
2134 |
"movq %7, %%mm1 \n\t" |
2134 |
"movq "MANGLE(ff_bgr2VCoeff)", %%mm3 \n\t" |
2135 |
"movq %7, %%mm3 \n\t" |
2135 |
|
2136 |
|
2136 |
"pmaddwd %%mm0, %%mm1 \n\t" |
2137 |
"pmaddwd %%mm0, %%mm1 \n\t" |
2137 |
"pmaddwd %%mm2, %%mm3 \n\t" |
2138 |
"pmaddwd %%mm2, %%mm3 \n\t" |
Lines 2188-2199
Link Here
|
2188 |
"paddw %%mm1, %%mm5 \n\t" |
2189 |
"paddw %%mm1, %%mm5 \n\t" |
2189 |
"paddw %%mm3, %%mm2 \n\t" |
2190 |
"paddw %%mm3, %%mm2 \n\t" |
2190 |
"paddw %%mm5, %%mm2 \n\t" |
2191 |
"paddw %%mm5, %%mm2 \n\t" |
2191 |
"movq "MANGLE(ff_w1111)", %%mm5 \n\t" |
2192 |
"movq %5, %%mm5 \n\t" |
2192 |
"psrlw $2, %%mm4 \n\t" |
2193 |
"psrlw $2, %%mm4 \n\t" |
2193 |
"psrlw $2, %%mm2 \n\t" |
2194 |
"psrlw $2, %%mm2 \n\t" |
2194 |
#endif |
2195 |
#endif |
2195 |
"movq "MANGLE(ff_bgr2VCoeff)", %%mm1 \n\t" |
2196 |
"movq %7, %%mm1 \n\t" |
2196 |
"movq "MANGLE(ff_bgr2VCoeff)", %%mm3 \n\t" |
2197 |
"movq %7, %%mm3 \n\t" |
2197 |
|
2198 |
|
2198 |
"pmaddwd %%mm4, %%mm1 \n\t" |
2199 |
"pmaddwd %%mm4, %%mm1 \n\t" |
2199 |
"pmaddwd %%mm2, %%mm3 \n\t" |
2200 |
"pmaddwd %%mm2, %%mm3 \n\t" |
Lines 2217-2229
Link Here
|
2217 |
"punpckldq %%mm4, %%mm0 \n\t" |
2218 |
"punpckldq %%mm4, %%mm0 \n\t" |
2218 |
"punpckhdq %%mm4, %%mm1 \n\t" |
2219 |
"punpckhdq %%mm4, %%mm1 \n\t" |
2219 |
"packsswb %%mm1, %%mm0 \n\t" |
2220 |
"packsswb %%mm1, %%mm0 \n\t" |
2220 |
"paddb "MANGLE(ff_bgr2UVOffset)", %%mm0 \n\t" |
2221 |
"paddb %8, %%mm0 \n\t" |
2221 |
"movd %%mm0, (%2, %%"REG_a") \n\t" |
2222 |
"movd %%mm0, (%2, %%"REG_a") \n\t" |
2222 |
"punpckhdq %%mm0, %%mm0 \n\t" |
2223 |
"punpckhdq %%mm0, %%mm0 \n\t" |
2223 |
"movd %%mm0, (%3, %%"REG_a") \n\t" |
2224 |
"movd %%mm0, (%3, %%"REG_a") \n\t" |
2224 |
"add $4, %%"REG_a" \n\t" |
2225 |
"add $4, %%"REG_a" \n\t" |
2225 |
" js 1b \n\t" |
2226 |
" js 1b \n\t" |
2226 |
: : "r" (src+chromWidth*6), "r" (src+srcStride+chromWidth*6), "r" (udst+chromWidth), "r" (vdst+chromWidth), "g" (-chromWidth) |
2227 |
: : "r" (src+chromWidth*6), "r" (src+srcStride+chromWidth*6), "r" (udst+chromWidth), "r" (vdst+chromWidth), "g" (-chromWidth), |
|
|
2228 |
"m"(ff_w1111), "m"(ff_bgr2UCoeff), "m"(ff_bgr2VCoeff), "m"(ff_bgr2UVOffset) |
2227 |
: "%"REG_a, "%"REG_d |
2229 |
: "%"REG_a, "%"REG_d |
2228 |
); |
2230 |
); |
2229 |
|
2231 |
|