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
|
// RUN: mlir-opt %s -test-lower-to-arm-sme -test-lower-to-llvm | \
// RUN: %mcr_aarch64_cmd \
// RUN: -e=main -entry-point-result=void \
// RUN: -march=aarch64 -mattr="+sve,+sme" \
// RUN: -shared-libs=%mlir_runner_utils,%mlir_c_runner_utils,%arm_sme_abi_shlib,%mlir_arm_runner_utils | \
// RUN: FileCheck %s
#transpose = affine_map<(d0, d1) -> (d1, d0)>
func.func @fill2DMemrefRows(%memref: memref<?x?xf32>) {
%c0 = arith.constant 0 : index
%c1 = arith.constant 1 : index
%rows = memref.dim %memref, %c0 : memref<?x?xf32>
%cols = memref.dim %memref, %c1 : memref<?x?xf32>
scf.for %i = %c0 to %rows step %c1 {
scf.for %j = %c0 to %cols step %c1 {
%n = arith.addi %i, %c1 : index
%val = arith.index_cast %n : index to i32
%val_f32 = arith.sitofp %val : i32 to f32
memref.store %val_f32, %memref[%i, %j] : memref<?x?xf32>
}
}
return
}
func.func @testTransposedReadWithMask(%maskRows: index, %maskCols: index) {
%in = memref.alloca() : memref<4x16xf32>
%out = memref.alloca() : memref<16x4xf32>
%inDyn = memref.cast %in : memref<4x16xf32> to memref<?x?xf32>
%outDyn = memref.cast %out : memref<16x4xf32> to memref<?x?xf32>
func.call @fill2DMemrefRows(%inDyn) : (memref<?x?xf32>) -> ()
/// A mask so we only read the first maskRows x maskCols portion of %in.
%mask = vector.create_mask %maskRows, %maskCols : vector<[4]x[16]xi1>
%pad = arith.constant 0.0 : f32
%c0 = arith.constant 0 : index
/// A vector.transfer_read with a transpose permutation map. So the input data
/// (and mask) have a [4]x[16] shape, but the output is [16]x[4].
%readTransposed = vector.transfer_read %inDyn[%c0, %c0], %pad, %mask
{permutation_map = #transpose, in_bounds = [true, true]} : memref<?x?xf32>, vector<[16]x[4]xf32>
/// Write the vector back to a memref (that also has a transposed shape).
vector.transfer_write %readTransposed, %outDyn[%c0, %c0] {in_bounds = [true, true]} : vector<[16]x[4]xf32>, memref<?x?xf32>
/// Print the input memref.
vector.print str "Input memref:"
%inUnranked = memref.cast %inDyn : memref<?x?xf32> to memref<*xf32>
call @printMemrefF32(%inUnranked) : (memref<*xf32>) -> ()
/// Print the result memref.
vector.print str "Masked transposed result:"
%outUnranked = memref.cast %outDyn : memref<?x?xf32> to memref<*xf32>
call @printMemrefF32(%outUnranked) : (memref<*xf32>) -> ()
return
}
func.func @testTransposedWriteWithMask(%maskRows: index, %maskCols: index) {
%in = memref.alloca() : memref<16x4xf32>
%out = memref.alloca() : memref<4x16xf32>
%c0_f32 = arith.constant 0.0 : f32
linalg.fill ins(%c0_f32 : f32) outs(%out : memref<4x16xf32>)
%inDyn = memref.cast %in : memref<16x4xf32> to memref<?x?xf32>
%outDyn = memref.cast %out : memref<4x16xf32> to memref<?x?xf32>
func.call @fill2DMemrefRows(%inDyn) : (memref<?x?xf32>) -> ()
/// A regular read.
%c0 = arith.constant 0 : index
%read = vector.transfer_read %inDyn[%c0, %c0], %c0_f32 {in_bounds = [true, true]}
: memref<?x?xf32>, vector<[16]x[4]xf32>
/// A mask so we only write the first maskRows x maskCols portion of transpose(%in).
