88 lines
2.4 KiB
C
88 lines
2.4 KiB
C
/* 16-bit signed integer dot product
|
|
* MMX assisted version; also for SSE
|
|
*
|
|
* Copyright 2004 Phil Karn
|
|
* May be used under the terms of the GNU Lesser General Public License (LGPL)
|
|
*/
|
|
#include <stdlib.h>
|
|
#include "fec.h"
|
|
|
|
struct dotprod {
|
|
int len; /* Number of coefficients */
|
|
|
|
/* On a MMX or SSE machine, these hold 4 copies of the coefficients,
|
|
* preshifted by 0,1,2,3 words to meet all possible input data
|
|
* alignments (see Intel ap559 on MMX dot products).
|
|
*/
|
|
signed short *coeffs[4];
|
|
};
|
|
long dotprod_mmx_assist(signed short *a, signed short *b, int cnt);
|
|
|
|
/* Create and return a descriptor for use with the dot product function */
|
|
void *initdp_mmx(signed short coeffs[], int len)
|
|
{
|
|
struct dotprod *dp;
|
|
int i, j;
|
|
|
|
|
|
if (len == 0) {
|
|
return NULL;
|
|
}
|
|
|
|
dp = (struct dotprod *)calloc(1, sizeof(struct dotprod));
|
|
dp->len = len;
|
|
|
|
/* Make 4 copies of coefficients, one for each data alignment */
|
|
for (i = 0; i < 4; i++) {
|
|
dp->coeffs[i] = (signed short *)calloc(1 + (len + i - 1) / 4,
|
|
4 * sizeof(signed short));
|
|
for (j = 0; j < len; j++) {
|
|
dp->coeffs[i][j + i] = coeffs[j];
|
|
}
|
|
}
|
|
return (void *)dp;
|
|
}
|
|
|
|
|
|
/* Free a dot product descriptor created earlier */
|
|
void freedp_mmx(void *p)
|
|
{
|
|
struct dotprod *dp = (struct dotprod *)p;
|
|
int i;
|
|
|
|
for (i = 0; i < 4; i++)
|
|
if (dp->coeffs[i] != NULL) {
|
|
free(dp->coeffs[i]);
|
|
}
|
|
free(dp);
|
|
}
|
|
|
|
/* Compute a dot product given a descriptor and an input array
|
|
* The length is taken from the descriptor
|
|
*/
|
|
long dotprod_mmx(void *p, signed short a[])
|
|
{
|
|
struct dotprod *dp = (struct dotprod *)p;
|
|
int al;
|
|
signed short *ar;
|
|
|
|
/* Round input data address down to 8 byte boundary
|
|
* NB: depending on the alignment of a[], memory
|
|
* before a[] will be accessed. The contents don't matter since they'll
|
|
* be multiplied by zero coefficients. I can't conceive of any
|
|
* situation where this could cause a segfault since memory protection
|
|
* in the x86 machines is done on much larger boundaries
|
|
*/
|
|
ar = (signed short *)((int)a & ~7);
|
|
|
|
/* Choose one of 4 sets of pre-shifted coefficients. al is both the
|
|
* index into dp->coeffs[] and the number of 0 words padded onto
|
|
* that coefficients array for alignment purposes
|
|
*/
|
|
al = a - ar;
|
|
|
|
/* Call assembler routine to do the work, passing number of 4-word blocks */
|
|
return dotprod_mmx_assist(ar, dp->coeffs[al], (dp->len + al - 1) / 4 + 1);
|
|
}
|
|
|