Mercurial > libavcodec.hg

comparison x86/vp8dsp.asm @ 12205:d38e8565ba05 libavcodec

Find changesets by keywords (author, files, the commit message), revision number or hash, or revset expression.
VP8 MBedge loopfilter MMX/MMX2/SSE2 functions for both luma (width=16) and chroma (width=8).
author rbultje
date Tue, 20 Jul 2010 22:58:56 +0000
parents 563339ea87aa
children 9eef00a43280
comparison
equal deleted inserted replaced
12204:563339ea87aa 12205:d38e8565ba05
148 cextern pb_1 148 cextern pb_1
149 cextern pw_3 149 cextern pw_3
150 cextern pb_3 150 cextern pb_3
151 cextern pw_4 151 cextern pw_4
152 cextern pb_4 152 cextern pb_4
153 cextern pw_9
154 cextern pw_18
155 cextern pw_27
156 cextern pw_63
153 cextern pw_64 157 cextern pw_64
154 cextern pb_80 158 cextern pb_80
155 cextern pb_F8 159 cextern pb_F8
156 cextern pb_FE 160 cextern pb_FE
157 161
1892 %else 1896 %else
1893 INNER_LOOPFILTER sse2, h, 6, 16, 13 1897 INNER_LOOPFILTER sse2, h, 6, 16, 13
1894 %endif 1898 %endif
1895 INNER_LOOPFILTER sse2, v, 6, 8, 13 1899 INNER_LOOPFILTER sse2, v, 6, 8, 13
1896 INNER_LOOPFILTER sse2, h, 6, 8, 13 1900 INNER_LOOPFILTER sse2, h, 6, 8, 13
1901
1902 ;-----------------------------------------------------------------------------
1903 ; void vp8_h/v_loop_filter<size>_mbedge_<opt>(uint8_t *dst, [uint8_t *v,] int stride,
1904 ; int flimE, int flimI, int hev_thr);
1905 ;-----------------------------------------------------------------------------
1906
1907 ; write 4 or 8 words in the mmx/xmm registers as 8 lines
1908 ; 1 and 2 are the registers to write, this can be the same (for SSE2)
1909 ; 3 is a general-purpose register that we will clobber
1910 ; 4 is a pointer to the destination's 4th line
1911 ; 5 is -stride and +stride
1912 %macro WRITE_8W 6
1913 movd %3, %1
1914 %if mmsize == 8
1915 punpckhdq %1, %1
1916 %else
1917 psrldq %1, 4
1918 %endif
1919 mov [%4+%5*4], %3w
1920 shr %3, 16
1921 add %4, %6
1922 mov [%4+%5*4], %3w
1923
1924 movd %3, %1
1925 %if mmsize == 16
1926 psrldq %1, 4
1927 %endif
1928 add %4, %5
1929 mov [%4+%5*2], %3w
1930 shr %3, 16
1931 mov [%4+%5 ], %3w
1932
1933 movd %3, %2
1934 %if mmsize == 8
1935 punpckhdq %2, %2
1936 %else
1937 psrldq %2, 4
1938 %endif
1939 mov [%4 ], %3w
1940 shr %3, 16
1941 mov [%4+%6 ], %3w
1942
1943 movd %3, %2
1944 add %4, %6
1945 mov [%4+%6 ], %3w
1946 shr %3, 16
1947 mov [%4+%6*2], %3w
1948 %if mmsize == 8
1949 add %4, %5
1950 %endif
