llvm-project

builtins-amdgcn-gfx12-wmma-w32.cl
156 строк · 7.1 Кб
Перенос по словам
1
// NOTE: Assertions have been autogenerated by utils/update_cc_test_checks.py
2
// REQUIRES: amdgpu-registered-target
3
// RUN: %clang_cc1 -triple amdgcn-unknown-unknown -target-cpu gfx1200 -target-feature +wavefrontsize32 -emit-llvm -o - %s | FileCheck %s --check-prefix=CHECK-GFX1200
4

5
typedef int    v2i   __attribute__((ext_vector_type(2)));
6
typedef float  v8f   __attribute__((ext_vector_type(8)));
7
typedef half   v8h   __attribute__((ext_vector_type(8)));
8
typedef short  v8s   __attribute__((ext_vector_type(8)));
9
typedef int    v8i   __attribute__((ext_vector_type(8)));
10

11
// Wave32
12

13
//
14
// amdgcn_wmma_f32_16x16x16_f16
15
//
16

17
// CHECK-GFX1200-LABEL: @test_amdgcn_wmma_f32_16x16x16_f16_w32(
18
// CHECK-GFX1200-NEXT:  entry:
19
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <8 x float> @llvm.amdgcn.wmma.f32.16x16x16.f16.v8f32.v8f16(<8 x half> [[A:%.*]], <8 x half> [[B:%.*]], <8 x float> [[C:%.*]])
20
// CHECK-GFX1200-NEXT:    store <8 x float> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 32, !tbaa [[TBAA4:![0-9]+]]
21
// CHECK-GFX1200-NEXT:    ret void
22
//
23
void test_amdgcn_wmma_f32_16x16x16_f16_w32(global v8f* out, v8h a, v8h b, v8f c)
24
{
25
  *out = __builtin_amdgcn_wmma_f32_16x16x16_f16_w32_gfx12(a, b, c);
26
}
27

28
//
29
// amdgcn_wmma_f32_16x16x16_bf16
30
//
31

32
// CHECK-GFX1200-LABEL: @test_amdgcn_wmma_f32_16x16x16_bf16_w32(
33
// CHECK-GFX1200-NEXT:  entry:
34
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <8 x float> @llvm.amdgcn.wmma.f32.16x16x16.bf16.v8f32.v8i16(<8 x i16> [[A:%.*]], <8 x i16> [[B:%.*]], <8 x float> [[C:%.*]])
35
// CHECK-GFX1200-NEXT:    store <8 x float> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 32, !tbaa [[TBAA4]]
36
// CHECK-GFX1200-NEXT:    ret void
37
//
38
void test_amdgcn_wmma_f32_16x16x16_bf16_w32(global v8f* out, v8s a, v8s b, v8f c)
39
{
40
  *out = __builtin_amdgcn_wmma_f32_16x16x16_bf16_w32_gfx12(a, b, c);
41
}
42

43
//
44
// amdgcn_wmma_f16_16x16x16_f16
45
//
46

47
// CHECK-GFX1200-LABEL: @test_amdgcn_wmma_f16_16x16x16_f16_w32(
48
// CHECK-GFX1200-NEXT:  entry:
49
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <8 x half> @llvm.amdgcn.wmma.f16.16x16x16.f16.v8f16.v8f16(<8 x half> [[A:%.*]], <8 x half> [[B:%.*]], <8 x half> [[C:%.*]], i1 false)
50
// CHECK-GFX1200-NEXT:    store <8 x half> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 16, !tbaa [[TBAA4]]
51
// CHECK-GFX1200-NEXT:    ret void
52
//
53
void test_amdgcn_wmma_f16_16x16x16_f16_w32(global v8h* out, v8h a, v8h b, v8h c)
54
{
55
  *out = __builtin_amdgcn_wmma_f16_16x16x16_f16_w32_gfx12(a, b, c);
56
}
57

