[FFmpeg-cvslog] vp9: add mxext versions of the single-block (w=4, npx=8) h/v loopfilters.
Ronald S. Bultje
git at videolan.org
Wed Jul 27 00:29:09 EEST 2016
ffmpeg | branch: master | Ronald S. Bultje <rsbultje at gmail.com> | Tue Jul 19 12:06:32 2016 -0400| [7ca422bb1b8f80f8412a229454a8ac6bd9655ae8] | committer: Ronald S. Bultje
vp9: add mxext versions of the single-block (w=4,npx=8) h/v loopfilters.
Each takes about 0.5% of runtime in my profiles, and they didn't have
any SIMD yet so far (we only had simd for npx=16 double-block versions).
> http://git.videolan.org/gitweb.cgi/ffmpeg.git/?a=commit;h=7ca422bb1b8f80f8412a229454a8ac6bd9655ae8
---
libavcodec/x86/vp9dsp_init.c | 3 +
libavcodec/x86/vp9lpf.asm | 149 +++++++++++++++++++++++++++++--------------
2 files changed, 103 insertions(+), 49 deletions(-)
diff --git a/libavcodec/x86/vp9dsp_init.c b/libavcodec/x86/vp9dsp_init.c
index 10751e5..359d38e 100644
--- a/libavcodec/x86/vp9dsp_init.c
+++ b/libavcodec/x86/vp9dsp_init.c
@@ -126,6 +126,7 @@ void ff_vp9_loop_filter_v_##size1##_##size2##_##opt(uint8_t *dst, ptrdiff_t stri
void ff_vp9_loop_filter_h_##size1##_##size2##_##opt(uint8_t *dst, ptrdiff_t stride, \
int E, int I, int H)
+lpf_funcs(4, 8, mmxext);
lpf_funcs(16, 16, sse2);
lpf_funcs(16, 16, ssse3);
lpf_funcs(16, 16, avx);
@@ -281,6 +282,8 @@ av_cold void ff_vp9dsp_init_x86(VP9DSPContext *dsp, int bpp, int bitexact)
}
if (EXTERNAL_MMXEXT(cpu_flags)) {
+ dsp->loop_filter_8[0][0] = ff_vp9_loop_filter_h_4_8_mmxext;
+ dsp->loop_filter_8[0][1] = ff_vp9_loop_filter_v_4_8_mmxext;
init_subpel2(4, 0, 4, put, 8, mmxext);
init_subpel2(4, 1, 4, avg, 8, mmxext);
init_fpel_func(4, 1, 4, avg, _8, mmxext);
diff --git a/libavcodec/x86/vp9lpf.asm b/libavcodec/x86/vp9lpf.asm
index 2c4fe21..56bba4d 100644
--- a/libavcodec/x86/vp9lpf.asm
+++ b/libavcodec/x86/vp9lpf.asm
@@ -52,7 +52,7 @@ mask_mix48: times 8 db 0x00
SECTION .text
%macro SCRATCH 3
-%if ARCH_X86_64
+%ifdef m8
SWAP %1, %2
%else
mova [%3], m%1
@@ -60,7 +60,7 @@ SECTION .text
%endmacro
%macro UNSCRATCH 3
-%if ARCH_X86_64
+%ifdef m8
SWAP %1, %2
%else
mova m%1, [%3]
@@ -69,7 +69,7 @@ SECTION .text
; %1 = abs(%2-%3)
%macro ABSSUB 4 ; dst, src1 (RO), src2 (RO), tmp
-%if ARCH_X86_64
+%ifdef m8
psubusb %1, %3, %2
psubusb %4, %2, %3
%else
@@ -102,7 +102,7 @@ SECTION .text
%endmacro
%macro UNPACK 4
-%if ARCH_X86_64
+%ifdef m8
punpck%1bw %2, %3, %4
%else
mova %2, %3
@@ -334,22 +334,24 @@ SECTION .text
%endmacro
%macro DEFINE_TRANSPOSED_P7_TO_Q7 0-1 0
-%define P3 rsp + 0 + %1
-%define P2 rsp + 16 + %1
-%define P1 rsp + 32 + %1
-%define P0 rsp + 48 + %1
-%define Q0 rsp + 64 + %1
-%define Q1 rsp + 80 + %1
-%define Q2 rsp + 96 + %1
-%define Q3 rsp + 112 + %1
-%define P7 rsp + 128 + %1
-%define P6 rsp + 144 + %1
-%define P5 rsp + 160 + %1
-%define P4 rsp + 176 + %1
