Debian Codegraph @

mlir/test/Dialect/Linalg/vectorization/linalg-ops-with-patterns.mlir

raw
1 // RUN: mlir-opt %s -transform-interpreter -split-input-file | FileCheck %s
2
3 // CHECK-LABEL: contraction_dot
4 func.func @contraction_dot(%A: memref<1584xf32>, %B: memref<1584xf32>, %C: memref<f32>) {
5
6 // CHECK: arith.mulf %{{.*}}, %{{.*}} : vector<1584xf32>
7 // CHECK: vector.multi_reduction <add>, %{{.*}}, {{.*}} [0] : vector<1584xf32> to f32
8 linalg.dot ins(%A, %B: memref<1584xf32>, memref<1584xf32>)
9 outs(%C: memref<f32>)
10 return
11 }
12
13 module attributes {transform.with_named_sequence} {
14 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
15 %0 = transform.structured.match ops{["linalg.dot"]} in %arg1 : (!transform.any_op) -> !transform.any_op
16 transform.structured.vectorize %0 : !transform.any_op
17 transform.yield
18 }
19 }
20
21 // -----
22
23 // CHECK-LABEL: contraction_matvec
24 func.func @contraction_matvec(%A: memref<1584x1584xf32>, %B: memref<1584xf32>, %C: memref<1584xf32>) {
25
26 // CHECK: arith.mulf %{{.*}}, %{{.*}} : vector<1584x1584xf32>
27 // CHECK: vector.multi_reduction <add>, %{{.*}}, {{.*}} [1] : vector<1584x1584xf32> to vector<1584xf32>
28 linalg.matvec ins(%A, %B: memref<1584x1584xf32>, memref<1584xf32>)
29 outs(%C: memref<1584xf32>)
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.matvec"]} in %arg1 : (!transform.any_op) -> !transform.any_op
36 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
37 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns } : (!transform.any_op) -> !transform.any_op
38 transform.yield
39 }
40 }
41
42 // -----
43
44 // CHECK-LABEL: contraction_matmul
45 func.func @contraction_matmul(%A: memref<1584x1584xf32>, %B: memref<1584x1584xf32>, %C: memref<1584x1584xf32>) {
46 // CHECK: arith.mulf %{{.*}}, %{{.*}} : vector<1584x1584x1584xf32>
47 // CHECK: vector.multi_reduction <add>, %{{.*}}, {{.*}} [2] : vector<1584x1584x1584xf32> to vector<1584x1584xf32>
48 linalg.matmul ins(%A, %B: memref<1584x1584xf32>, memref<1584x1584xf32>)
49 outs(%C: memref<1584x1584xf32>)
50 return
51 }
52
53 module attributes {transform.with_named_sequence} {
54 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
55 %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
56 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
57 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns } : (!transform.any_op) -> !transform.any_op
58 transform.yield
59 }
60 }
61
62 // -----
63
64 // CHECK-LABEL: contraction_batch_matmul
65 func.func @contraction_batch_matmul(%A: memref<1584x1584x1584xf32>, %B: memref<1584x1584x1584xf32>, %C: memref<1584x1584x1584xf32>) {
66 // CHECK: arith.mulf %{{.*}}, %{{.*}} : vector<1584x1584x1584x1584xf32>
67 // CHECK: vector.multi_reduction <add>, %{{.*}}, {{.*}} [3] : vector<1584x1584x1584x1584xf32> to vector<1584x1584x1584xf32>
68 linalg.batch_matmul
69 ins(%A, %B: memref<1584x1584x1584xf32>, memref<1584x1584x1584xf32>)
70 outs(%C: memref<1584x1584x1584xf32>)
71 return
72 }
73
74 module attributes {transform.with_named_sequence} {
75 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
76 %0 = transform.structured.match ops{["linalg.batch_matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
77 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
78 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns } : (!transform.any_op) -> !transform.any_op
79 transform.yield
80 }
81 }
82
83 // -----
84
85 // CHECK-LABEL: @matmul_as_contract
86 // CHECK-SAME: %[[A:.*]]: tensor<24x12xf32>
87 // CHECK-SAME: %[[B:.*]]: tensor<12x25xf32>
88 // CHECK-SAME: %[[C:.*]]: tensor<24x25xf32>