1951 %endmacro
1952
1953 %macro MBEDGE_LOOPFILTER 5
1954 %if %4 == 8 ; chroma
1955 cglobal vp8_%2_loop_filter8uv_mbedge_%1, 6, %3, %5
1956 %define dst8_reg r1
1957 %define mstride_reg r2
1958 %define E_reg r3
1959 %define I_reg r4
1960 %define hev_thr_reg r5
1961 %else ; luma
1962 cglobal vp8_%2_loop_filter16y_mbedge_%1, 5, %3, %5
1963 %define mstride_reg r1
1964 %define E_reg r2
1965 %define I_reg r3
1966 %define hev_thr_reg r4
1967 %ifdef m8 ; x86-64, sse2
1968 %define dst8_reg r4
1969 %elif mmsize == 16 ; x86-32, sse2
1970 %define dst8_reg r5
1971 %else ; x86-32, mmx/mmxext
1972 %define cnt_reg r5
1973 %endif
1974 %endif
1975 %define dst_reg r0
1976 %define stride_reg E_reg
1977 %define dst2_reg I_reg
1978 %ifndef m8
1979 %define stack_reg hev_thr_reg
1980 %endif
1981
1982 %ifndef m8 ; mmx/mmxext or sse2 on x86-32
1983 ; splat function arguments
1984 SPLATB_REG m0, E_reg, %1 ; E
1985 SPLATB_REG m1, I_reg, %1 ; I
1986 SPLATB_REG m2, hev_thr_reg, %1 ; hev_thresh
1987
1988 ; align stack
1989 mov stack_reg, rsp ; backup stack pointer
1990 and rsp, ~(mmsize-1) ; align stack
1991 sub rsp, mmsize * 8 ; stack layout: [0]=E, [1]=I, [2]=hev_thr
1992 ; [3]=hev() result
1993 ; [4]=filter tmp result
1994 ; [5]/[6] = p2/q2 backup
1995 ; [7]=lim_res sign result
1996
1997 %define flim_E [rsp]
1998 %define flim_I [rsp+mmsize]
1999 %define hev_thr [rsp+mmsize*2]
2000 %define mask_res [rsp+mmsize*3]
2001 %define lim_res [rsp+mmsize*4]
2002 %define p0backup [rsp+mmsize*3]
2003 %define q0backup [rsp+mmsize*4]
2004 %define p2backup [rsp+mmsize*5]
2005 %define q2backup [rsp+mmsize*6]
2006 %define lim_sign [rsp+mmsize*7]
2007
2008 mova flim_E, m0
2009 mova flim_I, m1
2010 mova hev_thr, m2
2011
2012 %else ; sse2 on x86-64
2013
2014 %define flim_E m9
2015 %define flim_I m10
2016 %define hev_thr m11
2017 %define mask_res m12
2018 %define lim_res m8
2019 %define p0backup m12
2020 %define q0backup m8
2021 %define p2backup m13
2022 %define q2backup m14
2023 %define lim_sign m15
2024
2025 ; splat function arguments
2026 SPLATB_REG flim_E, E_reg, %1 ; E
2027 SPLATB_REG flim_I, I_reg, %1 ; I
2028 SPLATB_REG hev_thr, hev_thr_reg, %1 ; hev_thresh
2029 %endif
2030
2031 %if mmsize == 8 && %4 == 16 ; mmx/mmxext
2032 mov cnt_reg, 2
2033 %endif
2034 mov stride_reg, mstride_reg
2035 neg mstride_reg
2036 %ifidn %2, h
2037 lea dst_reg, [dst_reg + stride_reg*4-4]
2038 %if %4 == 8
2039 lea dst8_reg, [dst8_reg+ stride_reg*4-4]
