1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
|
#include "stdlib.h"
#include "util.h"
#include "dataset.h"
#define REG_I 8
#define REG_J 2
//#define BLOCK_I 32
#define BLOCK_J 16
#define BLOCK_K 16
#define LDA 32
#define NCORES 2
#define MIN(X,Y) (X < Y ? X : Y)
void __attribute__((noinline)) matmul(const int coreid, const int ncores, const int lda, const data_t A[], const data_t B[], data_t C[] )
{
// ***************************** //
// **** ADD YOUR CODE HERE ***** //
// ***************************** //
//
// feel free to make a separate function for MI and MSI versions.
int i, j, k, ri, rj, ii, jj, kk;
data_t *Aj, *Cj, *Bi;
data_t c[REG_I][REG_J], a[REG_J], b[REG_I];
size_t start = coreid * (LDA / NCORES), end = (coreid == NCORES - 1 ? LDA : (coreid + 1) * (LDA / NCORES));
/* if (coreid > 0) { */
/* return; */
/* } */
/* start = 0, end = lda; */
if (ncores == NCORES && lda == LDA) {
for (jj = start; jj < end; jj += BLOCK_J)
for (kk = 0; kk < LDA; kk += BLOCK_K)
//for (ii = 0; ii < LDA; ii += BLOCK_I)
for (j = jj; j < MIN(end, jj + BLOCK_J); j += REG_J) {
Aj = A + j*LDA;
Cj = C + j*LDA;
for (i = 0; i < LDA; i += REG_I) {
/* Load C in register blocks. */
Bi = B + i;
for (ri = 0; ri < REG_I; ri++) {
for (rj = 0; rj < REG_J; rj++) {
c[ri][rj] = Cj[i + ri + ( rj)*LDA];
}
}
for (k = kk; k < MIN(LDA, kk + BLOCK_K); k++) {
/* Load a,b in register blocks. */
/* for (rj = 0; rj < REG_J; rj++) {
a[rj] = A[(j + rj)*LDA + k];
}*/
/* for (ri = 0; ri < REG_I; ri++) { */
/* b[ri] = Bi[k*LDA + ri]; */
/* } */
/* /\* Compute C in register blocks. *\/ */
/* for (rj = 0; rj < REG_J; rj++) { */
/* a[rj] = Aj[( rj)*LDA + k]; */
/* for (ri = 0; ri < REG_I; ri++) { */
/* c[ri][rj] += a[rj] * b[ri]; */
/* } */
/* } */
a[0] = Aj[k];
a[1] = Aj[k + LDA];
b[0] = Bi[k*LDA];
b[1] = Bi[k*LDA + 1];
b[2] = Bi[k*LDA + 2];
b[3] = Bi[k*LDA + 3];
b[4] = Bi[k*LDA + 4];
b[5] = Bi[k*LDA + 5];
b[6] = Bi[k*LDA + 6];
b[7] = Bi[k*LDA + 7];
c[0][0] += b[0] * a[0];
c[0][1] += b[0] * a[1];
c[1][0] += b[1] * a[0];
c[1][1] += b[1] * a[1];
c[2][0] += b[2] * a[0];
c[2][1] += b[2] * a[1];
c[3][0] += b[3] * a[0];
c[3][1] += b[3] * a[1];
c[4][0] += b[4] * a[0];
c[4][1] += b[4] * a[1];
c[5][0] += b[5] * a[0];
c[5][1] += b[5] * a[1];
c[6][0] += b[6] * a[0];
c[6][1] += b[6] * a[1];
c[7][0] += b[7] * a[0];
c[7][1] += b[7] * a[1];
/* c[0][0] += b[0] * a[0]; */
/* c[1][1] += b[1] * a[1]; */
/* c[2][0] += b[2] * a[0]; */
/* c[3][1] += b[3] * a[1]; */
/* c[4][0] += b[4] * a[0]; */
/* c[5][1] += b[5] * a[1]; */
/* c[6][0] += b[6] * a[0]; */
/* c[7][1] += b[7] * a[1]; */
/* c[0][0] += b[0] * a[0]; */
/* c[1][1] += b[1] * a[1]; */
/* c[2][0] += b[2] * a[0]; */
/* c[3][1] += b[3] * a[1]; */
/* c[4][0] += b[4] * a[0]; */
/* c[5][1] += b[5] * a[1]; */
/* c[6][0] += b[6] * a[0]; */
/* c[7][1] += b[7] * a[1]; */
}
/* store C in register blocks. */
for (ri = 0; ri < REG_I; ri++) {
for (rj = 0; rj < REG_J; rj++) {
Cj[i + ri + (rj)*LDA] = c[ri][rj];
}
}
}
}
/* We only care about performance for 32x32 matrices and 2 cores. Otherwise just naive mat_mul */
} else {
if (coreid > 0)
return;
for ( i = 0; i < lda; i++ )
for ( j = 0; j < lda; j++ )
for ( k = 0; k < lda; k++ )
C[i + j*lda] += A[j*lda + k] * B[k*lda + i];
}
}
|
