fmul.ll
4.99 KB
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
; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx900 -mattr=+half-rate-64-ops < %s | FileCheck -check-prefixes=FASTF64,FASTF16,ALL %s
; RUN: opt -cost-model -analyze -mtriple=amdgcn-unknown-amdhsa -mattr=-half-rate-64-ops < %s | FileCheck -check-prefixes=SLOWF64,SLOWF16,ALL %s
; RUN: opt -cost-model -cost-kind=code-size -analyze -mtriple=amdgcn-unknown-amdhsa -mcpu=gfx900 -mattr=+half-rate-64-ops < %s | FileCheck -check-prefixes=FASTF64,FASTF16,ALL %s
; RUN: opt -cost-model -cost-kind=code-size -analyze -mtriple=amdgcn-unknown-amdhsa -mattr=-half-rate-64-ops < %s | FileCheck -check-prefixes=SLOWF64,SLOWF16,ALL %s
; ALL: 'fmul_f32'
; ALL: estimated cost of 1 for {{.*}} fmul float
define amdgpu_kernel void @fmul_f32(float addrspace(1)* %out, float addrspace(1)* %vaddr, float %b) #0 {
%vec = load float, float addrspace(1)* %vaddr
%add = fmul float %vec, %b
store float %add, float addrspace(1)* %out
ret void
}
; ALL: 'fmul_v2f32'
; ALL: estimated cost of 2 for {{.*}} fmul <2 x float>
define amdgpu_kernel void @fmul_v2f32(<2 x float> addrspace(1)* %out, <2 x float> addrspace(1)* %vaddr, <2 x float> %b) #0 {
%vec = load <2 x float>, <2 x float> addrspace(1)* %vaddr
%add = fmul <2 x float> %vec, %b
store <2 x float> %add, <2 x float> addrspace(1)* %out
ret void
}
; ALL: 'fmul_v3f32'
; Allow for 4 when v3f32 is illegal and TargetLowering thinks it needs widening,
; and 3 when it is legal.
; ALL: estimated cost of {{[34]}} for {{.*}} fmul <3 x float>
define amdgpu_kernel void @fmul_v3f32(<3 x float> addrspace(1)* %out, <3 x float> addrspace(1)* %vaddr, <3 x float> %b) #0 {
%vec = load <3 x float>, <3 x float> addrspace(1)* %vaddr
%add = fmul <3 x float> %vec, %b
store <3 x float> %add, <3 x float> addrspace(1)* %out
ret void
}
; ALL: 'fmul_v5f32'
; Allow for 8 when v5f32 is illegal and TargetLowering thinks it needs widening,
; and 5 when it is legal.
; ALL: estimated cost of {{[58]}} for {{.*}} fmul <5 x float>
define amdgpu_kernel void @fmul_v5f32(<5 x float> addrspace(1)* %out, <5 x float> addrspace(1)* %vaddr, <5 x float> %b) #0 {
%vec = load <5 x float>, <5 x float> addrspace(1)* %vaddr
%add = fmul <5 x float> %vec, %b
store <5 x float> %add, <5 x float> addrspace(1)* %out
ret void
}
; ALL: 'fmul_f64'
; FASTF64: estimated cost of 2 for {{.*}} fmul double
; SLOWF64: estimated cost of 3 for {{.*}} fmul double
define amdgpu_kernel void @fmul_f64(double addrspace(1)* %out, double addrspace(1)* %vaddr, double %b) #0 {
%vec = load double, double addrspace(1)* %vaddr
%add = fmul double %vec, %b
store double %add, double addrspace(1)* %out
ret void
}
; ALL: 'fmul_v2f64'
; FASTF64: estimated cost of 4 for {{.*}} fmul <2 x double>
; SLOWF64: estimated cost of 6 for {{.*}} fmul <2 x double>
define amdgpu_kernel void @fmul_v2f64(<2 x double> addrspace(1)* %out, <2 x double> addrspace(1)* %vaddr, <2 x double> %b) #0 {
%vec = load <2 x double>, <2 x double> addrspace(1)* %vaddr
%add = fmul <2 x double> %vec, %b
store <2 x double> %add, <2 x double> addrspace(1)* %out
ret void
}
; ALL: 'fmul_v3f64'
; FASTF64: estimated cost of 6 for {{.*}} fmul <3 x double>
; SLOWF64: estimated cost of 9 for {{.*}} fmul <3 x double>
define amdgpu_kernel void @fmul_v3f64(<3 x double> addrspace(1)* %out, <3 x double> addrspace(1)* %vaddr, <3 x double> %b) #0 {
%vec = load <3 x double>, <3 x double> addrspace(1)* %vaddr
%add = fmul <3 x double> %vec, %b
store <3 x double> %add, <3 x double> addrspace(1)* %out
ret void
}
; ALL: 'fmul_f16'
; ALL: estimated cost of 1 for {{.*}} fmul half
define amdgpu_kernel void @fmul_f16(half addrspace(1)* %out, half addrspace(1)* %vaddr, half %b) #0 {
%vec = load half, half addrspace(1)* %vaddr
%add = fmul half %vec, %b
store half %add, half addrspace(1)* %out
ret void
}
; ALL: 'fmul_v2f16'
; SLOWF16 estimated cost of 2 for {{.*}} fmul <2 x half>
; FASTF16 estimated cost of 1 for {{.*}} fmul <2 x half>
define amdgpu_kernel void @fmul_v2f16(<2 x half> addrspace(1)* %out, <2 x half> addrspace(1)* %vaddr, <2 x half> %b) #0 {
%vec = load <2 x half>, <2 x half> addrspace(1)* %vaddr
%add = fmul <2 x half> %vec, %b
store <2 x half> %add, <2 x half> addrspace(1)* %out
ret void
}
; ALL: 'fmul_v3f16'
; SLOWF16 estimated cost of 4 for {{.*}} fmul <3 x half>
; FASTF16 estimated cost of 2 for {{.*}} fmul <3 x half>
define amdgpu_kernel void @fmul_v3f16(<3 x half> addrspace(1)* %out, <3 x half> addrspace(1)* %vaddr, <3 x half> %b) #0 {
%vec = load <3 x half>, <3 x half> addrspace(1)* %vaddr
%add = fmul <3 x half> %vec, %b
store <3 x half> %add, <3 x half> addrspace(1)* %out
ret void
}
; ALL: 'fmul_v4f16'
; SLOWF16: estimated cost of 4 for {{.*}} fmul <4 x half>
; FASTF16: estimated cost of 2 for {{.*}} fmul <4 x half>
define amdgpu_kernel void @fmul_v4f16(<4 x half> addrspace(1)* %out, <4 x half> addrspace(1)* %vaddr, <4 x half> %b) #0 {
%vec = load <4 x half>, <4 x half> addrspace(1)* %vaddr
%add = fmul <4 x half> %vec, %b
store <4 x half> %add, <4 x half> addrspace(1)* %out
ret void
}
attributes #0 = { nounwind }