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
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
|
// RUN: mlir-opt -pass-pipeline="builtin.module(gpu.module(convert-gpu-to-llvm-spv))" -split-input-file -verify-diagnostics %s \
// RUN: | FileCheck --check-prefixes=CHECK-64,CHECK %s
// RUN: mlir-opt -pass-pipeline="builtin.module(gpu.module(convert-gpu-to-llvm-spv{index-bitwidth=32}))" -split-input-file -verify-diagnostics %s \
// RUN: | FileCheck --check-prefixes=CHECK-32,CHECK %s
gpu.module @builtins {
// CHECK-64: llvm.func spir_funccc @_Z14get_num_groupsj(i32) -> i64
// CHECK-64: llvm.func spir_funccc @_Z12get_local_idj(i32) -> i64
// CHECK-64: llvm.func spir_funccc @_Z14get_local_sizej(i32) -> i64
// CHECK-64: llvm.func spir_funccc @_Z13get_global_idj(i32) -> i64
// CHECK-64: llvm.func spir_funccc @_Z12get_group_idj(i32) -> i64
// CHECK-32: llvm.func spir_funccc @_Z14get_num_groupsj(i32) -> i32
// CHECK-32: llvm.func spir_funccc @_Z12get_local_idj(i32) -> i32
// CHECK-32: llvm.func spir_funccc @_Z14get_local_sizej(i32) -> i32
// CHECK-32: llvm.func spir_funccc @_Z13get_global_idj(i32) -> i32
// CHECK-32: llvm.func spir_funccc @_Z12get_group_idj(i32) -> i32
// CHECK-LABEL: gpu_block_id
func.func @gpu_block_id() -> (index, index, index) {
// CHECK: [[C0:%.*]] = llvm.mlir.constant(0 : i32) : i32
// CHECK-64: llvm.call spir_funccc @_Z12get_group_idj([[C0]]) : (i32) -> i64
// CHECK-32: llvm.call spir_funccc @_Z12get_group_idj([[C0]]) : (i32) -> i32
%block_id_x = gpu.block_id x
// CHECK: [[C1:%.*]] = llvm.mlir.constant(1 : i32) : i32
// CHECK-64: llvm.call spir_funccc @_Z12get_group_idj([[C1]]) : (i32) -> i64
// CHECK-32: llvm.call spir_funccc @_Z12get_group_idj([[C1]]) : (i32) -> i32
%block_id_y = gpu.block_id y
// CHECK: [[C2:%.*]] = llvm.mlir.constant(2 : i32) : i32
// CHECK-64: llvm.call spir_funccc @_Z12get_group_idj([[C2]]) : (i32) -> i64
// CHECK-32: llvm.call spir_funccc @_Z12get_group_idj([[C2]]) : (i32) -> i32
%block_id_z = gpu.block_id z
return %block_id_x, %block_id_y, %block_id_z : index, index, index
}
// CHECK-LABEL: gpu_global_id
func.func @gpu_global_id() -> (index, index, index) {
// CHECK: [[C0:%.*]] = llvm.mlir.constant(0 : i32) : i32
// CHECK-64: llvm.call spir_funccc @_Z13get_global_idj([[C0]]) : (i32) -> i64
// CHECK-32: llvm.call spir_funccc @_Z13get_global_idj([[C0]]) : (i32) -> i32
%global_id_x = gpu.global_id x
// CHECK: [[C1:%.*]] = llvm.mlir.constant(1 : i32) : i32
// CHECK-64: llvm.call spir_funccc @_Z13get_global_idj([[C1]]) : (i32) -> i64
// CHECK-32: llvm.call spir_funccc @_Z13get_global_idj([[C1]]) : (i32) -> i32
%global_id_y = gpu.global_id y
// CHECK: [[C2:%.*]] = llvm.mlir.constant(2 : i32) : i32
// CHECK-64: llvm.call spir_funccc @_Z13get_global_idj([[C2]]) : (i32) -> i64
// CHECK-32: llvm.call spir_funccc @_Z13get_global_idj([[C2]]) : (i32) -> i32
%global_id_z = gpu.global_id z
return %global_id_x, %global_id_y, %global_id_z : index, index, index
}
// CHECK-LABEL: gpu_block_dim
