Bill Yi | 4e213d5 | 2015-06-23 13:53:11 -0700 | [diff] [blame] | 1 | /* 16-bit signed integer dot product |
| 2 | * MMX assisted version; also for SSE |
| 3 | * |
| 4 | * Copyright 2004 Phil Karn |
| 5 | * May be used under the terms of the GNU Lesser General Public License (LGPL) |
| 6 | */ |
| 7 | #include <stdlib.h> |
| 8 | #include "fec.h" |
| 9 | |
| 10 | struct dotprod { |
| 11 | int len; /* Number of coefficients */ |
| 12 | |
| 13 | /* On a MMX or SSE machine, these hold 4 copies of the coefficients, |
| 14 | * preshifted by 0,1,2,3 words to meet all possible input data |
| 15 | * alignments (see Intel ap559 on MMX dot products). |
| 16 | */ |
| 17 | signed short *coeffs[4]; |
| 18 | }; |
| 19 | long dotprod_mmx_assist(signed short *a,signed short *b,int cnt); |
| 20 | |
| 21 | /* Create and return a descriptor for use with the dot product function */ |
| 22 | void *initdp_mmx(signed short coeffs[],int len){ |
| 23 | struct dotprod *dp; |
| 24 | int i,j; |
| 25 | |
| 26 | |
| 27 | if(len == 0) |
| 28 | return NULL; |
| 29 | |
| 30 | dp = (struct dotprod *)calloc(1,sizeof(struct dotprod)); |
| 31 | dp->len = len; |
| 32 | |
| 33 | /* Make 4 copies of coefficients, one for each data alignment */ |
| 34 | for(i=0;i<4;i++){ |
| 35 | dp->coeffs[i] = (signed short *)calloc(1+(len+i-1)/4, |
| 36 | 4*sizeof(signed short)); |
| 37 | for(j=0;j<len;j++) |
| 38 | dp->coeffs[i][j+i] = coeffs[j]; |
| 39 | } |
| 40 | return (void *)dp; |
| 41 | } |
| 42 | |
| 43 | |
| 44 | /* Free a dot product descriptor created earlier */ |
| 45 | void freedp_mmx(void *p){ |
| 46 | struct dotprod *dp = (struct dotprod *)p; |
| 47 | int i; |
| 48 | |
| 49 | for(i=0;i<4;i++) |
| 50 | if(dp->coeffs[i] != NULL) |
| 51 | free(dp->coeffs[i]); |
| 52 | free(dp); |
| 53 | } |
| 54 | |
| 55 | /* Compute a dot product given a descriptor and an input array |
| 56 | * The length is taken from the descriptor |
| 57 | */ |
| 58 | long dotprod_mmx(void *p,signed short a[]){ |
| 59 | struct dotprod *dp = (struct dotprod *)p; |
| 60 | int al; |
| 61 | signed short *ar; |
| 62 | |
| 63 | /* Round input data address down to 8 byte boundary |
| 64 | * NB: depending on the alignment of a[], memory |
| 65 | * before a[] will be accessed. The contents don't matter since they'll |
| 66 | * be multiplied by zero coefficients. I can't conceive of any |
| 67 | * situation where this could cause a segfault since memory protection |
| 68 | * in the x86 machines is done on much larger boundaries |
| 69 | */ |
| 70 | ar = (signed short *)((int)a & ~7); |
| 71 | |
| 72 | /* Choose one of 4 sets of pre-shifted coefficients. al is both the |
| 73 | * index into dp->coeffs[] and the number of 0 words padded onto |
| 74 | * that coefficients array for alignment purposes |
| 75 | */ |
| 76 | al = a - ar; |
| 77 | |
| 78 | /* Call assembler routine to do the work, passing number of 4-word blocks */ |
| 79 | return dotprod_mmx_assist(ar,dp->coeffs[al],(dp->len+al-1)/4+1); |
| 80 | } |
| 81 | |