[Ffmpeg-devel] [RFC] AltiVec optimizations, try 2

Kostya kostya.shishkov
Wed Aug 2 06:01:26 CEST 2006


On Tue, Aug 01, 2006 at 10:04:45AM +0200, Luca Barbato wrote:
> Kostya wrote:
> > Here is my second attempt of writing optimized code. 
> > Please test how it works on your Macs (my rough test show ~6% speedup).
> > If there are no objections I'll commit it tomorrow.
> > 
> > 
> > ------------------------------------------------------------------------
> > 
> > diff -ru --exclude .svn ffmpeg/libavcodec/Makefile ffmpeg-vc1/libavcodec/Makefile
> 
> svn diff ?
> 
> > diff -ru --exclude .svn ffmpeg/libavcodec/ppc/dsputil_ppc.c ffmpeg-vc1/libavcodec/ppc/dsputil_ppc.c
> > --- ffmpeg/libavcodec/ppc/dsputil_ppc.c	2006-07-27 18:22:09.000000000 +0300
> > +++ ffmpeg-vc1/libavcodec/ppc/dsputil_ppc.c	2006-07-31 11:51:04.000000000 +0300
> > @@ -251,6 +251,10 @@
> >  
> >  void dsputil_h264_init_ppc(DSPContext* c, AVCodecContext *avctx);
> >  
> > +#ifdef HAVE_ALTIVEC
> > +void vc1dsp_init_altivec(DSPContext* c, AVCodecContext *avctx);
> > +#endif
> > +
> 
> I think we should reorder the init code to be more rational, I'll do
> something about it soonish
> 

If it will take more than a couple of days than I commit (modified) version. Otherwise I can wait.

> > +/* constants used in transform */
> > +static const vector int vec_64 = (vector int)64;
> > +static const vector int vec_7 = (vector int)7;
> > +static const vector int vec_5 = (vector int)5;
> > +static const vector int vec_4 = (vector int)4;
> > +static const vector int vec_3 = (vector int)3;
> > +static const vector int vec_2 = (vector int)2;
> > +static const vector int vec_1 = (vector int)1;
> 
> define it with the vec_splat_s32(value) and
> vec_sl(vec_splat_s32(4),vec_splat(4)) for 64

Here is single .c file attached for review. Now it compiles with GCC 3.3

> 
> Not sure about that many pack and unpack, Michael maybe has already a
> recipe to avoid some of them.

Yes, but that will be future optimizations...
> 
> lu
> 
> -- 
> 
> Luca Barbato
> 
> Gentoo/linux Gentoo/PPC
> http://dev.gentoo.org/~lu_zero
> 
> 
-------------- next part --------------
/*
 * VC-1 and WMV3 decoder - DSP functions AltiVec-optimized
 * Copyright (c) 2006 Konstantin Shishkov
 *
 * This library 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 of the License, or (at your option) any later version.
 *
 * This library 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 this library; if not, write to the Free Software
 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
 *
 */

#include "../dsputil.h"

#include "gcc_fixes.h"

#include "dsputil_altivec.h"

// Transpose 8x8 matrix of 16-bit elements. Borrowed from mpegvideo_altivec.c
#define TRANSPOSE8(a,b,c,d,e,f,g,h) \
do { \
    vector signed short A1, B1, C1, D1, E1, F1, G1, H1; \
    vector signed short A2, B2, C2, D2, E2, F2, G2, H2; \
 \
    A1 = vec_mergeh (a, e); \
    B1 = vec_mergel (a, e); \
    C1 = vec_mergeh (b, f); \
    D1 = vec_mergel (b, f); \
    E1 = vec_mergeh (c, g); \
    F1 = vec_mergel (c, g); \
    G1 = vec_mergeh (d, h); \
    H1 = vec_mergel (d, h); \
 \
    A2 = vec_mergeh (A1, E1); \
    B2 = vec_mergel (A1, E1); \
    C2 = vec_mergeh (B1, F1); \
    D2 = vec_mergel (B1, F1); \
    E2 = vec_mergeh (C1, G1); \
    F2 = vec_mergel (C1, G1); \
    G2 = vec_mergeh (D1, H1); \
    H2 = vec_mergel (D1, H1); \
 \
    a = vec_mergeh (A2, E2); \
    b = vec_mergel (A2, E2); \
    c = vec_mergeh (B2, F2); \
    d = vec_mergel (B2, F2); \
    e = vec_mergeh (C2, G2); \
    f = vec_mergel (C2, G2); \
    g = vec_mergeh (D2, H2); \
    h = vec_mergel (D2, H2); \
} while (0)

