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
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
|
// See LICENSE for license details.
#include "common.h"
#include <assert.h>
#include <math.h>
#include <stdint.h>
#include <alloca.h>
#define MIN(a, b) ((a) < (b) ? (a) : (b))
static void mm_naive(size_t m, size_t n, size_t p,
t* a, size_t lda, t* b, size_t ldb, t* c, size_t ldc)
{
for (size_t i = 0; i < m; i++)
{
for (size_t j = 0; j < n; j++)
{
t s0 = c[i*ldc+j], s1 = 0, s2 = 0, s3 = 0;
for (size_t k = 0; k < p/4*4; k+=4)
{
s0 = fma(a[i*lda+k+0], b[(k+0)*ldb+j], s0);
s1 = fma(a[i*lda+k+1], b[(k+1)*ldb+j], s1);
s2 = fma(a[i*lda+k+2], b[(k+2)*ldb+j], s2);
s3 = fma(a[i*lda+k+3], b[(k+3)*ldb+j], s3);
}
for (size_t k = p/4*4; k < p; k++)
s0 = fma(a[i*lda+k], b[k*ldb+j], s0);
c[i*ldc+j] = (s0 + s1) + (s2 + s3);
}
}
}
static inline void mm_rb(size_t m, size_t n, size_t p,
t* a, size_t lda, t* b, size_t ldb, t* c, size_t ldc)
{
size_t mb = m/RBM*RBM, nb = n/RBN*RBN;
for (size_t i = 0; i < mb; i += RBM)
{
for (size_t j = 0; j < nb; j += RBN)
kloop(p, a+i*lda, lda, b+j, ldb, c+i*ldc+j, ldc);
mm_naive(RBM, n - nb, p, a+i*lda, lda, b+nb, ldb, c+i*ldc+nb, ldc);
}
mm_naive(m - mb, n, p, a+mb*lda, lda, b, ldb, c+mb*ldc, ldc);
}
static inline void repack(t* a, size_t lda, const t* a0, size_t lda0, size_t m, size_t p)
{
for (size_t i = 0; i < m; i++)
{
for (size_t j = 0; j < p/8*8; j+=8)
{
t t0 = a0[i*lda0+j+0];
t t1 = a0[i*lda0+j+1];
t t2 = a0[i*lda0+j+2];
t t3 = a0[i*lda0+j+3];
t t4 = a0[i*lda0+j+4];
t t5 = a0[i*lda0+j+5];
t t6 = a0[i*lda0+j+6];
t t7 = a0[i*lda0+j+7];
a[i*lda+j+0] = t0;
a[i*lda+j+1] = t1;
a[i*lda+j+2] = t2;
a[i*lda+j+3] = t3;
a[i*lda+j+4] = t4;
a[i*lda+j+5] = t5;
a[i*lda+j+6] = t6;
a[i*lda+j+7] = t7;
}
for (size_t j = p/8*8; j < p; j++)
a[i*lda+j] = a0[i*lda0+j];
}
}
static void mm_cb(size_t m, size_t n, size_t p,
t* a, size_t lda, t* b, size_t ldb, t* c, size_t ldc)
{
size_t nmb = m/CBM, nnb = n/CBN, npb = p/CBK;
size_t mb = nmb*CBM, nb = nnb*CBN, pb = npb*CBK;
//t a1[mb*pb], b1[pb*nb], c1[mb*nb];
t* a1 = (t*)alloca_aligned(sizeof(t)*mb*pb, 8192);
t* b1 = (t*)alloca_aligned(sizeof(t)*pb*nb, 8192);
t* c1 = (t*)alloca_aligned(sizeof(t)*mb*nb, 8192);
for (size_t i = 0; i < mb; i += CBM)
for (size_t j = 0; j < pb; j += CBK)
repack(a1 + (npb*(i/CBM) + j/CBK)*(CBM*CBK), CBK, a + i*lda + j, lda, CBM, CBK);
for (size_t i = 0; i < pb; i += CBK)
for (size_t j = 0; j < nb; j += CBN)
repack(b1 + (nnb*(i/CBK) + j/CBN)*(CBK*CBN), CBN, b + i*ldb + j, ldb, CBK, CBN);
for (size_t i = 0; i < mb; i += CBM)
for (size_t j = 0; j < nb; j += CBN)
repack(c1 + (nnb*(i/CBM) + j/CBN)*(CBM*CBN), CBN, c + i*ldc + j, ldc, CBM, CBN);
for (size_t i = 0; i < mb; i += CBM)
{
for (size_t j = 0; j < nb; j += CBN)
{
for (size_t k = 0; k < pb; k += CBK)
{
mm_rb(CBM, CBN, CBK,
a1 + (npb*(i/CBM) + k/CBK)*(CBM*CBK), CBK,
b1 + (nnb*(k/CBK) + j/CBN)*(CBK*CBN), CBN,
c1 + (nnb*(i/CBM) + j/CBN)*(CBM*CBN), CBN);
}
if (pb < p)
{
mm_rb(CBM, CBN, p - pb,
a + i*lda + pb, lda,
b + pb*ldb + j, ldb,
c1 + (nnb*(i/CBM) + j/CBN)*(CBM*CBN), CBN);
}
}
if (nb < n)
{
for (size_t k = 0; k < p; k += CBK)
{
mm_rb(CBM, n - nb, MIN(p - k, CBK),
a + i*lda + k, lda,
b + k*ldb + nb, ldb,
c + i*ldc + nb, ldc);
}
}
}
if (mb < m)
{
for (size_t j = 0; j < n; j += CBN)
{
for (size_t k = 0; k < p; k += CBK)
{
mm_rb(m - mb, MIN(n - j, CBN), MIN(p - k, CBK),
a + mb*lda + k, lda,
b + k*ldb + j, ldb,
c + mb*ldc + j, ldc);
}
}
}
for (size_t i = 0; i < mb; i += CBM)
for (size_t j = 0; j < nb; j += CBN)
repack(c + i*ldc + j, ldc, c1 + (nnb*(i/CBM) + j/CBN)*(CBM*CBN), CBN, CBM, CBN);
}
void mm(size_t m, size_t n, size_t p,
t* a, size_t lda, t* b, size_t ldb, t* c, size_t ldc)
{
if (__builtin_expect(m <= 2*CBM && n <= 2*CBN && p <= 2*CBK, 1))
mm_rb(m, n, p, a, lda, b, ldb, c, ldc);
else
mm_cb(m, n, p, a, lda, b, ldb, c, ldc);
}
|
