[FFmpeg-devel] [PATCH V4 1/2] libswscale/x86/yuv2rgb: Change inline assembly into nasm code

Fu, Ting ting.fu at intel.com
Fri Jan 3 08:59:28 EET 2020



> -----Original Message-----
> From: ffmpeg-devel <ffmpeg-devel-bounces at ffmpeg.org> On Behalf Of
> Michael Niedermayer
> Sent: Friday, December 27, 2019 07:38 PM
> To: FFmpeg development discussions and patches <ffmpeg-devel at ffmpeg.org>
> Subject: Re: [FFmpeg-devel] [PATCH V4 1/2] libswscale/x86/yuv2rgb: Change
> inline assembly into nasm code
> 
> On Thu, Dec 19, 2019 at 11:35:51AM +0800, Ting Fu wrote:
> > Tested using this command:
> > ./ffmpeg -pix_fmt yuv420p -s 1920*1080 -i ArashRawYuv420.yuv \ -vcodec
> > rawvideo -s 1920*1080 -pix_fmt rgb24 -f null /dev/null
> >
> 
> > The fps increase from 151 to 389 on my local machine.
> 
> Thats nice but why is there such a difference from changing the way the code is
> assembled ?
> This should definitly be explained more detailedly in the commit message
> 
Hi, Michael

The fps increasing means mmx compared to C code, not inline compared nasm one. I will remove it from the commit message next patch version.

