From: Xu Jun <xuju...@sjtu.edu.cn>
Read 16 elements from memory, shuffle and parallally compute 4 rows at a time,
shuffle and parallelly write 16 results to memory.
Performance improves about 15% compared to v1.
Tested using this command:
./ffmpeg_g -s 1280*720 -pix_fmt yuv420p -i test.yuv -vf convolution="1 2 3 4 5
6 7 8 9:1 2 3 4 5 6 7 8 9:1 2 3 4 5 6 7 8 9:1 2 3 4 5 6 7 8
9:1/45:1/45:1/45:1/45:1:2:3:4:row:row:row:row" -an -vframes 5000 -f null
/dev/null -benchmark
after patch:
frame= 4317 fps=622 q=-0.0 Lsize=N/A time=00:02:52.68 bitrate=N/A speed=24.9x
video:2260kB audio:0kB subtitle:0kB other streams:0kB global headers:0kB muxing
overhead: unknown
bench: utime=20.539s stime=1.834s rtime=6.943s
before patch(c version):
frame= 4317 fps=306 q=-0.0 Lsize=N/A time=00:02:52.68 bitrate=N/A speed=12.2x
video:2260kB audio:0kB subtitle:0kB other streams:0kB global headers:0kB muxing
overhead: unknown
bench: utime=60.591s stime=1.787s rtime=14.100s
Signed-off-by: Xu Jun <xuju...@sjtu.edu.cn>
---
libavfilter/x86/vf_convolution.asm | 131 ++++++++++++++++++++++++++
libavfilter/x86/vf_convolution_init.c | 9 ++
2 files changed, 140 insertions(+)
mode change 100644 => 100755 libavfilter/x86/vf_convolution.asm
diff --git a/libavfilter/x86/vf_convolution.asm
b/libavfilter/x86/vf_convolution.asm
old mode 100644
new mode 100755
index 754d4d1064..2a09374b00
--- a/libavfilter/x86/vf_convolution.asm
+++ b/libavfilter/x86/vf_convolution.asm
@@ -154,3 +154,134 @@ cglobal filter_3x3, 4, 15, 7, dst, width, rdiv, bias,
matrix, ptr, c0, c1, c2, c
INIT_XMM sse4
FILTER_3X3
%endif
+
+; void filter_row_sse4(uint8_t *dst, int width,
+; float rdiv, float bias, const int *const matrix,
+; const uint8_t *c[], int peak, int radius,
+; int dstride, int stride)
+
+%macro COMPUTE_4ROW 1
+pshufb m7, m6, m4 ; get 4 uint8s from the 16 uint8s
+pmulld m7, m5
+paddd m1%1, m7
+%endmacro
+
+%macro CVT_PACK_ROW 1
+cvtdq2ps m1%1, m1%1
+mulps m1%1, m0 ; sum *= rdiv
+addps m1%1, m1 ; sum += bias
+addps m1%1, m3 ; sum += 0.5
+cvttps2dq m1%1, m1%1
+packssdw m1%1, m1%1
+packuswb m1%1, m1%1
+%endmacro
+
+%if ARCH_X86_64
+INIT_XMM sse4
+%if UNIX64
+cglobal filter_row, 6, 10, 14, dst, width, matrix, ptr, mult, rad, r, x, i, ci
+%else
+cglobal filter_row, 4, 10, 14, dst, width, rdiv, bias, matrix, ptr, mult, rad,
r, x, i, ci
+%endif
+
+%if WIN64
+ SWAP m0, m2
+ SWAP m1, m3
+ mov r2q, matrixmp
+ mov r3q, ptrmp
+ mov r5q, radmp
+ DEFINE_ARGS dst, width, matrix, ptr, mult, rad, r, x, i, ci
+%endif
+
+movsxdifnidn widthq, widthd
+movsxdifnidn radq, radd
+lea radq, [radq * 2 + 1]
+VBROADCASTSS m0, m0
+VBROADCASTSS m1, m1
+pxor m2, m2 ; zero
+movss m3, [half]
+VBROADCASTSS m3, m3 ; 0.5