// main steps of 8x8 transform
#define STEP8(s0, s1, s2, s3, s4, s5, s6, s7, vec_rnd) \
do { \
    t0 = vec_sl(vec_add(s0, s4), vec_2); \
    t0 = vec_add(vec_sl(t0, vec_1), t0); \
    t0 = vec_add(t0, vec_rnd); \
    t1 = vec_sl(vec_sub(s0, s4), vec_2); \
    t1 = vec_add(vec_sl(t1, vec_1), t1); \
    t1 = vec_add(t1, vec_rnd); \
    t2 = vec_add(vec_sl(s6, vec_2), vec_sl(s6, vec_1)); \
    t2 = vec_add(t2, vec_sl(s2, vec_4)); \
    t3 = vec_add(vec_sl(s2, vec_2), vec_sl(s2, vec_1)); \
    t3 = vec_sub(t3, vec_sl(s6, vec_4)); \
    t4 = vec_add(t0, t2); \
    t5 = vec_add(t1, t3); \
    t6 = vec_sub(t1, t3); \
    t7 = vec_sub(t0, t2); \
\
    t0 = vec_sl(vec_add(s1, s3), vec_4); \
    t0 = vec_add(t0, vec_sl(s5, vec_3)); \
    t0 = vec_add(t0, vec_sl(s7, vec_2)); \
    t0 = vec_add(t0, vec_sub(s5, s3)); \
\
    t1 = vec_sl(vec_sub(s1, s5), vec_4); \
    t1 = vec_sub(t1, vec_sl(s7, vec_3)); \
    t1 = vec_sub(t1, vec_sl(s3, vec_2)); \
    t1 = vec_sub(t1, vec_add(s1, s7)); \
\
    t2 = vec_sl(vec_sub(s7, s3), vec_4); \
    t2 = vec_add(t2, vec_sl(s1, vec_3)); \
    t2 = vec_add(t2, vec_sl(s5, vec_2)); \
    t2 = vec_add(t2, vec_sub(s1, s7)); \
\
    t3 = vec_sl(vec_sub(s5, s7), vec_4); \
    t3 = vec_sub(t3, vec_sl(s3, vec_3)); \
    t3 = vec_add(t3, vec_sl(s1, vec_2)); \
    t3 = vec_sub(t3, vec_add(s3, s5)); \
\
    s0 = vec_add(t4, t0); \
    s1 = vec_add(t5, t1); \
    s2 = vec_add(t6, t2); \
    s3 = vec_add(t7, t3); \
    s4 = vec_sub(t7, t3); \
    s5 = vec_sub(t6, t2); \
    s6 = vec_sub(t5, t1); \
    s7 = vec_sub(t4, t0); \
}while(0)

#define SHIFT_HOR8(s0, s1, s2, s3, s4, s5, s6, s7) \
do { \
    s0 = vec_sra(s0, vec_3); \
    s1 = vec_sra(s1, vec_3); \
    s2 = vec_sra(s2, vec_3); \
    s3 = vec_sra(s3, vec_3); \
    s4 = vec_sra(s4, vec_3); \
    s5 = vec_sra(s5, vec_3); \
    s6 = vec_sra(s6, vec_3); \
    s7 = vec_sra(s7, vec_3); \
}while(0)

