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
|
// RUN: mlir-opt %s -test-transform-dialect-interpreter -split-input-file -verify-diagnostics | FileCheck %s
// CHECK-LABEL: @vectorize_matmul
// CHECK-SAME: %[[A:.*]]: tensor<24x12xf32>
// CHECK-SAME: %[[B:.*]]: tensor<12x25xf32>
// CHECK-SAME: %[[C:.*]]: tensor<24x25xf32>
func.func @vectorize_matmul(%arg0: tensor<24x12xf32>,
%arg1: tensor<12x25xf32>,
%arg2: tensor<24x25xf32>) -> tensor<24x25xf32> {
// CHECK: %[[vA:.+]] = vector.transfer_read %[[A]]
// CHECK: %[[vB:.+]] = vector.transfer_read %[[B]]
// CHECK: %[[vC:.+]] = vector.transfer_read %[[C]]
// CHECK: %[[vR:.+]] = vector.contract {{.*}} %[[vA]], %[[vB]], %[[vC]]
// CHECK: vector.transfer_write %[[vR]], %[[C]]
%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
%1 = get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
%2 = transform.structured.vectorize %1 : (!transform.any_op) -> !transform.any_op
}
// -----
// CHECK-LABEL: @vectorize_matmul_memref
// CHECK-SAME: %[[A:.*]]: memref<24x12xf32>
// CHECK-SAME: %[[B:.*]]: memref<12x25xf32>
// CHECK-SAME: %[[C:.*]]: memref<24x25xf32>
func.func @vectorize_matmul_memref(%arg0: memref<24x12xf32>,
%arg1: memref<12x25xf32>,
%arg2: memref<24x25xf32>) {
// CHECK: %[[vA:.+]] = vector.transfer_read %[[A]]
// CHECK: %[[vB:.+]] = vector.transfer_read %[[B]]
// CHECK: %[[vC:.+]] = vector.transfer_read %[[C]]
// CHECK: %[[vR:.+]] = vector.contract {{.*}} %[[vA]], %[[vB]], %[[vC]]
// CHECK: vector.transfer_write %[[vR]], %[[C]]
linalg.matmul ins(%arg0, %arg1 : memref<24x12xf32>, memref<12x25xf32>) outs(%arg2 : memref<24x25xf32>)
return
}
transform.sequence failures(propagate) {
^bb1(%arg1: !transform.any_op):
%0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
%1 = get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
%2 = transform.structured.vectorize %1 : (!transform.any_op) -> !transform.any_op
}
// -----
// CHECK-LABEL: @vectorize_copy_memref
// CHECK-SAME: %[[A:.*]]: memref<100x100xf32>,
// CHECK-SAME: %[[B:.*]]: memref<100x100xf32>
func.func @vectorize_copy_memref(%arg0: memref<100x100xf32>,
%arg1: memref<100x100xf32>) {
// CHECK: %[[vA:.+]] = vector.transfer_read %[[A]]
// CHECK: vector.transfer_write %[[vA]], %[[B]]
linalg.copy ins(%arg0 : memref<100x100xf32>) outs(%arg1 : memref<100x100xf32>)
return
}
transform.sequence failures(propagate) {
^bb1(%arg1: !transform.any_op):
%0 = transform.structured.match ops{["linalg.copy"]} in %arg1 : (!transform.any_op) -> !transform.any_op
%1 = get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
%2 = transform.structured.vectorize %1 : (!transform.any_op) -> !transform.any_op
}
// -----
#map0 = affine_map<()[s0] -> (-s0 + 12, 7)>
#map1 = affine_map<()[s0] -> (-s0 + 7)>
// CHECK-LABEL: @vectorize_keep_pad
// CHECK-SAME: %[[C:[a-zA-Z0-9_]+]]: tensor<24x25xf32>
func.func @vectorize_keep_pad(
%arg0: tensor<24x12xf32>, %arg1: tensor<12x25xf32>,
%arg2: tensor<24x25xf32>, %arg3: index, %arg4: index,
%arg5: index) -> tensor<24x25xf32> {
%c0 = arith.constant 0 : index
%cst = arith.constant 0.000000e+00 : f32
%0 = affine.min #map0()[%arg5]
%1 = tensor.extract_slice %arg0[%arg3, %arg5] [4, %0] [1, 1] : tensor<24x12xf32> to tensor<4x?xf32>
%2 = tensor.extract_slice %arg1[%arg5, %arg4] [%0, 5] [1, 1] : tensor<12x25xf32> to tensor<?x5xf32>
%3 = tensor.extract_slice %arg2[%arg3, %arg4] [4, 5] [1, 1] : tensor<24x25xf32> to tensor<4x5xf32>
%4 = affine.apply #map1()[%0]
// CHECK: %[[pA:.*]] = tensor.pad
