-
Notifications
You must be signed in to change notification settings - Fork 19
/
snrm2_amd64.s
121 lines (91 loc) · 2.01 KB
/
snrm2_amd64.s
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
// func Snrm2(N int, X []float32, incX int) float32
TEXT ·Snrm2(SB), 7, $0
MOVQ N+0(FP), BP
MOVQ X_data+8(FP), SI
MOVQ incX+32(FP), AX
// Check data bounaries
MOVQ BP, CX
DECQ CX
IMULQ AX, CX // CX = incX * (N - 1)
CMPQ CX, X_len+16(FP)
JGE panic
// Clear accumulators
XORPS X0, X0
// Setup strides
SALQ $2, AX // AX = sizeof(float32) * incX
// Check that there are 4 or more pairs for SIMD calculations
SUBQ $4, BP
JL rest // There are less than 4 pairs to process
// Check if incX != 1 or incY != 1
CMPQ AX, $4
JNE with_stride
// Fully optimized loop (for incX == incY == 1)
full_simd_loop:
// Multiply four values
MOVUPS (SI), X1
MULPS X1, X1
// Update data pointer
ADDQ $16, SI
// Accumulate the results of multiplications
ADDPS X1, X0
SUBQ $4, BP
JGE full_simd_loop // There are 4 or more pairs to process
JMP hsum
with_stride:
// Setup long strides
MOVQ AX, CX
SALQ $1, CX // CX = 8 * incX
// Partially optimized loop
half_simd_loop:
// Load first two values
MOVSS (SI), X1
MOVSS (SI)(AX*1), X2
// Create half-vector
UNPCKLPS X2, X1
// Update data pointers using long strides
ADDQ CX, SI
// Load second two values
MOVSS (SI), X2
MOVSS (SI)(AX*1), X3
// Create half-vector
UNPCKLPS X3, X2
// Update data pointer using long strides
ADDQ CX, SI
// Create full-vector
MOVLHPS X2, X1
// Square it
MULPS X1, X1
// Accumulate the result of multiplication
ADDPS X1, X0
SUBQ $4, BP
JGE half_simd_loop // There are 4 or more values to process
hsum:
// Horizontal sum
MOVHLPS X0, X1
ADDPS X0, X1
MOVSS X1, X0
SHUFPS $0xe1, X1, X1
ADDSS X1, X0
rest:
// Undo last SUBQ
ADDQ $4, BP
// Check that are there any value to process
JE end
loop:
// Multiply one value
MOVSS (SI), X1
MULSS X1, X1
// Update data pointers
ADDQ AX, SI
// Accumulate the results of multiplication
ADDSS X1, X0
DECQ BP
JNE loop
end:
// Return the square root of sum
SQRTSS X0, X0
MOVSS X0, r+40(FP)
RET
panic:
CALL ·panicIndex(SB)
RET