| 1 |
// RUN: mlir-opt %s -transform-interpreter -split-input-file -verify-diagnostics | FileCheck %s |
| 2 |
|
| 3 |
func.func @conv1d_nwc_wcf_dyn_ch_dim(%input: memref<4x6x?xf32>, %filter: memref<1x?x8xf32>, %output: memref<4x2x8xf32>) { |
| 4 |
// expected-error @+1 {{Attempted to vectorize, but failed}} |
| 5 |
linalg.conv_1d_nwc_wcf |
| 6 |
{dilations = dense<1> : tensor<1xi64>, strides = dense<3> : tensor<1xi64>} |
| 7 |
ins(%input, %filter : memref<4x6x?xf32>, memref<1x?x8xf32>) |
| 8 |
outs(%output : memref<4x2x8xf32>) |
| 9 |
return |
| 10 |
} |
| 11 |
|
| 12 |
module attributes {transform.with_named_sequence} { |
| 13 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 14 |
%0 = transform.structured.match ops{["linalg.conv_1d_nwc_wcf"]} in %arg1 : (!transform.any_op) -> !transform.any_op |
| 15 |
transform.structured.vectorize %0 : !transform.any_op |
| 16 |
transform.yield |
| 17 |
} |
| 18 |
} |
| 19 |
|
| 20 |
// ----- |
| 21 |
|
| 22 |
// Masked vectorisation of 1D depthwise CW convs is not yet supported |
| 23 |
|
| 24 |
func.func @depthwise_conv1d_ncw_cw(%input: memref<3x?x4xf32>, %filter: memref<?x1xf32>, %output: memref<3x?x4xf32>) { |
| 25 |
// expected-error @+1 {{Attempted to vectorize, but failed}} |
| 26 |
linalg.depthwise_conv_1d_ncw_cw |
| 27 |
{dilations = dense<2> : tensor<1xi64>, strides = dense<1> : tensor<1xi64>} |
| 28 |
ins(%input, %filter : memref<3x?x4xf32>, memref<?x1xf32>) |
| 29 |
outs(%output : memref<3x?x4xf32>) |
| 30 |
return |
| 31 |
} |
| 32 |
|
| 33 |
module attributes {transform.with_named_sequence} { |
| 34 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 35 |
%0 = transform.structured.match ops{["linalg.depthwise_conv_1d_ncw_cw"]} in %arg1 : (!transform.any_op) -> !transform.any_op |
| 36 |
transform.structured.vectorize %0 vector_sizes [3, 4, 5, 1] : !transform.any_op |
| 37 |
transform.yield |
| 38 |
} |
| 39 |
} |
| 40 |
|
| 41 |
// ----- |
| 42 |
|
| 43 |
func.func @depthwise_conv1d_nwc_wc_dyn_w_dim(%input: memref<3x?x4xf32>, %filter: memref<?x4xf32>, %output: memref<3x?x4xf32>) { |
| 44 |
// expected-error @+1 {{Attempted to vectorize, but failed}} |
| 45 |
linalg.depthwise_conv_1d_nwc_wc |
| 46 |
{dilations = dense<2> : tensor<1xi64>, strides = dense<1> : tensor<1xi64>} |
| 47 |
ins(%input, %filter : memref<3x?x4xf32>, memref<?x4xf32>) |
| 48 |
outs(%output : memref<3x?x4xf32>) |
| 49 |
return |
| 50 |
} |
| 51 |
|
| 52 |
module attributes {transform.with_named_sequence} { |
| 53 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 54 |
%0 = transform.structured.match ops{["linalg.depthwise_conv_1d_nwc_wc"]} in %arg1 : (!transform.any_op) -> !transform.any_op |
| 55 |
transform.structured.vectorize %0 vector_sizes [3, 2, 4, 2] : !transform.any_op |
| 56 |
transform.yield |
| 57 |
} |
| 58 |
} |
| 59 |
|
| 60 |
// ----- |
| 61 |
|
| 62 |
func.func @depthwise_conv1d_nwc_wc_dyn_ch_dim(%input: memref<3x5x?xf32>, %filter: memref<2x?xf32>, %output: memref<3x2x?xf32>) { |
| 63 |
// expected-error @+1 {{Attempted to vectorize, but failed}} |
| 64 |
linalg.depthwise_conv_1d_nwc_wc |
| 65 |
ins(%input, %filter : memref<3x5x?xf32>, memref<2x?xf32>) |
| 66 |
outs(%output : memref<3x2x?xf32>) |
| 67 |
return |
| 68 |
} |
| 69 |
|
| 70 |
module attributes {transform.with_named_sequence} { |
| 71 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 72 |
