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
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
|
// RUN: mlir-opt --test-transform-dialect-interpreter -split-input-file -verify-diagnostics %s | FileCheck %s
#map = affine_map<()[s0] -> (-s0 + 12, 7)>
// CHECK-LABEL: @static_sizes_output_divisible
func.func @static_sizes_output_divisible(%arg0: tensor<24x12xf32>,
%arg1: tensor<12x25xf32>,
%arg2: tensor<24x25xf32>,
%iv0 : index, %iv1 : index, %iv2 : index) -> tensor<24x25xf32> {
%0 = affine.min #map()[%iv2]
// CHECK: %[[T0:.*]] = tensor.extract_slice %
// CHECK: %[[T1:.*]] = tensor.extract_slice %
// CHECK: %[[T2:.*]] = tensor.extract_slice %
%1 = tensor.extract_slice %arg0[%iv0, %iv2] [4, %0] [1, 1] : tensor<24x12xf32> to tensor<4x?xf32>
%2 = tensor.extract_slice %arg1[%iv2, %iv1] [%0, 5] [1, 1] : tensor<12x25xf32> to tensor<?x5xf32>
%3 = tensor.extract_slice %arg2[%iv0, %iv1] [4, 5] [1, 1] : tensor<24x25xf32> to tensor<4x5xf32>
// CHECK-DAG: %[[CST:.*]] = arith.constant 0.
// CHECK: %[[T3:.*]] = tensor.pad %[[T0]] nofold
// CHECK: tensor.yield %[[CST]]
// CHECK: %[[T4:.*]] = tensor.pad %[[T1]] nofold
// CHECK: %[[T5:.*]] = linalg.matmul
// CHECK-SAME: ins(%[[T3]], %[[T4]] : tensor<4x7xf32>, tensor<7x5xf32>)
// CHECK-SAME: outs(%[[T2]] : tensor<4x5xf32>)
%4 = linalg.matmul ins(%1, %2 : tensor<4x?xf32>, tensor<?x5xf32>) outs(%3 : tensor<4x5xf32>) -> tensor<4x5xf32>
%5 = tensor.insert_slice %4 into %arg2[%iv0, %iv1] [4, 5] [1, 1] : tensor<4x5xf32> into tensor<24x25xf32>
func.return %5 : tensor<24x25xf32>
}
transform.sequence failures(propagate) {
^bb1(%arg1: !transform.any_op):
%0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
%padded, %pad = transform.structured.pad %0 {
padding_values=[0.0 : f32, 0.0 : f32, 0.0 : f32],
padding_dimensions=[0, 1, 2],
pack_paddings=[1, 1, 0]
} : (!transform.any_op) -> (!transform.any_op, !transform.any_op)
}
// -----
#map = affine_map<()[s0] -> (-s0 + 12, 7)>
// CHECK-LABEL: @pad_to_multiple
func.func @pad_to_multiple(%arg0: tensor<24x12xf32>,
%arg1: tensor<12x25xf32>,
%arg2: tensor<24x25xf32>,
%iv0 : index, %iv1 : index, %iv2 : index) -> tensor<24x25xf32> {
%0 = affine.min #map()[%iv2]
%1 = tensor.extract_slice %arg0[%iv0, %iv2] [4, %0] [1, 1] : tensor<24x12xf32> to tensor<4x?xf32>
%2 = tensor.extract_slice %arg1[%iv2, %iv1] [%0, 5] [1, 1] : tensor<12x25xf32> to tensor<?x5xf32>
%3 = tensor.extract_slice %arg2[%iv0, %iv1] [4, 5] [1, 1] : tensor<24x25xf32> to tensor<4x5xf32>
// CHECK: linalg.matmul
// CHECK-SAME: ins(%{{.*}}, %{{.*}} : tensor<4x7xf32>, tensor<7x6xf32>)
// CHECK-SAME: outs(%{{.*}} : tensor<4x6xf32>)
%4 = linalg.matmul ins(%1, %2 : tensor<4x?xf32>, tensor<?x5xf32>) outs(%3 : tensor<4x5xf32>) -> tensor<4x5xf32>
%5 = tensor.insert_slice %4 into %arg2[%iv0, %iv1] [4, 5] [1, 1] : tensor<4x5xf32> into tensor<24x25xf32>
func.return %5 : tensor<24x25xf32>
}
transform.sequence failures(propagate) {
^bb1(%arg1: !transform.any_op):
%0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
%padded, %pad = transform.structured.pad %0 {
