Message ID | 20230307090806.2003-1-zhujunxian@oss.cipunited.com |
---|---|
State | New |
Headers | show |
Series | [FFmpeg-devel,v3] avcodec/mathops: Optimize generic mid_pred function | expand |
Context | Check | Description |
---|---|---|
yinshiyou/make_loongarch64 | fail | Make failed |
andriy/make_x86 | success | Make finished |
andriy/make_fate_x86 | success | Make fate finished |
On Tue, Mar 07, 2023 at 05:08:27PM +0800, Junxian Zhu wrote: > From: Junxian Zhu <zhujunxian@oss.cipunited.com> > > Rewrite mid_pred function in generic mathops.h, reduce branch jump to improve performance. And because nowadays new version compiler can compile enough short asmbbely code as handwritting in these function, so remove specified optimized mips inline asmbbely mathops.h. as you write, that it improves performance what speed effect does this have exactly? thx [...]
Michael Niedermayer <michael@niedermayer.cc> 于2023年3月8日周三 04:45写道: > > On Tue, Mar 07, 2023 at 05:08:27PM +0800, Junxian Zhu wrote: > > From: Junxian Zhu <zhujunxian@oss.cipunited.com> > > > > Rewrite mid_pred function in generic mathops.h, reduce branch jump to improve performance. And because nowadays new version compiler can compile enough short asmbbely code as handwritting in these function, so remove specified optimized mips inline asmbbely mathops.h. > > as you write, that it improves performance > what speed effect does this have exactly? > thx > I tested the performance, using this code ``` #include <stdio.h> #include <time.h> #include <stdlib.h> #define FFMIN(a, b) ( a>b ? b : a ) #define FFMAX(a, b) ( a>b ? a : b ) int mid_pred(int a, int b, int c) { #if OLD if(a>b){ if(c>b){ if(c>a) b=a; else b=c; } }else{ if(b>c){ if(c>a) b=c; else b=a; } } return b; #else int t0,t1,t2,t3; t0 = (a > b) ? b : a ; t1 = (a > b) ? a : b ; t2 = (t0 > c) ? t0 : c; t3 = (t1 > t2) ? t2 : t1; return t3; #endif } int main() { int a[1024], b[1024], c[1024], d[1024]; srand(time(NULL)); for(int i=0; i<1024; i++) { a[i] = rand(); b[i] = rand(); c[i] = rand(); } for (int j=0; j<1e7+rand()%2; j++) for(int i=0; i<1024; i++) d[i] = mid_pred(a[i], b[i], c[i]); printf("%d, %d\n", d[rand()%1024], j); } ``` On MacOS 13.2 with Apple M1: The old code the new code 2.1s 2.3s On Cavium ThunderX / arm64 (GCC 10.2.1 -O3) The old code the new code 52.7s 37.8s On Loongson 3A4000/mips64el (GCC 10.2.1 -O3) The old code the new code 90s 5s On Intel(R) Xeon(R) CPU E7-4820 v4 @ 2.00GHz (GCC 10.2.1 -O3) The old code the new code 14.4s 15.4s On SF19A2890/MIPS interAptiv (GCC 10.2.1 -O3) The old code the new code 314s 39.3s On Intel(R) Xeon(R) CPU E7-4820 v4 @ 2.00GHz (GCC 12.2.0 -O3) The old code the new code 14.4s 8.8s On sifive,bullet0/rv64imafdc (GCC 12.2.0 -O3, 1e6 times instead of 1e7) The old code the new code 11.9s 15.2s On Freescale i.MX53/ARMv7 Processor rev 5 (v7l) (GCC 12.2.0 -O3, 1e6 times instead of 1e7) The old code the new code 24.1s 15.7s On POWER8 (architected), altivec supported, BIG ENDIAN, ppc64 (GCC 12.2.0 -O3) The old code the new code 43.1s 50.8s On POWER8 (architected), altivec supported, LITTLE ENDIAN, ppc64el (GCC 12.2.0 -O3) The old code the new code 7.8s 4.7s On PA8900 (Shortfin) PA-RISC (GCC 12.2.0 -O3 1e6 times instead of 1e7) The old code the new code 39.9s 47.2s On IBM/S390 aka s390x (GCC 12.2.0 -O3) The old code the new code 82.2s 30.8s On Intel(R) Itanium(R) Processor 9320 (GCC 12.2.0 -O3) The old code the new code 89.5s 78.1s Cavium Octeon III V0.2 FPU V0.0 /mipsel (GCC 12.2.0 -O3) The old code the new code 117.5s 118.5s > [...] > -- > Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB > > It is dangerous to be right in matters on which the established authorities > are wrong. -- Voltaire > _______________________________________________ > ffmpeg-devel mailing list > ffmpeg-devel@ffmpeg.org > https://ffmpeg.org/mailman/listinfo/ffmpeg-devel > > To unsubscribe, visit link above, or email > ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