#define SHIFT_VERT8(s0, s1, s2, s3, s4, s5, s6, s7) \
do { \
    s0 = vec_sra(s0, vec_7); \
    s1 = vec_sra(s1, vec_7); \
    s2 = vec_sra(s2, vec_7); \
    s3 = vec_sra(s3, vec_7); \
    s4 = vec_sra(vec_add(s4, vec_1s), vec_7); \
    s5 = vec_sra(vec_add(s5, vec_1s), vec_7); \
    s6 = vec_sra(vec_add(s6, vec_1s), vec_7); \
    s7 = vec_sra(vec_add(s7, vec_1s), vec_7); \
}while(0)

/* main steps of 4x4 transform */
#define STEP4(s0, s1, s2, s3, vec_rnd) \
do { \
    t1 = vec_add(vec_sl(s0, vec_4), s0); \
    t1 = vec_add(t1, vec_rnd); \
    t2 = vec_add(vec_sl(s2, vec_4), s2); \
    t0 = vec_add(t1, t2); \
    t1 = vec_sub(t1, t2); \
    t3 = vec_sl(vec_sub(s3, s1), vec_1); \
    t3 = vec_add(t3, vec_sl(t3, vec_2)); \
    t2 = vec_add(t3, vec_sl(s1, vec_5)); \
    t3 = vec_add(t3, vec_sl(s3, vec_3)); \
    t3 = vec_add(t3, vec_sl(s3, vec_2)); \
    s0 = vec_add(t0, t2); \
    s1 = vec_sub(t1, t3); \
    s2 = vec_add(t1, t3); \
    s3 = vec_sub(t0, t2); \
}while (0)

#define SHIFT_HOR4(s0, s1, s2, s3) \
    s0 = vec_sra(s0, vec_3); \
    s1 = vec_sra(s1, vec_3); \
    s2 = vec_sra(s2, vec_3); \
    s3 = vec_sra(s3, vec_3);

#define SHIFT_VERT4(s0, s1, s2, s3) \
    s0 = vec_sra(s0, vec_7); \
    s1 = vec_sra(s1, vec_7); \
    s2 = vec_sra(s2, vec_7); \
    s3 = vec_sra(s3, vec_7);

/** Do inverse transform on 8x8 block
*/
static void vc1_inv_trans_8x8_altivec(DCTELEM block[64])
{
    vector signed short src0, src1, src2, src3, src4, src5, src6, src7;
    vector signed int s0, s1, s2, s3, s4, s5, s6, s7;
    vector signed int s8, s9, sA, sB, sC, sD, sE, sF;
    vector signed int t0, t1, t2, t3, t4, t5, t6, t7;
    const vector signed int vec_64 = vec_sl(vec_splat_s32(4), vec_splat_u32(4));
    const vector unsigned int vec_7 = vec_splat_u32(7);
    const vector unsigned int vec_5 = vec_splat_u32(5);
    const vector unsigned int vec_4 = vec_splat_u32(4);
    const vector  signed int vec_4s = vec_splat_s32(4);
    const vector unsigned int vec_3 = vec_splat_u32(3);
    const vector unsigned int vec_2 = vec_splat_u32(2);
    const vector  signed int vec_1s = vec_splat_s32(1);
    const vector unsigned int vec_1 = vec_splat_u32(1);


    src0 = vec_ld(  0, block);
    src1 = vec_ld( 16, block);
    src2 = vec_ld( 32, block);
    src3 = vec_ld( 48, block);
    src4 = vec_ld( 64, block);
    src5 = vec_ld( 80, block);
    src6 = vec_ld( 96, block);
    src7 = vec_ld(112, block);