2040 %endif
2041 %endif
2042
2043 %if mmsize == 8
2044 .next8px
2045 %endif
2046 ; read
2047 lea dst2_reg, [dst_reg + stride_reg]
2048 %ifidn %2, v
2049 %if %4 == 8 && mmsize == 16
2050 %define movrow movh
2051 %else
2052 %define movrow mova
2053 %endif
2054 movrow m0, [dst_reg +mstride_reg*4] ; p3
2055 movrow m1, [dst2_reg+mstride_reg*4] ; p2
2056 movrow m2, [dst_reg +mstride_reg*2] ; p1
2057 movrow m5, [dst2_reg] ; q1
2058 movrow m6, [dst2_reg+ stride_reg] ; q2
2059 movrow m7, [dst2_reg+ stride_reg*2] ; q3
2060 %if mmsize == 16 && %4 == 8
2061 movhps m0, [dst8_reg+mstride_reg*4]
2062 movhps m2, [dst8_reg+mstride_reg*2]
2063 add dst8_reg, stride_reg
2064 movhps m1, [dst8_reg+mstride_reg*4]
2065 movhps m5, [dst8_reg]
2066 movhps m6, [dst8_reg+ stride_reg]
2067 movhps m7, [dst8_reg+ stride_reg*2]
2068 add dst8_reg, mstride_reg
2069 %endif
2070 %elif mmsize == 8 ; mmx/mmxext (h)
2071 ; read 8 rows of 8px each
2072 movu m0, [dst_reg +mstride_reg*4]
2073 movu m1, [dst2_reg+mstride_reg*4]
2074 movu m2, [dst_reg +mstride_reg*2]
2075 movu m3, [dst_reg +mstride_reg]
2076 movu m4, [dst_reg]
2077 movu m5, [dst2_reg]
2078 movu m6, [dst2_reg+ stride_reg]
2079
2080 ; 8x8 transpose
2081 TRANSPOSE4x4B 0, 1, 2, 3, 7
2082 mova q0backup, m1
2083 movu m7, [dst2_reg+ stride_reg*2]
2084 TRANSPOSE4x4B 4, 5, 6, 7, 1
2085 SBUTTERFLY dq, 0, 4, 1 ; p3/p2
2086 SBUTTERFLY dq, 2, 6, 1 ; q0/q1
2087 SBUTTERFLY dq, 3, 7, 1 ; q2/q3
2088 mova m1, q0backup
2089 mova q0backup, m2 ; store q0
2090 SBUTTERFLY dq, 1, 5, 2 ; p1/p0
2091 mova p0backup, m5 ; store p0
2092 SWAP 1, 4
2093 SWAP 2, 4
2094 SWAP 6, 3
2095 SWAP 5, 3
2096 %else ; sse2 (h)
2097 %if %4 == 16
2098 lea dst8_reg, [dst_reg + stride_reg*8]
2099 %endif
2100
2101 ; read 16 rows of 8px each, interleave
2102 movh m0, [dst_reg +mstride_reg*4]
2103 movh m1, [dst8_reg+mstride_reg*4]
2104 movh m2, [dst_reg +mstride_reg*2]
2105 movh m5, [dst8_reg+mstride_reg*2]
2106 movh m3, [dst_reg +mstride_reg]
2107 movh m6, [dst8_reg+mstride_reg]
2108 movh m4, [dst_reg]
2109 movh m7, [dst8_reg]
2110 punpcklbw m0, m1 ; A/I
2111 punpcklbw m2, m5 ; C/K
2112 punpcklbw m3, m6 ; D/L
2113 punpcklbw m4, m7 ; E/M
2114
2115 add dst8_reg, stride_reg
2116 movh m1, [dst2_reg+mstride_reg*4]
2117 movh m6, [dst8_reg+mstride_reg*4]
2118 movh m5, [dst2_reg]
2119 movh m7, [dst8_reg]
2120 punpcklbw m1, m6 ; B/J
2121 punpcklbw m5, m7 ; F/N
2122 movh m6, [dst2_reg+ stride_reg]
