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
|
#include "atlas_asm.h"
/*
* function is: void DoFlops(size_t nflops);
*/
#define nflops %rdi
#define N %rax
#define ZR %xmm0
#define A0 %xmm1
#define B0 %xmm2
#define B1 %xmm3
#define B2 %xmm4
#define B3 %xmm5
#define B4 %xmm6
#define C0 %xmm7
#define C1 %xmm8
#define C2 %xmm9
#define C3 %xmm10
#define C4 %xmm11
#define C5 %xmm12
#define C6 %xmm13
#define C7 %xmm14
.text
.globl ATL_asmdecor(DoFlops)
ATL_asmdecor(DoFlops):
/*
* Zero all xmm regs
*/
xorps %xmm0, %xmm0
xorps %xmm1, %xmm1
xorps %xmm2, %xmm2
xorps %xmm3, %xmm3
xorps %xmm4, %xmm4
xorps %xmm5, %xmm5
xorps %xmm6, %xmm6
xorps %xmm7, %xmm7
xorps %xmm8, %xmm8
xorps %xmm9, %xmm9
xorps %xmm10, %xmm10
xorps %xmm11, %xmm11
xorps %xmm12, %xmm12
xorps %xmm13, %xmm13
xorps %xmm14, %xmm14
xorps %xmm15, %xmm15
/*
* This loop adds into 8 different accumulators, after doing a chained
* multiplication. The number of flops is therefore:
* (veclen)*(vecflops)*(naccum) = 4 * 2 * 8 = 64 flops/iteration
*/
movq nflops, N
shr $6, N /* N = nflops / 64 */
LOOPN:
/*
* On Intel chips, you need to write all read registers once every loop
* iteration, or you cannot achieve peak.
*/
mulps A0, B0
addps B0, C0
mulps A0, B1
addps B1, C1
mulps A0, B2
addps B2, C2
mulps A0, B3
addps B3, C3
mulps A0, B0
addps B0, C4
mulps A0, B1
addps B1, C5
mulps A0, B2
addps B2, C6
mulps A0, B3
addps B3, C7
#ifdef Intel
xorps A0, A0
#endif
sub $1, N
jnz LOOPN
ret
|