%5 = tensor.pad %1 nofold low[%c0, %c0] high[%c0, %4] {
^bb0(%arg6: index, %arg7: index):
tensor.yield %cst : f32
} : tensor<4x?xf32> to tensor<4x7xf32>
%6 = affine.apply #map1()[%0]
// CHECK: %[[pB:.*]] = tensor.pad
%7 = tensor.pad %2 nofold low[%c0, %c0] high[%6, %c0] {
^bb0(%arg6: index, %arg7: index):
tensor.yield %cst : f32
} : tensor<?x5xf32> to tensor<7x5xf32>
// CHECK: %[[vA:.+]] = vector.transfer_read %[[pA]]
// CHECK: %[[vB:.+]] = vector.transfer_read %[[pB]]
// CHECK: %[[vC:.+]] = vector.transfer_read %[[C]]
// CHECK: %[[vR:.+]] = vector.contract {{.*}} %[[vA]], %[[vB]], %[[vC]]
// CHECK: vector.transfer_write %[[vR]], %[[C]]
%8 = linalg.matmul ins(%5, %7 : tensor<4x7xf32>, tensor<7x5xf32>) outs(%3 : tensor<4x5xf32>) -> tensor<4x5xf32>
%9 = tensor.insert_slice %8 into %arg2[%arg3, %arg4] [4, 5] [1, 1] : tensor<4x5xf32> into tensor<24x25xf32>
return %9 : 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
%1 = get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
%2 = transform.structured.vectorize %1 : (!transform.any_op) -> !transform.any_op
}
// -----
#map0 = affine_map<()[s0] -> (-s0 + 12, 7)>
#map1 = affine_map<()[s0] -> (-s0 + 7)>
// CHECK-LABEL: @vectorize_pad
// CHECK-SAME: %[[A:.+]]: tensor<24x12xf32>
// CHECK-SAME: %[[B:.+]]: tensor<12x25xf32>
// CHECK-SAME: %[[C:.+]]: tensor<24x25xf32>
func.func @vectorize_pad(
%arg0: tensor<24x12xf32>, %arg1: tensor<12x25xf32>,
%arg2: tensor<24x25xf32>, %arg3: index, %arg4: index,
%arg5: index) -> tensor<24x25xf32> {
%c0 = arith.constant 0 : index
%cst = arith.constant 0.000000e+00 : f32
%0 = affine.min #map0()[%arg5]
// CHECK: %[[sA:.+]] = tensor.extract_slice %[[A]]
// CHECK: %[[sB:.+]] = tensor.extract_slice %[[B]]
%1 = tensor.extract_slice %arg0[%arg3, %arg5] [4, %0] [1, 1] : tensor<24x12xf32> to tensor<4x?xf32>
%2 = tensor.extract_slice %arg1[%arg5, %arg4] [%0, 5] [1, 1] : tensor<12x25xf32> to tensor<?x5xf32>
%3 = tensor.extract_slice %arg2[%arg3, %arg4] [4, 5] [1, 1] : tensor<24x25xf32> to tensor<4x5xf32>
// CHECK: %[[vA:.+]] = vector.transfer_read %[[sA]]
%4 = affine.apply #map1()[%0]
%5 = tensor.pad %1 nofold low[%c0, %c0] high[%c0, %4] {
^bb0(%arg6: index, %arg7: index):
tensor.yield %cst : f32
} : tensor<4x?xf32> to tensor<4x7xf32>
%6 = affine.apply #map1()[%0]
// CHECK: %[[vB:.+]] = vector.transfer_read %[[sB]]
%7 = tensor.pad %2 nofold low[%c0, %c0] high[%6, %c0] {
^bb0(%arg6: index, %arg7: index):
tensor.yield %cst : f32
} : tensor<?x5xf32> to tensor<7x5xf32>
// CHECK: %[[vC:.+]] = vector.transfer_read %[[C]]
// CHECK: %[[vR:.+]] = vector.contract {{.*}} %[[vA]], %[[vB]], %[[vC]]
// CHECK: vector.transfer_write %[[vR]], %[[C]]
%8 = linalg.matmul ins(%5, %7 : tensor<4x7xf32>, tensor<7x5xf32>) outs(%3 : tensor<4x5xf32>) -> tensor<4x5xf32>
%9 = tensor.insert_slice %8 into %arg2[%arg3, %arg4] [4, 5] [1, 1] : tensor<4x5xf32> into tensor<24x25xf32>
return %9 : 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
%1 = get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
%2 = transform.structured.vectorize %1 {vectorize_padding} : (!transform.any_op) -> !transform.any_op
}
// -----
func.func @vectorize(%arg0: tensor<24x12xf32>,
%arg1: tensor<12x25xf32>,
%arg2: tensor<24x25xf32>) -> tensor<24x25xf32> {
// expected-note @below {{non-isolated target}}
%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 requires isolated-from-above targets}}
%2 = transform.structured.vectorize %0 : (!transform.any_op) -> !transform.any_op
}
|