%mask = vector.create_mask %maskRows, %maskCols : vector<[4]x[16]xi1>
/// Write out the data with a transpose. Here (like the read test) the mask
/// matches the shape of the memory, not the vector.
vector.transfer_write %read, %outDyn[%c0, %c0], %mask {permutation_map = #transpose, in_bounds = [true, true]}
: vector<[16]x[4]xf32>, memref<?x?xf32>
/// Print the input memref.
vector.print str "Input memref:"
%inUnranked = memref.cast %inDyn : memref<?x?xf32> to memref<*xf32>
call @printMemrefF32(%inUnranked) : (memref<*xf32>) -> ()
/// Print the result memref.
vector.print str "Masked transposed result:"
%outUnranked = memref.cast %outDyn : memref<?x?xf32> to memref<*xf32>
call @printMemrefF32(%outUnranked) : (memref<*xf32>) -> ()
return
}
func.func @main() {
/// Set the SVL to 128-bits (i.e. vscale = 1).
/// This test is for the use of multiple tiles rather than scalability.
%c128 = arith.constant 128 : i32
func.call @setArmSVLBits(%c128) : (i32) -> ()
// CHECK: Input memref:
// CHECK: [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]
// CHECK-NEXT: [2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2, 2]
// CHECK-NEXT: [3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3, 3]
// CHECK-NEXT: [4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4, 4]
//
// CHECK: Masked transposed result:
// CHECK: [1, 2, 0, 0]
// CHECK-NEXT: [1, 2, 0, 0]
// CHECK-NEXT: [1, 2, 0, 0]
// CHECK-NEXT: [1, 2, 0, 0]
// CHECK-NEXT: [1, 2, 0, 0]
// CHECK-NEXT: [1, 2, 0, 0]
// CHECK-NEXT: [1, 2, 0, 0]
// CHECK-NEXT: [1, 2, 0, 0]
// CHECK-NEXT: [1, 2, 0, 0]
// CHECK-NEXT: [1, 2, 0, 0]
// CHECK-NEXT: [1, 2, 0, 0]
// CHECK-NEXT: [1, 2, 0, 0]
// CHECK-NEXT: [1, 2, 0, 0]
// CHECK-NEXT: [1, 2, 0, 0]
// CHECK-NEXT: [1, 2, 0, 0]
// CHECK-NEXT: [0, 0, 0, 0]
%readMaskRows = arith.constant 2 : index
%readMaskCols = arith.constant 15 : index
func.call @testTransposedReadWithMask(%readMaskRows, %readMaskCols) : (index, index) -> ()
// CHECK: Input memref:
// CHECK: [1, 1, 1, 1]
// CHECK-NEXT: [2, 2, 2, 2]
// CHECK-NEXT: [3, 3, 3, 3]
// CHECK-NEXT: [4, 4, 4, 4]
// CHECK-NEXT: [5, 5, 5, 5]
// CHECK-NEXT: [6, 6, 6, 6]
// CHECK-NEXT: [7, 7, 7, 7]
// CHECK-NEXT: [8, 8, 8, 8]
// CHECK-NEXT: [9, 9, 9, 9]
// CHECK-NEXT: [10, 10, 10, 10]
// CHECK-NEXT: [11, 11, 11, 11]
// CHECK-NEXT: [12, 12, 12, 12]
// CHECK-NEXT: [13, 13, 13, 13]
// CHECK-NEXT: [14, 14, 14, 14]
// CHECK-NEXT: [15, 15, 15, 15]
// CHECK-NEXT: [16, 16, 16, 16]
//
// CHECK: Masked transposed result:
// CHECK: [1, 2, 3, 4, 5, 6, 7, 8, 0, 0, 0, 0, 0, 0, 0, 0]
// CHECK-NEXT: [1, 2, 3, 4, 5, 6, 7, 8, 0, 0, 0, 0, 0, 0, 0, 0]
// CHECK-NEXT: [1, 2, 3, 4, 5, 6, 7, 8, 0, 0, 0, 0, 0, 0, 0, 0]
// CHECK-NEXT: [0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
%writeMaskRows = arith.constant 3 : index
%writeMaskCols = arith.constant 8 : index
func.call @testTransposedWriteWithMask(%writeMaskRows, %writeMaskCols) : (index, index) -> ()
return
}
func.func private @printMemrefF32(%ptr : memref<*xf32>)
func.func private @setArmSVLBits(%bits : i32)
|