comparison ppc/gmc_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 "dsputil_ppc.h"
#include "util_altivec.h"

/*
  altivec-enhanced gmc1. ATM this code assume stride is a multiple of 8,
  to preserve proper dst alignment.
*/
#define GMC1_PERF_COND (h==8)
void gmc1_altivec(uint8_t *dst /* align 8 */, uint8_t *src /* align1 */, int stride, int h, int x16, int y16, int rounder)
{
POWERPC_PERF_DECLARE(altivec_gmc1_num, GMC1_PERF_COND);
    const DECLARE_ALIGNED_16(unsigned short, rounder_a[8]) =
        {rounder, rounder, rounder, rounder,
         rounder, rounder, rounder, rounder};
    const DECLARE_ALIGNED_16(unsigned short, ABCD[8]) =
        {
            (16-x16)*(16-y16), /* A */
            (   x16)*(16-y16), /* B */
            (16-x16)*(   y16), /* C */

    Av = vec_splat(tempA, 0);
    Bv = vec_splat(tempA, 1);
    Cv = vec_splat(tempA, 2);
    Dv = vec_splat(tempA, 3);

    rounderV = vec_ld(0, (unsigned short*)rounder_a);

    // we'll be able to pick-up our 9 char elements
    // at src from those 32 bytes
    // we load the first batch here, as inside the loop
    // we can re-use 'src+stride' from one iteration

10079:71ead14665e3

 */

#include "libavcodec/dsputil.h"
#include "dsputil_ppc.h"
#include "util_altivec.h"
#include "types_altivec.h"

/*
  altivec-enhanced gmc1. ATM this code assume stride is a multiple of 8,
  to preserve proper dst alignment.
*/
#define GMC1_PERF_COND (h==8)
void gmc1_altivec(uint8_t *dst /* align 8 */, uint8_t *src /* align1 */, int stride, int h, int x16, int y16, int rounder)
{
POWERPC_PERF_DECLARE(altivec_gmc1_num, GMC1_PERF_COND);
    const DECLARE_ALIGNED_16(unsigned short, rounder_a) = rounder;
    const DECLARE_ALIGNED_16(unsigned short, ABCD[8]) =
        {
            (16-x16)*(16-y16), /* A */
            (   x16)*(16-y16), /* B */
            (16-x16)*(   y16), /* C */

    Av = vec_splat(tempA, 0);
    Bv = vec_splat(tempA, 1);
    Cv = vec_splat(tempA, 2);
    Dv = vec_splat(tempA, 3);

    rounderV = vec_splat((vec_u16)vec_lde(0, &rounder_a), 0);

    // we'll be able to pick-up our 9 char elements
    // at src from those 32 bytes
    // we load the first batch here, as inside the loop
    // we can re-use 'src+stride' from one iteration