| 1 |
// RUN: mlir-opt %s -transform-interpreter -split-input-file | FileCheck %s |
| 2 |
|
| 3 |
///---------------------------------------------------------------------------------------- |
| 4 |
/// Tests for vectorization patterns for tensor.pad, i.e. |
| 5 |
/// * transform.apply_patterns.linalg.pad_vectorization |
| 6 |
/// |
| 7 |
/// These tests are meant os a lower granule than tests in |
| 8 |
/// * pad-with-patterns.mlir. |
| 9 |
/// The goal is to test specific patterns. To this end, some inputs already |
| 10 |
/// contain Vector ops (on top of `tensor.pad`). |
| 11 |
///---------------------------------------------------------------------------------------- |
| 12 |
|
| 13 |
///---------------------------------------------------------------------------------------- |
| 14 |
/// [Pattern: PadOpVectorizationWithTransferReadPattern] |
| 15 |
///---------------------------------------------------------------------------------------- |
| 16 |
// CHECK-LABEL: func @pad_and_transfer_read |
| 17 |
// CHECK-SAME: %[[ARG0:.*]]: tensor<5x6xf32> |
| 18 |
// CHECK-NOT: tensor.pad |
| 19 |
// CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index |
| 20 |
// CHECK-DAG: %[[C5:.*]] = arith.constant 5.0 |
| 21 |
// CHECK: %[[RESULT:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], %[[C5]] : tensor<5x6xf32>, vector<7x9xf32> |
| 22 |
// CHECK: return %[[RESULT]] |
| 23 |
func.func @pad_and_transfer_read(%arg0: tensor<5x6xf32>) -> vector<7x9xf32> { |
| 24 |
%c0 = arith.constant 0 : index |
| 25 |
%c5 = arith.constant 5.0 : f32 |
| 26 |
%c6 = arith.constant 6.0 : f32 |
| 27 |
%0 = tensor.pad %arg0 low[0, 0] high[5, 7] { |
| 28 |
^bb0(%arg1: index, %arg2: index): |
| 29 |
tensor.yield %c5 : f32 |
| 30 |
} : tensor<5x6xf32> to tensor<10x13xf32> |
| 31 |
%1 = vector.transfer_read %0[%c0, %c0], %c6 |
| 32 |
: tensor<10x13xf32>, vector<7x9xf32> |
| 33 |
return %1 : vector<7x9xf32> |
| 34 |
} |
| 35 |
|
| 36 |
module attributes {transform.with_named_sequence} { |
| 37 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 38 |
%func_op = transform.structured.match ops{["func.func"]} in %arg1 : (!transform.any_op) -> !transform.op<"func.func"> |
| 39 |
|
| 40 |
transform.apply_patterns to %func_op { |
| 41 |
transform.apply_patterns.linalg.pad_vectorization |
| 42 |
} : !transform.op<"func.func"> |
| 43 |
transform.yield |
| 44 |
} |
| 45 |
} |
| 46 |
|
| 47 |
// ----- |
| 48 |
|
| 49 |
///---------------------------------------------------------------------------------------- |
| 50 |
/// [Pattern: PadOpVectorizationWithTransferWritePattern] |
| 51 |
///---------------------------------------------------------------------------------------- |
| 52 |
func.func private @make_vector() -> vector<7x9xf32> |
| 53 |
|
| 54 |
// CHECK-LABEL: func @pad_and_transfer_write_static_low_and_high |
| 55 |
// CHECK-SAME: %[[ARG0:.*]]: tensor<5x6xf32> |
| 56 |
// CHECK-NOT: tensor.pad |
| 57 |
// CHECK: %[[C0:.*]] = arith.constant 0 : index |
| 58 |
// CHECK: %[[VEC0:.*]] = call @make_vector() : () -> vector<7x9xf32> |
| 59 |
// CHECK: %[[RESULT:.*]] = vector.transfer_write %[[VEC0]], %[[ARG0]][%[[C0]], %[[C0]]] : vector<7x9xf32>, tensor<5x6xf32> |
| 60 |
// CHECK: return %[[RESULT]] |
| 61 |
