Commit 222e6da6 authored by Timothy Gu's avatar Timothy Gu

x86/vf_blend: Add SSE2 optimization for divide

 4.5x faster than C float version with autovectorization
10  x faster than C int version
25  x faster than C float version without autovectorization
parent 1c9215e5
......@@ -24,6 +24,7 @@
SECTION_RODATA
ps_255: times 4 dd 255.0
pw_1: times 8 dw 1
pw_128: times 8 dw 128
pw_255: times 8 dw 255
......@@ -218,6 +219,35 @@ BLEND_INIT hardmix, 5
jl .loop
BLEND_END
BLEND_INIT divide, 4
pxor m2, m2
mova m3, [ps_255]
.nextrow:
mov xq, widthq
.loop:
movd m0, [topq + xq] ; 000000xx
movd m1, [bottomq + xq]
punpcklbw m0, m2 ; 00000x0x
punpcklbw m1, m2
punpcklwd m0, m2 ; 000x000x
punpcklwd m1, m2
cvtdq2ps m0, m0
cvtdq2ps m1, m1
divps m0, m1 ; a / b
mulps m0, m3 ; a / b * 255
minps m0, m3
cvttps2dq m0, m0
packssdw m0, m0 ; 00000x0x
packuswb m0, m0 ; 000000xx
movd [dstq + xq], m0
add xq, mmsize / 4
jl .loop
BLEND_END
BLEND_INIT phoenix, 4
mova m3, [pb_255]
.nextrow:
......
......@@ -39,6 +39,7 @@ BLEND_FUNC(difference128, sse2)
BLEND_FUNC(multiply, sse2)
BLEND_FUNC(screen, sse2)
BLEND_FUNC(hardmix, sse2)
BLEND_FUNC(divide, sse2)
BLEND_FUNC(lighten, sse2)
BLEND_FUNC(or, sse2)
BLEND_FUNC(phoenix, sse2)
......@@ -61,6 +62,7 @@ av_cold void ff_blend_init_x86(FilterParams *param, int is_16bit)
case BLEND_AVERAGE: param->blend = ff_blend_average_sse2; break;
case BLEND_DARKEN: param->blend = ff_blend_darken_sse2; break;
case BLEND_DIFFERENCE128: param->blend = ff_blend_difference128_sse2; break;
case BLEND_DIVIDE: param->blend = ff_blend_divide_sse2; break;
case BLEND_HARDMIX: param->blend = ff_blend_hardmix_sse2; break;
case BLEND_LIGHTEN: param->blend = ff_blend_lighten_sse2; break;
case BLEND_MULTIPLY: param->blend = ff_blend_multiply_sse2; break;
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment