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
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
|
// RUN: mlir-opt --transform-interpreter %s -split-input-file -verify-diagnostics | FileCheck %s
// Test One-Shot Bufferize.
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
%0 = transform.structured.match ops{["func.func"]} in %arg1 : (!transform.any_op) -> !transform.any_op
%1 = transform.bufferization.one_shot_bufferize %0 : (!transform.any_op) -> !transform.any_op
transform.yield
}
}
// CHECK-LABEL: func @test_function(
// CHECK-SAME: %[[A:.*]]: tensor<?xf32>
func.func @test_function(%A : tensor<?xf32>, %v : vector<4xf32>) -> (tensor<?xf32>) {
%c0 = arith.constant 0 : index
// CHECK: %[[A_memref:.*]] = bufferization.to_memref %[[A]]
// CHECK: %[[dim:.*]] = memref.dim %[[A_memref]]
// CHECK: %[[alloc:.*]] = memref.alloc(%[[dim]])
// CHECK: memref.copy %[[A_memref]], %[[alloc]]
// CHECK: vector.transfer_write %{{.*}}, %[[alloc]]
// CHECK: %[[res_tensor:.*]] = bufferization.to_tensor %[[alloc]]
%0 = vector.transfer_write %v, %A[%c0] : vector<4xf32>, tensor<?xf32>
// CHECK: return %[[res_tensor]]
return %0 : tensor<?xf32>
}
// -----
// Emit linalg.copy instead of memref.copy.
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
%0 = transform.structured.match ops{["func.func"]} in %arg1 : (!transform.any_op) -> !transform.any_op
%1 = transform.bufferization.one_shot_bufferize %0 {memcpy_op = "linalg.copy"} : (!transform.any_op) -> !transform.any_op
transform.yield
}
}
// CHECK-LABEL: func @test_function(
// CHECK-SAME: %[[A:.*]]: tensor<?xf32>
// CHECK-NOT: memref.copy
func.func @test_function(%A : tensor<?xf32>, %v : vector<4xf32>) -> (tensor<?xf32>) {
%c0 = arith.constant 0 : index
// CHECK: %[[A_memref:.*]] = bufferization.to_memref %[[A]]
// CHECK: %[[dim:.*]] = memref.dim %[[A_memref]]
// CHECK: %[[alloc:.*]] = memref.alloc(%[[dim]])
// CHECK: linalg.copy ins(%[[A_memref]] : memref<{{.*}}>) outs(%[[alloc]]
// CHECK: vector.transfer_write %{{.*}}, %[[alloc]]
// CHECK: %[[res_tensor:.*]] = bufferization.to_tensor %[[alloc]]
%0 = vector.transfer_write %v, %A[%c0] : vector<4xf32>, tensor<?xf32>
// CHECK: return %[[res_tensor]]
return %0 : tensor<?xf32>
}
// -----
// Test analysis of One-Shot Bufferize only.
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
%0 = transform.structured.match ops{["func.func"]} in %arg1 : (!transform.any_op) -> !transform.any_op
%1 = transform.bufferization.one_shot_bufferize %0
{test_analysis_only = true} : (!transform.any_op) -> !transform.any_op
transform.yield
}
}
// CHECK-LABEL: func @test_function_analysis(
// CHECK-SAME: %[[A:.*]]: tensor<?xf32>
func.func @test_function_analysis(%A : tensor<?xf32>, %v : vector<4xf32>) -> (tensor<?xf32>) {
%c0 = arith.constant 0 : index
// CHECK: vector.transfer_write
// CHECK-SAME: {__inplace_operands_attr__ = ["none", "false", "none"]}
// CHECK-SAME: tensor<?xf32>
%0 = vector.transfer_write %v, %A[%c0] : vector<4xf32>, tensor<?xf32>
return %0 : tensor<?xf32>
}
// -----
// Test One-Shot Bufferize transform failure with an unknown op. This would be
// allowed with `allow_unknown_ops`.
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
%0 = transform.structured.match ops{["func.func"]} in %arg1 : (!transform.any_op) -> !transform.any_op
// expected-error @+1 {{bufferization failed}}
%1 = transform.bufferization.one_shot_bufferize %0 : (!transform.any_op) -> !transform.any_op
transform.yield
}
}
func.func @test_unknown_op_failure() -> (tensor<?xf32>) {
// expected-error @+1 {{op was not bufferized}}
%0 = "test.dummy_op"() : () -> (tensor<?xf32>)
return %0 : tensor<?xf32>
}
// -----
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.consumed}) {
// %arg1 is the module
%0 = transform.bufferization.one_shot_bufferize %arg1 : (!transform.any_op) -> !transform.any_op
transform.yield
}
}
module {
// CHECK-LABEL: func @test_function(
// CHECK-SAME: %[[A:.*]]: tensor<?xf32>
func.func @test_function(%A : tensor<?xf32>, %v : vector<4xf32>) -> (tensor<?xf32>) {
%c0 = arith.constant 0 : index
// CHECK: %[[A_memref:.*]] = bufferization.to_memref %[[A]]
// CHECK: %[[dim:.*]] = memref.dim %[[A_memref]]
// CHECK: %[[alloc:.*]] = memref.alloc(%[[dim]])
// CHECK: memref.copy %[[A_memref]], %[[alloc]]
// CHECK: vector.transfer_write %{{.*}}, %[[alloc]]
// CHECK: %[[res_tensor:.*]] = bufferization.to_tensor %[[alloc]]
%0 = vector.transfer_write %v, %A[%c0] : vector<4xf32>, tensor<?xf32>
// CHECK: return %[[res_tensor]]
return %0 : tensor<?xf32>
}
}
// -----
// Test we use identity layout at function boundaries.