89 func.func @matmul_as_contract(%A: tensor<24x12xf32>,
90 %B: tensor<12x25xf32>,
91 %C: tensor<24x25xf32>) -> tensor<24x25xf32> {
92 // CHECK: %[[vA:.+]] = vector.transfer_read %[[A]]
93 // CHECK: %[[vB:.+]] = vector.transfer_read %[[B]]
94 // CHECK: %[[vC:.+]] = vector.transfer_read %[[C]]
95 // CHECK: %[[vR:.+]] = vector.contract {{.*}} %[[vA]], %[[vB]], %[[vC]]
96 // CHECK: vector.transfer_write %[[vR]], %[[C]]
97 %0 = linalg.contract
98 indexing_maps = [affine_map<(m, n, k) -> (m, k)>,
99 affine_map<(m, n, k) -> (k, n)>,
100 affine_map<(m, n, k) -> (m, n)>]
101 ins(%A, %B : tensor<24x12xf32>, tensor<12x25xf32>)
102 outs(%C : tensor<24x25xf32>) -> tensor<24x25xf32>
103 func.return %0 : tensor<24x25xf32>
104 }
105
106 module attributes {transform.with_named_sequence} {
107 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
108 %0 = transform.structured.match ops{["linalg.contract"]} in %arg1 : (!transform.any_op) -> !transform.any_op
109 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
110 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
111 // TODO: also tests the other available vectorization strategies
112 transform.yield
113 }
114 }
115
116 // -----
117
118 #matmul_trait = {
119 indexing_maps = [
120 affine_map<(m, n, k) -> (m, k)>,
121 affine_map<(m, n, k) -> (k, n)>,
122 affine_map<(m, n, k) -> (m, n)>
123 ],
124 iterator_types = ["parallel", "parallel", "reduction"]
125 }
126
127 // CHECK-LABEL: func @vectorization_test
128 func.func @vectorization_test(%A: memref<8x16xf32>, %B: memref<16x32xf32>,
129 %C: memref<8x32xf32>) {
130 // CHECK: vector.transfer_read %{{.*}} : memref<8x16xf32>, vector<8x32x16xf32>
131 // CHECK: vector.transfer_read %{{.*}} : memref<16x32xf32>, vector<8x32x16xf32>
132 // CHECK: %[[ACC:.*]] = vector.transfer_read %{{.*}} : memref<8x32xf32>, vector<8x32xf32>
133 // CHECK: %[[MUL:.*]] = arith.mulf %{{.*}}, %{{.*}} : vector<8x32x16xf32>
134 // CHECK: %[[R:.*]] = vector.multi_reduction <add>, %[[MUL]], %[[ACC]] [2] : vector<8x32x16xf32> to vector<8x32xf32>
135 // CHECK: vector.transfer_write %{{.*}}, %{{.*}} : vector<8x32xf32>, memref<8x32xf32>
136 linalg.generic #matmul_trait
137 ins(%A, %B : memref<8x16xf32>, memref<16x32xf32>)
138 outs(%C : memref<8x32xf32>) {
139 ^bb(%a: f32, %b: f32, %c: f32) :
140 %d = arith.mulf %a, %b: f32
141 %e = arith.addf %c, %d: f32
142 linalg.yield %e : f32
143 }
144 return
145 }
146
147 module attributes {transform.with_named_sequence} {
148 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
149 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
150 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
151 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns, disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op
152 transform.yield
153 }
154 }
155
156 // -----
157
158 #map = affine_map<() -> ()>
159
160 // CHECK-LABEL: func.func @generic_0d(
161 // CHECK-SAME: %[[ARG_0:.*]]: tensor<f32>, %[[ARG_1:.*]]: tensor<f32>, %[[ARG_2:.*]]: tensor<f32>)
162 func.func @generic_0d(%arg0: tensor<f32>, %arg1: tensor<f32>,
163 %arg2: tensor<f32>) -> tensor<f32> {
164 // CHECK: %[[PAD:.*]] = ub.poison : f32
165 // CHECK: %[[READ_0:.*]] = vector.transfer_read %[[ARG_0]][], %[[PAD]] : tensor<f32>, vector<f32>
166 // CHECK: %[[ARG_0_AS_SCALAR:.*]] = vector.extract %[[READ_0]][] : f32 from vector<f32>
167 // CHECK: %[[READ_1:.*]] = vector.transfer_read %[[ARG_1]][], %[[PAD]] : tensor<f32>, vector<f32>
168 // CHECK: %[[ARG_1_AS_SCALAR:.*]] = vector.extract %[[READ_1]][] : f32 from vector<f32>
169 // CHECK: %[[READ_2:.*]] = vector.transfer_read %[[ARG_2]][], %[[PAD]] : tensor<f32>, vector<f32>
170 // CHECK: %[[ARG_2_AS_SCALAR:.*]] = vector.extract %[[READ_2]][] : f32 from vector<f32>
171 // CHECK: %[[MULF:.*]] = arith.mulf %[[ARG_0_AS_SCALAR]], %[[ARG_1_AS_SCALAR]] : f32
172 // CHECK: %[[ADDF:.*]] = arith.addf %[[ARG_2_AS_SCALAR]], %[[MULF]] : f32
173 // CHECK: %[[ADDF_BCAST:.*]] = vector.broadcast %[[ADDF]] : f32 to vector<f32>
174 // CHECK: vector.transfer_write %[[ADDF_BCAST]], %[[ARG_2]][] : vector<f32>, tensor<f32>
175 %res = linalg.generic {
176 indexing_maps = [#map, #map, #map],
177 iterator_types = []
178 } ins(%arg0, %arg1 : tensor<f32>, tensor<f32>)
179 outs(%arg2 : tensor<f32>) {
180 ^bb(%a: f32, %b: f32, %c: f32) :
181 %d = arith.mulf %a, %b: f32
182 %e = arith.addf %c, %d: f32
183 linalg.yield %e : f32
184 } -> tensor<f32>
185
186 return %res : tensor<f32>
187 }
188
189 module attributes {transform.with_named_sequence} {
190 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
191 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
192 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
193 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns, disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op
194 transform.yield
195 }
196 }
197
198 // -----
199
200 #matmul_transpose_out_trait = {
201 indexing_maps = [
202 affine_map<(m, n, k) -> (m, k)>,
203 affine_map<(m, n, k) -> (k, n)>,
204 affine_map<(m, n, k) -> (n, m)>
205 ],
206 iterator_types = ["parallel", "parallel", "reduction"]
207 }
208
209 // CHECK-LABEL: func @generic_output_transpose
210 func.func @generic_output_transpose(%A: memref<8x16xf32>, %B: memref<16x32xf32>,
211 %C: memref<32x8xf32>) {
212 // CHECK: vector.transfer_read %{{.*}} : memref<8x16xf32>, vector<8x32x16xf32>
213 // CHECK: vector.transfer_read %{{.*}} : memref<16x32xf32>, vector<8x32x16xf32>
214 // CHECK: %[[ACC:.*]] = vector.transfer_read %{{.*}} : memref<32x8xf32>, vector<8x32xf32>
215 // CHECK: %[[MUL:.*]] = arith.mulf %{{.*}}, %{{.*}} : vector<8x32x16xf32>
216 // CHECK: %[[R:.*]] = vector.multi_reduction <add>, %[[MUL]], %[[ACC]] [2] : vector<8x32x16xf32> to vector<8x32xf32>
217 // CHECK: vector.transfer_write %{{.*}}, %{{.*}} : vector<8x32xf32>, memref<32x8xf32>
218 linalg.generic #matmul_transpose_out_trait
219 ins(%A, %B : memref<8x16xf32>, memref<16x32xf32>)
220 outs(%C : memref<32x8xf32>) {
221 ^bb(%a: f32, %b: f32, %c: f32) :
222 %d = arith.mulf %a, %b: f32
223 %e = arith.addf %c, %d: f32
224 linalg.yield %e : f32
225 }
226 return
227 }
228
229 module attributes {transform.with_named_sequence} {
230 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
231 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
232 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
233 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns, disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op
234 transform.yield
235 }
236 }
237
238 // -----
239
240 #map0 = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
241 #map1 = affine_map<(d0, d1, d2) -> (d1, d0, d2)>
242 // CHECK: #[[MAP:.+]] = affine_map<(d0, d1, d2) -> (d1, d0, d2)>
243 // CHECK: func @generic_interchanged_transpose
244 func.func @generic_interchanged_transpose(%arg0: tensor<12x128x32xf32>) -> tensor<128x12x32xf32> {
245 // CHECK: %[[IN:.+]] = vector.transfer_read
246 // CHECK: vector.transfer_write %[[IN]], {{.+}} permutation_map = #[[MAP]]
247 %0 = tensor.empty() : tensor<128x12x32xf32>
248 %1 = linalg.generic {indexing_maps = [#map0, #map1],
249 iterator_types = ["parallel", "parallel", "parallel"]}
250 ins(%arg0 : tensor<12x128x32xf32>)
251 outs(%0 : tensor<128x12x32xf32>) {
252 ^bb0(%arg1: f32, %arg2: f32):