-%define Q4 rsp + 192 + %1
-%define Q5 rsp + 208 + %1
-%define Q6 rsp + 224 + %1
-%define Q7 rsp + 240 + %1
+%define P3 rsp + 0*mmsize + %1
+%define P2 rsp + 1*mmsize + %1
+%define P1 rsp + 2*mmsize + %1
+%define P0 rsp + 3*mmsize + %1
+%define Q0 rsp + 4*mmsize + %1
+%define Q1 rsp + 5*mmsize + %1
+%define Q2 rsp + 6*mmsize + %1
+%define Q3 rsp + 7*mmsize + %1
+%if mmsize == 16
+%define P7 rsp + 8*mmsize + %1
+%define P6 rsp + 9*mmsize + %1
+%define P5 rsp + 10*mmsize + %1
+%define P4 rsp + 11*mmsize + %1
+%define Q4 rsp + 12*mmsize + %1
+%define Q5 rsp + 13*mmsize + %1
+%define Q6 rsp + 14*mmsize + %1
+%define Q7 rsp + 15*mmsize + %1
+%endif
%endmacro
; ..............AB -> AAAAAAAABBBBBBBB
@@ -365,12 +367,12 @@ SECTION .text
%macro LOOPFILTER 5 ; %1=v/h %2=size1 %3+%4=stack, %5=32bit stack only
%if UNIX64
-cglobal vp9_loop_filter_%1_%2_16, 5, 9, 16, %3 + %4, dst, stride, E, I, H, mstride, dst2, stride3, mstride3
+cglobal vp9_loop_filter_%1_%2_ %+ mmsize, 5, 9, 16, %3 + %4, dst, stride, E, I, H, mstride, dst2, stride3, mstride3
%else
%if WIN64
-cglobal vp9_loop_filter_%1_%2_16, 4, 8, 16, %3 + %4, dst, stride, E, I, mstride, dst2, stride3, mstride3
+cglobal vp9_loop_filter_%1_%2_ %+ mmsize, 4, 8, 16, %3 + %4, dst, stride, E, I, mstride, dst2, stride3, mstride3
%else
-cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride, dst2, stride3, mstride3
+cglobal vp9_loop_filter_%1_%2_ %+ mmsize, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride, dst2, stride3, mstride3
%define Ed dword r2m
%define Id dword r3m
%endif
@@ -384,18 +386,22 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
lea mstride3q, [mstrideq*3]
%ifidn %1, h
-%if %2 > 16
+%if %2 != 16
+%if mmsize == 16
%define movx movh
+%else
+%define movx mova
+%endif
lea dstq, [dstq + 4*strideq - 4]
%else
%define movx movu
lea dstq, [dstq + 4*strideq - 8] ; go from top center (h pos) to center left (v pos)
%endif
- lea dst2q, [dstq + 8*strideq]
%else
lea dstq, [dstq + 4*mstrideq]
- lea dst2q, [dstq + 8*strideq]
%endif
+ ; FIXME we shouldn't need two dts registers if mmsize == 8
+ lea dst2q, [dstq + 8*strideq]
DEFINE_REAL_P7_TO_Q7
@@ -406,11 +412,11 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
movx m3, [P4]
movx m4, [P3]
movx m5, [P2]
-%if ARCH_X86_64 || %2 != 16
+%if (ARCH_X86_64 && mmsize == 16) || %2 > 16
movx m6, [P1]
%endif
movx m7, [P0]
-%if ARCH_X86_64
+%ifdef m8
movx m8, [Q0]
movx m9, [Q1]
movx m10, [Q2]
@@ -502,7 +508,7 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
movhps [Q5], m6
movhps [Q7], m7
DEFINE_TRANSPOSED_P7_TO_Q7
-%else ; %2 == 44/48/84/88
+%elif %2 > 16 ; %2 == 44/48/84/88
punpcklbw m0, m1
punpcklbw m2, m3
punpcklbw m4, m5
@@ -529,12 +535,31 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
mova [Q1], m5
mova [Q2], m7
mova [Q3], m3
+%else ; %2 == 4 || %2 == 8