58
//
59
// amdgcn_wmma_bf16_16x16x16_bf16
60
//
61

62
// CHECK-GFX1200-LABEL: @test_amdgcn_wmma_bf16_16x16x16_bf16_w32(
63
// CHECK-GFX1200-NEXT:  entry:
64
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <8 x i16> @llvm.amdgcn.wmma.bf16.16x16x16.bf16.v8i16.v8i16(<8 x i16> [[A:%.*]], <8 x i16> [[B:%.*]], <8 x i16> [[C:%.*]], i1 false)
65
// CHECK-GFX1200-NEXT:    store <8 x i16> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 16, !tbaa [[TBAA4]]
66
// CHECK-GFX1200-NEXT:    ret void
67
//
68
void test_amdgcn_wmma_bf16_16x16x16_bf16_w32(global v8s* out, v8s a, v8s b, v8s c)
69
{
70
  *out = __builtin_amdgcn_wmma_bf16_16x16x16_bf16_w32_gfx12(a, b, c);
71
}
72

73
//
74
// amdgcn_wmma_i32_16x16x16_iu8
75
//
76

77
// CHECK-GFX1200-LABEL: @test_amdgcn_wmma_i32_16x16x16_iu8_w32(
78
// CHECK-GFX1200-NEXT:  entry:
79
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <8 x i32> @llvm.amdgcn.wmma.i32.16x16x16.iu8.v8i32.v2i32(i1 true, <2 x i32> [[A:%.*]], i1 true, <2 x i32> [[B:%.*]], <8 x i32> [[C:%.*]], i1 false)
80
// CHECK-GFX1200-NEXT:    store <8 x i32> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 32, !tbaa [[TBAA4]]
81
// CHECK-GFX1200-NEXT:    ret void
82
//
83
void test_amdgcn_wmma_i32_16x16x16_iu8_w32(global v8i* out, v2i a, v2i b, v8i c)
84
{
85
  *out = __builtin_amdgcn_wmma_i32_16x16x16_iu8_w32_gfx12(true, a, true, b, c, false);
86
}
87

88
//
89
// amdgcn_wmma_i32_16x16x16_iu4
90
//
91

92
// CHECK-GFX1200-LABEL: @test_amdgcn_wmma_i32_16x16x16_iu4_w32(
93
// CHECK-GFX1200-NEXT:  entry:
94
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <8 x i32> @llvm.amdgcn.wmma.i32.16x16x16.iu4.v8i32.i32(i1 true, i32 [[A:%.*]], i1 true, i32 [[B:%.*]], <8 x i32> [[C:%.*]], i1 false)
95
// CHECK-GFX1200-NEXT:    store <8 x i32> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 32, !tbaa [[TBAA4]]
96
// CHECK-GFX1200-NEXT:    ret void
97
//
98
void test_amdgcn_wmma_i32_16x16x16_iu4_w32(global v8i* out, int a, int b, v8i c)
99
{
100
  *out = __builtin_amdgcn_wmma_i32_16x16x16_iu4_w32_gfx12(true, a, true, b, c, false);
101
}
102

103
// CHECK-GFX1200-LABEL: @test_amdgcn_wmma_f32_16x16x16_fp8_fp8_w32(
104
// CHECK-GFX1200-NEXT:  entry:
105
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <8 x float> @llvm.amdgcn.wmma.f32.16x16x16.fp8.fp8.v8f32.v2i32(<2 x i32> [[A:%.*]], <2 x i32> [[B:%.*]], <8 x float> [[C:%.*]])
106
// CHECK-GFX1200-NEXT:    store <8 x float> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 32, !tbaa [[TBAA4]]
107
// CHECK-GFX1200-NEXT:    ret void
108
//
109
void test_amdgcn_wmma_f32_16x16x16_fp8_fp8_w32(global v8f* out, v2i a, v2i b, v8f c)
110
{
111
  *out = __builtin_amdgcn_wmma_f32_16x16x16_fp8_fp8_w32_gfx12(a, b, c);
112
}
113