%0 = transform.structured.match ops{["linalg.depthwise_conv_1d_nwc_wc"]} in %arg1 : (!transform.any_op) -> !transform.any_op |
| 73 |
transform.structured.vectorize %0 : !transform.any_op |
| 74 |
transform.yield |
| 75 |
} |
| 76 |
} |
| 77 |
|
| 78 |
// ----- |
| 79 |
|
| 80 |
func.func @depthwise_conv1d_nwc_wc_dyn_w_dim(%input: memref<3x?x3xf32>, %filter: memref<2x3xf32>, %output: memref<3x?x3xf32>) { |
| 81 |
// expected-error @+1 {{Attempted to vectorize, but failed}} |
| 82 |
linalg.depthwise_conv_1d_nwc_wc |
| 83 |
ins(%input, %filter : memref<3x?x3xf32>, memref<2x3xf32>) |
| 84 |
outs(%output : memref<3x?x3xf32>) |
| 85 |
return |
| 86 |
} |
| 87 |
|
| 88 |
module attributes {transform.with_named_sequence} { |
| 89 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 90 |
%0 = transform.structured.match ops{["linalg.depthwise_conv_1d_nwc_wc"]} in %arg1 : (!transform.any_op) -> !transform.any_op |
| 91 |
transform.structured.vectorize %0 : !transform.any_op |
| 92 |
transform.yield |
| 93 |
} |
| 94 |
} |
| 95 |
|
| 96 |
// ----- |
| 97 |
|
| 98 |
func.func @conv1d_dyn_w_dim(%input: tensor<?xf32>, %filter: tensor<4xf32>, %output: tensor<?xf32>) -> tensor<?xf32> { |
| 99 |
// expected-error @+1 {{Attempted to vectorize, but failed}} |
| 100 |
%0 = linalg.conv_1d ins(%input, %filter : tensor<?xf32>, tensor<4xf32>) |
| 101 |
outs(%output : tensor<?xf32>) -> tensor<?xf32> |
| 102 |
return %0 : tensor<?xf32> |
| 103 |
} |
| 104 |
|
| 105 |
module attributes {transform.with_named_sequence} { |
| 106 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 107 |
%0 = transform.structured.match ops{["linalg.conv_1d"]} in %arg1 : (!transform.any_op) -> !transform.any_op |
| 108 |
transform.structured.vectorize %0 : !transform.any_op |
| 109 |
transform.yield |
| 110 |
} |
| 111 |
} |
| 112 |
|
| 113 |
// ----- |
| 114 |
|
| 115 |
func.func @conv2d_nchw_fchw(%input: tensor<1x5x8x8xf32>, %filter: tensor<4x5x3x3xf32>, %output: tensor<1x4x6x6xf32>) { |
| 116 |
// expected-error @+1 {{Attempted to vectorize, but failed}} |
| 117 |
linalg.conv_2d_nchw_fchw {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%input, %filter : tensor<1x5x8x8xf32>, tensor<4x5x3x3xf32>) outs(%output : tensor<1x4x6x6xf32>) -> tensor<1x4x6x6xf32> |
| 118 |
return |
| 119 |
} |
| 120 |
|
| 121 |
module attributes {transform.with_named_sequence} { |
| 122 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 123 |
%0 = transform.structured.match ops{["linalg.conv_2d_nchw_fchw"]} in %arg1 : (!transform.any_op) -> !transform.any_op |
| 124 |
transform.structured.vectorize %0 : !transform.any_op |
| 125 |
transform.yield |
| 126 |
} |
| 127 |
} |
| 128 |
|
| 129 |
// ----- |
| 130 |
|
| 131 |
func.func @conv2d_nhwc_fhwc(%input: tensor<1x8x8x5xf32>, %filter: tensor<4x3x3x5xf32>, %output: tensor<1x6x6x4xf32>) { |
| 132 |
// expected-error @+1 {{Attempted to vectorize, but failed}} |
| 133 |
linalg.conv_2d_nhwc_fhwc {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%input, %filter : tensor<1x8x8x5xf32>, tensor<4x3x3x5xf32>) outs(%output : tensor<1x6x6x4xf32>) -> tensor<1x6x6x4xf32> |
| 134 |
return |
| 135 |
} |
| 136 |
|
| 137 |
|
| 138 |
module attributes {transform.with_named_sequence} { |
| 139 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 140 |
