Skip to content

GitLab

D
dav1d
Commit aec3d25c authored by Ronald S. Bultje's avatar Ronald S. Bultje
Browse files
Options

Add horizontal loopfilter AVX2 SIMD

parent 24effd47
Hide whitespace changes
Inline Side-by-side
Showing with 846 additions and 70 deletions
src/lf_apply.c
View file @ aec3d25c
......@@ -49,7 +49,7 @@ static inline void filter_plane_cols_y(const Dav1dFrameContext *const f,
for (int x = 0; x < w; x++) {
if (!have_left && !x) continue;
dsp->lf.loop_filter_sb[0][0](&dst[x * 4], ls,
starty4 ? (const uint32_t[3]) {
starty4 ? (const uint32_t[4]) {
mask[x][0] >> starty4,
mask[x][1] >> starty4,
mask[x][2] >> starty4,
......@@ -98,14 +98,14 @@ static inline void filter_plane_cols_uv(const Dav1dFrameContext *const f,
for (int x = 0; x < w; x++) {
if (!have_left && !x) continue;
dsp->lf.loop_filter_sb[1][0](&u[x * 4], ls,
starty4 ? (const uint32_t[2]) {
starty4 ? (const uint32_t[3]) {
mask[x][0] >> starty4,
mask[x][1] >> starty4,
} : mask[x],
(const uint8_t(*)[4]) &lvl[x][2], b4_stride,
&f->lf.lim_lut, endy4 - starty4);
dsp->lf.loop_filter_sb[1][0](&v[x * 4], ls,
starty4 ? (const uint32_t[2]) {
starty4 ? (const uint32_t[3]) {
mask[x][0] >> starty4,
mask[x][1] >> starty4,
} : mask[x],
......
src/x86/loopfilter.asm
View file @ aec3d25c
......@@ -63,22 +63,496 @@ SECTION .text
por %1, %4
%endmacro
%macro FILTER 1 ; width
%macro TRANSPOSE_16x4_AND_WRITE_4x32 5
; transpose 16x4
punpcklbw m%5, m%1, m%2
punpckhbw m%1, m%2
punpcklbw m%2, m%3, m%4
punpckhbw m%3, m%4
punpcklwd m%4, m%5, m%2
punpckhwd m%5, m%2
punpcklwd m%2, m%1, m%3
punpckhwd m%1, m%3
; write out
movd [dstq+strideq*0-2], xm%4
pextrd [dstq+strideq*1-2], xm%4, 1
pextrd [dstq+strideq*2-2], xm%4, 2
pextrd [dstq+stride3q-2], xm%4, 3
lea dstq, [dstq+strideq*4]
movd [dstq+strideq*0-2], xm%5
pextrd [dstq+strideq*1-2], xm%5, 1
pextrd [dstq+strideq*2-2], xm%5, 2
pextrd [dstq+stride3q-2], xm%5, 3
lea dstq, [dstq+strideq*4]
movd [dstq+strideq*0-2], xm%2
pextrd [dstq+strideq*1-2], xm%2, 1
pextrd [dstq+strideq*2-2], xm%2, 2
pextrd [dstq+stride3q-2], xm%2, 3
lea dstq, [dstq+strideq*4]
movd [dstq+strideq*0-2], xm%1
pextrd [dstq+strideq*1-2], xm%1, 1
pextrd [dstq+strideq*2-2], xm%1, 2
pextrd [dstq+stride3q-2], xm%1, 3