padding_values=[0.0 : f32, 0.0 : f32, 0.0 : f32],
padding_dimensions=[0, 1, 2],
pad_to_multiple_of=[2, 2, 1],
pack_paddings=[1, 1, 0]
} : (!transform.any_op) -> (!transform.any_op, !transform.any_op)
}
// -----
#map = affine_map<()[s0] -> (-s0 + 12, 7)>
// CHECK-LABEL: @static_sizes_output_divisible_on_empty_op
func.func @static_sizes_output_divisible_on_empty_op(%arg0: tensor<24x12xf32>,
%arg1: tensor<12x25xf32>, %arg2: tensor<24x25xf32>, %iv0: index,
%iv1: index, %iv2: index) -> tensor<24x25xf32> {
%0 = affine.min #map()[%iv2]
// CHECK: %[[T0:.*]] = tensor.empty
// CHECK: %[[T1:.*]] = tensor.empty
// CHECK: %[[T2:.*]] = tensor.empty
%1 = tensor.empty(%0) : tensor<4x?xf32>
%2 = tensor.empty(%0) : tensor<?x5xf32>
%3 = tensor.empty() : tensor<4x5xf32>
// CHECK-DAG: %[[CST:.*]] = arith.constant 0.
// CHECK: %[[T3:.*]] = tensor.pad %[[T0]] nofold
// CHECK: tensor.yield %[[CST]]
// CHECK: %[[T4:.*]] = tensor.pad %[[T1]] nofold
// CHECK: %[[T5:.*]] = linalg.matmul
// CHECK-SAME: ins(%[[T3]], %[[T4]] : tensor<4x7xf32>, tensor<7x5xf32>)
// CHECK-SAME: outs(%[[T2]] : tensor<4x5xf32>)
%4 = linalg.matmul ins(%1, %2 : tensor<4x?xf32>, tensor<?x5xf32>) outs(%3 : tensor<4x5xf32>) -> tensor<4x5xf32>
%5 = tensor.insert_slice %4 into %arg2[%iv0, %iv1] [4, 5] [1, 1] : tensor<4x5xf32> into tensor<24x25xf32>
func.return %5 : tensor<24x25xf32>
}
transform.sequence failures(propagate) {
^bb1(%arg1: !transform.any_op):
%0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
%padded, %pad = transform.structured.pad %0 {
padding_values=[0.0 : f32, 0.0 : f32, 0.0 : f32],
padding_dimensions=[0, 1, 2],
pack_paddings=[1, 1, 0]
} : (!transform.any_op) -> (!transform.any_op, !transform.any_op)
}
// -----
func.func @pad(%arg0: tensor<24x12xf32>,
%arg1: tensor<12x25xf32>,
%arg2: tensor<24x25xf32>) -> tensor<24x25xf32> {
// expected-note @below {{when applied to this op}}
%0 = linalg.matmul ins(%arg0, %arg1 : tensor<24x12xf32>, tensor<12x25xf32>) outs(%arg2 : tensor<24x25xf32>) -> tensor<24x25xf32>
func.return %0 : tensor<24x25xf32>
}
transform.sequence failures(propagate) {
^bb1(%arg1: !transform.any_op):
%0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
// expected-error @below {{op expects a padding value of type 'f32', got 0 : i32}}
%padded, %pad = transform.structured.pad %0 {
padding_values=[0: i32, 0.0 : f32, 0.0 : f32],
padding_dimensions=[0, 1, 2],
pack_paddings=[1, 1, 0]
} : (!transform.any_op) -> (!transform.any_op, !transform.any_op)
}
// -----
func.func @pad(%arg0: tensor<24x12xf32>,
%arg1: tensor<12x25xf32>,
%arg2: tensor<24x25xf32>) -> tensor<24x25xf32> {
// expected-note @below {{when applied to this op}}
%0 = linalg.matmul ins(%arg0, %arg1 : tensor<24x12xf32>, tensor<12x25xf32>) outs(%arg2 : tensor<24x25xf32>) -> tensor<24x25xf32>
func.return %0 : tensor<24x25xf32>
}
transform.sequence failures(propagate) {
^bb1(%arg1: !transform.any_op):
%0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
// expected-error @below {{expects a padding that parses to 'f32', got "{foo}"}}
