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
|
// RUN: mlir-opt --arith-int-narrowing="int-bitwidths-supported=1,8,16,32" \
// RUN: --verify-diagnostics %s | FileCheck %s
// Check that we can calculate `linalg.index` value bounds and use them to
// optimize index casts.
//===----------------------------------------------------------------------===//
// arith.index_cast
//===----------------------------------------------------------------------===//
// CHECK-LABEL: func @linalg_indexcast_dim_0_i8
// CHECK: %[[IDX:.+]] = linalg.index 0 : index
// CHECK-NEXT: %[[INT:.+]] = arith.index_cast %[[IDX]] : index to i8
// CHECK-NEXT: %[[FP:.+]] = arith.sitofp %[[INT]] : i8 to f16
// CHECK-NEXT: linalg.yield %[[FP]] : f16
func.func @linalg_indexcast_dim_0_i8(%arg0: tensor<f16>) -> tensor<128xf16> {
%init = tensor.empty() : tensor<128xf16>
%res = linalg.generic {
indexing_maps = [affine_map<(d0) -> ()>, affine_map<(d0) -> (d0)>],
iterator_types = ["parallel"]
}
ins(%arg0 : tensor<f16>)
outs(%init : tensor<128xf16>) {
^bb0(%in: f16, %out: f16):
%idx = linalg.index 0 : index
%int = arith.index_cast %idx : index to i64
%fp = arith.sitofp %int : i64 to f16
linalg.yield %fp : f16
} -> tensor<128xf16>
return %res : tensor<128xf16>
}
// CHECK-LABEL: func @linalg_indexcast_dim_1_i16
// CHECK: %[[IDX:.+]] = linalg.index 1 : index
// CHECK-NEXT: %[[INT:.+]] = arith.index_cast %[[IDX]] : index to i16
// CHECK-NEXT: %[[FP:.+]] = arith.sitofp %[[INT]] : i16 to f16
// CHECK-NEXT: linalg.yield %[[FP]] : f16
func.func @linalg_indexcast_dim_1_i16(%arg0: tensor<f16>, %arg1: tensor<?x129xf16>) -> tensor<?x129xf16> {
%res = linalg.generic {
indexing_maps = [affine_map<(d0, d1) -> ()>, affine_map<(d0, d1) -> (d0, d1)>],
iterator_types = ["parallel", "parallel"]
}
ins(%arg0 : tensor<f16>)
outs(%arg1 : tensor<?x129xf16>) {
^bb0(%in: f16, %out: f16):
%idx = linalg.index 1 : index
%int = arith.index_cast %idx : index to i64
%fp = arith.sitofp %int : i64 to f16
linalg.yield %fp : f16
} -> tensor<?x129xf16>
return %res : tensor<?x129xf16>
}
// CHECK-LABEL: func @linalg_indexcast_dynamic_dim_i64
// CHECK: %[[IDX:.+]] = linalg.index 0 : index
// CHECK-NEXT: %[[INT:.+]] = arith.index_cast %[[IDX]] : index to i64
// CHECK-NEXT: %[[FP:.+]] = arith.sitofp %[[INT]] : i64 to f16
// CHECK-NEXT: linalg.yield %[[FP]] : f16
func.func @linalg_indexcast_dynamic_dim_i64(%arg0: tensor<f16>, %arg1: tensor<?xf16>) -> tensor<?xf16> {
%res = linalg.generic {
indexing_maps = [affine_map<(d0) -> ()>, affine_map<(d0) -> (d0)>],
iterator_types = ["parallel"]
}
ins(%arg0 : tensor<f16>)
outs(%arg1 : tensor<?xf16>) {
^bb0(%in: f16, %out: f16):
%idx = linalg.index 0 : index
%int = arith.index_cast %idx : index to i64
%fp = arith.sitofp %int : i64 to f16
linalg.yield %fp : f16
} -> tensor<?xf16>
return %res : tensor<?xf16>
}
//===----------------------------------------------------------------------===//
// arith.index_castui
//===----------------------------------------------------------------------===//
// CHECK-LABEL: func @linalg_indexcastui_dim_0_i8
// CHECK: %[[IDX:.+]] = linalg.index 0 : index
// CHECK-NEXT: %[[INT:.+]] = arith.index_castui %[[IDX]] : index to i8
// CHECK-NEXT: %[[FP:.+]] = arith.uitofp %[[INT]] : i8 to f16
// CHECK-NEXT: linalg.yield %[[FP]] : f16
func.func @linalg_indexcastui_dim_0_i8(%arg0: tensor<f16>) -> tensor<256xf16> {
%init = tensor.empty() : tensor<256xf16>
%res = linalg.generic {
indexing_maps = [affine_map<(d0) -> ()>, affine_map<(d0) -> (d0)>],
iterator_types = ["parallel"]
}
ins(%arg0 : tensor<f16>)
outs(%init : tensor<256xf16>) {
^bb0(%in: f16, %out: f16):
%idx = linalg.index 0 : index
%int = arith.index_castui %idx : index to i64
%fp = arith.uitofp %int : i64 to f16
linalg.yield %fp : f16
} -> tensor<256xf16>
return %res : tensor<256xf16>
}
// CHECK-LABEL: func @linalg_indexcastui_dim_1_i16
// CHECK: %[[IDX:.+]] = linalg.index 1 : index
// CHECK-NEXT: %[[INT:.+]] = arith.index_castui %[[IDX]] : index to i16
// CHECK-NEXT: %[[FP:.+]] = arith.uitofp %[[INT]] : i16 to f16
// CHECK-NEXT: linalg.yield %[[FP]] : f16
func.func @linalg_indexcastui_dim_1_i16(%arg0: tensor<f16>, %arg1: tensor<?x257xf16>) -> tensor<?x257xf16> {
%res = linalg.generic {
indexing_maps = [affine_map<(d0, d1) -> ()>, affine_map<(d0, d1) -> (d0, d1)>],
iterator_types = ["parallel", "parallel"]
}
ins(%arg0 : tensor<f16>)
outs(%arg1 : tensor<?x257xf16>) {
^bb0(%in: f16, %out: f16):
%idx = linalg.index 1 : index
%int = arith.index_castui %idx : index to i64
%fp = arith.uitofp %int : i64 to f16
linalg.yield %fp : f16
} -> tensor<?x257xf16>
return %res : tensor<?x257xf16>
}
// CHECK-LABEL: func @linalg_indexcastui_dynamic_dim_i64
// CHECK: %[[IDX:.+]] = linalg.index 0 : index
// CHECK-NEXT: %[[INT:.+]] = arith.index_castui %[[IDX]] : index to i64
// CHECK-NEXT: %[[FP:.+]] = arith.uitofp %[[INT]] : i64 to f16
// CHECK-NEXT: linalg.yield %[[FP]] : f16
func.func @linalg_indexcastui_dynamic_dim_i64(%arg0: tensor<f16>, %arg1: tensor<?xf16>) -> tensor<?xf16> {
%res = linalg.generic {
indexing_maps = [affine_map<(d0) -> ()>, affine_map<(d0) -> (d0)>],
iterator_types = ["parallel"]
}
ins(%arg0 : tensor<f16>)
outs(%arg1 : tensor<?xf16>) {
^bb0(%in: f16, %out: f16):
%idx = linalg.index 0 : index
%int = arith.index_castui %idx : index to i64
%fp = arith.uitofp %int : i64 to f16
linalg.yield %fp : f16
} -> tensor<?xf16>
return %res : tensor<?xf16>
}
|