lea dstq, [dstq+strideq*4]
vextracti128 xm%4, m%4, 1
vextracti128 xm%5, m%5, 1
vextracti128 xm%2, m%2, 1
vextracti128 xm%1, m%1, 1
movd [dstq+strideq*0-2], xm%4
pextrd [dstq+strideq*1-2], xm%4, 1
pextrd [dstq+strideq*2-2], xm%4, 2
pextrd [dstq+stride3q-2], xm%4, 3
lea dstq, [dstq+strideq*4]
movd [dstq+strideq*0-2], xm%5
pextrd [dstq+strideq*1-2], xm%5, 1
pextrd [dstq+strideq*2-2], xm%5, 2
pextrd [dstq+stride3q-2], xm%5, 3
lea dstq, [dstq+strideq*4]
movd [dstq+strideq*0-2], xm%2
pextrd [dstq+strideq*1-2], xm%2, 1
pextrd [dstq+strideq*2-2], xm%2, 2
pextrd [dstq+stride3q-2], xm%2, 3
lea dstq, [dstq+strideq*4]
movd [dstq+strideq*0-2], xm%1
pextrd [dstq+strideq*1-2], xm%1, 1
pextrd [dstq+strideq*2-2], xm%1, 2
pextrd [dstq+stride3q-2], xm%1, 3
lea dstq, [dstq+strideq*4]
%endmacro
%macro TRANSPOSE_16X16B 3 ; in_load_15_from_mem, out_store_0_in_mem, mem
%if %1 == 0
mova %3, m15
%endif
; input in m0-15
punpcklbw m15, m0, m1
punpckhbw m0, m1
punpcklbw m1, m2, m3
punpckhbw m2, m3
punpcklbw m3, m4, m5
punpckhbw m4, m5
punpcklbw m5, m6, m7
punpckhbw m6, m7
punpcklbw m7, m8, m9
punpckhbw m8, m9
punpcklbw m9, m10, m11
punpckhbw m10, m11
punpcklbw m11, m12, m13
punpckhbw m12, m13
mova m13, %3
mova %3, m12
punpcklbw m12, m14, m13
punpckhbw m13, m14, m13
; interleaved in m15,0,1,2,3,4,5,6,7,8,9,10,11,rsp%3,12,13
punpcklwd m14, m15, m1
punpckhwd m15, m1
punpcklwd m1, m0, m2
punpckhwd m0, m2
punpcklwd m2, m3, m5
punpckhwd m3, m5
punpcklwd m5, m4, m6
punpckhwd m4, m6
punpcklwd m6, m7, m9
punpckhwd m7, m9
punpcklwd m9, m8, m10
punpckhwd m8, m10
punpcklwd m10, m11, m12
punpckhwd m11, m12
mova m12, %3
mova %3, m11
punpcklwd m11, m12, m13
punpckhwd m12, m13
; interleaved in m14,15,1,0,2,3,5,4,6,7,9,8,10,rsp%3,11,12
punpckldq m13, m14, m2
punpckhdq m14, m2
punpckldq m2, m15, m3
punpckhdq m15, m3
punpckldq m3, m1, m5
punpckhdq m1, m5
punpckldq m5, m0, m4
punpckhdq m0, m4
punpckldq m4, m6, m10
punpckhdq m6, m10
punpckldq m10, m9, m11
punpckhdq m9, m11
punpckldq m11, m8, m12
punpckhdq m8, m12
mova m12, %3
mova %3, m8
punpckldq m8, m7, m12
punpckhdq m7, m12
; interleaved in m13,14,2,15,3,1,5,0,4,6,8,7,10,9,11,rsp%3
punpcklqdq m12, m13, m4
punpckhqdq m13, m4
punpcklqdq m4, m14, m6
punpckhqdq m14, m6
