comparison ppc/mpegvideo_altivec.c @ 10079:71ead14665e3 libavcodec

PPC: simplify loading some values into altivec registers Instead of filling a local array with the desired value and loading it, load a single element and vec_splat() it to fill the vector.

10078:57f034d80624

#include "libavcodec/dsputil.h"
#include "libavcodec/mpegvideo.h"

#include "dsputil_ppc.h"
#include "util_altivec.h"
// Swaps two variables (used for altivec registers)
#define SWAP(a,b) \
do { \
    __typeof__(a) swap_temp=a; \
    a=b; \

        nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
    }

    {
        register const vector signed short vczero = (const vector signed short)vec_splat_s16(0);
        DECLARE_ALIGNED_16(short, qmul8[]) =
            {
              qmul, qmul, qmul, qmul,
              qmul, qmul, qmul, qmul
            };
        DECLARE_ALIGNED_16(short, qadd8[]) =
            {
              qadd, qadd, qadd, qadd,
              qadd, qadd, qadd, qadd
            };
        DECLARE_ALIGNED_16(short, nqadd8[]) =
            {
              -qadd, -qadd, -qadd, -qadd,
              -qadd, -qadd, -qadd, -qadd
            };
        register vector signed short blockv, qmulv, qaddv, nqaddv, temp1;
        register vector bool short blockv_null, blockv_neg;
        register short backup_0 = block[0];
        register int j = 0;

        qmulv = vec_ld(0, qmul8);
        qaddv = vec_ld(0, qadd8);
        nqaddv = vec_ld(0, nqadd8);

#if 0   // block *is* 16 bytes-aligned, it seems.
        // first make sure block[j] is 16 bytes-aligned
        for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) {
            level = block[j];

10079:71ead14665e3

#include "libavcodec/dsputil.h"
#include "libavcodec/mpegvideo.h"

#include "dsputil_ppc.h"
#include "util_altivec.h"
#include "types_altivec.h"

// Swaps two variables (used for altivec registers)
#define SWAP(a,b) \
do { \
    __typeof__(a) swap_temp=a; \
    a=b; \

        nCoeffs= s->intra_scantable.raster_end[ s->block_last_index[n] ];
    }

    {
        register const vector signed short vczero = (const vector signed short)vec_splat_s16(0);
        DECLARE_ALIGNED_16(short, qmul8) = qmul;
        DECLARE_ALIGNED_16(short, qadd8) = qadd;
        register vector signed short blockv, qmulv, qaddv, nqaddv, temp1;
        register vector bool short blockv_null, blockv_neg;
        register short backup_0 = block[0];
        register int j = 0;

        qmulv = vec_splat((vec_s16)vec_lde(0, &qmul8), 0);
        qaddv = vec_splat((vec_s16)vec_lde(0, &qadd8), 0);
        nqaddv = vec_sub(vczero, qaddv);

#if 0   // block *is* 16 bytes-aligned, it seems.
        // first make sure block[j] is 16 bytes-aligned
        for(j = 0; (j <= nCoeffs) && ((((unsigned long)block) + (j << 1)) & 0x0000000F) ; j++) {
            level = block[j];