vector-transfer-full-partial-split.mlir
10.9 KB
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
// RUN: mlir-opt %s -test-vector-transfer-full-partial-split | FileCheck %s
// RUN: mlir-opt %s -test-vector-transfer-full-partial-split=use-linalg-copy | FileCheck %s --check-prefix=LINALG
// CHECK-DAG: #[[$map_p4:.*]] = affine_map<()[s0] -> (s0 + 4)>
// CHECK-DAG: #[[$map_p8:.*]] = affine_map<()[s0] -> (s0 + 8)>
// CHECK-DAG: #[[$map_2d_stride_1:.*]] = affine_map<(d0, d1)[s0, s1] -> (d0 * s1 + s0 + d1)>
// LINALG-DAG: #[[$map_p4:.*]] = affine_map<()[s0] -> (s0 + 4)>
// LINALG-DAG: #[[$map_p8:.*]] = affine_map<()[s0] -> (s0 + 8)>
// LINALG-DAG: #[[$map_2d_stride_1:.*]] = affine_map<(d0, d1)[s0, s1] -> (d0 * s1 + s0 + d1)>
// LINALG-DAG: #[[$map_2d_dynamic:.*]] = affine_map<(d0, d1)[s0, s1, s2] -> (d0 * s1 + s0 + d1 * s2)>
// LINALG-DAG: #[[$bounds_map_4:.*]] = affine_map<(d0, d1, d2) -> (d0 - d1, 4)>
// LINALG-DAG: #[[$bounds_map_8:.*]] = affine_map<(d0, d1, d2) -> (d0 - d1, 8)>
// CHECK-LABEL: split_vector_transfer_read_2d(
// CHECK-SAME: %[[A:[a-zA-Z0-9]*]]: memref
// CHECK-SAME: %[[i:[a-zA-Z0-9]*]]: index
// CHECK-SAME: %[[j:[a-zA-Z0-9]*]]: index
// LINALG-LABEL: split_vector_transfer_read_2d(
// LINALG-SAME: %[[A:[a-zA-Z0-9]*]]: memref
// LINALG-SAME: %[[i:[a-zA-Z0-9]*]]: index
// LINALG-SAME: %[[j:[a-zA-Z0-9]*]]: index
func @split_vector_transfer_read_2d(%A: memref<?x8xf32>, %i: index, %j: index) -> vector<4x8xf32> {
%c0 = constant 0 : index
%f0 = constant 0.0 : f32
// CHECK-DAG: %[[c0:.*]] = constant 0 : index
// CHECK-DAG: %[[c8:.*]] = constant 8 : index
// CHECK-DAG: %[[cst:.*]] = constant 0.000000e+00 : f32
// alloca for boundary full tile
// CHECK: %[[alloc:.*]] = alloca() {alignment = 32 : i64} : memref<4x8xf32>
// %i + 4 <= dim(%A, 0)
// CHECK: %[[idx0:.*]] = affine.apply #[[$map_p4]]()[%[[i]]]
// CHECK: %[[d0:.*]] = dim %[[A]], %[[c0]] : memref<?x8xf32>
// CHECK: %[[cmp0:.*]] = cmpi "sle", %[[idx0]], %[[d0]] : index
// %j + 8 <= dim(%A, 1)
// CHECK: %[[idx1:.*]] = affine.apply #[[$map_p8]]()[%[[j]]]
// CHECK: %[[cmp1:.*]] = cmpi "sle", %[[idx1]], %[[c8]] : index
// are both conds true
// CHECK: %[[cond:.*]] = and %[[cmp0]], %[[cmp1]] : i1
// CHECK: %[[ifres:.*]]:3 = scf.if %[[cond]] -> (memref<?x8xf32>, index, index) {
// inBounds, just yield %A
// CHECK: scf.yield %[[A]], %[[i]], %[[j]] : memref<?x8xf32>, index, index
// CHECK: } else {
// slow path, fill tmp alloc and yield a memref_casted version of it
// CHECK: %[[slow:.*]] = vector.transfer_read %[[A]][%[[i]], %[[j]]], %cst :
// CHECK-SAME: memref<?x8xf32>, vector<4x8xf32>