2123 movh m7, [dst8_reg+ stride_reg]
2124 punpcklbw m6, m7 ; G/O
2125
2126 ; 8x16 transpose
2127 TRANSPOSE4x4B 0, 1, 2, 3, 7
2128 %ifdef m8
2129 SWAP 1, 8
2130 %else
2131 mova q0backup, m1
2132 %endif
2133 movh m7, [dst2_reg+ stride_reg*2]
2134 movh m1, [dst8_reg+ stride_reg*2]
2135 punpcklbw m7, m1 ; H/P
2136 TRANSPOSE4x4B 4, 5, 6, 7, 1
2137 SBUTTERFLY dq, 0, 4, 1 ; p3/p2
2138 SBUTTERFLY dq, 2, 6, 1 ; q0/q1
2139 SBUTTERFLY dq, 3, 7, 1 ; q2/q3
2140 %ifdef m8
2141 SWAP 1, 8
2142 SWAP 2, 8
2143 %else
2144 mova m1, q0backup
2145 mova q0backup, m2 ; store q0
2146 %endif
2147 SBUTTERFLY dq, 1, 5, 2 ; p1/p0
2148 %ifdef m12
2149 SWAP 5, 12
2150 %else
2151 mova p0backup, m5 ; store p0
2152 %endif
2153 SWAP 1, 4
2154 SWAP 2, 4
2155 SWAP 6, 3
2156 SWAP 5, 3
2157 %endif
2158
2159 ; normal_limit for p3-p2, p2-p1, q3-q2 and q2-q1
2160 mova m4, m1
2161 SWAP 4, 1
2162 psubusb m4, m0 ; p2-p3
2163 psubusb m0, m1 ; p3-p2
2164 por m0, m4 ; abs(p3-p2)
2165
2166 mova m4, m2
2167 SWAP 4, 2
2168 psubusb m4, m1 ; p1-p2
2169 mova p2backup, m1
2170 psubusb m1, m2 ; p2-p1
2171 por m1, m4 ; abs(p2-p1)
2172
2173 mova m4, m6
2174 SWAP 4, 6
2175 psubusb m4, m7 ; q2-q3
2176 psubusb m7, m6 ; q3-q2
2177 por m7, m4 ; abs(q3-q2)
2178
2179 mova m4, m5
2180 SWAP 4, 5
2181 psubusb m4, m6 ; q1-q2
2182 mova q2backup, m6
2183 psubusb m6, m5 ; q2-q1
2184 por m6, m4 ; abs(q2-q1)
2185
2186 %ifidn %1, mmx
2187 mova m4, flim_I
2188 pxor m3, m3
2189 psubusb m0, m4
2190 psubusb m1, m4
2191 psubusb m7, m4
2192 psubusb m6, m4
2193 pcmpeqb m0, m3 ; abs(p3-p2) <= I
2194 pcmpeqb m1, m3 ; abs(p2-p1) <= I
2195 pcmpeqb m7, m3 ; abs(q3-q2) <= I
2196 pcmpeqb m6, m3 ; abs(q2-q1) <= I
2197 pand m0, m1
2198 pand m7, m6
2199 pand m0, m7
2200 %else ; mmxext/sse2
2201 pmaxub m0, m1
2202 pmaxub m6, m7
2203 pmaxub m0, m6
2204 %endif
2205
2206 ; normal_limit and high_edge_variance for p1-p0, q1-q0
2207 SWAP 7, 3 ; now m7 is zero
2208 %ifidn %2, v
2209 movrow m3, [dst_reg +mstride_reg] ; p0
2210 %if mmsize == 16 && %4 == 8
2211 movhps m3, [dst8_reg+mstride_reg]
2212 %endif
2213 %elifdef m12
2214 SWAP 3, 12
2215 %else
2216 mova m3, p0backup
2217 %endif
2218
2219 mova m1, m2
2220 SWAP 1, 2
2221 mova m6, m3
2222 SWAP 3, 6
2223 psubusb m1, m3 ; p1-p0
2224 psubusb m6, m2 ; p0-p1
2225 por m1, m6 ; abs(p1-p0)
2226 %ifidn %1, mmx
2227 mova m6, m1
2228 psubusb m1, m4
2229 psubusb m6, hev_thr
2230 pcmpeqb m1, m7 ; abs(p1-p0) <= I
2231 pcmpeqb m6, m7 ; abs(p1-p0) <= hev_thresh