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.consumed}) {
%0 = transform.bufferization.one_shot_bufferize layout{IdentityLayoutMap} %arg1
{ bufferize_function_boundaries = true } : (!transform.any_op) -> !transform.any_op
transform.yield
}
}
// CHECK: func.func @matmul(
// CHECK-SAME: %[[A:.*]]: memref<12x9xf32>,
// CHECK-SAME: %[[B:.*]]: memref<9x6xf32>,
// CHECK-SAME: %[[C:.*]]: memref<12x6xf32>) -> memref<12x6xf32> {
func.func @matmul(%A: tensor<12x9xf32>, %B: tensor<9x6xf32>, %C: tensor<12x6xf32>) -> tensor<12x6xf32> {
// CHECK: linalg.matmul ins(%[[A]], %[[B]] : memref<12x9xf32>, memref<9x6xf32>) outs(%[[C]] : memref<12x6xf32>)
%D = linalg.matmul ins(%A, %B: tensor<12x9xf32>, tensor<9x6xf32>) outs(%C: tensor<12x6xf32>) -> tensor<12x6xf32>
// CHECK: return %[[C]] : memref<12x6xf32>
return %D : tensor<12x6xf32>
}
// -----
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
%0 = transform.structured.match ops{["tensor.empty"]} in %arg1 : (!transform.any_op) -> !transform.any_op
%1 = transform.cast %0 : !transform.any_op to !transform.op<"tensor.empty">
transform.bufferization.empty_tensor_to_alloc_tensor %1 : (!transform.op<"tensor.empty">) -> !transform.op<"bufferization.alloc_tensor">
transform.yield
}
}
// Expect `bufferization.empty_tensor_to_alloc_tensor` to replace the tensor.empty.
func.func @empty_to_tensor_alloc() -> tensor<2x2xf32> {
// CHECK: bufferization.alloc_tensor
%0 = tensor.empty() : tensor<2x2xf32>
return %0 : tensor<2x2xf32>
}
// -----
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
%0 = transform.structured.match ops{["func.func"]} in %arg1 : (!transform.any_op) -> !transform.any_op
transform.bufferization.eliminate_empty_tensors %0 : !transform.any_op
transform.yield
}
}
// CHECK-LABEL: func @empty_tensor_elimination(
// CHECK: tensor.extract_slice
// CHECK: linalg.fill
// CHECK: tensor.insert_slice
func.func @empty_tensor_elimination(
%t: tensor<10xf32>, %f: f32) -> tensor<10xf32> {
%0 = tensor.empty() : tensor<5xf32>
%1 = linalg.fill ins(%f : f32) outs(%0 : tensor<5xf32>) -> tensor<5xf32>
%2 = tensor.insert_slice %1 into %t [1][5][1]
: tensor<5xf32> into tensor<10xf32>
return %2 : tensor<10xf32>
}
// -----
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
%0 = transform.structured.match ops{["func.func"]} in %arg1 : (!transform.any_op) -> !transform.any_op
transform.bufferization.buffer_loop_hoisting %0 : !transform.any_op
transform.yield
}
}
// CHECK-LABEL: func @buffer_loop_hoisting(
// CHECK: memref.alloca
// CHECK: scf.for
// CHECK: memref.store
func.func @buffer_loop_hoisting(%lb: index, %ub: index, %step: index, %f: f32, %pos: index) {
scf.for %iv = %lb to %ub step %step {
%0 = memref.alloca() : memref<5xf32>
memref.store %f, %0[%pos] : memref<5xf32>
}
return
}
// -----
module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
%alloc_tensor = transform.structured.match ops{["bufferization.alloc_tensor"]} in %arg1
: (!transform.any_op) -> !transform.op<"bufferization.alloc_tensor">
%2, %new = transform.structured.bufferize_to_allocation %alloc_tensor
{alloc_op = "memref.alloca"}
: !transform.op<"bufferization.alloc_tensor">
transform.yield
}
}
// Expect `bufferization.bufferize_to_allocation` to create an alloc.
// CHECK-LABEL: func.func @empty_to_tensor_alloc()
func.func @empty_to_tensor_alloc() -> tensor<2x2xf32> {
// CHECK-NEXT: %[[alloca:.*]] = memref.alloca() : memref<2x2xf32>
// CHECK-NEXT: %[[tensor:.*]] = bufferization.to_tensor %[[alloca]] restrict writable : memref<2x2xf32>
// CHECK-NEXT: return %[[tensor]] : tensor<2x2xf32>
%0 = bufferization.alloc_tensor() : tensor<2x2xf32>
return %0 : tensor<2x2xf32>
}
|