%0 = transform.structured.match ops{["linalg.conv_2d_nhwc_fhwc"]} in %arg1 : (!transform.any_op) -> !transform.any_op |
| 141 |
transform.structured.vectorize %0 : !transform.any_op |
| 142 |
transform.yield |
| 143 |
} |
| 144 |
} |
| 145 |
|
| 146 |
// ----- |
| 147 |
|
| 148 |
func.func @conv3d_ncdhw_fcdhw(%input: tensor<1x5x8x8x8xf32>, %filter: tensor<4x5x3x3x3xf32>, %output: tensor<1x4x6x6x6xf32>) { |
| 149 |
// expected-error @+1 {{Attempted to vectorize, but failed}} |
| 150 |
linalg.conv_3d_ncdhw_fcdhw {dilations = dense<1> : vector<3xi64>, strides = dense<1> : vector<3xi64>} ins(%input, %filter : tensor<1x5x8x8x8xf32>, tensor<4x5x3x3x3xf32>) outs(%output : tensor<1x4x6x6x6xf32>) -> tensor<1x4x6x6x6xf32> |
| 151 |
return |
| 152 |
} |
| 153 |
|
| 154 |
module attributes {transform.with_named_sequence} { |
| 155 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 156 |
%0 = transform.structured.match ops{["linalg.conv_3d_ncdhw_fcdhw"]} in %arg1 : (!transform.any_op) -> !transform.any_op |
| 157 |
transform.structured.vectorize %0 : !transform.any_op |
| 158 |
transform.yield |
| 159 |
} |
| 160 |
} |
| 161 |
|
| 162 |
// ----- |
| 163 |
|
| 164 |
func.func @test_pack_no_vectorize_dynamic_shape(%arg0: tensor<?xf32>, %arg1: tensor<4x16xf32>) -> tensor<4x16xf32> { |
| 165 |
%pad = arith.constant 0.000000e+00 : f32 |
| 166 |
// expected-error @+1 {{Attempted to vectorize, but failed}} |
| 167 |
%pack = linalg.pack %arg0 padding_value(%pad : f32) inner_dims_pos = [0] inner_tiles = [16] into %arg1 : tensor<?xf32> -> tensor<4x16xf32> |
| 168 |
return %pack : tensor<4x16xf32> |
| 169 |
} |
| 170 |
|
| 171 |
module attributes {transform.with_named_sequence} { |
| 172 |
transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) { |
| 173 |
%0 = transform.structured.match ops{["linalg.pack"]} in %arg0 : (!transform.any_op) -> !transform.any_op |
| 174 |
transform.structured.vectorize %0 : !transform.any_op |
| 175 |
transform.yield |
| 176 |
} |
| 177 |
} |
| 178 |
|
| 179 |
// ----- |
| 180 |
|
| 181 |
func.func @linalg_reduce_scalable_leading_dim(%input: tensor<?x?xf32>, |
| 182 |
%acc: tensor<?xf32>) -> tensor<?xf32> { |
| 183 |
|
| 184 |
// expected-error @+1 {{Attempted to vectorize, but failed}} |
| 185 |
%0 = linalg.reduce ins(%input : tensor<?x?xf32>) outs(%acc : tensor<?xf32>) dimensions = [0] |
| 186 |
(%in: f32, %init: f32) { |
| 187 |
%0 = arith.addf %in, %init : f32 |
| 188 |
linalg.yield %0 : f32 |
| 189 |
} |
| 190 |
return %0 : tensor<?xf32> |
| 191 |
} |
| 192 |
|
| 193 |
module attributes {transform.with_named_sequence} { |
| 194 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 195 |
%0 = transform.structured.match ops{["linalg.reduce"]} in %arg1 : (!transform.any_op) -> !transform.any_op |
| 196 |
transform.structured.vectorize %0 vector_sizes [[4], 1] : !transform.any_op |
| 197 |
transform.yield |
| 198 |
} |
| 199 |
} |
| 200 |
|
| 201 |
// ----- |
| 202 |
|
| 203 |
func.func @linalg_generic_reduction_scalable_leading_dim(%input: tensor<?x?xf32>, |
| 204 |
%acc: tensor<?xf32>) -> tensor<?xf32> { |
| 205 |
|
| 206 |
// expected-error @+1 {{Attempted to vectorize, but failed}} |
| 207 |
%0 = linalg.generic { indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>, |
| 208 |
affine_map<(d0, d1) -> (d1)>], |
| 209 |
iterator_types = ["reduction", "parallel"] } |
| 210 |
ins(%input : tensor<?x?xf32>) |