func.func @gpu_block_dim() -> (index, index, index) {
// CHECK: [[C0:%.*]] = llvm.mlir.constant(0 : i32) : i32
// CHECK-64: llvm.call spir_funccc @_Z14get_local_sizej([[C0]]) : (i32) -> i64
// CHECK-32: llvm.call spir_funccc @_Z14get_local_sizej([[C0]]) : (i32) -> i32
%block_dim_x = gpu.block_dim x
// CHECK: [[C1:%.*]] = llvm.mlir.constant(1 : i32) : i32
// CHECK-64: llvm.call spir_funccc @_Z14get_local_sizej([[C1]]) : (i32) -> i64
// CHECK-32: llvm.call spir_funccc @_Z14get_local_sizej([[C1]]) : (i32) -> i32
%block_dim_y = gpu.block_dim y
// CHECK: [[C2:%.*]] = llvm.mlir.constant(2 : i32) : i32
// CHECK-64: llvm.call spir_funccc @_Z14get_local_sizej([[C2]]) : (i32) -> i64
// CHECK-32: llvm.call spir_funccc @_Z14get_local_sizej([[C2]]) : (i32) -> i32
%block_dim_z = gpu.block_dim z
return %block_dim_x, %block_dim_y, %block_dim_z : index, index, index
}
// CHECK-LABEL: gpu_thread_id
func.func @gpu_thread_id() -> (index, index, index) {
// CHECK: [[C0:%.*]] = llvm.mlir.constant(0 : i32) : i32
// CHECK-64: llvm.call spir_funccc @_Z12get_local_idj([[C0]]) : (i32) -> i64
// CHECK-32: llvm.call spir_funccc @_Z12get_local_idj([[C0]]) : (i32) -> i32
%thread_id_x = gpu.thread_id x
// CHECK: [[C1:%.*]] = llvm.mlir.constant(1 : i32) : i32
// CHECK-64: llvm.call spir_funccc @_Z12get_local_idj([[C1]]) : (i32) -> i64
// CHECK-32: llvm.call spir_funccc @_Z12get_local_idj([[C1]]) : (i32) -> i32
%thread_id_y = gpu.thread_id y
// CHECK: [[C2:%.*]] = llvm.mlir.constant(2 : i32) : i32
// CHECK-64: llvm.call spir_funccc @_Z12get_local_idj([[C2]]) : (i32) -> i64
// CHECK-32: llvm.call spir_funccc @_Z12get_local_idj([[C2]]) : (i32) -> i32
%thread_id_z = gpu.thread_id z
return %thread_id_x, %thread_id_y, %thread_id_z : index, index, index
}
// CHECK-LABEL: gpu_grid_dim
func.func @gpu_grid_dim() -> (index, index, index) {
// CHECK: [[C0:%.*]] = llvm.mlir.constant(0 : i32) : i32
// CHECK-64: llvm.call spir_funccc @_Z14get_num_groupsj([[C0]]) : (i32) -> i64
// CHECK-32: llvm.call spir_funccc @_Z14get_num_groupsj([[C0]]) : (i32) -> i32
%grid_dim_x = gpu.grid_dim x
// CHECK: [[C1:%.*]] = llvm.mlir.constant(1 : i32) : i32
// CHECK-64: llvm.call spir_funccc @_Z14get_num_groupsj([[C1]]) : (i32) -> i64
// CHECK-32: llvm.call spir_funccc @_Z14get_num_groupsj([[C1]]) : (i32) -> i32
%grid_dim_y = gpu.grid_dim y
// CHECK: [[C2:%.*]] = llvm.mlir.constant(2 : i32) : i32
// CHECK-64: llvm.call spir_funccc @_Z14get_num_groupsj([[C2]]) : (i32) -> i64
// CHECK-32: llvm.call spir_funccc @_Z14get_num_groupsj([[C2]]) : (i32) -> i32
%grid_dim_z = gpu.grid_dim z
return %grid_dim_x, %grid_dim_y, %grid_dim_z : index, index, index
}
}
// -----
gpu.module @barriers {
// CHECK: llvm.func spir_funccc @_Z7barrierj(i32) attributes {convergent}
// CHECK-LABEL: gpu_barrier
func.func @gpu_barrier() {
// CHECK: [[FLAGS:%.*]] = llvm.mlir.constant(1 : i32) : i32
// CHECK: llvm.call spir_funccc @_Z7barrierj([[FLAGS]]) : (i32) -> ()
gpu.barrier
return
}
}
// -----
// Check `gpu.shuffle` conversion with default subgroup size.