    TRANSPOSE8(src0, src1, src2, src3, src4, src5, src6, src7);
    s0 = vec_unpackl(src0);
    s1 = vec_unpackl(src1);
    s2 = vec_unpackl(src2);
    s3 = vec_unpackl(src3);
    s4 = vec_unpackl(src4);
    s5 = vec_unpackl(src5);
    s6 = vec_unpackl(src6);
    s7 = vec_unpackl(src7);
    s8 = vec_unpackh(src0);
    s9 = vec_unpackh(src1);
    sA = vec_unpackh(src2);
    sB = vec_unpackh(src3);
    sC = vec_unpackh(src4);
    sD = vec_unpackh(src5);
    sE = vec_unpackh(src6);
    sF = vec_unpackh(src7);
    STEP8(s0, s1, s2, s3, s4, s5, s6, s7, vec_4s);
    SHIFT_HOR8(s0, s1, s2, s3, s4, s5, s6, s7);
    STEP8(s8, s9, sA, sB, sC, sD, sE, sF, vec_4s);
    SHIFT_HOR8(s8, s9, sA, sB, sC, sD, sE, sF);
    src0 = vec_pack(s8, s0);
    src1 = vec_pack(s9, s1);
    src2 = vec_pack(sA, s2);
    src3 = vec_pack(sB, s3);
    src4 = vec_pack(sC, s4);
    src5 = vec_pack(sD, s5);
    src6 = vec_pack(sE, s6);
    src7 = vec_pack(sF, s7);
    TRANSPOSE8(src0, src1, src2, src3, src4, src5, src6, src7);

    s0 = vec_unpackl(src0);
    s1 = vec_unpackl(src1);
    s2 = vec_unpackl(src2);
    s3 = vec_unpackl(src3);
    s4 = vec_unpackl(src4);
    s5 = vec_unpackl(src5);
    s6 = vec_unpackl(src6);
    s7 = vec_unpackl(src7);
    s8 = vec_unpackh(src0);
    s9 = vec_unpackh(src1);
    sA = vec_unpackh(src2);
    sB = vec_unpackh(src3);
    sC = vec_unpackh(src4);
    sD = vec_unpackh(src5);
    sE = vec_unpackh(src6);
    sF = vec_unpackh(src7);
    STEP8(s0, s1, s2, s3, s4, s5, s6, s7, vec_64);
    SHIFT_VERT8(s0, s1, s2, s3, s4, s5, s6, s7);
    STEP8(s8, s9, sA, sB, sC, sD, sE, sF, vec_64);
    SHIFT_VERT8(s8, s9, sA, sB, sC, sD, sE, sF);
    src0 = vec_pack(s8, s0);
    src1 = vec_pack(s9, s1);
    src2 = vec_pack(sA, s2);
    src3 = vec_pack(sB, s3);
    src4 = vec_pack(sC, s4);
    src5 = vec_pack(sD, s5);
    src6 = vec_pack(sE, s6);
    src7 = vec_pack(sF, s7);

    vec_st(src0,  0, block);
    vec_st(src1, 16, block);
    vec_st(src2, 32, block);
    vec_st(src3, 48, block);
    vec_st(src4, 64, block);
    vec_st(src5, 80, block);
    vec_st(src6, 96, block);
    vec_st(src7,112, block);
}

/** Do inverse transform on 8x4 part of block
*/
static void vc1_inv_trans_8x4_altivec(DCTELEM block[64], int n)
{
    vector signed short src0, src1, src2, src3, src4, src5, src6, src7;
    vector signed int s0, s1, s2, s3, s4, s5, s6, s7;
    vector signed int s8, s9, sA, sB, sC, sD, sE, sF;
    vector signed int t0, t1, t2, t3, t4, t5, t6, t7;
    const vector signed int vec_64 = vec_sl(vec_splat_s32(4), vec_splat_u32(4));
    const vector unsigned int vec_7 = vec_splat_u32(7);
    const vector unsigned int vec_5 = vec_splat_u32(5);
    const vector unsigned int vec_4 = vec_splat_u32(4);
    const vector  signed int vec_4s = vec_splat_s32(4);
    const vector unsigned int vec_3 = vec_splat_u32(3);
    const vector unsigned int vec_2 = vec_splat_u32(2);
    const vector unsigned int vec_1 = vec_splat_u32(1);

    src0 = vec_ld(  0, block);
    src1 = vec_ld( 16, block);
    src2 = vec_ld( 32, block);
    src3 = vec_ld( 48, block);
    src4 = vec_ld( 64, block);
    src5 = vec_ld( 80, block);
    src6 = vec_ld( 96, block);
    src7 = vec_ld(112, block);