punpcklqdq m6, m2, m8
punpckhqdq m2, m8
punpcklqdq m8, m15, m7
punpckhqdq m15, m7
punpcklqdq m7, m3, m10
punpckhqdq m3, m10
punpcklqdq m10, m1, m9
punpckhqdq m1, m9
punpcklqdq m9, m5, m11
punpckhqdq m5, m11
mova m11, %3
mova %3, m12
punpcklqdq m12, m0, m11
punpckhqdq m0, m11
%if %2 == 0
mova m11, %3
%endif
; interleaved m11,13,4,14,6,2,8,15,7,3,10,1,9,5,12,0
SWAP 0, 11, 1, 13, 5, 2, 4, 6, 8, 7, 15
SWAP 3, 14, 12, 9
%endmacro
%macro FILTER 2 ; width [4/6/8/16], dir [h/v]
; load data
%ifidn %2, v
%if %1 == 4
lea tmpq, [dstq+mstrideq*2]
mova m3, [tmpq+strideq*0] ; p1
mova m4, [tmpq+strideq*1] ; p0
mova m5, [tmpq+strideq*2] ; q0
mova m6, [tmpq+stride3q] ; q1
%else
; load 6-8 pixels, remainder (for wd=16) will be read inline
lea tmpq, [dstq+mstrideq*4]
%if %1 != 6
mova m12, [tmpq+strideq*0]
%endif
mova m13, [tmpq+strideq*1]
mova m3, [tmpq+strideq*2]
mova m4, [tmpq+stride3q]
mova m5, [dstq+strideq*0]
mova m6, [dstq+strideq*1]
mova m14, [dstq+strideq*2]
%if %1 != 6
mova m15, [dstq+stride3q]
%endif
%endif
%else
; load lines
%if %1 == 4
movd xm3, [dstq+strideq*0-2]
movd xm4, [dstq+strideq*1-2]
movd xm5, [dstq+strideq*2-2]
movd xm6, [dstq+stride3q -2]
lea tmpq, [dstq+strideq*4]
pinsrd xm3, [tmpq+strideq*0-2], 2
pinsrd xm4, [tmpq+strideq*1-2], 2
pinsrd xm5, [tmpq+strideq*2-2], 2
pinsrd xm6, [tmpq+stride3q -2], 2
lea tmpq, [tmpq+strideq*4]
pinsrd xm3, [tmpq+strideq*0-2], 1
pinsrd xm4, [tmpq+strideq*1-2], 1
pinsrd xm5, [tmpq+strideq*2-2], 1
pinsrd xm6, [tmpq+stride3q -2], 1
lea tmpq, [tmpq+strideq*4]
pinsrd xm3, [tmpq+strideq*0-2], 3
pinsrd xm4, [tmpq+strideq*1-2], 3
pinsrd xm5, [tmpq+strideq*2-2], 3
pinsrd xm6, [tmpq+stride3q -2], 3
lea tmpq, [tmpq+strideq*4]
movd xm12, [tmpq+strideq*0-2]
movd xm13, [tmpq+strideq*1-2]
movd xm14, [tmpq+strideq*2-2]
movd xm15, [tmpq+stride3q -2]
lea tmpq, [tmpq+strideq*4]
pinsrd xm12, [tmpq+strideq*0-2], 2
pinsrd xm13, [tmpq+strideq*1-2], 2
pinsrd xm14, [tmpq+strideq*2-2], 2
pinsrd xm15, [tmpq+stride3q -2], 2
lea tmpq, [tmpq+strideq*4]
pinsrd xm12, [tmpq+strideq*0-2], 1
pinsrd xm13, [tmpq+strideq*1-2], 1
pinsrd xm14, [tmpq+strideq*2-2], 1
pinsrd xm15, [tmpq+stride3q -2], 1
lea tmpq, [tmpq+strideq*4]
pinsrd xm12, [tmpq+strideq*0-2], 3
pinsrd xm13, [tmpq+strideq*1-2], 3