gpu.module @shuffles {
// CHECK: llvm.func spir_funccc @_Z22sub_group_shuffle_downdj(f64, i32) -> f64 attributes {convergent}
// CHECK: llvm.func spir_funccc @_Z20sub_group_shuffle_upfj(f32, i32) -> f32 attributes {convergent}
// CHECK: llvm.func spir_funccc @_Z21sub_group_shuffle_xorlj(i64, i32) -> i64 attributes {convergent}
// CHECK: llvm.func spir_funccc @_Z17sub_group_shuffleij(i32, i32) -> i32 attributes {convergent}
// CHECK-LABEL: gpu_shuffles
// CHECK-SAME: (%[[VAL_0:.*]]: i32, %[[VAL_1:.*]]: i32, %[[VAL_2:.*]]: i64, %[[VAL_3:.*]]: i32, %[[VAL_4:.*]]: f32, %[[VAL_5:.*]]: i32, %[[VAL_6:.*]]: f64, %[[VAL_7:.*]]: i32)
func.func @gpu_shuffles(%val0: i32, %id: i32,
%val1: i64, %mask: i32,
%val2: f32, %delta_up: i32,
%val3: f64, %delta_down: i32) {
%width = arith.constant 32 : i32
// CHECK: llvm.call spir_funccc @_Z17sub_group_shuffleij(%[[VAL_0]], %[[VAL_1]]) : (i32, i32) -> i32
// CHECK: llvm.mlir.constant(true) : i1
// CHECK: llvm.call spir_funccc @_Z21sub_group_shuffle_xorlj(%[[VAL_2]], %[[VAL_3]]) : (i64, i32) -> i64
// CHECK: llvm.mlir.constant(true) : i1
// CHECK: llvm.call spir_funccc @_Z20sub_group_shuffle_upfj(%[[VAL_4]], %[[VAL_5]]) : (f32, i32) -> f32
// CHECK: llvm.mlir.constant(true) : i1
// CHECK: llvm.call spir_funccc @_Z22sub_group_shuffle_downdj(%[[VAL_6]], %[[VAL_7]]) : (f64, i32) -> f64
// CHECK: llvm.mlir.constant(true) : i1
%shuffleResult0, %valid0 = gpu.shuffle idx %val0, %id, %width : i32
%shuffleResult1, %valid1 = gpu.shuffle xor %val1, %mask, %width : i64
%shuffleResult2, %valid2 = gpu.shuffle up %val2, %delta_up, %width : f32
%shuffleResult3, %valid3 = gpu.shuffle down %val3, %delta_down, %width : f64
return
}
}
// -----
// Check `gpu.shuffle` conversion with explicit subgroup size.
gpu.module @shuffles attributes {
spirv.target_env = #spirv.target_env<#spirv.vce<v1.4, [Kernel, Addresses, GroupNonUniformShuffle, Int64], []>, #spirv.resource_limits<subgroup_size = 16>>
} {
// CHECK: llvm.func spir_funccc @_Z22sub_group_shuffle_downdj(f64, i32) -> f64 attributes {convergent}
// CHECK: llvm.func spir_funccc @_Z20sub_group_shuffle_upfj(f32, i32) -> f32 attributes {convergent}
// CHECK: llvm.func spir_funccc @_Z21sub_group_shuffle_xorlj(i64, i32) -> i64 attributes {convergent}
// CHECK: llvm.func spir_funccc @_Z17sub_group_shuffleij(i32, i32) -> i32 attributes {convergent}
// CHECK-LABEL: gpu_shuffles
// CHECK-SAME: (%[[VAL_0:.*]]: i32, %[[VAL_1:.*]]: i32, %[[VAL_2:.*]]: i64, %[[VAL_3:.*]]: i32, %[[VAL_4:.*]]: f32, %[[VAL_5:.*]]: i32, %[[VAL_6:.*]]: f64, %[[VAL_7:.*]]: i32)
func.func @gpu_shuffles(%val0: i32, %id: i32,
%val1: i64, %mask: i32,
%val2: f32, %delta_up: i32,
%val3: f64, %delta_down: i32) {
%width = arith.constant 16 : i32
// CHECK: llvm.call spir_funccc @_Z17sub_group_shuffleij(%[[VAL_0]], %[[VAL_1]]) : (i32, i32) -> i32
// CHECK: llvm.mlir.constant(true) : i1
// CHECK: llvm.call spir_funccc @_Z21sub_group_shuffle_xorlj(%[[VAL_2]], %[[VAL_3]]) : (i64, i32) -> i64
// CHECK: llvm.mlir.constant(true) : i1
// CHECK: llvm.call spir_funccc @_Z20sub_group_shuffle_upfj(%[[VAL_4]], %[[VAL_5]]) : (f32, i32) -> f32
// CHECK: llvm.mlir.constant(true) : i1
// CHECK: llvm.call spir_funccc @_Z22sub_group_shuffle_downdj(%[[VAL_6]], %[[VAL_7]]) : (f64, i32) -> f64
// CHECK: llvm.mlir.constant(true) : i1
%shuffleResult0, %valid0 = gpu.shuffle idx %val0, %id, %width : i32
%shuffleResult1, %valid1 = gpu.shuffle xor %val1, %mask, %width : i64
%shuffleResult2, %valid2 = gpu.shuffle up %val2, %delta_up, %width : f32
%shuffleResult3, %valid3 = gpu.shuffle down %val3, %delta_down, %width : f64
return
}
}
// -----
// Cannot convert due to shuffle width and target subgroup size mismatch
gpu.module @shuffles_mismatch {
func.func @gpu_shuffles(%val: i32, %id: i32) {
%width = arith.constant 16 : i32
// expected-error@below {{failed to legalize operation 'gpu.shuffle' that was explicitly marked illegal}}
%shuffleResult, %valid = gpu.shuffle idx %val, %id, %width : i32
return
}
}
// -----
// Cannot convert due to variable shuffle width
gpu.module @shuffles_mismatch {
func.func @gpu_shuffles(%val: i32, %id: i32, %width: i32) {
// expected-error@below {{failed to legalize operation 'gpu.shuffle' that was explicitly marked illegal}}
%shuffleResult, %valid = gpu.shuffle idx %val, %id, %width : i32
return
}
}
|