+ SBUTTERFLY bw, 0, 1, 6
+ SBUTTERFLY bw, 2, 3, 6
+ SBUTTERFLY bw, 4, 5, 6
+ mova [rsp+4*mmsize], m5
+ mova m6, [P1]
+ SBUTTERFLY bw, 6, 7, 5
+ DEFINE_TRANSPOSED_P7_TO_Q7
+ TRANSPOSE4x4W 0, 2, 4, 6, 5
+ mova [P3], m0
+ mova [P2], m2
+ mova [P1], m4
+ mova [P0], m6
+ mova m5, [rsp+4*mmsize]
+ TRANSPOSE4x4W 1, 3, 5, 7, 0
+ mova [Q0], m1
+ mova [Q1], m3
+ mova [Q2], m5
+ mova [Q3], m7
%endif ; %2
%endif ; x86-32/64
%endif ; %1 == h
; calc fm mask
-%if %2 == 16
+%if %2 == 16 || mmsize == 8
%if cpuflag(ssse3)
pxor m0, m0
%endif
@@ -552,7 +577,7 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
mova m0, [pb_80]
pxor m2, m0
pxor m3, m0
-%if ARCH_X86_64
+%ifdef m8
%ifidn %1, v
mova m8, [P3]
mova m9, [P2]
@@ -613,10 +638,10 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
; (m3: fm, m8..15: p3 p2 p1 p0 q0 q1 q2 q3)
; calc flat8in (if not 44_16) and hev masks
-%if %2 != 44
+%if %2 != 44 && %2 != 4
mova m6, [pb_81] ; [1 1 1 1 ...] ^ 0x80
ABSSUB_GT m2, rp3, rp0, m6, m5 ; abs(p3 - p0) <= 1
-%if ARCH_X86_64
+%ifdef m8
mova m8, [pb_80]
%define rb80 m8
%else
@@ -655,8 +680,15 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
%endif
%else
mova m6, [pb_80]
+%if %2 == 44
movd m7, Hd
SPLATB_MIX m7
+%else
+%if cpuflag(ssse3)
+ pxor m0, m0
+%endif
+ SPLATB_REG m7, H, m0 ; H H H H ...
+%endif
pxor m7, m6
ABSSUB m4, rp1, rp0, m1 ; abs(p1 - p0)
pxor m4, m6
@@ -670,7 +702,7 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
%if %2 == 16
; (m0: hev, m2: flat8in, m3: fm, m6: pb_81, m9..15: p2 p1 p0 q0 q1 q2 q3)
; calc flat8out mask
-%if ARCH_X86_64
+%ifdef m8
mova m8, [P7]
mova m9, [P6]
%define rp7 m8
@@ -682,7 +714,7 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
ABSSUB_GT m1, rp7, rp0, m6, m5 ; abs(p7 - p0) <= 1
ABSSUB_GT m7, rp6, rp0, m6, m5 ; abs(p6 - p0) <= 1
por m1, m7
-%if ARCH_X86_64
+%ifdef m8
mova m8, [P5]
mova m9, [P4]
%define rp5 m8
@@ -695,7 +727,7 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
por m1, m7
ABSSUB_GT m7, rp4, rp0, m6, m5 ; abs(p4 - p0) <= 1
por m1, m7
-%if ARCH_X86_64
+%ifdef m8
mova m14, [Q4]
mova m15, [Q5]
%define rq4 m14
@@ -708,7 +740,7 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
por m1, m7
ABSSUB_GT m7, rq5, rq0, m6, m5 ; abs(q5 - q0) <= 1
por m1, m7
-%if ARCH_X86_64
+%ifdef m8
mova m14, [Q6]
mova m15, [Q7]
%define rq6 m14
@@ -738,7 +770,7 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
; (m0: hev, [m1: flat8out], [m2: flat8in], m3: fm, m8..15: p5 p4 p1 p0 q0 q1 q6 q7)
; filter2()
-%if %2 != 44
+%if %2 != 44 && %2 != 4
mova m6, [pb_80] ; already in m6 if 44_16
SCRATCH 2, 15, rsp+%3+%4
%if %2 == 16
@@ -756,7 +788,7 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
paddsb m4, m2 ; 3*(q0 - p0) + (p1 - q1)
paddsb m6, m4, [pb_4] ; m6: f1 = clip(f + 4, 127)
paddsb m4, [pb_3] ; m4: f2 = clip(f + 3, 127)