| 211 |
outs(%acc : tensor<?xf32>) { |
| 212 |
^bb(%in: f32, %out: f32) : |
| 213 |
%0 = arith.addf %in, %out : f32 |
| 214 |
linalg.yield %0 : f32 |
| 215 |
} -> tensor<?xf32> |
| 216 |
return %0 : tensor<?xf32> |
| 217 |
} |
| 218 |
|
| 219 |
module attributes {transform.with_named_sequence} { |
| 220 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 221 |
%0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op |
| 222 |
transform.structured.vectorize %0 vector_sizes [[4], 1] : !transform.any_op |
| 223 |
transform.yield |
| 224 |
} |
| 225 |
} |
| 226 |
|
| 227 |
// ----- |
| 228 |
|
| 229 |
func.func @linalg_matvec_scalable_two_dims(%A: memref<?x?xf32>, %B: memref<?xf32>, %C: memref<?xf32>) { |
| 230 |
// expected-error @+1 {{Attempted to vectorize, but failed}} |
| 231 |
linalg.matvec ins(%A, %B: memref<?x?xf32>, memref<?xf32>) |
| 232 |
outs(%C: memref<?xf32>) |
| 233 |
return |
| 234 |
} |
| 235 |
|
| 236 |
module attributes {transform.with_named_sequence} { |
| 237 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 238 |
%matmul = transform.structured.match ops{["linalg.matvec"]} in %arg1 : (!transform.any_op) -> !transform.any_op |
| 239 |
transform.structured.vectorize %matmul vector_sizes [[4], [4]] : !transform.any_op |
| 240 |
transform.yield |
| 241 |
} |
| 242 |
} |
| 243 |
|
| 244 |
// ----- |
| 245 |
|
| 246 |
func.func @linalg_matmul_scalable_leading_parallel_dim(%A: memref<?x?xf32>, %B: memref<?x?xf32>, %C: memref<?x?xf32>) { |
| 247 |
// expected-error @+1 {{Attempted to vectorize, but failed}} |
| 248 |
linalg.matmul ins(%A, %B: memref<?x?xf32>, memref<?x?xf32>) |
| 249 |
outs(%C: memref<?x?xf32>) |
| 250 |
return |
| 251 |
} |
| 252 |
|
| 253 |
module attributes {transform.with_named_sequence} { |
| 254 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 255 |
%matmul = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op |
| 256 |
transform.structured.vectorize %matmul vector_sizes [[8], 16, 4] : !transform.any_op |
| 257 |
transform.yield |
| 258 |
} |
| 259 |
} |
| 260 |
|
| 261 |
// ----- |
| 262 |
|
| 263 |
func.func @linalg_matmul_scalable_trailing_reduction_dim(%A: memref<?x?xf32>, %B: memref<?x?xf32>, %C: memref<?x?xf32>) { |
| 264 |
// expected-error @+1 {{Attempted to vectorize, but failed}} |
| 265 |
linalg.matmul ins(%A, %B: memref<?x?xf32>, memref<?x?xf32>) |
| 266 |
outs(%C: memref<?x?xf32>) |
| 267 |
return |
| 268 |
} |
| 269 |
|
| 270 |
module attributes {transform.with_named_sequence} { |
| 271 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 272 |
%matmul = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op |
| 273 |
transform.structured.vectorize %matmul vector_sizes [8, 16, [4]] : !transform.any_op |
| 274 |
transform.yield |
| 275 |
} |
| 276 |
} |
| 277 |
|
| 278 |
// ----- |
| 279 |
|
| 280 |
func.func @linalg_generic_matmul_scalable_two_trailing_dims(%A: tensor<?x64xf32>, %B: tensor<64x?xf32>, |
| 281 |
%C: tensor<?x?xf32>) -> tensor<?x?xf32> { |
| 282 |
|
| 283 |
// expected-error @+1 {{Attempted to vectorize, but failed}} |
| 284 |
%0 = linalg.generic { indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d2)>, |
| 285 |
affine_map<(d0, d1, d2) -> (d2, d1)>, |
| 286 |
affine_map<(d0, d1, d2) -> (d0, d1)>], |
| 287 |
iterator_types = ["parallel", "parallel", "reduction"] } |
| 288 |