// CHECK: %[[cast_alloc:.*]] = vector.type_cast %[[alloc]] :
// CHECK-SAME: memref<4x8xf32> to memref<vector<4x8xf32>>
// CHECK: store %[[slow]], %[[cast_alloc]][] : memref<vector<4x8xf32>>
// CHECK: %[[yielded:.*]] = memref_cast %[[alloc]] :
// CHECK-SAME: memref<4x8xf32> to memref<?x8xf32>
// CHECK: scf.yield %[[yielded]], %[[c0]], %[[c0]] :
// CHECK-SAME: memref<?x8xf32>, index, index
// CHECK: }
// CHECK: %[[res:.*]] = vector.transfer_read %[[ifres]]#0[%[[ifres]]#1, %[[ifres]]#2], %[[cst]]
// CHECK_SAME: {masked = [false, false]} : memref<?x8xf32>, vector<4x8xf32>
// LINALG-DAG: %[[c0:.*]] = constant 0 : index
// LINALG-DAG: %[[c1:.*]] = constant 1 : index
// LINALG-DAG: %[[c4:.*]] = constant 4 : index
// LINALG-DAG: %[[c8:.*]] = constant 8 : index
// LINALG-DAG: %[[cst:.*]] = constant 0.000000e+00 : f32
// alloca for boundary full tile
// LINALG: %[[alloc:.*]] = alloca() {alignment = 32 : i64} : memref<4x8xf32>
// %i + 4 <= dim(%A, 0)
// LINALG: %[[idx0:.*]] = affine.apply #[[$map_p4]]()[%[[i]]]
// LINALG: %[[d0:.*]] = dim %[[A]], %[[c0]] : memref<?x8xf32>
// LINALG: %[[cmp0:.*]] = cmpi "sle", %[[idx0]], %[[d0]] : index
// %j + 8 <= dim(%A, 1)
// LINALG: %[[idx1:.*]] = affine.apply #[[$map_p8]]()[%[[j]]]
// LINALG: %[[cmp1:.*]] = cmpi "sle", %[[idx1]], %[[c8]] : index
// are both conds true
// LINALG: %[[cond:.*]] = and %[[cmp0]], %[[cmp1]] : i1
// LINALG: %[[ifres:.*]]:3 = scf.if %[[cond]] -> (memref<?x8xf32>, index, index) {
// inBounds, just yield %A
// LINALG: scf.yield %[[A]], %[[i]], %[[j]] : memref<?x8xf32>, index, index
// LINALG: } else {
// slow path, fill tmp alloc and yield a memref_casted version of it
// LINALG: linalg.fill(%[[alloc]], %[[cst]]) : memref<4x8xf32>, f32
// LINALG: %[[d0:.*]] = dim %[[A]], %[[c0]] : memref<?x8xf32>
// LINALG: %[[sv0:.*]] = affine.min #[[$bounds_map_4]](%[[d0]], %[[i]], %[[c4]])
// LINALG: %[[sv1:.*]] = affine.min #[[$bounds_map_8]](%[[c8]], %[[j]], %[[c8]])
// LINALG: %[[sv:.*]] = subview %[[A]][%[[i]], %[[j]]] [%[[sv0]], %[[sv1]]] [%[[c1]], %[[c1]]]
// LINALG-SAME: memref<?x8xf32> to memref<?x?xf32, #[[$map_2d_dynamic]]>
// LINALG: linalg.copy(%[[sv]], %[[alloc]]) : memref<?x?xf32, #[[$map_2d_dynamic]]>, memref<4x8xf32>
// LINALG: %[[yielded:.*]] = memref_cast %[[alloc]] :
// LINALG-SAME: memref<4x8xf32> to memref<?x8xf32>
// LINALG: scf.yield %[[yielded]], %[[c0]], %[[c0]] :
// LINALG-SAME: memref<?x8xf32>, index, index
// LINALG: }
// LINALG: %[[res:.*]] = vector.transfer_read %[[ifres]]#0[%[[ifres]]#1, %[[ifres]]#2], %[[cst]]