-%if ARCH_X86_64
+%ifdef m8
mova m14, [pb_10] ; will be reused in filter4()
%define rb10 m14
%else
@@ -765,8 +797,8 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
SRSHIFT3B_2X m6, m4, rb10, m7 ; f1 and f2 sign byte shift by 3
SIGN_SUB m7, rq0, m6, m5 ; m7 = q0 - f1
SIGN_ADD m1, rp0, m4, m5 ; m1 = p0 + f2
-%if %2 != 44
-%if ARCH_X86_64
+%if %2 != 44 && %2 != 4
+%ifdef m8
pandn m6, m15, m3 ; ~mask(in) & mask(fm)
%else
mova m6, [rsp+%3+%4]
@@ -787,8 +819,8 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
paddsb m6, m2, [pb_4] ; m6: f1 = clip(f + 4, 127)
paddsb m2, [pb_3] ; m2: f2 = clip(f + 3, 127)
SRSHIFT3B_2X m6, m2, rb10, m4 ; f1 and f2 sign byte shift by 3
-%if %2 != 44
-%if ARCH_X86_64
+%if %2 != 44 && %2 != 4
+%ifdef m8
pandn m5, m15, m3 ; ~mask(in) & mask(fm)
%else
mova m5, [rsp+%3+%4]
@@ -815,26 +847,26 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
mova [P1], m1
mova [Q1], m4
-%if %2 != 44
+%if %2 != 44 && %2 != 4
UNSCRATCH 2, 15, rsp+%3+%4
%endif
; ([m1: flat8out], m2: flat8in, m3: fm, m10..13: p1 p0 q0 q1)
; filter6()
-%if %2 != 44
+%if %2 != 44 && %2 != 4
pxor m0, m0
%if %2 > 16
pand m3, m2
%else
pand m2, m3 ; mask(fm) & mask(in)
-%if ARCH_X86_64
+%ifdef m8
pandn m3, m8, m2 ; ~mask(out) & (mask(fm) & mask(in))
%else
mova m3, [rsp+%3+%4+16]
pandn m3, m2
%endif
%endif
-%if ARCH_X86_64
+%ifdef m8
mova m14, [P3]
mova m9, [Q3]
%define rp3 m14
@@ -882,7 +914,7 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
pand m1, m2 ; mask(out) & (mask(fm) & mask(in))
mova m2, [P7]
mova m3, [P6]
-%if ARCH_X86_64
+%ifdef m8
mova m8, [P5]
mova m9, [P4]
%define rp5 m8
@@ -1008,7 +1040,7 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
movhps [Q5], m6
movhps [Q7], m7
%endif
-%elif %2 == 44
+%elif %2 == 44 || %2 == 4
SWAP 0, 1 ; m0 = p1
SWAP 1, 7 ; m1 = p0
SWAP 2, 5 ; m2 = q0
@@ -1018,6 +1050,7 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
SBUTTERFLY bw, 2, 3, 4
SBUTTERFLY wd, 0, 2, 4
SBUTTERFLY wd, 1, 3, 4
+%if mmsize == 16
movd [P7], m0
movd [P3], m2
movd [Q0], m1
@@ -1047,6 +1080,20 @@ cglobal vp9_loop_filter_%1_%2_16, 2, 6, 16, %3 + %4 + %5, dst, stride, mstride,
movd [Q3], m1
movd [Q7], m3
%else
+ movd [P7], m0
+ movd [P5], m2
+ movd [P3], m1
+ movd [P1], m3
+ psrlq m0, 32
+ psrlq m2, 32
+ psrlq m1, 32
+ psrlq m3, 32
+ movd [P6], m0
+ movd [P4], m2
+ movd [P2], m1
+ movd [P0], m3
+%endif
+%else
; the following code do a transpose of 8 full lines to 16 half
; lines (high part). It is inlined to avoid the need of a staging area
mova m0, [P3]
@@ -1137,3 +1184,7 @@ LPF_16_VH_ALL_OPTS 44, 0, 128, 0
LPF_16_VH_ALL_OPTS 48, 256, 128, 16
LPF_16_VH_ALL_OPTS 84, 256, 128, 16
LPF_16_VH_ALL_OPTS 88, 256, 128, 16
+
+INIT_MMX mmxext
+LOOPFILTER v, 4, 0, 0, 0
+LOOPFILTER h, 4, 0, 64, 0
More information about the ffmpeg-cvslog
mailing list