114
// CHECK-GFX1200-LABEL: @test_amdgcn_wmma_f32_16x16x16_fp8_bf8_w32(
115
// CHECK-GFX1200-NEXT:  entry:
116
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <8 x float> @llvm.amdgcn.wmma.f32.16x16x16.fp8.bf8.v8f32.v2i32(<2 x i32> [[A:%.*]], <2 x i32> [[B:%.*]], <8 x float> [[C:%.*]])
117
// CHECK-GFX1200-NEXT:    store <8 x float> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 32, !tbaa [[TBAA4]]
118
// CHECK-GFX1200-NEXT:    ret void
119
//
120
void test_amdgcn_wmma_f32_16x16x16_fp8_bf8_w32(global v8f* out, v2i a, v2i b, v8f c)
121
{
122
  *out = __builtin_amdgcn_wmma_f32_16x16x16_fp8_bf8_w32_gfx12(a, b, c);
123
}
124

125
// CHECK-GFX1200-LABEL: @test_amdgcn_wmma_f32_16x16x16_bf8_fp8_w32(
126
// CHECK-GFX1200-NEXT:  entry:
127
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <8 x float> @llvm.amdgcn.wmma.f32.16x16x16.bf8.fp8.v8f32.v2i32(<2 x i32> [[A:%.*]], <2 x i32> [[B:%.*]], <8 x float> [[C:%.*]])
128
// CHECK-GFX1200-NEXT:    store <8 x float> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 32, !tbaa [[TBAA4]]
129
// CHECK-GFX1200-NEXT:    ret void
130
//
131
void test_amdgcn_wmma_f32_16x16x16_bf8_fp8_w32(global v8f* out, v2i a, v2i b, v8f c)
132
{
133
  *out = __builtin_amdgcn_wmma_f32_16x16x16_bf8_fp8_w32_gfx12(a, b, c);
134
}
135

136
// CHECK-GFX1200-LABEL: @test_amdgcn_wmma_f32_16x16x16_bf8_bf8_w32(
137
// CHECK-GFX1200-NEXT:  entry:
138
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <8 x float> @llvm.amdgcn.wmma.f32.16x16x16.bf8.bf8.v8f32.v2i32(<2 x i32> [[A:%.*]], <2 x i32> [[B:%.*]], <8 x float> [[C:%.*]])
139
// CHECK-GFX1200-NEXT:    store <8 x float> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 32, !tbaa [[TBAA4]]
140
// CHECK-GFX1200-NEXT:    ret void
141
//
142
void test_amdgcn_wmma_f32_16x16x16_bf8_bf8_w32(global v8f* out, v2i a, v2i b, v8f c)
143
{
144
  *out = __builtin_amdgcn_wmma_f32_16x16x16_bf8_bf8_w32_gfx12(a, b, c);
145
}
146

147
// CHECK-GFX1200-LABEL: @test_amdgcn_wmma_i32_16x16x32_iu4_w32(
148
// CHECK-GFX1200-NEXT:  entry:
149
// CHECK-GFX1200-NEXT:    [[TMP0:%.*]] = tail call <8 x i32> @llvm.amdgcn.wmma.i32.16x16x32.iu4.v8i32.v2i32(i1 true, <2 x i32> [[A:%.*]], i1 true, <2 x i32> [[B:%.*]], <8 x i32> [[C:%.*]], i1 false)
150
// CHECK-GFX1200-NEXT:    store <8 x i32> [[TMP0]], ptr addrspace(1) [[OUT:%.*]], align 32, !tbaa [[TBAA4]]
151
// CHECK-GFX1200-NEXT:    ret void
152
//
153
void test_amdgcn_wmma_i32_16x16x32_iu4_w32(global v8i* out, v2i a, v2i b, v8i c)
154
{
155
  *out = __builtin_amdgcn_wmma_i32_16x16x32_iu4_w32_gfx12(true, a, true, b, c, false);
156
}
157
llvm-project

Использование cookies
