unsigned long
dev16_altivec(const vector unsigned char *cur,
unsigned long stride)
{
vector unsigned char t2, t3, t4, mn;
vector unsigned int mean, dev;
vector signed int sumdiffs;
vector unsigned char c0, c1, c2, c3, c4, c5, c6, c7, c8, c9, c10, c11, c12,
c13, c14, c15;
unsigned long result;
ZERODEF;
mean = (vector unsigned int) (ZEROVEC);
dev = (vector unsigned int) (ZEROVEC);
stride >>= 4;
MEAN16(0);
MEAN16(1);
MEAN16(2);
MEAN16(3);
MEAN16(4);
MEAN16(5);
MEAN16(6);
MEAN16(7);
MEAN16(8);
MEAN16(9);
MEAN16(10);
MEAN16(11);
MEAN16(12);
MEAN16(13);
MEAN16(14);
MEAN16(15);
sumdiffs = vec_sums((vector signed int) mean, (vector signed int) ZEROVEC);
mn = vec_perm((vector unsigned char) sumdiffs,
(vector unsigned char) sumdiffs, (vector unsigned char) (14,
14,
14,
14,
14,
14,
14,
14,
14,
14,
14,
14,
14,
14,
14,
14));
DEV16(0);
DEV16(1);
DEV16(2);
DEV16(3);
DEV16(4);
DEV16(5);
DEV16(6);
DEV16(7);
DEV16(8);
DEV16(9);
DEV16(10);
DEV16(11);
DEV16(12);
DEV16(13);
DEV16(14);
DEV16(15);
/* sum all parts of difference into one 32 bit quantity */
sumdiffs = vec_sums((vector signed int) dev, (vector signed int) ZEROVEC);
/* copy vector sum into unaligned result */
sumdiffs = vec_splat(sumdiffs, 3);
vec_ste(sumdiffs, 0, (int *) &result);
return (result);
}
uint32_t
dev16_altivec_c(vector unsigned char *cur,
uint32_t stride)
{
vector unsigned char t2, t3, mn;
vector unsigned int mean, dev;
vector unsigned int sumdiffs;
vector unsigned char *ptr;
uint32_t result;
#ifdef DEBUG
/* print alignment errors if DEBUG is on */
if(((unsigned long)cur) & 0x7)
fprintf(stderr, "dev16_altivec:incorrect align, cur: %lx\n", (long)cur);
if(stride & 0xf)
fprintf(stderr, "dev16_altivec:incorrect align, stride: %lu\n", stride);
#endif
dev = mean = vec_splat_u32(0);
stride >>= 4;
/* set pointer to iterate through cur */
ptr = cur;
MEAN16();
MEAN16();
MEAN16();
MEAN16();
MEAN16();
MEAN16();
MEAN16();
MEAN16();
MEAN16();
MEAN16();
MEAN16();
MEAN16();
MEAN16();
MEAN16();
MEAN16();
MEAN16();
/* Add all together in sumdiffs */
sumdiffs = (vector unsigned int)vec_sums((vector signed int) mean, vec_splat_s32(0));
/* teilen durch 16 * 16 */
mn = vec_perm((vector unsigned char)sumdiffs, (vector unsigned char)sumdiffs, vec_splat_u8(14));
/* set pointer to iterate through cur */
ptr = cur;
DEV16();
DEV16();
DEV16();
DEV16();
DEV16();
DEV16();
DEV16();
DEV16();
DEV16();
DEV16();
DEV16();
DEV16();
DEV16();
DEV16();
DEV16();
DEV16();
/* sum all parts of difference into one 32 bit quantity */
sumdiffs = (vector unsigned int)vec_sums((vector signed int) dev, vec_splat_s32(0));
/* copy vector sum into unaligned result */
sumdiffs = vec_splat(sumdiffs, 3);
vec_ste(sumdiffs, 0, (uint32_t*) &result);
return result;
}