%padded, %pad = transform.structured.pad %0 {
padding_values=["{foo}", 0.0 : f32, 0.0 : f32],
padding_dimensions=[0, 1, 2],
pack_paddings=[1, 1, 0]
} : (!transform.any_op) -> (!transform.any_op, !transform.any_op)
}
// -----
// CHECK-LABEL: @pad(
func.func @pad(%arg0: tensor<24x12xf32>,
%arg1: tensor<12x25xf32>,
%arg2: tensor<24x25xf32>) -> tensor<24x25xf32> {
// This is attached to an error that is silenceable and is not reported by this transform
// {{when applied to this op}}
%0 = linalg.matmul ins(%arg0, %arg1 : tensor<24x12xf32>, tensor<12x25xf32>) outs(%arg2 : tensor<24x25xf32>) -> tensor<24x25xf32>
func.return %0 : tensor<24x25xf32>
}
transform.sequence failures(suppress) {
^bb1(%arg1: !transform.any_op):
%0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
// This error is silenceable and is not reported by this transform
// {{transform.structured.pad failed to apply}}
%padded, %pad = transform.structured.pad %0 {
padding_values=[0.0 : f32, 0.0 : f32, 0.0 : f32],
padding_dimensions=[0, 1, 2],
pack_paddings=[1, 1, 0]
} : (!transform.any_op) -> (!transform.any_op, !transform.any_op)
}
// -----
// Check that the padding can be applied even when the output argument of the
// linalg op is not produced by an empty op or an extract_slice op.
// CHECK-DAG: #[[$MAP_MIN:.*]] = affine_map<(d0) -> (-d0 + 2044, 16)>
// CHECK-DAG: #[[$MAP_TO_16:.*]] = affine_map<(d0) -> (-d0 + 16)>
// CHECK-LABEL: @outs_not_produced_by_empty_or_extract_slice(
// CHECK-SAME: %[[A:[^: ]*]]: tensor<128x2044xf32>,
// CHECK-SAME: %[[B:[^: ]*]]: tensor<2044x128xf32>)
func.func @outs_not_produced_by_empty_or_extract_slice(%a : tensor<128x2044xf32>, %b : tensor<2044x128xf32>) -> tensor<128x128xf32> {
%cst = arith.constant 0.000000e+00 : f32
%0 = tensor.empty() : tensor<128x128xf32>
%9 = linalg.fill ins(%cst : f32) outs(%0 : tensor<128x128xf32>) -> tensor<128x128xf32>
%c0 = arith.constant 0 : index
%c16 = arith.constant 16 : index
%c2044 = arith.constant 2044 : index
// CHECK: scf.for %[[ARG3:.*]] = {{.*}} iter_args(%[[ARG4:.*]] = %{{.*}})
%10 = scf.for %arg3 = %c0 to %c2044 step %c16 iter_args(%arg4 = %9) -> (tensor<128x128xf32>) {
// CHECK: %[[MIN:.*]] = affine.min #[[$MAP_MIN]](%[[ARG3]])
%11 = affine.min affine_map<(d0) -> (-d0 + 2044, 16)>(%arg3)
// CHECK: %[[A_SLICE:.*]] = tensor.extract_slice %[[A]]
// CHECK: %[[B_SLICE:.*]] = tensor.extract_slice %[[B]]
%extracted_slice_2 = tensor.extract_slice %a[0, %arg3] [128, %11] [1, 1] : tensor<128x2044xf32> to tensor<128x?xf32>
%extracted_slice_3 = tensor.extract_slice %b[%arg3, 0] [%11, 128] [1, 1] : tensor<2044x128xf32> to tensor<?x128xf32>
// CHECK-DAG: %[[CST:.*]] = arith.constant 0.
// CHECK-DAG: %[[TO_16:.*]] = affine.apply #[[$MAP_TO_16]](%[[MIN]])
// CHECK: %[[PADDED_A_SLICE:.*]] = tensor.pad %[[A_SLICE]] nofold low[0, 0] high[0, %[[TO_16]]]
// CHECK: tensor.yield %[[CST]]
// CHECK: %[[PADDED_B_SLICE:.*]] = tensor.pad %[[B_SLICE]] nofold
// The output shape is already padded, so actually we shouldn't
// add anything to the upper bound.
