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
|
static void
dgemm_micro_kernel(long kc,
double alpha, const double *A, const double *B,
double beta,
double *C, long incRowC, long incColC)
{
double AB[MR*NR] __attribute__ ((aligned (16)));
int i, j, l;
//
// Compute AB = A*B
//
register __m128d ab_00_11, ab_20_31;
register __m128d ab_01_10, ab_21_30;
register __m128d ab_02_13, ab_22_33;
register __m128d ab_03_12, ab_23_32;
register __m128d tmp0, tmp1, tmp2, tmp3;
register __m128d tmp4, tmp5, tmp6, tmp7;
ab_00_11 = _mm_setzero_pd(); ab_20_31 = _mm_setzero_pd();
ab_01_10 = _mm_setzero_pd(); ab_21_30 = _mm_setzero_pd();
ab_02_13 = _mm_setzero_pd(); ab_22_33 = _mm_setzero_pd();
ab_03_12 = _mm_setzero_pd(); ab_23_32 = _mm_setzero_pd();
tmp0 = _mm_load_pd(A);
tmp1 = _mm_load_pd(A+2);
tmp2 = _mm_load_pd(B);
tmp3 = _mm_setzero_pd();
tmp4 = _mm_setzero_pd();
tmp5 = _mm_setzero_pd();
tmp6 = _mm_setzero_pd();
tmp7 = _mm_setzero_pd();
for (l=0; l<kc; ++l) {
ab_02_13 = _mm_add_pd(ab_02_13, tmp3);
tmp3 = _mm_load_pd(B+2);
ab_22_33 = _mm_add_pd(ab_22_33, tmp6);
tmp6 = tmp2;
tmp4 = _mm_shuffle_pd(tmp2, tmp2, _MM_SHUFFLE2(0, 1));
tmp2 = _mm_mul_pd(tmp2, tmp0);
tmp6 = _mm_mul_pd(tmp6, tmp1);
ab_03_12 = _mm_add_pd(ab_03_12, tmp5);
ab_23_32 = _mm_add_pd(ab_23_32, tmp7);
tmp7 = tmp4;
tmp4 = _mm_mul_pd(tmp4, tmp0);
tmp7 = _mm_mul_pd(tmp7, tmp1);
ab_00_11 = _mm_add_pd(ab_00_11, tmp2);
tmp2 = _mm_load_pd(B+4);
ab_20_31 = _mm_add_pd(ab_20_31, tmp6);
tmp6 = tmp3;
tmp5 = _mm_shuffle_pd(tmp3, tmp3, _MM_SHUFFLE2(0, 1));
tmp3 = _mm_mul_pd(tmp3, tmp0);
tmp6 = _mm_mul_pd(tmp6, tmp1);
ab_01_10 = _mm_add_pd(ab_01_10, tmp4);
ab_21_30 = _mm_add_pd(ab_21_30, tmp7);
tmp7 = tmp5;
tmp5 = _mm_mul_pd(tmp5, tmp0);
tmp0 = _mm_load_pd(A+4);
tmp7 = _mm_mul_pd(tmp7, tmp1);
tmp1 = _mm_load_pd(A+6);
A += 4;
B += 4;
}
ab_02_13 = _mm_add_pd(ab_02_13, tmp3);
ab_22_33 = _mm_add_pd(ab_22_33, tmp6);
ab_03_12 = _mm_add_pd(ab_03_12, tmp5);
ab_23_32 = _mm_add_pd(ab_23_32, tmp7);
_mm_storel_pd(&AB[0+0*4], ab_00_11);
_mm_storeh_pd(&AB[1+0*4], ab_01_10);
_mm_storel_pd(&AB[2+0*4], ab_20_31);
_mm_storeh_pd(&AB[3+0*4], ab_21_30);
_mm_storel_pd(&AB[0+1*4], ab_01_10);
_mm_storeh_pd(&AB[1+1*4], ab_00_11);
_mm_storel_pd(&AB[2+1*4], ab_21_30);
_mm_storeh_pd(&AB[3+1*4], ab_20_31);
_mm_storel_pd(&AB[0+2*4], ab_02_13);
_mm_storeh_pd(&AB[1+2*4], ab_03_12);
_mm_storel_pd(&AB[2+2*4], ab_22_33);
_mm_storeh_pd(&AB[3+2*4], ab_23_32);
_mm_storel_pd(&AB[0+3*4], ab_03_12);
_mm_storeh_pd(&AB[1+3*4], ab_02_13);
_mm_storel_pd(&AB[2+3*4], ab_23_32);
_mm_storeh_pd(&AB[3+3*4], ab_22_33);
//
// Update C <- beta*C
//
if (beta==0.0) {
for (j=0; j<NR; ++j) {
for (i=0; i<MR; ++i) {
C[i*incRowC+j*incColC] = 0.0;
}
}
} else if (beta!=1.0) {
for (j=0; j<NR; ++j) {
for (i=0; i<MR; ++i) {
C[i*incRowC+j*incColC] *= beta;
}
}
}
//
// Update C <- C + alpha*AB (note: the case alpha==0.0 was already treated in
// the above layer dgemm_nn)
//
if (alpha==1.0) {
for (j=0; j<NR; ++j) {
for (i=0; i<MR; ++i) {
C[i*incRowC+j*incColC] += AB[i+j*MR];
}
}
} else {
for (j=0; j<NR; ++j) {
for (i=0; i<MR; ++i) {
C[i*incRowC+j*incColC] += alpha*AB[i+j*MR];
}
}
}
}
|