ins(%A, %B : tensor<?x64xf32>, tensor<64x?xf32>) |
| 289 |
outs(%C: tensor<?x?xf32>) { |
| 290 |
^bb(%in1: f32, %in2: f32, %out: f32) : |
| 291 |
%0 = arith.mulf %in1, %in2 : f32 |
| 292 |
%1 = arith.addf %0, %out : f32 |
| 293 |
linalg.yield %1 : f32 |
| 294 |
} -> tensor<?x?xf32> |
| 295 |
return %0 : tensor<?x?xf32> |
| 296 |
} |
| 297 |
|
| 298 |
module attributes {transform.with_named_sequence} { |
| 299 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 300 |
%0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op |
| 301 |
transform.structured.vectorize %0 vector_sizes [2, [4], [4]] : !transform.any_op |
| 302 |
transform.yield |
| 303 |
} |
| 304 |
} |
| 305 |
|
| 306 |
// ----- |
| 307 |
|
| 308 |
// Padding with non-zero low pad values is not supported, unless the corresponding |
| 309 |
// result dim is 1. Here `%l0` being a non-zero low pad applied to a |
| 310 |
// non-unit result dimension makes this case unsupported. |
| 311 |
func.func @tensor_pad_non_zero_low_pad( |
| 312 |
%0 : tensor<?x?xf32>, %h0 : index, %h1 : index, %l0 : index) |
| 313 |
-> tensor<2x4xf32> { |
| 314 |
// expected-error @+3 {{Attempted to vectorize, but failed}} |
| 315 |
%cst = arith.constant 42.43 : f32 |
| 316 |
%c0 = arith.constant 0 : index |
| 317 |
%1 = tensor.pad %0 low[%l0, %c0] high[%h0, %h1] { |
| 318 |
^bb0(%hh1: index, %hh2: index): |
| 319 |
tensor.yield %cst : f32 |
| 320 |
} : tensor<?x?xf32> to tensor<2x4xf32> |
| 321 |
return %1: tensor<2x4xf32> |
| 322 |
} |
| 323 |
|
| 324 |
module attributes {transform.with_named_sequence} { |
| 325 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 326 |
%0 = transform.structured.match ops{["tensor.pad"]} in %arg1 |
| 327 |
: (!transform.any_op) -> !transform.any_op |
| 328 |
transform.structured.vectorize %0 vector_sizes [2, 4] : !transform.any_op |
| 329 |
transform.yield |
| 330 |
} |
| 331 |
} |
| 332 |
|
| 333 |
// ----- |
| 334 |
|
| 335 |
// With dynamically shaped source, the vectorizer infers the vector size for |
| 336 |
// xfer Ops from the destination tensor and, conservatively, assumes |
| 337 |
// out-of-bounds accesses. Out-of-bounds accesses require a pad value, but |
| 338 |
// that's impossible to recover in this example. Hence no vectorization. |
| 339 |
|
| 340 |
// TODO: Use diagnostics once we can vectorize tensor.insert_slice with |
| 341 |
// transform.structured.vectorize |
| 342 |
|
| 343 |
// CHECK-LABEL: @insert_dynamic_slice_unknown_pad |
| 344 |
// CHECK-NOT: vector |
| 345 |
// CHECK: tensor.insert_slice |
| 346 |
func.func @insert_dynamic_slice_unknown_pad(%arg0: tensor<1x?x3xf32>, %arg1: tensor<9x8x7x1x2x3xf32>, %size: index) -> tensor<9x8x7x1x2x3xf32> { |
| 347 |
%res = tensor.insert_slice %arg0 into %arg1[0, 0, 0, 0, 0, 0] [1, 1, 1, 1, %size, 3][1, 1, 1, 1, 1, 1] : tensor<1x?x3xf32> into tensor<9x8x7x1x2x3xf32> |
| 348 |
return %res : tensor<9x8x7x1x2x3xf32> |
| 349 |
} |
| 350 |
|
| 351 |
module attributes {transform.with_named_sequence} { |
| 352 |
transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) { |
| 353 |
%0 = transform.structured.match ops{["tensor.insert_slice"]} in %arg1 : (!transform.any_op) -> !transform.any_op |
| 354 |
%1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op |
| 355 |
%2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op |
| 356 |
transform.yield |
| 357 |
} |
| 358 |
} |