// LINALG_SAME: {masked = [false, false]} : memref<?x8xf32>, vector<4x8xf32>
%1 = vector.transfer_read %A[%i, %j], %f0 : memref<?x8xf32>, vector<4x8xf32>
// LINALG: return %[[res]] : vector<4x8xf32>
return %1: vector<4x8xf32>
}
// CHECK-LABEL: split_vector_transfer_read_strided_2d(
// CHECK-SAME: %[[A:[a-zA-Z0-9]*]]: memref
// CHECK-SAME: %[[i:[a-zA-Z0-9]*]]: index
// CHECK-SAME: %[[j:[a-zA-Z0-9]*]]: index
// LINALG-LABEL: split_vector_transfer_read_strided_2d(
// LINALG-SAME: %[[A:[a-zA-Z0-9]*]]: memref
// LINALG-SAME: %[[i:[a-zA-Z0-9]*]]: index
// LINALG-SAME: %[[j:[a-zA-Z0-9]*]]: index
func @split_vector_transfer_read_strided_2d(
%A: memref<7x8xf32, offset:?, strides:[?, 1]>,
%i: index, %j: index) -> vector<4x8xf32> {
%c0 = constant 0 : index
%f0 = constant 0.0 : f32
// CHECK-DAG: %[[c0:.*]] = constant 0 : index
// CHECK-DAG: %[[c7:.*]] = constant 7 : index
// CHECK-DAG: %[[c8:.*]] = constant 8 : index
// CHECK-DAG: %[[cst:.*]] = constant 0.000000e+00 : f32
// alloca for boundary full tile
// CHECK: %[[alloc:.*]] = alloca() {alignment = 32 : i64} : memref<4x8xf32>
// %i + 4 <= dim(%A, 0)
// CHECK: %[[idx0:.*]] = affine.apply #[[$map_p4]]()[%[[i]]]
// CHECK: %[[cmp0:.*]] = cmpi "sle", %[[idx0]], %[[c7]] : index
// %j + 8 <= dim(%A, 1)
// CHECK: %[[idx1:.*]] = affine.apply #[[$map_p8]]()[%[[j]]]
// CHECK: %[[cmp1:.*]] = cmpi "sle", %[[idx1]], %[[c8]] : index
// are both conds true
// CHECK: %[[cond:.*]] = and %[[cmp0]], %[[cmp1]] : i1
// CHECK: %[[ifres:.*]]:3 = scf.if %[[cond]] -> (memref<?x8xf32, #[[$map_2d_stride_1]]>, index, index) {
// inBounds but not cast-compatible: yield a memref_casted form of %A
// CHECK: %[[casted:.*]] = memref_cast %arg0 :
// CHECK-SAME: memref<7x8xf32, #[[$map_2d_stride_1]]> to memref<?x8xf32, #[[$map_2d_stride_1]]>
// CHECK: scf.yield %[[casted]], %[[i]], %[[j]] :
// CHECK-SAME: memref<?x8xf32, #[[$map_2d_stride_1]]>, index, index
// CHECK: } else {
// slow path, fill tmp alloc and yield a memref_casted version of it
// CHECK: %[[slow:.*]] = vector.transfer_read %[[A]][%[[i]], %[[j]]], %cst :
// CHECK-SAME: memref<7x8xf32, #[[$map_2d_stride_1]]>, vector<4x8xf32>
// CHECK: %[[cast_alloc:.*]] = vector.type_cast %[[alloc]] :
// CHECK-SAME: memref<4x8xf32> to memref<vector<4x8xf32>>
// CHECK: store %[[slow]], %[[cast_alloc]][] :
// CHECK-SAME: memref<vector<4x8xf32>>
// CHECK: %[[yielded:.*]] = memref_cast %[[alloc]] :
// CHECK-SAME: memref<4x8xf32> to memref<?x8xf32, #[[$map_2d_stride_1]]>
// CHECK: scf.yield %[[yielded]], %[[c0]], %[[c0]] :
// CHECK-SAME: memref<?x8xf32, #[[$map_2d_stride_1]]>, index, index
// CHECK: }
