[FFmpeg-devel] [RFC] RGB32 / BGR32 ethernal bug
Stefano Sabatini
stefano.sabatini-lala
Wed Feb 3 20:50:53 CET 2010
On date Tuesday 2010-02-02 21:37:31 +0100, Michael Niedermayer encoded:
> On Sun, Jan 31, 2010 at 01:22:53PM +0100, Stefano Sabatini wrote:
> > On date Sunday 2010-01-31 02:06:12 +0100, Michael Niedermayer encoded:
> > > On Sun, Jan 31, 2010 at 01:48:23AM +0100, Stefano Sabatini wrote:
> > > > On date Tuesday 2010-01-26 09:17:12 +0100, Michael Niedermayer encoded:
> > [...]
> > > > > you write byte shuffling code, id suggest to call it accordingly
> > > > > theres nothing rgb specific on this, it could be usefull to do packed yuv
> > > > > shuffling as well.
> > > > >
> > > > > like maybe:
> > > > > byte_shuffle_0123_to_0321
> > > > > but before you convert your code to this, what is the speed gain from
> > > > > a/b/c/d being constants instead of just passed into the function?
> > > > > there should be enough registers for this in theory
> > > > > also
> > > > > dst+=4; src+=4
> > > > > for(dst<dstend
> > > > > would reduce the register need
> > > >
> > > > Patch+diff of the regression test attached, maybe rgb2rgb.{h,c} is not
> > > > the right place for the shuffle functions and you can suggest a better
> > > > place.
> > >
> > > it can be moved later
> > >
> > >
> > > >
> > > > I have yet to try/benchmark with a generic shuffle_byte_4() function.
> > >
> > > then ill wait with a review
> >
> > Attached patches with the two different approaches, code with ad-hoc
> > custom shuffle functions and with the generic shuffle function
> > shuffle_bytes_abcd, first solution looks simpler and faster, but I may
> > miss something.
> >
> > Regards.
> > --
> > FFmpeg = Fostering Fostering Miracolous Purposeless Eager Geisha
>
> > rgb2rgb.c | 25 +++++++++++++++++++++++++
> > rgb2rgb.h | 6 ++++++
> > swscale.c | 43 ++++++++++++++++++++++++++++++++++++-------
> > 3 files changed, 67 insertions(+), 7 deletions(-)
> > 175a71db692d965f98bce9a98c7df61277ee653c lsws-fix-rgba32-scaling.patch
> > Index: ffmpeg/libswscale/rgb2rgb.c
> > ===================================================================
> > --- ffmpeg.orig/libswscale/rgb2rgb.c 2010-01-31 01:50:52.000000000 +0100
> > +++ ffmpeg/libswscale/rgb2rgb.c 2010-01-31 11:43:22.000000000 +0100
> > @@ -442,3 +442,28 @@
> > dst[i] = ((b<<1)&0x07) | ((g&0x07)<<3) | ((r&0x03)<<6);
> > }
> > }
> > +
> > +void shuffle_bytes_0123(const uint8_t *src, uint8_t *dst, long src_size)
> > +{
> > + memcpy(dst, src, src_size);
> > +}
> > +
> > +#define DEFINE_SHUFFLE_BYTES(a, b, c, d) \
> > +void shuffle_bytes_##a##b##c##d(const uint8_t *src, uint8_t *dst, long src_size) \
> > +{ \
> > + long i; \
> > + \
> > + for (i = 0; i < src_size; i+=4) { \
> > + dst[i + 0] = src[i + a]; \
> > + dst[i + 1] = src[i + b]; \
> > + dst[i + 2] = src[i + c]; \
> > + dst[i + 3] = src[i + d]; \
> > + } \
> > +}
> > +
> > +DEFINE_SHUFFLE_BYTES(0, 3, 2, 1);
> > +DEFINE_SHUFFLE_BYTES(1, 2, 3, 0);
> > +DEFINE_SHUFFLE_BYTES(2, 1, 0, 3);
> > +DEFINE_SHUFFLE_BYTES(3, 0, 1, 2);
> > +DEFINE_SHUFFLE_BYTES(3, 2, 1, 0);
> > +
> > Index: ffmpeg/libswscale/rgb2rgb.h
> > ===================================================================
> > --- ffmpeg.orig/libswscale/rgb2rgb.h 2010-01-31 01:50:52.000000000 +0100
> > +++ ffmpeg/libswscale/rgb2rgb.h 2010-01-31 11:43:22.000000000 +0100
> > @@ -60,6 +60,12 @@
> > void rgb15tobgr15(const uint8_t *src, uint8_t *dst, long src_size);
> > void bgr8torgb8 (const uint8_t *src, uint8_t *dst, long src_size);
> >
> > +void shuffle_bytes_0123(const uint8_t *src, uint8_t *dst, long src_size);
> > +void shuffle_bytes_0321(const uint8_t *src, uint8_t *dst, long src_size);
> > +void shuffle_bytes_1230(const uint8_t *src, uint8_t *dst, long src_size);
> > +void shuffle_bytes_2103(const uint8_t *src, uint8_t *dst, long src_size);
> > +void shuffle_bytes_3012(const uint8_t *src, uint8_t *dst, long src_size);
> > +void shuffle_bytes_3210(const uint8_t *src, uint8_t *dst, long src_size);
> >
> > void palette8topacked32(const uint8_t *src, uint8_t *dst, long num_pixels, const uint8_t *palette);
> > void palette8topacked24(const uint8_t *src, uint8_t *dst, long num_pixels, const uint8_t *palette);
>
> > Index: ffmpeg/libswscale/swscale.c
> > ===================================================================
> > --- ffmpeg.orig/libswscale/swscale.c 2010-01-31 01:50:52.000000000 +0100
> > +++ ffmpeg/libswscale/swscale.c 2010-01-31 13:18:22.000000000 +0100
> > @@ -66,6 +66,7 @@
> > #include "libavutil/intreadwrite.h"
> > #include "libavutil/x86_cpu.h"
> > #include "libavutil/avutil.h"
> > +#include "libavutil/timer.h"
> > #include "libavutil/bswap.h"
> > #include "libavutil/pixdesc.h"
> >
>
> ehm
Yes I left that to show you how I was benchmarking.
> also the memcpy case should be redundant
Indeed.
Patch and updated diff in attachment.
Regards.
--
FFmpeg = Fabulous Fast Maxi Prodigious Exuberant Game
-------------- next part --------------
A non-text attachment was scrubbed...
Name: lsws-fix-rgba32-scaling.patch
Type: text/x-diff
Size: 5987 bytes
Desc: not available
URL: <http://lists.mplayerhq.hu/pipermail/ffmpeg-devel/attachments/20100203/c9fa0484/attachment.patch>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: swscale.test.diff
Type: text/x-diff
Size: 28261 bytes
Desc: not available
URL: <http://lists.mplayerhq.hu/pipermail/ffmpeg-devel/attachments/20100203/c9fa0484/attachment.diff>
More information about the ffmpeg-devel
mailing list