pinsrd xm14, [tmpq+strideq*2-2], 3
pinsrd xm15, [tmpq+stride3q -2], 3
vinserti128 m3, xm12, 1
vinserti128 m4, xm13, 1
vinserti128 m5, xm14, 1
vinserti128 m6, xm15, 1
; transpose 4x16
; xm3: A-D0,A-D8,A-D4,A-D12
; xm4: A-D1,A-D9,A-D5,A-D13
; xm5: A-D2,A-D10,A-D6,A-D14
; xm6: A-D3,A-D11,A-D7,A-D15
punpcklbw m7, m3, m4
punpckhbw m3, m4
punpcklbw m4, m5, m6
punpckhbw m5, m6
; xm7: A0-1,B0-1,C0-1,D0-1,A8-9,B8-9,C8-9,D8-9
; xm3: A4-5,B4-5,C4-5,D4-5,A12-13,B12-13,C12-13,D12-13
; xm4: A2-3,B2-3,C2-3,D2-3,A10-11,B10-11,C10-11,D10-11
; xm5: A6-7,B6-7,C6-7,D6-7,A14-15,B14-15,C14-15,D14-15
punpcklwd m6, m7, m4
punpckhwd m7, m4
punpcklwd m4, m3, m5
punpckhwd m3, m5
; xm6: A0-3,B0-3,C0-3,D0-3
; xm7: A8-11,B8-11,C8-11,D8-11
; xm4: A4-7,B4-7,C4-7,D4-7
; xm3: A12-15,B12-15,C12-15,D12-15
punpckldq m5, m6, m4
punpckhdq m6, m4
punpckldq m4, m7, m3
punpckhdq m7, m3
; xm5: A0-7,B0-7
; xm6: C0-7,D0-7
; xm4: A8-15,B8-15
; xm7: C8-15,D8-15
punpcklqdq m3, m5, m4
punpckhqdq m4, m5, m4
punpcklqdq m5, m6, m7
punpckhqdq m6, m7
; xm3: A0-15
; xm5: B0-15
; xm4: C0-15
; xm6: D0-15
%elif %1 == 6 || %1 == 8
movq xm3, [dstq+strideq*0-%1/2]
movq xm4, [dstq+strideq*1-%1/2]
movq xm5, [dstq+strideq*2-%1/2]
movq xm6, [dstq+stride3q -%1/2]
lea tmpq, [dstq+strideq*8]
movhps xm3, [tmpq+strideq*0-%1/2]
movhps xm4, [tmpq+strideq*1-%1/2]
movhps xm5, [tmpq+strideq*2-%1/2]
movhps xm6, [tmpq+stride3q -%1/2]
lea tmpq, [tmpq+strideq*8]
movq xm7, [tmpq+strideq*0-%1/2]
movq xm8, [tmpq+strideq*1-%1/2]
movq xm9, [tmpq+strideq*2-%1/2]
movq xm11, [tmpq+stride3q -%1/2]
lea tmpq, [tmpq+strideq*8]
movhps xm7, [tmpq+strideq*0-%1/2]
movhps xm8, [tmpq+strideq*1-%1/2]
movhps xm9, [tmpq+strideq*2-%1/2]
movhps xm11, [tmpq+stride3q -%1/2]
vinserti128 m3, xm7, 1
vinserti128 m4, xm8, 1
vinserti128 m5, xm9, 1
vinserti128 m6, xm11, 1
lea tmpq, [dstq+strideq*4]
movq xm12, [tmpq+strideq*0-%1/2]
movq xm13, [tmpq+strideq*1-%1/2]
movq xm14, [tmpq+strideq*2-%1/2]
movq xm15, [tmpq+stride3q -%1/2]
lea tmpq, [tmpq+strideq*8]
movhps xm12, [tmpq+strideq*0-%1/2]
movhps xm13, [tmpq+strideq*1-%1/2]
movhps xm14, [tmpq+strideq*2-%1/2]
movhps xm15, [tmpq+stride3q -%1/2]
lea tmpq, [tmpq+strideq*8]
movq xm7, [tmpq+strideq*0-%1/2]
movq xm8, [tmpq+strideq*1-%1/2]