2232 pand m0, m1
2233 mova mask_res, m6
2234 %else ; mmxext/sse2
2235 pmaxub m0, m1 ; max_I
2236 SWAP 1, 4 ; max_hev_thresh
2237 %endif
2238
2239 SWAP 6, 4 ; now m6 is I
2240 %ifidn %2, v
2241 movrow m4, [dst_reg] ; q0
2242 %if mmsize == 16 && %4 == 8
2243 movhps m4, [dst8_reg]
2244 %endif
2245 %elifdef m8
2246 SWAP 4, 8
2247 %else
2248 mova m4, q0backup
2249 %endif
2250 mova m1, m4
2251 SWAP 1, 4
2252 mova m7, m5
2253 SWAP 7, 5
2254 psubusb m1, m5 ; q0-q1
2255 psubusb m7, m4 ; q1-q0
2256 por m1, m7 ; abs(q1-q0)
2257 %ifidn %1, mmx
2258 mova m7, m1
2259 psubusb m1, m6
2260 psubusb m7, hev_thr
2261 pxor m6, m6
2262 pcmpeqb m1, m6 ; abs(q1-q0) <= I
2263 pcmpeqb m7, m6 ; abs(q1-q0) <= hev_thresh
2264 mova m6, mask_res
2265 pand m0, m1 ; abs([pq][321]-[pq][210]) <= I
2266 pand m6, m7
2267 %else ; mmxext/sse2
2268 pxor m7, m7
2269 pmaxub m0, m1
2270 pmaxub m6, m1
2271 psubusb m0, flim_I
2272 psubusb m6, hev_thr
2273 pcmpeqb m0, m7 ; max(abs(..)) <= I
2274 pcmpeqb m6, m7 ; !(max(abs..) > thresh)
2275 %endif
2276 %ifdef m12
2277 SWAP 6, 12
2278 %else
2279 mova mask_res, m6 ; !(abs(p1-p0) > hev_t || abs(q1-q0) > hev_t)
2280 %endif
2281
2282 ; simple_limit
2283 mova m1, m3
2284 SWAP 1, 3
2285 mova m6, m4 ; keep copies of p0/q0 around for later use
2286 SWAP 6, 4
2287 psubusb m1, m4 ; p0-q0
2288 psubusb m6, m3 ; q0-p0
2289 por m1, m6 ; abs(q0-p0)
2290 paddusb m1, m1 ; m1=2*abs(q0-p0)
2291
2292 mova m7, m2
2293 SWAP 7, 2
2294 mova m6, m5
2295 SWAP 6, 5
2296 psubusb m7, m5 ; p1-q1
2297 psubusb m6, m2 ; q1-p1
2298 por m7, m6 ; abs(q1-p1)
2299 pxor m6, m6
2300 pand m7, [pb_FE]
2301 psrlq m7, 1 ; abs(q1-p1)/2
2302 paddusb m7, m1 ; abs(q0-p0)*2+abs(q1-p1)/2
2303 psubusb m7, flim_E
2304 pcmpeqb m7, m6 ; abs(q0-p0)*2+abs(q1-p1)/2 <= E
2305 pand m0, m7 ; normal_limit result
2306
2307 ; filter_common; at this point, m2-m5=p1-q1 and m0 is filter_mask
2308 %ifdef m8 ; x86-64 && sse2
2309 mova m8, [pb_80]
2310 %define pb_80_var m8
2311 %else ; x86-32 or mmx/mmxext
2312 %define pb_80_var [pb_80]
2313 %endif
2314 mova m1, m4
2315 mova m7, m3
2316 pxor m1, pb_80_var
2317 pxor m7, pb_80_var
2318 psubsb m1, m7 ; (signed) q0-p0
2319 mova m6, m2
2320 mova m7, m5
2321 pxor m6, pb_80_var
2322 pxor m7, pb_80_var
2323 psubsb m6, m7 ; (signed) p1-q1
2324 mova m7, mask_res
2325 paddsb m6, m1
2326 paddsb m6, m1
2327 paddsb m6, m1
2328 pand m6, m0
2329 %ifdef m8
2330 mova lim_res, m6 ; 3*(qp-p0)+(p1-q1) masked for filter_mbedge