> 
> >
> > Signed-off-by: Ting Fu <ting.fu at intel.com>
> > ---
> >  libswscale/x86/Makefile           |   1 +
> >  libswscale/x86/swscale.c          |  16 +-
> >  libswscale/x86/yuv2rgb.c          |  81 +++---
> >  libswscale/x86/yuv2rgb_template.c | 441 ++++++------------------------
> >  libswscale/x86/yuv_2_rgb.asm      | 270 ++++++++++++++++++
> >  5 files changed, 395 insertions(+), 414 deletions(-)  create mode
> > 100644 libswscale/x86/yuv_2_rgb.asm
> >
> > diff --git a/libswscale/x86/Makefile b/libswscale/x86/Makefile index
> > f317d5dd9b..831d5359aa 100644
> > --- a/libswscale/x86/Makefile
> > +++ b/libswscale/x86/Makefile
> > @@ -12,3 +12,4 @@ X86ASM-OBJS                     += x86/input.o                          \
> >                                     x86/output.o                         \
> >                                     x86/scale.o                          \
> >                                     x86/rgb_2_rgb.o                      \
> > +                                   x86/yuv_2_rgb.o                      \
> > diff --git a/libswscale/x86/swscale.c b/libswscale/x86/swscale.c index
> > 0eed4f18d5..e9d474a1e8 100644
> > --- a/libswscale/x86/swscale.c
> > +++ b/libswscale/x86/swscale.c
> > @@ -29,6 +29,14 @@
> >  #include "libavutil/cpu.h"
> >  #include "libavutil/pixdesc.h"
> >
> > +const DECLARE_ALIGNED(8, uint64_t, ff_dither4)[2] = {
> > +    0x0103010301030103LL,
> > +    0x0200020002000200LL,};
> > +
> > +const DECLARE_ALIGNED(8, uint64_t, ff_dither8)[2] = {
> > +    0x0602060206020602LL,
> > +    0x0004000400040004LL,};
> > +
> >  #if HAVE_INLINE_ASM
> >
> >  #define DITHER1XBPP
> > @@ -38,14 +46,6 @@ DECLARE_ASM_CONST(8, uint64_t, bFC)=
> 0xFCFCFCFCFCFCFCFCLL;
> >  DECLARE_ASM_CONST(8, uint64_t, w10)=       0x0010001000100010LL;
> >  DECLARE_ASM_CONST(8, uint64_t, w02)=       0x0002000200020002LL;
> >
> > -const DECLARE_ALIGNED(8, uint64_t, ff_dither4)[2] = {
> > -    0x0103010301030103LL,
> > -    0x0200020002000200LL,};
> > -
> > -const DECLARE_ALIGNED(8, uint64_t, ff_dither8)[2] = {
> > -    0x0602060206020602LL,
> > -    0x0004000400040004LL,};
> > -
> >  DECLARE_ASM_CONST(8, uint64_t, b16Mask)=   0x001F001F001F001FLL;
> >  DECLARE_ASM_CONST(8, uint64_t, g16Mask)=   0x07E007E007E007E0LL;
> >  DECLARE_ASM_CONST(8, uint64_t, r16Mask)=   0xF800F800F800F800LL;
> > diff --git a/libswscale/x86/yuv2rgb.c b/libswscale/x86/yuv2rgb.c index
> > 5e2f77c20f..ed9b613cab 100644
> > --- a/libswscale/x86/yuv2rgb.c
> > +++ b/libswscale/x86/yuv2rgb.c
> > @@ -37,7 +37,7 @@
> >  #include "libavutil/x86/cpu.h"
> >  #include "libavutil/cpu.h"
> >
> > -#if HAVE_INLINE_ASM
> > +#if HAVE_X86ASM
> >
> >  #define DITHER1XBPP // only for MMX
> >
> > @@ -50,70 +50,51 @@ DECLARE_ASM_CONST(8, uint64_t, pb_03) =
> > 0x0303030303030303ULL;  DECLARE_ASM_CONST(8, uint64_t, pb_07) =
> > 0x0707070707070707ULL;
> >
> >  //MMX versions
> > -#if HAVE_MMX_INLINE && HAVE_6REGS
> > -#undef RENAME
> > +#if HAVE_MMX
> >  #undef COMPILE_TEMPLATE_MMXEXT
> >  #define COMPILE_TEMPLATE_MMXEXT 0
> > -#define RENAME(a) a ## _mmx
> > -#include "yuv2rgb_template.c"
> > -#endif /* HAVE_MMX_INLINE && HAVE_6REGS */
> > +#endif /* HAVE_MMX */
> >
> >  // MMXEXT versions
> > -#if HAVE_MMXEXT_INLINE && HAVE_6REGS
> > -#undef RENAME
> > +#if HAVE_MMXEXT
> >  #undef COMPILE_TEMPLATE_MMXEXT
> >  #define COMPILE_TEMPLATE_MMXEXT 1
> > -#define RENAME(a) a ## _mmxext
> > -#include "yuv2rgb_template.c"
> > -#endif /* HAVE_MMXEXT_INLINE && HAVE_6REGS */
> > +#endif /* HAVE_MMXEXT */
> >
> > -#endif /* HAVE_INLINE_ASM */
> > +#include "yuv2rgb_template.c"
> >
> >  av_cold SwsFunc ff_yuv2rgb_init_x86(SwsContext *c)  { -#if
> > HAVE_MMX_INLINE && HAVE_6REGS
> >      int cpu_flags = av_get_cpu_flags();
> >
> > -#if HAVE_MMXEXT_INLINE
> > -    if (INLINE_MMXEXT(cpu_flags)) {
> > -        switch (c->dstFormat) {
> > -        case AV_PIX_FMT_RGB24:
> > -            return yuv420_rgb24_mmxext;
> > -        case AV_PIX_FMT_BGR24:
> > -            return yuv420_bgr24_mmxext;
> > -        }
> > -    }
> > -#endif
> > -
> > -    if (INLINE_MMX(cpu_flags)) {
> 
> > +    if (EXTERNAL_MMX(cpu_flags) || EXTERNAL_MMXEXT(cpu_flags)) {
> 
> i would expect EXTERNAL_MMXEXT to imply EXTERNAL_MMX
> 

I was thinking the mmx-only processor. Under this circumstance, the mmx-only processor will not be accelerated. Should that be OK? Or it means I will be OK for not care much about old mmx-only processor in following patches?

> 
> [...]
> > +        case AV_PIX_FMT_RGB24:
> > +            return yuv420_rgb24;
> 
> [...]
> 
> > +static inline int yuv420_rgb24(SwsContext *c, const uint8_t *src[],
> > +                                       int srcStride[],
> > +                                       int srcSliceY, int srcSliceH,
> > +                                       uint8_t *dst[], int
> > +dstStride[]) {
> 
> This looks a bit odd, a inline function which only use is taking a pointer to it and
> returning it

The 'inline' keyword will be removed in next patch. Did I get you correctly? 
And this function is used for verifying CPU SIMD version and some init work.

Thank you,
Ting Fu

> 
> [...]
> --
> Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
> 
> Whats the most studid thing your enemy could do ? Blow himself up Whats the
> most studid thing you could do ? Give up your rights and freedom because your
> enemy blew himself up.



More information about the ffmpeg-devel mailing list