Hi,
2015-01-25 2:05 GMT+01:00 James Almer <jamr...@gmail.com>:
> 2 to 2.5 times faster.
>
> Signed-off-by: James Almer <jamr...@gmail.com>
> ---
> libavcodec/x86/sbrdsp.asm | 114
> +++++++++++++++++++++++++++++++++++++++++++
Not the first time that I notice that, but memmoves are often
suboptimal using old SSE ones.
While movlhps is fine, movlps isn't, on my old core i5. You may want
to validate this with the attached patch, where storing ps_mask3 in m8
is a gain in Win64 (the gain does not match the number of loops, but
it is still there).
Benchmarks:
x64: 6023 decicycles in g, 262108 runs, 36 skips
SSE: 3049 decicycles in g, 262130 runs, 14 skips
SSE3: 2843 decicycles in g, 262086 runs, 58 skips
movq: 2693 decicycles in g, 262117 runs, 27 skips
m8: 2648 decicycles in g, 262083 runs, 61 skips
Thanks for doing it, I had only 3yo scraps left and no further
motivation to tackle the start/tail parts.
--
Christophe
From 49e41dd86eb65a774f3561420dd5e9de83f328f2 Mon Sep 17 00:00:00 2001
From: Christophe Gisquet <christophe.gisq...@gmail.com>
Date: Sun, 25 Jan 2015 13:52:16 +0100
Subject: [PATCH 2/2] Use different mem moves.
---
libavcodec/x86/sbrdsp.asm | 28 ++++++++++++++++++----------
1 file changed, 18 insertions(+), 10 deletions(-)
diff --git a/libavcodec/x86/sbrdsp.asm b/libavcodec/x86/sbrdsp.asm
index c9f2d88..955d6cc 100644
--- a/libavcodec/x86/sbrdsp.asm
+++ b/libavcodec/x86/sbrdsp.asm
@@ -448,19 +448,27 @@ cglobal sbr_qmf_deint_neg, 2,4,4,v,src,vrev,c
REP_RET
%macro SBR_AUTOCORRELATE 0
-cglobal sbr_autocorrelate, 2,3,8,32, x, phi, cnt
+cglobal sbr_autocorrelate, 2,3,8+ARCH_X86_64,32, x, phi, cnt
mov cntq, 37*8
add xq, cntq
neg cntq
+%if ARCH_X86_64
+ mova m8, [ps_mask3]
+%define MASK m8
+%else
+%define MASK [ps_mask3]
+%endif
%if cpuflag(sse3)
+%define MOVH movq
movddup m5, [xq+cntq]
%else
+%define MOVH movlps
movlps m5, [xq+cntq]
movlhps m5, m5
%endif
- movlps m7, [xq+cntq+8 ]
- movlps m1, [xq+cntq+16]
+ MOVH m7, [xq+cntq+8 ]
+ MOVH m1, [xq+cntq+16]
shufps m7, m7, q0110
shufps m1, m1, q0110
mulps m3, m5, m7 ; x[0][0] * x[1][0], x[0][1] * x[1][1], x[0][0] * x[1][1], x[0][1] * x[1][0]
@@ -470,7 +478,7 @@ cglobal sbr_autocorrelate, 2,3,8,32, x, phi, cnt
movaps [rsp+16], m4
add cntq, 8
- movlps m2, [xq+cntq+16]
+ MOVH m2, [xq+cntq+16]
movlhps m7, m7
shufps m2, m2, q0110
mulps m6, m7, m1 ; real_sum1 = x[1][0] * x[2][0], x[1][1] * x[2][1], x[1][0] * x[2][1], x[1][1] * x[2][0]
@@ -481,7 +489,7 @@ cglobal sbr_autocorrelate, 2,3,8,32, x, phi, cnt
align 16
.loop:
add cntq, 8
- movlps m0, [xq+cntq+16]
+ MOVH m0, [xq+cntq+16]
movlhps m1, m1
shufps m0, m0, q0110
mulps m3, m1, m2
@@ -491,7 +499,7 @@ align 16
addps m5, m4 ; real_sum2 += x[i][0] * x[i + 2][0], x[i][1] * x[i + 2][1], x[i][0] * x[i + 2][1], x[i][1] * x[i + 2][0];
addps m7, m1 ; real_sum0 += x[i][0] * x[i][0], x[i][1] * x[i][1];
add cntq, 8
- movlps m1, [xq+cntq+16]
+ MOVH m1, [xq+cntq+16]
movlhps m2, m2
shufps m1, m1, q0110
mulps m3, m2, m0
@@ -501,7 +509,7 @@ align 16
addps m5, m4 ; real_sum2 += x[i][0] * x[i + 2][0], x[i][1] * x[i + 2][1], x[i][0] * x[i + 2][1], x[i][1] * x[i + 2][0];
addps m7, m2 ; real_sum0 += x[i][0] * x[i][0], x[i][1] * x[i][1];
add cntq, 8
- movlps m2, [xq+cntq+16]
+ MOVH m2, [xq+cntq+16]
movlhps m0, m0
shufps m2, m2, q0110
mulps m3, m0, m1
@@ -520,9 +528,9 @@ align 16
addps m6, [rsp ] ; real_sum1 + x[ 0][0] * x[ 1][0] + x[ 0][1] * x[ 1][1]; imag_sum1 + x[ 0][0] * x[ 1][1] - x[ 0][1] * x[ 1][0];
addps m7, [rsp+16] ; real_sum0 + x[ 0][0] * x[ 0][0] + x[ 0][1] * x[ 0][1];
- xorps m4, [ps_mask3]
- xorps m5, [ps_mask3]
- xorps m6, [ps_mask3]
+ xorps m4, MASK
+ xorps m5, MASK
+ xorps m6, MASK
%if cpuflag(sse3)
movshdup m2, m1
haddps m4, m5
--
1.9.2.msysgit.0
_______________________________________________
ffmpeg-devel mailing list
ffmpeg-devel@ffmpeg.org
http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