func.func @pad_and_transfer_write_static_low_and_high( |
| 62 |
%arg0: tensor<5x6xf32>) -> tensor<5x6xf32> { |
| 63 |
%c0 = arith.constant 0 : index |
| 64 |
%c5 = arith.constant 5.0 : f32 |
| 65 |
%0 = tensor.pad %arg0 low[0, 0] high[5, 7] { |
| 66 |
^bb0(%arg2: index, %arg3: index): |
| 67 |
tensor.yield %c5 : f32 |
| 68 |
} : tensor<5x6xf32> to tensor<10x13xf32> |
| 69 |
%1 = call @make_vector() : () -> vector<7x9xf32> |
| 70 |
%2 = vector.transfer_write %1, %0[%c0, %c0] |
| 71 |
: vector<7x9xf32>, tensor<10x13xf32> |
| 72 |
%3 = tensor.extract_slice %2[0, 0] [5, 6] [1, 1] : tensor<10x13xf32> to tensor<5x6xf32> |
| 73 |
return %3 : tensor<5x6xf32> |
| 74 |
} |
| 75 |
|
| 76 |
module attributes {transform.with_named_sequence} { |
| 77 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 78 |
%func_op = transform.structured.match ops{["func.func"]} in %arg1 : (!transform.any_op) -> !transform.op<"func.func"> |
| 79 |
|
| 80 |
transform.apply_patterns to %func_op { |
| 81 |
transform.apply_patterns.linalg.pad_vectorization |
| 82 |
} : !transform.op<"func.func"> |
| 83 |
transform.yield |
| 84 |
} |
| 85 |
} |
| 86 |
|
| 87 |
// ----- |
| 88 |
|
| 89 |
func.func private @make_vector() -> vector<7x9xf32> |
| 90 |
|
| 91 |
// CHECK-LABEL: func @pad_and_transfer_write_static_low_dynamic_high |
| 92 |
// CHECK-SAME: %[[ARG0:.*]]: tensor<?x?xf32>, %[[SIZE:.*]]: index, %[[PADDING:.*]]: index |
| 93 |
// CHECK-NOT: tensor.pad |
| 94 |
// CHECK: %[[C0:.*]] = arith.constant 0 : index |
| 95 |
// CHECK: %[[SUB:.*]] = tensor.extract_slice %[[ARG0]][0, 0] [%[[SIZE]], 6] [1, 1] : tensor<?x?xf32> to tensor<?x6xf32> |
| 96 |
// CHECK: %[[VEC0:.*]] = call @make_vector() : () -> vector<7x9xf32> |
| 97 |
// CHECK: %[[RESULT:.*]] = vector.transfer_write %[[VEC0]], %[[SUB]][%[[C0]], %[[C0]]] : vector<7x9xf32>, tensor<?x6xf32> |
| 98 |
// CHECK: return %[[RESULT]] |
| 99 |
func.func @pad_and_transfer_write_static_low_dynamic_high( |
| 100 |
%arg0: tensor<?x?xf32>, %size: index, %padding: index) -> tensor<?x6xf32> { |
| 101 |
%c0 = arith.constant 0 : index |
| 102 |
%c5 = arith.constant 5.0 : f32 |
| 103 |
%s = tensor.extract_slice %arg0[0, 0] [%size, 6] [1, 1] |
| 104 |
: tensor<?x?xf32> to tensor<?x6xf32> |
| 105 |
%0 = tensor.pad %s low[0, 0] high[%padding, 7] { |
| 106 |
^bb0(%arg2: index, %arg3: index): |
| 107 |
tensor.yield %c5 : f32 |
| 108 |
} : tensor<?x6xf32> to tensor<?x13xf32> |
| 109 |
%1 = call @make_vector() : () -> vector<7x9xf32> |
| 110 |
%2 = vector.transfer_write %1, %0[%c0, %c0] |
| 111 |
: vector<7x9xf32>, tensor<?x13xf32> |
| 112 |
%3 = tensor.extract_slice %2[0, 0] [%size, 6] [1, 1] : tensor<?x13xf32> to tensor<?x6xf32> |
| 113 |
return %3 : tensor<?x6xf32> |
| 114 |
} |
| 115 |
|
| 116 |
module attributes {transform.with_named_sequence} { |
| 117 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 118 |
%func_op = transform.structured.match ops{["func.func"]} in %arg1 : (!transform.any_op) -> !transform.op<"func.func"> |
| 119 |
|
| 120 |
transform.apply_patterns to %func_op { |
| 121 |
transform.apply_patterns.linalg.pad_vectorization |
| 122 |
} : !transform.op<"func.func"> |
| 123 |
transform.yield |
| 124 |
} |
| 125 |
} |
| 126 |
|
| 127 |
// ----- |
| 128 |
|
| 129 |
func.func private @make_vector() -> vector<7x9xf32> |