253 linalg.yield %arg1 : f32
254 } -> tensor<128x12x32xf32>
255 return %1 : tensor<128x12x32xf32>
256 }
257
258 module attributes {transform.with_named_sequence} {
259 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
260 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
261 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
262 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns, disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op
263 transform.yield
264 }
265 }
266
267 // -----
268
269 #matmul_trait = {
270 indexing_maps = [
271 affine_map<(m, n, k) -> (m, k)>,
272 affine_map<(m, n, k) -> (k, n)>,
273 affine_map<(m, n, k) -> (m, n)>
274 ],
275 iterator_types = ["parallel", "parallel", "reduction"]
276 }
277
278 // CHECK-LABEL: func @vectorization_test_integer
279 func.func @vectorization_test_integer(%A: memref<8x16xi32>, %B: memref<16x32xi32>,
280 %C: memref<8x32xi32>) {
281 // CHECK: vector.transfer_read %{{.*}} : memref<8x16xi32>, vector<8x32x16xi32>
282 // CHECK: vector.transfer_read %{{.*}} : memref<16x32xi32>, vector<8x32x16xi32>
283 // CHECK: %[[ACC:.*]] = vector.transfer_read %{{.*}} : memref<8x32xi32>, vector<8x32xi32>
284 // CHECK: %[[MUL:.*]] = arith.muli %{{.*}}, %{{.*}} : vector<8x32x16xi32>
285 // CHECK: vector.multi_reduction <add>, %[[MUL]], %[[ACC]] [2] : vector<8x32x16xi32> to vector<8x32xi32>
286 // CHECK: vector.transfer_write %{{.*}}, %{{.*}} : vector<8x32xi32>, memref<8x32xi32>
287 linalg.generic #matmul_trait
288 ins(%A, %B : memref<8x16xi32>, memref<16x32xi32>)
289 outs(%C : memref<8x32xi32>) {
290 ^bb(%a: i32, %b: i32, %c: i32) :
291 %d = arith.muli %a, %b: i32
292 %e = arith.addi %c, %d: i32
293 linalg.yield %e : i32
294 }
295 return
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 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
302 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns, disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op
303 transform.yield
304 }
305 }
306
307 // -----
308
309 // CHECK-LABEL: func @vectorization_test_2
310 func.func @vectorization_test_2(%A: memref<8x16xf32>, %B: memref<16x32xf32>,
311 %C: memref<8x32xf32>) {
312 // CHECK: arith.mulf %{{.*}}, %{{.*}} : vector<8x32x16xf32>
313 // CHECK: vector.multi_reduction <add>, %{{.*}}, {{.*}} [2] : vector<8x32x16xf32> to vector<8x32xf32>
314 linalg.matmul
315 ins(%A, %B: memref<8x16xf32>, memref<16x32xf32>)
316 outs(%C: memref<8x32xf32>)
317 return
318 }
319
320 module attributes {transform.with_named_sequence} {
321 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
322 %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
323 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
324 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns } : (!transform.any_op) -> !transform.any_op
325 transform.yield
326 }
327 }
328
329 // -----
330
331 // CHECK-LABEL: func @test_vectorize_scalar_input
332 func.func @test_vectorize_scalar_input(%A : memref<8x16xf32>, %arg0 : f32) {
333 // CHECK: %[[V:.*]] = vector.broadcast {{.*}} : f32 to vector<8x16xf32>
334 // CHECK: vector.transfer_write %[[V]], {{.*}} : vector<8x16xf32>, memref<8x16xf32>
335 linalg.generic {
336 indexing_maps = [affine_map<(m, n) -> ()>, affine_map<(m, n) -> (m, n)>],
337 iterator_types = ["parallel", "parallel"]}
338 ins(%arg0 : f32)
339 outs(%A: memref<8x16xf32>) {
340 ^bb(%0: f32, %1: f32) :
341 linalg.yield %0 : f32
342 }
343 return
344 }
345
346 module attributes {transform.with_named_sequence} {
347 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
348 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
349 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