movq xm9, [tmpq+strideq*2-%1/2]
movq xm11, [tmpq+stride3q -%1/2]
lea tmpq, [tmpq+strideq*8]
movhps xm7, [tmpq+strideq*0-%1/2]
movhps xm8, [tmpq+strideq*1-%1/2]
movhps xm9, [tmpq+strideq*2-%1/2]
movhps xm11, [tmpq+stride3q -%1/2]
vinserti128 m12, xm7, 1
vinserti128 m13, xm8, 1
vinserti128 m14, xm9, 1
vinserti128 m15, xm11, 1
; transpose 8x16
; xm3: A-H0,A-H8
; xm4: A-H1,A-H9
; xm5: A-H2,A-H10
; xm6: A-H3,A-H11
; xm12: A-H4,A-H12
; xm13: A-H5,A-H13
; xm14: A-H6,A-H14
; xm15: A-H7,A-H15
punpcklbw m7, m3, m4
punpckhbw m3, m4
punpcklbw m4, m5, m6
punpckhbw m5, m6
punpcklbw m6, m12, m13
punpckhbw m12, m13
punpcklbw m13, m14, m15
punpckhbw m14, m15
; xm7: A0-1,B0-1,C0-1,D0-1,E0-1,F0-1,G0-1,H0-1
; xm3: A8-9,B8-9,C8-9,D8-9,E8-9,F8-9,G8-9,H8-9
; xm4: A2-3,B2-3,C2-3,D2-3,E2-3,F2-3,G2-3,H2-3
; xm5: A10-11,B10-11,C10-11,D10-11,E10-11,F10-11,G10-11,H10-11
; xm6: A4-5,B4-5,C4-5,D4-5,E4-5,F4-5,G4-5,H4-5
; xm12: A12-13,B12-13,C12-13,D12-13,E12-13,F12-13,G12-13,H12-13
; xm13: A6-7,B6-7,C6-7,D6-7,E6-7,F6-7,G6-7,H6-7
; xm14: A14-15,B14-15,C14-15,D14-15,E14-15,F14-15,G14-15,H14-15
punpcklwd m15, m7, m4
punpckhwd m7, m4
punpcklwd m4, m3, m5
punpckhwd m3, m5
punpcklwd m5, m6, m13
punpckhwd m6, m13
punpcklwd m13, m12, m14
punpckhwd m12, m14
; xm15: A0-3,B0-3,C0-3,D0-3
; xm7: E0-3,F0-3,G0-3,H0-3
; xm4: A8-11,B8-11,C8-11,D8-11
; xm3: E8-11,F8-11,G8-11,H8-11
; xm5: A4-7,B4-7,C4-7,D4-7
; xm6: E4-7,F4-7,G4-7,H4-7
; xm13: A12-15,B12-15,C12-15,D12-15
; xm12: E12-15,F12-15,G12-15,H12-15
punpckldq m14, m15, m5
punpckhdq m15, m5
punpckldq m5, m7, m6
%if %1 != 6
punpckhdq m7, m6
%endif
punpckldq m6, m4, m13
punpckhdq m4, m13
punpckldq m13, m3, m12
%if %1 != 6
punpckhdq m12, m3, m12
%endif
; xm14: A0-7,B0-7
; xm15: C0-7,D0-7
; xm5: E0-7,F0-7
; xm7: G0-7,H0-7
; xm6: A8-15,B8-15
; xm4: C8-15,D8-15
; xm13: E8-15,F8-15
; xm12: G8-15,H8-15
punpcklqdq m3, m14, m6
punpckhqdq m14, m6
punpckhqdq m6, m15, m4
punpcklqdq m15, m4
punpcklqdq m4, m5, m13
punpckhqdq m13, m5, m13
%if %1 == 8
punpcklqdq m5, m7, m12
punpckhqdq m12, m7, m12
; xm3: A0-15
; xm14: B0-15
; xm15: C0-15
; xm6: D0-15
; xm4: E0-15
; xm13: F0-15
; xm5: G0-15
; xm12: H0-15
SWAP 12, 3, 15
SWAP 13, 14, 5, 4, 6
; 3,14,15,6,4,13,5,12 -> 12,13,3,4,5,6,14,15