2331 pand lim_res, m7
2332 %else
2333 mova m0, m6
2334 pand m0, m7
2335 mova lim_res, m0
2336 %endif
2337 pandn m7, m6 ; 3*(q0-p0)+(p1-q1) masked for filter_common
2338
2339 mova m1, [pb_F8]
2340 mova m6, m7
2341 paddsb m7, [pb_3]
2342 paddsb m6, [pb_4]
2343 pand m7, m1
2344 pand m6, m1
2345
2346 pxor m1, m1
2347 pxor m0, m0
2348 pcmpgtb m1, m7
2349 psubb m0, m7
2350 psrlq m7, 3 ; +f2
2351 psrlq m0, 3 ; -f2
2352 pand m0, m1
2353 pandn m1, m7
2354 psubusb m3, m0
2355 paddusb m3, m1 ; p0+f2
2356
2357 pxor m1, m1
2358 pxor m0, m0
2359 pcmpgtb m0, m6
2360 psubb m1, m6
2361 psrlq m6, 3 ; +f1
2362 psrlq m1, 3 ; -f1
2363 pand m1, m0
2364 pandn m0, m6
2365 psubusb m4, m0
2366 paddusb m4, m1 ; q0-f1
2367
2368 ; filter_mbedge (m2-m5 = p1-q1; lim_res carries w)
2369 mova m7, [pw_63]
2370 %ifdef m8
2371 SWAP 1, 8
2372 %else
2373 mova m1, lim_res
2374 %endif
2375 pxor m0, m0
2376 mova m6, m1
2377 pcmpgtb m0, m1 ; which are negative
2378 punpcklbw m6, m0 ; signed byte->word
2379 punpckhbw m1, m0
2380 mova lim_sign, m0
2381 mova mask_res, m6 ; backup for later in filter
2382 mova lim_res, m1
2383 pmullw m6, [pw_27]
2384 pmullw m1, [pw_27]
2385 paddw m6, m7
2386 paddw m1, m7
2387 psraw m6, 7
2388 psraw m1, 7
2389 packsswb m6, m1 ; a0
2390 pxor m1, m1
2391 psubb m1, m6
2392 pand m1, m0 ; -a0
2393 pandn m0, m6 ; +a0
2394 psubusb m3, m1
2395 paddusb m4, m1
2396 paddusb m3, m0 ; p0+a0
2397 psubusb m4, m0 ; q0-a0
2398
2399 mova m6, mask_res
2400 mova m1, lim_res
2401 mova m0, lim_sign
2402 pmullw m6, [pw_18]
2403 pmullw m1, [pw_18]
2404 paddw m6, m7
2405 paddw m1, m7
2406 psraw m6, 7
2407 psraw m1, 7
2408 packsswb m6, m1 ; a1
2409 pxor m1, m1
2410 psubb m1, m6
2411 pand m1, m0 ; -a1
2412 pandn m0, m6 ; +a1
2413 psubusb m2, m1
2414 paddusb m5, m1
2415 paddusb m2, m0 ; p1+a1
2416 psubusb m5, m0 ; q1-a1
2417
2418 %ifdef m8
2419 SWAP 6, 12
2420 SWAP 1, 8
2421 %else
2422 mova m6, mask_res
2423 mova m1, lim_res
2424 %endif
2425 pmullw m6, [pw_9]
2426 pmullw m1, [pw_9]
2427 paddw m6, m7
2428 paddw m1, m7
2429 %ifdef m15
2430 SWAP 7, 15
2431 %else
2432 mova m7, lim_sign
2433 %endif
2434 psraw m6, 7
2435 psraw m1, 7
2436 packsswb m6, m1 ; a1
2437 pxor m0, m0
2438 psubb m0, m6
2439 pand m0, m7 ; -a1
2440 pandn m7, m6 ; +a1
2441 %ifdef m8
2442 SWAP 1, 13
2443 SWAP 6, 14
2444 %else
2445 mova m1, p2backup
2446 mova m6, q2backup
2447 %endif
2448 psubusb m1, m0
2449 paddusb m6, m0