On Wed, Mar 15, 2023 at 06:09:13PM +0800, YunQiang Su wrote: > Michael Niedermayer <michael@niedermayer.cc> 于2023年3月8日周三 04:45写道: > > > > On Tue, Mar 07, 2023 at 05:08:27PM +0800, Junxian Zhu wrote: > > > From: Junxian Zhu <zhujunxian@oss.cipunited.com> > > > > > > Rewrite mid_pred function in generic mathops.h, reduce branch jump to improve performance. And because nowadays new version compiler can compile enough short asmbbely code as handwritting in these function, so remove specified optimized mips inline asmbbely mathops.h. > > > > as you write, that it improves performance > > what speed effect does this have exactly? > > thx > > > > I tested the performance, using this code [...] > On MacOS 13.2 with Apple M1: > The old code the new code > 2.1s 2.3s > > On Cavium ThunderX / arm64 (GCC 10.2.1 -O3) > The old code the new code > 52.7s 37.8s > > On Loongson 3A4000/mips64el (GCC 10.2.1 -O3) > The old code the new code > 90s 5s > > On Intel(R) Xeon(R) CPU E7-4820 v4 @ 2.00GHz (GCC 10.2.1 -O3) > The old code the new code > 14.4s 15.4s > > On SF19A2890/MIPS interAptiv (GCC 10.2.1 -O3) > The old code the new code > 314s 39.3s > > On Intel(R) Xeon(R) CPU E7-4820 v4 @ 2.00GHz (GCC 12.2.0 -O3) > The old code the new code > 14.4s 8.8s > > On sifive,bullet0/rv64imafdc (GCC 12.2.0 -O3, 1e6 times instead of 1e7) > The old code the new code > 11.9s 15.2s > > On Freescale i.MX53/ARMv7 Processor rev 5 (v7l) (GCC 12.2.0 -O3, 1e6 > times instead of 1e7) > The old code the new code > 24.1s 15.7s > > On POWER8 (architected), altivec supported, BIG ENDIAN, ppc64 (GCC 12.2.0 -O3) > The old code the new code > 43.1s 50.8s > > On POWER8 (architected), altivec supported, LITTLE ENDIAN, ppc64el > (GCC 12.2.0 -O3) > The old code the new code > 7.8s 4.7s > > On PA8900 (Shortfin) PA-RISC (GCC 12.2.0 -O3 1e6 times instead of 1e7) > The old code the new code > 39.9s 47.2s > > On IBM/S390 aka s390x (GCC 12.2.0 -O3) > The old code the new code > 82.2s 30.8s > > On Intel(R) Itanium(R) Processor 9320 (GCC 12.2.0 -O3) > The old code the new code > 89.5s 78.1s > > Cavium Octeon III V0.2 FPU V0.0 /mipsel (GCC 12.2.0 -O3) > The old code the new code > 117.5s 118.5s These cover a quite extensive set of hw, impressive thx [...]
diff --git a/libavcodec/mathops.h b/libavcodec/mathops.h index c89054d6ed..f2ba4fabce 100644 --- a/libavcodec/mathops.h +++ b/libavcodec/mathops.h @@ -41,8 +41,6 @@ extern const uint8_t ff_zigzag_scan[16+1]; # include "arm/mathops.h" #elif ARCH_AVR32 # include "avr32/mathops.h" -#elif ARCH_MIPS -# include "mips/mathops.h" #elif ARCH_PPC # include "ppc/mathops.h" #elif ARCH_X86 @@ -98,18 +96,11 @@ static av_always_inline unsigned UMULH(unsigned a, unsigned b){ #define mid_pred mid_pred static inline av_const int mid_pred(int a, int b, int c) { - if(a>b){ - if(c>b){ - if(c>a) b=a; - else b=c; - } - }else{ - if(b>c){ - if(c>a) b=c; - else b=a; - } - } - return b; + int t0,t1,t2,t3; + int t0 = FFMIN(a, b); + int t1 = FFMAX(a, b); + int t2 = FFMAX(t0, c); + return FFMIN(t1, t2); } #endif diff --git a/libavcodec/mips/mathops.h b/libavcodec/mips/mathops.h deleted file mode 100644 index bb9dc8375a..0000000000 --- a/libavcodec/mips/mathops.h +++ /dev/null @@ -1,67 +0,0 @@ -/* - * Copyright (c) 2009 Mans Rullgard <mans@mansr.com> - * Copyright (c) 2015 Zhou Xiaoyong <zhouxiaoyong@loongson.cn> - * - * This file is part of FFmpeg. - * - * FFmpeg is free software; you can redistribute it and/or - * modify it under the terms of the GNU Lesser General Public - * License as published by the Free Software Foundation; either - * version 2.1 of the License, or (at your option) any later version. - * - * FFmpeg is distributed in the hope that it will be useful, - * but WITHOUT ANY WARRANTY; without even the implied warranty of - * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU - * Lesser General Public License for more details. - * - * You should have received a copy of the GNU Lesser General Public - * License along with FFmpeg; if not, write to the Free Software - * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA - */ - -#ifndef AVCODEC_MIPS_MATHOPS_H -#define AVCODEC_MIPS_MATHOPS_H - -#include <stdint.h> -#include "config.h" -#include "libavutil/common.h" - -#if HAVE_INLINE_ASM - -#if HAVE_LOONGSON3 - -#define MULH MULH -static inline av_const int MULH(int a, int b) -{ - int c; - __asm__ ("dmult %1, %2 \n\t" - "mflo %0 \n\t" - "dsrl %0, %0, 32 \n\t" - : "=r"(c) - : "r"(a),"r"(b) - : "hi", "lo"); - return c; -} - -#define mid_pred mid_pred -static inline av_const int mid_pred(int a, int b, int c) -{ - int t = b; - __asm__ ("sgt $8, %1, %2 \n\t" - "movn %0, %1, $8 \n\t" - "movn %1, %2, $8 \n\t" - "sgt $8, %1, %3 \n\t" - "movz %1, %3, $8 \n\t" - "sgt $8, %0, %1 \n\t" - "movn %0, %1, $8 \n\t" - : "+&r"(t),"+&r"(a) - : "r"(b),"r"(c) - : "$8"); - return t; -} - -#endif /* HAVE_LOONGSON3 */ - -#endif /* HAVE_INLINE_ASM */ - -#endif /* AVCODEC_MIPS_MATHOPS_H */