+movdqu m8, [shuf_init] ; shuffle initialization
+movdqu m9, [shuf_step] ; shuffle step
+
+xor xq, xq
+cmp widthq, mmsize
+jl .less16
+
+mov rq, widthq
+and rq, mmsize-1
+sub widthq, rq
+
+.equal16:
+ pxor m10, m10
+ pxor m11, m11
+ pxor m12, m12
+ pxor m13, m13
+ ; m10-13 hold sums
+
+ lea iq, [radq - 1]
+ .loopi:
+ movd m5, [matrixq + 4*iq] ; matrix[i]
+ VBROADCASTSS m5, m5
+ mov ciq, [ptrq + iq * gprsize]
+ movdqu m6, [ciq + xq] ; c[i][y*stride] 16 uint8s
+
+ ;m4 controls shuffle
+ movdqa m4, m8
+ COMPUTE_4ROW 0 ; process 0-3 rows, sum in m10
+ paddd m4, m9
+ COMPUTE_4ROW 1 ; process 4-7 rows, sum in m11
+ paddd m4, m9
+ COMPUTE_4ROW 2 ; process 8-11 rows, sum in m12
+ paddd m4, m9
+ COMPUTE_4ROW 3 ; process 12-15 rows, sum in m13
+
+ sub iq, 1
+ jns .loopi
+
+ CVT_PACK_ROW 0 ; process 0-3 rows, result in m10's low 32bit
+ CVT_PACK_ROW 1 ; process 4-7 rows, result in m11's low 32bit
+ CVT_PACK_ROW 2 ; process 8-11 rows, result in m12's low 32bit
+ CVT_PACK_ROW 3 ; process 12-15 rows, result in m13's low 32bit
+ punpckldq m10, m11
+ punpckldq m12, m13
+ punpcklqdq m10, m12 ; pack 16 results in m10
+ movdqu [dstq+xq], m10
+
+ add xq, mmsize
+ cmp xq, widthq
+ jl .equal16
+
+ add widthq, rq
+ cmp xq, widthq
+ jge .end
+
+.less16:
+ xor rd, rd
+ lea iq, [radq - 1]
+ .loopr_i:
+ mov ciq, [ptrq + iq * gprsize]
+ movzx multd, byte [ciq + xq]
+ imul multd, [matrixq + 4*iq]
+ add rd, multd
+
+ sub iq, 1
+ jns .loopr_i
+
+ pxor m7, m7
+ cvtsi2ss m7, rd
+ mulss m7, m0 ; sum *= rdiv
+ addss m7, m1 ; sum += bias
+ addss m7, m3 ; sum += 0.5
+ cvttps2dq m7, m7
+ packssdw m7, m7
+ packuswb m7, m7
+ movd rd, m7
+ mov [dstq + xq], rb
+
+ add xq, 1
+ cmp xq, widthq
+ jl .less16
+.end:
+ RET
+%endif
diff --git a/libavfilter/x86/vf_convolution_init.c
b/libavfilter/x86/vf_convolution_init.c
index 51432406ed..5eb3b3bee1 100644
--- a/libavfilter/x86/vf_convolution_init.c
+++ b/libavfilter/x86/vf_convolution_init.c
@@ -29,6 +29,11 @@ void ff_filter_3x3_sse4(uint8_t *dst, int width,
const uint8_t *c[], int peak, int radius,
int dstride, int stride);
+void ff_filter_row_sse4(uint8_t *dst, int width,
+ float rdiv, float bias, const int *const matrix,
+ const uint8_t *c[], int peak, int radius,
+ int dstride, int stride);
+
av_cold void ff_convolution_init_x86(ConvolutionContext *s)
{
#if ARCH_X86_64
@@ -41,6 +46,10 @@ av_cold void ff_convolution_init_x86(ConvolutionContext *s)
s->filter[i] = ff_filter_3x3_sse4;
}
}
+ if (s->mode[i] == MATRIX_ROW) {
+ if (EXTERNAL_SSE4(cpu_flags))
+ s->filter[i] = ff_filter_row_sse4;
+ }
}
#endif
}
--
2.17.1
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
To unsubscribe, visit link above, or email
ffmpeg-devel-requ...@ffmpeg.org with subject "unsubscribe".
