[FFmpeg-devel] [PATCH] PPC64: Add IBM POWER8 SIMD Implementation
Michael Niedermayer
michael at niedermayer.cc
Tue Jun 21 01:06:39 CEST 2016
On Mon, Jun 20, 2016 at 05:55:47PM -0500, Dan Parrot wrote:
> On Tue, 2016-06-21 at 00:45 +0200, Michael Niedermayer wrote:
> > On Sun, Jun 19, 2016 at 09:57:42PM +0000, Dan Parrot wrote:
> > > First commit addressing Trac ticket #5570. Functions defined in libswscale/input.c
> > > have corresponding SIMD definitions in libswscale/ppc/input_vsx.c
> > > ---
> > > libswscale/ppc/Makefile | 1 +
> > > libswscale/ppc/input_vsx.c | 1070 +++++++++++++++++++++++++++++++++++++++++
> > > libswscale/swscale.c | 3 +
> > > libswscale/swscale_internal.h | 1 +
> > > 4 files changed, 1075 insertions(+)
> > > create mode 100644 libswscale/ppc/input_vsx.c
> > >
> > > diff --git a/libswscale/ppc/Makefile b/libswscale/ppc/Makefile
> > > index d1b596e..2482893 100644
> > > --- a/libswscale/ppc/Makefile
> > > +++ b/libswscale/ppc/Makefile
> > > @@ -1,3 +1,4 @@
> > > OBJS += ppc/swscale_altivec.o \
> > > + ppc/input_vsx.o \
> > > ppc/yuv2rgb_altivec.o \
> > > ppc/yuv2yuv_altivec.o \
> > > diff --git a/libswscale/ppc/input_vsx.c b/libswscale/ppc/input_vsx.c
> > > new file mode 100644
> > > index 0000000..adb0e38
> > > --- /dev/null
> > > +++ b/libswscale/ppc/input_vsx.c
> > > @@ -0,0 +1,1070 @@
> > > +/*
> > > + * Copyright (C) 2016 Dan Parrot <dan.parrot at mail.com>
> > > + *
> > > + * 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
> > > + */
> > > +
> > > +#include <math.h>
> > > +#include <stdint.h>
> > > +#include <stdio.h>
> > > +#include <string.h>
> > > +
> > > +#include "libavutil/avutil.h"
> > > +#include "libavutil/bswap.h"
> > > +#include "libavutil/cpu.h"
> > > +#include "libavutil/intreadwrite.h"
> > > +#include "libavutil/mathematics.h"
> > > +#include "libavutil/pixdesc.h"
> > > +#include "libavutil/avassert.h"
> > > +#include "config.h"
> > > +#include "libswscale/rgb2rgb.h"
> > > +#include "libswscale/swscale.h"
> > > +#include "libswscale/swscale_internal.h"
> > > +
> > > +#define input_pixel(pos) (isBE(origin) ? AV_RB16(pos) : AV_RL16(pos))
> > > +
> > > +#define r ((origin == AV_PIX_FMT_BGR48BE || origin == AV_PIX_FMT_BGR48LE || origin == AV_PIX_FMT_BGRA64BE || origin == AV_PIX_FMT_BGRA64LE) ? b_r : r_b)
> > > +#define b ((origin == AV_PIX_FMT_BGR48BE || origin == AV_PIX_FMT_BGR48LE || origin == AV_PIX_FMT_BGRA64BE || origin == AV_PIX_FMT_BGRA64LE) ? r_b : b_r)
> > > +
> > > +#if HAVE_VSX
> > > +
> > > +// This is a SIMD version for IBM POWER8 of function rgb64ToY_c_template
> > > +// in file libswscale/input.c
> > > +static av_always_inline void
> > > +rgb64ToY_c_template_vsx(uint16_t *dst, const uint16_t *src, int width,
> > > + enum AVPixelFormat origin, int32_t *rgb2yuv)
> > > +{
> > > + int32_t ry = rgb2yuv[RY_IDX], gy = rgb2yuv[GY_IDX], by = rgb2yuv[BY_IDX];
> > > + int i, j;
> > > + int num_vec, frag;
> > > +
> > > + num_vec = width / 8;
> > > + frag = width % 8;
> > > +
> > > + vector int v_ry = vec_splats((int)ry);
> > > + vector int v_gy = vec_splats((int)gy);
> > > + vector int v_by = vec_splats((int)by);
> > > +
> > > + int s_opr2;
> > > + s_opr2 = (int)(0x2001 << (RGB2YUV_SHIFT-1));
> > > +
> > > + vector int v_opr1 = vec_splats((int)RGB2YUV_SHIFT);
> > > + vector int v_opr2 = vec_splats((int)s_opr2);
> > > +
> > > + vector int v_r, v_g, v_b, v_tmp;
> > > + vector short v_tmpi, v_dst;
> > > +
> > > + for (i = 0; i < num_vec; i++) {
> > > + for (j = 7; j >= 0 ; j--) {
> > > + int r_b = input_pixel(&src[(i*8+j)*4+0]);
> > > + int g = input_pixel(&src[(i*8+j)*4+1]);
> > > + int b_r = input_pixel(&src[(i*8+j)*4+2]);
> > > +
> > > + v_r[j % 4] = r;
> > > + v_g[j % 4] = g;
> > > + v_b[j % 4] = b;
> > > +
> > > + if (!(j % 4)) {
> > ^
> >
> > > + v_tmp = v_ry * v_r;
> > > + v_tmp = v_tmp + v_gy * v_g;
> > > + v_tmp = v_tmp + v_by * v_b;
> > > + v_tmp = v_tmp + v_opr2;
> > > + v_tmp = vec_sr(v_tmp, (vector unsigned int)v_opr1);
> > > +
> > > + v_tmpi = (vector short)v_tmp;
> > > + v_dst[(j / 4) * 4 + 3] = v_tmpi[6];
> > ^
> > what is the speed of a division and modulo on PPC compared to a
> > bitwise and ?
> >
> > its also not trivial for the compiler to optimize then out as it
> > has to proof the varables are never negative
> >
> >
> > [...]
> > _______________________________________________
> > ffmpeg-devel mailing list
> > ffmpeg-devel at ffmpeg.org
> > http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
> I don't know the answer to those questions. But, I see your point.
> Bitwise operations should be faster than multiplications and divisions.
> So, I shall change the code to use bitwise ops and compare the execution
> time against its present value (well, average value over multiple runs)
it might not make a difference if the compiler does optmize them out
but you should know the approximate speed of the instructions
that your code is likely/potentially going to use. I mean this is
code optimized for PPC.
[...]
--
Michael GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
Its not that you shouldnt use gotos but rather that you should write
readable code and code with gotos often but not always is less readable
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 181 bytes
Desc: Digital signature
URL: <http://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20160621/834cdee2/attachment.sig>
More information about the ffmpeg-devel
mailing list