/*
* This function assumes cur and stride are 16 bytes aligned and ref is unaligned
*/
unsigned long
sad16_altivec(const vector unsigned char *cur,
const vector unsigned char *ref,
unsigned long stride,
const unsigned long best_sad)
{
vector unsigned char perm;
vector unsigned char t1, t2, t3, t4;
vector unsigned int sad;
vector signed int sumdiffs, best_vec;
unsigned long result;
ZERODEF;
#ifdef DEBUG
if (((unsigned long) cur) & 0xf)
fprintf(stderr, "sad16_altivec:incorrect align, cur: %x\n", cur);
// if (((unsigned long)ref) & 0xf)
// fprintf(stderr, "sad16_altivec:incorrect align, ref: %x\n", ref);
if (stride & 0xf)
fprintf(stderr, "sad16_altivec:incorrect align, stride: %x\n", stride);
#endif
/* initialization */
sad = (vector unsigned int) (ZEROVEC);
stride >>= 4;
perm = vec_lvsl(0, (unsigned char *) ref);
*((unsigned long *) &best_vec) = best_sad;
best_vec = vec_splat(best_vec, 0);
/* perform sum of differences between current and previous */
SAD16();
SAD16();
SAD16();
SAD16();
/* Temp sum for exit */
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) ZEROVEC);
if (vec_all_ge(sumdiffs, best_vec))
goto bail;
SAD16();
SAD16();
SAD16();
SAD16();
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) ZEROVEC);
if (vec_all_ge(sumdiffs, best_vec))
goto bail;
SAD16();
SAD16();
SAD16();
SAD16();
SAD16();
SAD16();
SAD16();
SAD16();
/* sum all parts of difference into one 32 bit quantity */
sumdiffs = vec_sums((vector signed int) sad, (vector signed int) ZEROVEC);
bail:
/* copy vector sum into unaligned result */
sumdiffs = vec_splat(sumdiffs, 3);
vec_ste(sumdiffs, 0, (int *) &result);
return (result);
}
uint32_t
sad16_altivec_c(vector unsigned char *cur,
vector unsigned char *ref,
uint32_t stride,
const uint32_t best_sad)
{
vector unsigned char perm;
vector unsigned char t1, t2;
vector unsigned int sad;
vector unsigned int sumdiffs;
vector unsigned int best_vec;
uint32_t result;
#ifdef DEBUG
/* print alignment errors if DEBUG is on */
if (((unsigned long) cur) & 0xf)
fprintf(stderr, "sad16_altivec:incorrect align, cur: %lx\n", (long)cur);
if (stride & 0xf)
fprintf(stderr, "sad16_altivec:incorrect align, stride: %lu\n", stride);
#endif
/* initialization */
sad = vec_splat_u32(0);
sumdiffs = sad;
stride >>= 4;
perm = vec_lvsl(0, (unsigned char *) ref);
*((uint32_t*)&best_vec) = best_sad;
best_vec = vec_splat(best_vec, 0);
/* perform sum of differences between current and previous */
SAD16();
SAD16();
SAD16();
SAD16();
SAD16();
SAD16();
SAD16();
SAD16();
SAD16();
SAD16();
SAD16();
SAD16();
SAD16();
SAD16();
SAD16();
SAD16();
bail:
/* copy vector sum into unaligned result */
sumdiffs = vec_splat(sumdiffs, 3);
vec_ste(sumdiffs, 0, (uint32_t*) &result);
return result;
}