| 130 |
|
| 131 |
// Negative test - low pad is non-zero |
| 132 |
|
| 133 |
// CHECK-LABEL: func @pad_and_transfer_write_static_non_zero_low_pad |
| 134 |
// CHECK: tensor.pad |
| 135 |
func.func @pad_and_transfer_write_static_non_zero_low_pad( |
| 136 |
%arg0: tensor<5x6xf32>) -> tensor<5x6xf32> { |
| 137 |
%c0 = arith.constant 0 : index |
| 138 |
%c5 = arith.constant 5.0 : f32 |
| 139 |
%0 = tensor.pad %arg0 low[0, 1] high[5, 6] { |
| 140 |
^bb0(%arg2: index, %arg3: index): |
| 141 |
tensor.yield %c5 : f32 |
| 142 |
} : tensor<5x6xf32> to tensor<10x13xf32> |
| 143 |
%1 = call @make_vector() : () -> vector<7x9xf32> |
| 144 |
%2 = vector.transfer_write %1, %0[%c0, %c0] |
| 145 |
: vector<7x9xf32>, tensor<10x13xf32> |
| 146 |
%3 = tensor.extract_slice %2[0, 0] [5, 6] [1, 1] : tensor<10x13xf32> to tensor<5x6xf32> |
| 147 |
return %3 : tensor<5x6xf32> |
| 148 |
} |
| 149 |
|
| 150 |
module attributes {transform.with_named_sequence} { |
| 151 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 152 |
%func_op = transform.structured.match ops{["func.func"]} in %arg1 : (!transform.any_op) -> !transform.op<"func.func"> |
| 153 |
|
| 154 |
transform.apply_patterns to %func_op { |
| 155 |
transform.apply_patterns.linalg.pad_vectorization |
| 156 |
} : !transform.op<"func.func"> |
| 157 |
transform.yield |
| 158 |
} |
| 159 |
} |
| 160 |
|
| 161 |
// ----- |
| 162 |
|
| 163 |
// Negative test - TransferWriteOp result is not _directly_ consumed by an |
| 164 |
// ExtractSliceOp (noet the non-zero offset). |
| 165 |
|
| 166 |
func.func private @make_vector() -> vector<7x9xf32> |
| 167 |
|
| 168 |
// CHECK-LABEL: func @pad_and_transfer_write_static_non_zero_offset |
| 169 |
// CHECK: tensor.pad |
| 170 |
func.func @pad_and_transfer_write_static_non_zero_offset( |
| 171 |
%arg0: tensor<5x6xf32>) -> tensor<5x6xf32> { |
| 172 |
%c0 = arith.constant 0 : index |
| 173 |
%c5 = arith.constant 5.0 : f32 |
| 174 |
%0 = tensor.pad %arg0 low[0, 0] high[5, 7] { |
| 175 |
^bb0(%arg2: index, %arg3: index): |
| 176 |
tensor.yield %c5 : f32 |
| 177 |
} : tensor<5x6xf32> to tensor<10x13xf32> |
| 178 |
%1 = call @make_vector() : () -> vector<7x9xf32> |
| 179 |
%2 = vector.transfer_write %1, %0[%c0, %c0] |
| 180 |
: vector<7x9xf32>, tensor<10x13xf32> |
| 181 |
%3 = tensor.extract_slice %2[0, 1] [5, 6] [1, 1] : tensor<10x13xf32> to tensor<5x6xf32> |
| 182 |
return %3 : tensor<5x6xf32> |
| 183 |
} |
| 184 |
|
| 185 |
module attributes {transform.with_named_sequence} { |
| 186 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 187 |
%func_op = transform.structured.match ops{["func.func"]} in %arg1 : (!transform.any_op) -> !transform.op<"func.func"> |
| 188 |
|
| 189 |
transform.apply_patterns to %func_op { |
| 190 |
transform.apply_patterns.linalg.pad_vectorization |
| 191 |
} : !transform.op<"func.func"> |
| 192 |
transform.yield |
| 193 |
} |
| 194 |
} |
| 195 |
|
| 196 |
// ----- |
| 197 |
|
| 198 |
///---------------------------------------------------------------------------------------- |
| 199 |
/// [Pattern: PadOpVectorizationWithInsertSlicePattern] |
| 200 |
///---------------------------------------------------------------------------------------- |
| 201 |
|
| 202 |