%else
SWAP 13, 3, 14
SWAP 6, 4, 15, 5
; 3,14,15,6,4,13 -> 13,3,4,5,6,14
%endif
%else
; load and 16x16 transpose. We only use 14 pixels but we'll need the
; remainder at the end for the second transpose
movu xm0, [dstq+strideq*0-8]
movu xm1, [dstq+strideq*1-8]
movu xm2, [dstq+strideq*2-8]
movu xm3, [dstq+stride3q -8]
lea tmpq, [dstq+strideq*4]
movu xm4, [tmpq+strideq*0-8]
movu xm5, [tmpq+strideq*1-8]
movu xm6, [tmpq+strideq*2-8]
movu xm7, [tmpq+stride3q -8]
lea tmpq, [tmpq+strideq*4]
movu xm8, [tmpq+strideq*0-8]
movu xm9, [tmpq+strideq*1-8]
movu xm10, [tmpq+strideq*2-8]
movu xm11, [tmpq+stride3q -8]
lea tmpq, [tmpq+strideq*4]
movu xm12, [tmpq+strideq*0-8]
movu xm13, [tmpq+strideq*1-8]
movu xm14, [tmpq+strideq*2-8]
movu xm15, [tmpq+stride3q -8]
lea tmpq, [tmpq+strideq*4]
vinserti128 m0, [tmpq+strideq*0-8], 1
vinserti128 m1, [tmpq+strideq*1-8], 1
vinserti128 m2, [tmpq+strideq*2-8], 1
vinserti128 m3, [tmpq+stride3q -8], 1
lea tmpq, [tmpq+strideq*4]
vinserti128 m4, [tmpq+strideq*0-8], 1
vinserti128 m5, [tmpq+strideq*1-8], 1
vinserti128 m6, [tmpq+strideq*2-8], 1
vinserti128 m7, [tmpq+stride3q -8], 1
lea tmpq, [tmpq+strideq*4]
vinserti128 m8, [tmpq+strideq*0-8], 1
vinserti128 m9, [tmpq+strideq*1-8], 1
vinserti128 m10, [tmpq+strideq*2-8], 1
vinserti128 m11, [tmpq+stride3q -8], 1
lea tmpq, [tmpq+strideq*4]
vinserti128 m12, [tmpq+strideq*0-8], 1
vinserti128 m13, [tmpq+strideq*1-8], 1
vinserti128 m14, [tmpq+strideq*2-8], 1
vinserti128 m15, [tmpq+stride3q -8], 1
TRANSPOSE_16X16B 0, 1, [rsp+11*32]
mova [rsp+12*32], m1
mova [rsp+13*32], m2
mova [rsp+14*32], m3
mova [rsp+15*32], m12
mova [rsp+16*32], m13
mova [rsp+17*32], m14
mova [rsp+18*32], m15
; 4,5,6,7,8,9,10,11 -> 12,13,3,4,5,6,14,15
SWAP 12, 4, 7
SWAP 13, 5, 8
SWAP 3, 6, 9
SWAP 10, 14
SWAP 11, 15
%endif
%endif
; load L/E/I/H
%ifidn %2, v
movu m1, [lq]
movu m0, [lq+l_strideq]
%else
movq xm1, [lq]
movq xm2, [lq+l_strideq*2]
movhps xm1, [lq+l_strideq]
movhps xm2, [lq+l_stride3q]
lea lq, [lq+l_strideq*4]
movq xm10, [lq]
movq xm0, [lq+l_strideq*2]
movhps xm10, [lq+l_strideq]
movhps xm0, [lq+l_stride3q]
lea lq, [lq+l_strideq*4]
vinserti128 m1, xm10, 1
vinserti128 m2, xm0, 1
shufps m0, m1, m2, q3131
shufps m1, m2, q2020
%endif