2450 paddusb m1, m7 ; p1+a1
2451 psubusb m6, m7 ; q1-a1
2452
2453 ; store
2454 %ifidn %2, v
2455 movrow [dst2_reg+mstride_reg*4], m1
2456 movrow [dst_reg +mstride_reg*2], m2
2457 movrow [dst_reg +mstride_reg ], m3
2458 movrow [dst_reg], m4
2459 movrow [dst2_reg], m5
2460 movrow [dst2_reg+ stride_reg ], m6
2461 %if mmsize == 16 && %4 == 8
2462 add dst8_reg, mstride_reg
2463 movhps [dst8_reg+mstride_reg*2], m1
2464 movhps [dst8_reg+mstride_reg ], m2
2465 movhps [dst8_reg], m3
2466 add dst8_reg, stride_reg
2467 movhps [dst8_reg], m4
2468 movhps [dst8_reg+ stride_reg ], m5
2469 movhps [dst8_reg+ stride_reg*2], m6
2470 %endif
2471 %else ; h
2472 inc dst_reg
2473 inc dst2_reg
2474
2475 ; 4x8/16 transpose
2476 TRANSPOSE4x4B 1, 2, 3, 4, 0
2477 SBUTTERFLY bw, 5, 6, 0
2478
2479 %if mmsize == 8 ; mmx/mmxext (h)
2480 WRITE_4x2D 1, 2, 3, 4, dst_reg, dst2_reg, mstride_reg, stride_reg
2481 add dst_reg, 4
2482 WRITE_8W m5, m6, dst2_reg, dst_reg, mstride_reg, stride_reg
2483 %else ; sse2 (h)
2484 lea dst8_reg, [dst8_reg+mstride_reg+1]
2485 WRITE_4x4D 1, 2, 3, 4, dst_reg, dst2_reg, dst8_reg, mstride_reg, stride_reg, %4
2486 add dst_reg, 4
2487 add dst8_reg, 4
2488 WRITE_8W m5, m5, dst2_reg, dst_reg, mstride_reg, stride_reg
2489 WRITE_8W m6, m6, dst2_reg, dst8_reg, mstride_reg, stride_reg
2490 %endif
2491 %endif
2492
2493 %if mmsize == 8
2494 %if %4 == 8 ; chroma
2495 %ifidn %2, h
2496 sub dst_reg, 5
2497 %endif
2498 cmp dst_reg, dst8_reg
2499 mov dst_reg, dst8_reg
2500 jnz .next8px
2501 %else
2502 %ifidn %2, h
2503 lea dst_reg, [dst_reg + stride_reg*8-5]
2504 %else ; v
2505 add dst_reg, 8
2506 %endif
2507 dec cnt_reg
2508 jg .next8px
2509 %endif
2510 %endif
2511
2512 %ifndef m8 ; sse2 on x86-32 or mmx/mmxext
2513 mov rsp, stack_reg ; restore stack pointer
2514 %endif
2515 RET
2516 %endmacro
2517
2518 INIT_MMX
2519 MBEDGE_LOOPFILTER mmx, v, 6, 16, 8
2520 MBEDGE_LOOPFILTER mmx, h, 6, 16, 8
2521 MBEDGE_LOOPFILTER mmxext, v, 6, 16, 8
2522 MBEDGE_LOOPFILTER mmxext, h, 6, 16, 8
2523
2524 MBEDGE_LOOPFILTER mmx, v, 6, 8, 8
2525 MBEDGE_LOOPFILTER mmx, h, 6, 8, 8
2526 MBEDGE_LOOPFILTER mmxext, v, 6, 8, 8
2527 MBEDGE_LOOPFILTER mmxext, h, 6, 8, 8
2528
2529 INIT_XMM
2530 MBEDGE_LOOPFILTER sse2, v, 5, 16, 16
2531 %ifdef m8
2532 MBEDGE_LOOPFILTER sse2, h, 5, 16, 16
2533 %else
2534 MBEDGE_LOOPFILTER sse2, h, 6, 16, 16
2535 %endif
2536 MBEDGE_LOOPFILTER sse2, v, 6, 8, 16
2537 MBEDGE_LOOPFILTER sse2, h, 6, 8, 16