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
|
; RUN: llc -mtriple=amdgcn -mcpu=gfx1200 -mattr=-flat-for-global -verify-machineinstrs < %s | FileCheck -enable-var-scope -check-prefixes=GCN %s
; GCN-LABEL: {{^}}test_fmaximum3_olt_0_f32:
; GCN: buffer_load_b32 [[REGC:v[0-9]+]]
; GCN: buffer_load_b32 [[REGB:v[0-9]+]]
; GCN: buffer_load_b32 [[REGA:v[0-9]+]]
; GCN: v_maximum3_f32 [[RESULT:v[0-9]+]], [[REGC]], [[REGB]], [[REGA]]
; GCN: buffer_store_b32 [[RESULT]],
define amdgpu_kernel void @test_fmaximum3_olt_0_f32(ptr addrspace(1) %out, ptr addrspace(1) %aptr, ptr addrspace(1) %bptr, ptr addrspace(1) %cptr) {
%a = load volatile float, ptr addrspace(1) %aptr, align 4
%b = load volatile float, ptr addrspace(1) %bptr, align 4
%c = load volatile float, ptr addrspace(1) %cptr, align 4
%f0 = call float @llvm.maximum.f32(float %a, float %b)
%f1 = call float @llvm.maximum.f32(float %f0, float %c)
store float %f1, ptr addrspace(1) %out, align 4
ret void
}
; Commute operand of second fmaximum
; GCN-LABEL: {{^}}test_fmaximum3_olt_1_f32:
; GCN: buffer_load_b32 [[REGB:v[0-9]+]]
; GCN: buffer_load_b32 [[REGA:v[0-9]+]]
; GCN: buffer_load_b32 [[REGC:v[0-9]+]]
; GCN: v_maximum3_f32 [[RESULT:v[0-9]+]], [[REGC]], [[REGB]], [[REGA]]
; GCN: buffer_store_b32 [[RESULT]],
define amdgpu_kernel void @test_fmaximum3_olt_1_f32(ptr addrspace(1) %out, ptr addrspace(1) %aptr, ptr addrspace(1) %bptr, ptr addrspace(1) %cptr) {
%a = load volatile float, ptr addrspace(1) %aptr, align 4
%b = load volatile float, ptr addrspace(1) %bptr, align 4
%c = load volatile float, ptr addrspace(1) %cptr, align 4
%f0 = call float @llvm.maximum.f32(float %a, float %b)
%f1 = call float @llvm.maximum.f32(float %c, float %f0)
store float %f1, ptr addrspace(1) %out, align 4
ret void
}
; GCN-LABEL: {{^}}test_fmaximum3_olt_0_f16:
; GCN: buffer_load_u16 [[REGC:v[0-9]+]]
; GCN: buffer_load_u16 [[REGB:v[0-9]+]]
; GCN: buffer_load_u16 [[REGA:v[0-9]+]]
; GCN: v_maximum3_f16 [[RESULT:v[0-9]+]], [[REGC]], [[REGB]], [[REGA]]
; GCN: buffer_store_b16 [[RESULT]],
define amdgpu_kernel void @test_fmaximum3_olt_0_f16(ptr addrspace(1) %out, ptr addrspace(1) %aptr, ptr addrspace(1) %bptr, ptr addrspace(1) %cptr) {
%a = load volatile half, ptr addrspace(1) %aptr, align 2
%b = load volatile half, ptr addrspace(1) %bptr, align 2
%c = load volatile half, ptr addrspace(1) %cptr, align 2
%f0 = call half @llvm.maximum.f16(half %a, half %b)
%f1 = call half @llvm.maximum.f16(half %f0, half %c)
store half %f1, ptr addrspace(1) %out, align 2
ret void
}
; GCN-LABEL: {{^}}test_fmaximum3_olt_1_f16:
; GCN: buffer_load_u16 [[REGA:v[0-9]+]]
; GCN: buffer_load_u16 [[REGB:v[0-9]+]]
; GCN: buffer_load_u16 [[REGC:v[0-9]+]]
; GCN: v_maximum3_f16 [[RESULT:v[0-9]+]], [[REGC]], [[REGA]], [[REGB]]
; GCN: buffer_store_b16 [[RESULT]],
define amdgpu_kernel void @test_fmaximum3_olt_1_f16(ptr addrspace(1) %out, ptr addrspace(1) %aptr, ptr addrspace(1) %bptr, ptr addrspace(1) %cptr) {
%a = load volatile half, ptr addrspace(1) %aptr, align 2
%b = load volatile half, ptr addrspace(1) %bptr, align 2
%c = load volatile half, ptr addrspace(1) %cptr, align 2
%f0 = call half @llvm.maximum.f16(half %a, half %b)
%f1 = call half @llvm.maximum.f16(half %c, half %f0)
store half %f1, ptr addrspace(1) %out, align 2
ret void
}
; Checks whether the test passes; performMinMaxCombine() should not optimize vector patterns of maximum3
; since there are no pack instructions for fmaximum3.
; GCN-LABEL: {{^}}no_fmaximum3_v2f16:
; GCN: v_pk_maximum_f16 v0, v0, v1
; GCN: v_pk_maximum_f16 v0, v2, v0
; GCN: v_pk_maximum_f16 v0, v0, v3
; GCN-NEXT: s_setpc_b64
define <2 x half> @no_fmaximum3_v2f16(<2 x half> %a, <2 x half> %b, <2 x half> %c, <2 x half> %d) {
entry:
%max = call <2 x half> @llvm.maximum.v2f16(<2 x half> %a, <2 x half> %b)
%max1 = call <2 x half> @llvm.maximum.v2f16(<2 x half> %c, <2 x half> %max)
%res = call <2 x half> @llvm.maximum.v2f16(<2 x half> %max1, <2 x half> %d)
ret <2 x half> %res
}
; GCN-LABEL: {{^}}no_fmaximum3_olt_0_f64:
; GCN-COUNT-2: v_maximum_f64
define amdgpu_kernel void @no_fmaximum3_olt_0_f64(ptr addrspace(1) %out, ptr addrspace(1) %aptr, ptr addrspace(1) %bptr, ptr addrspace(1) %cptr) {
%a = load volatile double, ptr addrspace(1) %aptr, align 4
%b = load volatile double, ptr addrspace(1) %bptr, align 4
%c = load volatile double, ptr addrspace(1) %cptr, align 4
%f0 = call double @llvm.maximum.f64(double %a, double %b)
%f1 = call double @llvm.maximum.f64(double %f0, double %c)
store double %f1, ptr addrspace(1) %out, align 4
ret void
}
declare double @llvm.maximum.f64(double, double)
declare float @llvm.maximum.f32(float, float)
declare half @llvm.maximum.f16(half, half)
declare <2 x half> @llvm.maximum.v2f16(<2 x half>, <2 x half>)
|
