1*638691a0SAndroid Build Coastguard Worker /* 16-bit signed integer dot product
2*638691a0SAndroid Build Coastguard Worker * SSE2 version
3*638691a0SAndroid Build Coastguard Worker * Copyright 2004 Phil Karn
4*638691a0SAndroid Build Coastguard Worker * May be used under the terms of the GNU Lesser General Public License (LGPL)
5*638691a0SAndroid Build Coastguard Worker */
6*638691a0SAndroid Build Coastguard Worker #define _XOPEN_SOURCE 600
7*638691a0SAndroid Build Coastguard Worker #include <stdlib.h>
8*638691a0SAndroid Build Coastguard Worker #include <memory.h>
9*638691a0SAndroid Build Coastguard Worker #include "fec.h"
10*638691a0SAndroid Build Coastguard Worker
11*638691a0SAndroid Build Coastguard Worker struct dotprod {
12*638691a0SAndroid Build Coastguard Worker int len; /* Number of coefficients */
13*638691a0SAndroid Build Coastguard Worker
14*638691a0SAndroid Build Coastguard Worker /* On a SSE2 machine, these hold 8 copies of the coefficients,
15*638691a0SAndroid Build Coastguard Worker * preshifted by 0,1,..7 words to meet all possible input data
16*638691a0SAndroid Build Coastguard Worker * alignments (see Intel ap559 on MMX dot products).
17*638691a0SAndroid Build Coastguard Worker */
18*638691a0SAndroid Build Coastguard Worker signed short *coeffs[8];
19*638691a0SAndroid Build Coastguard Worker };
20*638691a0SAndroid Build Coastguard Worker
21*638691a0SAndroid Build Coastguard Worker long dotprod_sse2_assist(signed short *a,signed short *b,int cnt);
22*638691a0SAndroid Build Coastguard Worker
23*638691a0SAndroid Build Coastguard Worker /* Create and return a descriptor for use with the dot product function */
initdp_sse2(signed short coeffs[],int len)24*638691a0SAndroid Build Coastguard Worker void *initdp_sse2(signed short coeffs[],int len){
25*638691a0SAndroid Build Coastguard Worker struct dotprod *dp;
26*638691a0SAndroid Build Coastguard Worker int i,j,blksize;
27*638691a0SAndroid Build Coastguard Worker
28*638691a0SAndroid Build Coastguard Worker if(len == 0)
29*638691a0SAndroid Build Coastguard Worker return NULL;
30*638691a0SAndroid Build Coastguard Worker
31*638691a0SAndroid Build Coastguard Worker dp = (struct dotprod *)calloc(1,sizeof(struct dotprod));
32*638691a0SAndroid Build Coastguard Worker dp->len = len;
33*638691a0SAndroid Build Coastguard Worker
34*638691a0SAndroid Build Coastguard Worker /* Make 8 copies of coefficients, one for each data alignment,
35*638691a0SAndroid Build Coastguard Worker * each aligned to 16-byte boundary
36*638691a0SAndroid Build Coastguard Worker */
37*638691a0SAndroid Build Coastguard Worker for(i=0;i<8;i++){
38*638691a0SAndroid Build Coastguard Worker blksize = (1+(len+i-1)/8) * 8*sizeof(signed short);
39*638691a0SAndroid Build Coastguard Worker posix_memalign((void **)&dp->coeffs[i],16,blksize);
40*638691a0SAndroid Build Coastguard Worker memset(dp->coeffs[i],0,blksize);
41*638691a0SAndroid Build Coastguard Worker for(j=0;j<len;j++)
42*638691a0SAndroid Build Coastguard Worker dp->coeffs[i][j+i] = coeffs[j];
43*638691a0SAndroid Build Coastguard Worker }
44*638691a0SAndroid Build Coastguard Worker return (void *)dp;
45*638691a0SAndroid Build Coastguard Worker }
46*638691a0SAndroid Build Coastguard Worker
47*638691a0SAndroid Build Coastguard Worker
48*638691a0SAndroid Build Coastguard Worker /* Free a dot product descriptor created earlier */
freedp_sse2(void * p)49*638691a0SAndroid Build Coastguard Worker void freedp_sse2(void *p){
50*638691a0SAndroid Build Coastguard Worker struct dotprod *dp = (struct dotprod *)p;
51*638691a0SAndroid Build Coastguard Worker int i;
52*638691a0SAndroid Build Coastguard Worker
53*638691a0SAndroid Build Coastguard Worker for(i=0;i<8;i++)
54*638691a0SAndroid Build Coastguard Worker if(dp->coeffs[i] != NULL)
55*638691a0SAndroid Build Coastguard Worker free(dp->coeffs[i]);
56*638691a0SAndroid Build Coastguard Worker free(dp);
57*638691a0SAndroid Build Coastguard Worker }
58*638691a0SAndroid Build Coastguard Worker
59*638691a0SAndroid Build Coastguard Worker /* Compute a dot product given a descriptor and an input array
60*638691a0SAndroid Build Coastguard Worker * The length is taken from the descriptor
61*638691a0SAndroid Build Coastguard Worker */
dotprod_sse2(void * p,signed short a[])62*638691a0SAndroid Build Coastguard Worker long dotprod_sse2(void *p,signed short a[]){
63*638691a0SAndroid Build Coastguard Worker struct dotprod *dp = (struct dotprod *)p;
64*638691a0SAndroid Build Coastguard Worker int al;
65*638691a0SAndroid Build Coastguard Worker signed short *ar;
66*638691a0SAndroid Build Coastguard Worker
67*638691a0SAndroid Build Coastguard Worker ar = (signed short *)((int)a & ~15);
68*638691a0SAndroid Build Coastguard Worker al = a - ar;
69*638691a0SAndroid Build Coastguard Worker
70*638691a0SAndroid Build Coastguard Worker /* Call assembler routine to do the work, passing number of 8-word blocks */
71*638691a0SAndroid Build Coastguard Worker return dotprod_sse2_assist(ar,dp->coeffs[al],(dp->len+al-1)/8+1);
72*638691a0SAndroid Build Coastguard Worker }
73