// CHECK: %[[res:.*]] = vector.transfer_read {{.*}} {masked = [false, false]} :
// CHECK-SAME: memref<?x8xf32, #[[$map_2d_stride_1]]>, vector<4x8xf32>
// LINALG-DAG: %[[c0:.*]] = constant 0 : index
// LINALG-DAG: %[[c1:.*]] = constant 1 : index
// LINALG-DAG: %[[c4:.*]] = constant 4 : index
// LINALG-DAG: %[[c7:.*]] = constant 7 : index
// LINALG-DAG: %[[c8:.*]] = constant 8 : index
// LINALG-DAG: %[[cst:.*]] = constant 0.000000e+00 : f32
// alloca for boundary full tile
// LINALG: %[[alloc:.*]] = alloca() {alignment = 32 : i64} : memref<4x8xf32>
// %i + 4 <= dim(%A, 0)
// LINALG: %[[idx0:.*]] = affine.apply #[[$map_p4]]()[%[[i]]]
// LINALG: %[[cmp0:.*]] = cmpi "sle", %[[idx0]], %[[c7]] : index
// %j + 8 <= dim(%A, 1)
// LINALG: %[[idx1:.*]] = affine.apply #[[$map_p8]]()[%[[j]]]
// LINALG: %[[cmp1:.*]] = cmpi "sle", %[[idx1]], %[[c8]] : index
// are both conds true
// LINALG: %[[cond:.*]] = and %[[cmp0]], %[[cmp1]] : i1
// LINALG: %[[ifres:.*]]:3 = scf.if %[[cond]] -> (memref<?x8xf32, #[[$map_2d_stride_1]]>, index, index) {
// inBounds but not cast-compatible: yield a memref_casted form of %A
// LINALG: %[[casted:.*]] = memref_cast %arg0 :
// LINALG-SAME: memref<7x8xf32, #[[$map_2d_stride_1]]> to memref<?x8xf32, #[[$map_2d_stride_1]]>
// LINALG: scf.yield %[[casted]], %[[i]], %[[j]] :
// LINALG-SAME: memref<?x8xf32, #[[$map_2d_stride_1]]>, index, index
// LINALG: } else {
// slow path, fill tmp alloc and yield a memref_casted version of it
// LINALG: linalg.fill(%[[alloc]], %[[cst]]) : memref<4x8xf32>, f32
// LINALG: %[[sv0:.*]] = affine.min #[[$bounds_map_4]](%[[c7]], %[[i]], %[[c4]])
// LINALG: %[[sv1:.*]] = affine.min #[[$bounds_map_8]](%[[c8]], %[[j]], %[[c8]])
// LINALG: %[[sv:.*]] = subview %[[A]][%[[i]], %[[j]]] [%[[sv0]], %[[sv1]]] [%[[c1]], %[[c1]]]
// LINALG-SAME: memref<7x8xf32, #[[$map_2d_stride_1]]> to memref<?x?xf32, #[[$map_2d_dynamic]]>
// LINALG: linalg.copy(%[[sv]], %[[alloc]]) : memref<?x?xf32, #[[$map_2d_dynamic]]>, memref<4x8xf32>
// LINALG: %[[yielded:.*]] = memref_cast %[[alloc]] :
// LINALG-SAME: memref<4x8xf32> to memref<?x8xf32, #[[$map_2d_stride_1]]>
// LINALG: scf.yield %[[yielded]], %[[c0]], %[[c0]] :
// LINALG-SAME: memref<?x8xf32, #[[$map_2d_stride_1]]>, index, index
// LINALG: }
// LINALG: %[[res:.*]] = vector.transfer_read {{.*}} {masked = [false, false]} :
// LINALG-SAME: memref<?x8xf32, #[[$map_2d_stride_1]]>, vector<4x8xf32>
%1 = vector.transfer_read %A[%i, %j], %f0 :
memref<7x8xf32, offset:?, strides:[?, 1]>, vector<4x8xf32>
// CHECK: return %[[res]] : vector<4x8xf32>
return %1 : vector<4x8xf32>
}