    TRANSPOSE8(src0, src1, src2, src3, src4, src5, src6, src7);
    s0 = vec_unpackl(src0);
    s1 = vec_unpackl(src1);
    s2 = vec_unpackl(src2);
    s3 = vec_unpackl(src3);
    s4 = vec_unpackl(src4);
    s5 = vec_unpackl(src5);
    s6 = vec_unpackl(src6);
    s7 = vec_unpackl(src7);
    s8 = vec_unpackh(src0);
    s9 = vec_unpackh(src1);
    sA = vec_unpackh(src2);
    sB = vec_unpackh(src3);
    sC = vec_unpackh(src4);
    sD = vec_unpackh(src5);
    sE = vec_unpackh(src6);
    sF = vec_unpackh(src7);
    STEP8(s0, s1, s2, s3, s4, s5, s6, s7, vec_4s);
    SHIFT_HOR8(s0, s1, s2, s3, s4, s5, s6, s7);
    STEP8(s8, s9, sA, sB, sC, sD, sE, sF, vec_4s);
    SHIFT_HOR8(s8, s9, sA, sB, sC, sD, sE, sF);
    src0 = vec_pack(s8, s0);
    src1 = vec_pack(s9, s1);
    src2 = vec_pack(sA, s2);
    src3 = vec_pack(sB, s3);
    src4 = vec_pack(sC, s4);
    src5 = vec_pack(sD, s5);
    src6 = vec_pack(sE, s6);
    src7 = vec_pack(sF, s7);
    TRANSPOSE8(src0, src1, src2, src3, src4, src5, src6, src7);

    if(!n){ // upper half of block
        s0 = vec_unpackh(src0);
        s1 = vec_unpackh(src1);
        s2 = vec_unpackh(src2);
        s3 = vec_unpackh(src3);
        s8 = vec_unpackl(src0);
        s9 = vec_unpackl(src1);
        sA = vec_unpackl(src2);
        sB = vec_unpackl(src3);
        STEP4(s0, s1, s2, s3, vec_64);
        SHIFT_VERT4(s0, s1, s2, s3);
        STEP4(s8, s9, sA, sB, vec_64);
        SHIFT_VERT4(s8, s9, sA, sB);
        src0 = vec_pack(s0, s8);
        src1 = vec_pack(s1, s9);
        src2 = vec_pack(s2, sA);
        src3 = vec_pack(s3, sB);

        vec_st(src0,  0, block);
        vec_st(src1, 16, block);
        vec_st(src2, 32, block);
        vec_st(src3, 48, block);
    } else { //lower half of block
        s0 = vec_unpackh(src4);
        s1 = vec_unpackh(src5);
        s2 = vec_unpackh(src6);
        s3 = vec_unpackh(src7);
        s8 = vec_unpackl(src4);
        s9 = vec_unpackl(src5);
        sA = vec_unpackl(src6);
        sB = vec_unpackl(src7);
        STEP4(s0, s1, s2, s3, vec_64);
        SHIFT_VERT4(s0, s1, s2, s3);
        STEP4(s8, s9, sA, sB, vec_64);
        SHIFT_VERT4(s8, s9, sA, sB);
        src4 = vec_pack(s0, s8);
        src5 = vec_pack(s1, s9);
        src6 = vec_pack(s2, sA);
        src7 = vec_pack(s3, sB);

        vec_st(src4, 64, block);
        vec_st(src5, 80, block);
        vec_st(src6, 96, block);
        vec_st(src7,112, block);
    }
}


void vc1dsp_init_altivec(DSPContext* dsp, AVCodecContext *avctx) {
    dsp->vc1_inv_trans_8x8 = vc1_inv_trans_8x8_altivec;
    dsp->vc1_inv_trans_8x4 = vc1_inv_trans_8x4_altivec;
}



More information about the ffmpeg-devel mailing list