func.func private @make_vector() -> tensor<12x13xf32> |
| 203 |
|
| 204 |
// CHECK-LABEL: func @pad_and_insert_slice_source |
| 205 |
// CHECK-SAME: %[[ARG0:.*]]: tensor<5x6xf32> |
| 206 |
// CHECK-NOT: tensor.pad |
| 207 |
// CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index |
| 208 |
// CHECK-DAG: %[[C5:.*]] = arith.constant 5.0 |
| 209 |
// CHECK: %[[VEC0:.*]] = call @make_vector() : () -> tensor<12x13xf32> |
| 210 |
// CHECK: %[[READ:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], %[[C5]] : tensor<5x6xf32>, vector<7x9xf32> |
| 211 |
// CHECK: %[[WRITE:.*]] = vector.transfer_write %[[READ]], %[[VEC0]][%[[C0]], %[[C0]]] {in_bounds = [true, true]} : vector<7x9xf32>, tensor<12x13xf32> |
| 212 |
// CHECK: return %[[WRITE]] |
| 213 |
func.func @pad_and_insert_slice_source( |
| 214 |
%arg0: tensor<5x6xf32>) -> tensor<12x13xf32> { |
| 215 |
%c0 = arith.constant 0 : index |
| 216 |
%c5 = arith.constant 5.0 : f32 |
| 217 |
%0 = tensor.pad %arg0 low[0, 0] high[2, 3] { |
| 218 |
^bb0(%arg2: index, %arg3: index): |
| 219 |
tensor.yield %c5 : f32 |
| 220 |
} : tensor<5x6xf32> to tensor<7x9xf32> |
| 221 |
%1 = call @make_vector() : () -> tensor<12x13xf32> |
| 222 |
%r = tensor.insert_slice %0 into %1[0, 0][7, 9][1, 1] : tensor<7x9xf32> into tensor<12x13xf32> |
| 223 |
return %r : tensor<12x13xf32> |
| 224 |
} |
| 225 |
|
| 226 |
module attributes {transform.with_named_sequence} { |
| 227 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 228 |
%func_op = transform.structured.match ops{["func.func"]} in %arg1 : (!transform.any_op) -> !transform.op<"func.func"> |
| 229 |
|
| 230 |
transform.apply_patterns to %func_op { |
| 231 |
transform.apply_patterns.linalg.pad_vectorization |
| 232 |
} : !transform.op<"func.func"> |
| 233 |
transform.yield |
| 234 |
} |
| 235 |
} |
| 236 |
|
| 237 |
// ----- |
| 238 |
|
| 239 |
func.func private @make_vector() -> tensor<12x13xf32> |
| 240 |
|
| 241 |
// The destination of tensor.insert_slice matches the result of tensor.pad - |
| 242 |
// not supported. |
| 243 |
|
| 244 |
// CHECK-LABEL: func.func @pad_and_insert_slice_dest( |
| 245 |
// CHECK-NOT: vector.transfer_read |
| 246 |
// CHECK-NOT: vector.transfer_write |
| 247 |
|
| 248 |
func.func @pad_and_insert_slice_dest( |
| 249 |
%arg0: tensor<1x5x6xf32>) -> tensor<1x12x13xf32> { |
| 250 |
%c5 = arith.constant 5.0 : f32 |
| 251 |
%0 = tensor.pad %arg0 low[0, 0, 0] high[0, 7, 7] { |
| 252 |
^bb0(%arg2: index, %arg3: index, %arg4: index): |
| 253 |
tensor.yield %c5 : f32 |
| 254 |
} : tensor<1x5x6xf32> to tensor<1x12x13xf32> |
| 255 |
%1 = call @make_vector() : () -> tensor<12x13xf32> |
| 256 |
%r = tensor.insert_slice %1 into %0[0, 0, 0][1, 12, 13][1, 1, 1] : tensor<12x13xf32> into tensor<1x12x13xf32> |
| 257 |
return %r : tensor<1x12x13xf32> |
| 258 |
} |
| 259 |
|
| 260 |
module attributes {transform.with_named_sequence} { |
| 261 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 262 |
%func_op = transform.structured.match ops{["func.func"]} in %arg1 : (!transform.any_op) -> !transform.op<"func.func"> |
| 263 |
|
| 264 |
transform.apply_patterns to %func_op { |
| 265 |
transform.apply_patterns.linalg.pad_vectorization |
| 266 |
} : !transform.op<"func.func"> |
| 267 |
transform.yield |
| 268 |
} |
| 269 |
} |