pxor m2, m2
pcmpeqb m3, m2, m0
pand m1, m3
pcmpeqb m10, m2, m0
pand m1, m10
por m0, m1 ; l[x][] ? l[x][] : l[x-stride][]
pshufb m0, [pb_4x1_4x5_4x9_4x13] ; l[x][1]
pcmpeqb m10, m2, m0 ; !L
pand m1, m0, [pb_240]
psrlq m1, 4 ; H
psrlq m2, m0, [lutq+128]
pand m2, [pb_63]
vpbroadcastb m4, [lutq+136]
pminub m2, m4
vpbroadcastb m1, [lutq+136]
pminub m2, m1
pmaxub m2, [pb_1] ; I
pand m1, m0, [pb_240]
psrlq m1, 4 ; H
paddb m0, [pb_2]
paddb m0, m0
paddb m0, m2 ; E
......@@ -86,20 +560,6 @@ SECTION .text
pxor m2, [pb_128]
pxor m0, [pb_128]
%if %1 == 4
lea tmpq, [dstq+mstrideq*2]
mova m3, [tmpq+strideq*0] ; p1
mova m4, [tmpq+strideq*1] ; p0
mova m5, [tmpq+strideq*2] ; q0
mova m6, [tmpq+stride3q] ; q1
%else
lea tmpq, [dstq+mstrideq*4]
mova m3, [tmpq+strideq*2]
mova m4, [tmpq+stride3q]
mova m5, [dstq+strideq*0]
mova m6, [dstq+strideq*1]
%endif
ABSSUB m8, m3, m4, m9 ; abs(p1-p0)
pmaxub m8, m10
ABSSUB m9, m5, m6, m10 ; abs(q1-q0)
......@@ -111,15 +571,6 @@ SECTION .text
pxor m7, m8, [pb_128]
pcmpgtb m7, m1 ; hev
%if %1 != 6
mova m12, [tmpq+strideq*0]
%endif
mova m13, [tmpq+strideq*1]
mova m14, [dstq+strideq*2]
%if %1 != 6
mova m15, [dstq+stride3q]
%endif
%if %1 == 6
ABSSUB m9, m13, m4, m10 ; abs(p2-p0)
pmaxub m9, m8
......@@ -179,23 +630,47 @@ SECTION .text
por m8, m10
%if %1 == 16
%ifidn %2, v
lea tmpq, [dstq+mstrideq*8]
mova m0, [tmpq+strideq*1]
%else
mova m0, [rsp+12*32]
%endif
ABSSUB m1, m0, m4, m2
%ifidn %2, v
mova m0, [tmpq+strideq*2]
%else
mova m0, [rsp+13*32]
%endif
ABSSUB m2, m0, m4, m10
pmaxub m1, m2
%ifidn %2, v
mova m0, [tmpq+stride3q]
%else
mova m0, [rsp+14*32]
%endif
ABSSUB m2, m0, m4, m10
pmaxub m1, m2
%ifidn %2, v
lea tmpq, [dstq+strideq*4]
mova m0, [tmpq+strideq*0]
%else
mova m0, [rsp+15*32]
%endif
ABSSUB m2, m0, m5, m10
pmaxub m1, m2
%ifidn %2, v
mova m0, [tmpq+strideq*1]
%else
mova m0, [rsp+16*32]
%endif
ABSSUB m2, m0, m5, m10
pmaxub m1, m2
%ifidn %2, v
mova m0, [tmpq+strideq*2]
%else
mova m0, [rsp+17*32]
%endif
ABSSUB m2, m0, m5, m10
pmaxub m1, m2
pxor m1, [pb_128]
......@@ -279,10 +754,16 @@ SECTION .text
%if %1 == 16
; flat16 filter
%ifidn %2, v
lea tmpq, [dstq+mstrideq*8]
mova m0, [tmpq+strideq*1] ; p6
mova m2, [tmpq+strideq*2] ; p5
mova m7, [tmpq+stride3q] ; p4