// CHECK: %[[PADDED_ARG4:.*]] = tensor.pad %[[ARG4]] nofold low[{{.*}}] high[0, 0]
// CHECK: %[[T5:.*]] = linalg.matmul
// CHECK-SAME: ins(%[[PADDED_A_SLICE]], %[[PADDED_B_SLICE]] : tensor<128x16xf32>, tensor<16x128xf32>)
// CHECK-SAME: outs(%[[PADDED_ARG4]] : tensor<128x128xf32>)
%res = linalg.matmul ins(%extracted_slice_2, %extracted_slice_3 : tensor<128x?xf32>, tensor<?x128xf32>) outs(%arg4 : tensor<128x128xf32>) -> tensor<128x128xf32>
scf.yield %res : tensor<128x128xf32>
}
return %10 : tensor<128x128xf32>
}
transform.sequence failures(propagate) {
^bb1(%arg1: !transform.any_op):
%0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
%padded, %pad = transform.structured.pad %0 {
padding_values=[0.0 : f32, 0.0 : f32, 0.0 : f32],
padding_dimensions=[0, 1, 2],
pack_paddings=[1, 1, 1]
} : (!transform.any_op) -> (!transform.any_op, !transform.any_op)
}
// -----
#map = affine_map<()[s0] -> (-s0 + 12, 7)>
// CHECK-LABEL: @pack_everything
func.func @pack_everything(%arg0: tensor<24x12xf32>,
%arg1: tensor<12x25xf32>,
%arg2: tensor<24x25xf32>,
%iv0 : index, %iv1 : index, %iv2 : index) -> tensor<24x25xf32> {
%0 = affine.min #map()[%iv2]
// CHECK: %[[T0:.*]] = tensor.extract_slice %
// CHECK: %[[T1:.*]] = tensor.extract_slice %
// CHECK: %[[T2:.*]] = tensor.extract_slice %
%1 = tensor.extract_slice %arg0[%iv0, %iv2] [4, %0] [1, 1] : tensor<24x12xf32> to tensor<4x?xf32>
%2 = tensor.extract_slice %arg1[%iv2, %iv1] [%0, 5] [1, 1] : tensor<12x25xf32> to tensor<?x5xf32>
%3 = tensor.extract_slice %arg2[%iv0, %iv1] [4, 5] [1, 1] : tensor<24x25xf32> to tensor<4x5xf32>
// CHECK-DAG: %[[CST:.*]] = arith.constant 0.
// CHECK: %[[PAD0:.*]] = tensor.pad %[[T0]] nofold
// CHECK: %[[PAD1:.*]] = tensor.pad %[[T1]] nofold
// CHECK: %[[PAD2:.*]] = tensor.pad %[[T2]] nofold
// CHECK: %[[T5:.*]] = linalg.matmul
// CHECK-SAME: ins(%[[PAD0]], %[[PAD1]] : tensor<4x7xf32>, tensor<7x5xf32>)
// CHECK-SAME: outs(%[[PAD2]] : tensor<4x5xf32>)
// Get unpadded result (no-op in this example).
// CHECK: %[[T6:.*]] = tensor.extract_slice %[[T5]]
// Copy back result to the original buffer, so that the destination of the
// computation does not change.
// CHECK: %[[T7:.*]] = bufferization.copy_tensor %[[T6]], %[[T2]]
%4 = linalg.matmul ins(%1, %2 : tensor<4x?xf32>, tensor<?x5xf32>) outs(%3 : tensor<4x5xf32>) -> tensor<4x5xf32>
// CHECK: %[[T8:.*]] = tensor.insert_slice %[[T7]] into %{{.*}}
%5 = tensor.insert_slice %4 into %arg2[%iv0, %iv1] [4, 5] [1, 1] : tensor<4x5xf32> into tensor<24x25xf32>
func.return %5 : tensor<24x25xf32>
}
transform.sequence failures(propagate) {
^bb1(%arg1: !transform.any_op):
%0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
%padded, %pad = transform.structured.pad %0 {
padding_values=[0.0 : f32, 0.0 : f32, 0.0 : f32],
padding_dimensions=[0, 1, 2],
pack_paddings=[1, 1, 1]
} : (!transform.any_op) -> (!transform.any_op, !transform.any_op)
}
|