350 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
351 transform.yield
352 }
353 }
354
355 // -----
356
357 // CHECK-LABEL: func @test_do_not_vectorize_unsupported_element_types
358 func.func @test_do_not_vectorize_unsupported_element_types(%A : memref<8x16xcomplex<f32>>, %arg0 : complex<f32>) {
359 // CHECK-NOT: vector.broadcast
360 // CHECK-NOT: vector.transfer_write
361 linalg.generic {
362 indexing_maps = [affine_map<(m, n) -> ()>, affine_map<(m, n) -> (m, n)>],
363 iterator_types = ["parallel", "parallel"]}
364 ins(%arg0 : complex<f32>)
365 outs(%A: memref<8x16xcomplex<f32>>) {
366 ^bb(%0: complex<f32>, %1: complex<f32>) :
367 linalg.yield %0 : complex<f32>
368 }
369 return
370 }
371
372 module attributes {transform.with_named_sequence} {
373 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
374 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
375 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
376 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
377 transform.yield
378 }
379 }
380
381 // -----
382
383 #map0 = affine_map<(d0) -> (d0)>
384
385 func.func @vectorize_affine_apply(%arg0: tensor<5xf32>, %arg3: index) -> tensor<5xi32> {
386 %0 = tensor.empty() : tensor<5xi32>
387 %1 = linalg.generic {indexing_maps = [#map0, #map0],
388 iterator_types = ["parallel"]}
389 ins(%arg0 : tensor<5xf32>)
390 outs(%0 : tensor<5xi32>) {
391 ^bb0(%arg1: f32, %arg2: i32):
392 %2 = linalg.index 0 : index
393 %11 = affine.apply affine_map<() -> (123)>()
394 %12 = affine.apply affine_map<(d0, d1) -> (d0 + d1)>(%2, %11)
395 %13 = affine.apply affine_map<(d0)[s0] -> (d0 + s0)>(%12)[%arg3]
396 %14 = affine.apply affine_map<(d0) -> (d0 + 1)>(%13)
397 %15 = affine.apply affine_map<(d0, d1, d2) -> (d0 + d1 + d2)>(%13, %14, %12)
398 %3 = arith.index_cast %15 : index to i32
399 linalg.yield %3 : i32
400 } -> tensor<5xi32>
401 return %1 : tensor<5xi32>
402 }
403
404 // CHECK-LABEL: func.func @vectorize_affine_apply
405 // CHECK-SAME: %arg0: tensor<5xf32>
406 // CHECK-SAME: %[[ARG1:.*]]: index
407 // CHECK-DAG: %[[CST:.*]] = arith.constant dense<[123, 124, 125, 126, 127]> : vector<5xindex>
408 // CHECK-DAG: %[[CST_0:.*]] = arith.constant dense<1> : vector<5xindex>
409 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index
410 // CHECK: %[[EMPTY:.*]] = tensor.empty() : tensor<5xi32>
411 // CHECK: %[[BCAST:.*]] = vector.broadcast %[[ARG1]] : index to vector<5xindex>
412 // CHECK: %[[ADDI_1:.*]] = arith.addi %[[BCAST]], %[[CST]] : vector<5xindex>
413 // CHECK: %[[ADDI_2:.*]] = arith.addi %[[ADDI_1]], %[[CST_0]] : vector<5xindex>
414 // CHECK: %[[ADDI_3:.*]] = arith.addi %[[ADDI_1]], %[[ADDI_2]] : vector<5xindex>
415 // CHECK: %[[ADDI_4:.*]] = arith.addi %[[ADDI_3]], %[[CST]] : vector<5xindex>
416 // CHECK: %[[CAST:.*]] = arith.index_cast %[[ADDI_4]] : vector<5xindex> to vector<5xi32>
417 // CHECK: vector.transfer_write %[[CAST]], %[[EMPTY]][%[[C0:.*]]] {in_bounds = [true]} : vector<5xi32>, tensor<5xi32>
418
419 module attributes {transform.with_named_sequence} {
420 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
421 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
422 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
423 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { vectorize_nd_extract } : (!transform.any_op) -> !transform.any_op
424 transform.yield
425 }
426 }
427
428 // -----
429
430 // CHECK-LABEL: func @test_vectorize_fill
431 func.func @test_vectorize_fill(%A : memref<8x16xf32>, %arg0 : f32) {
432 // CHECK: %[[V:.*]] = vector.broadcast {{.*}} : f32 to vector<8x16xf32>
433 // CHECK: vector.transfer_write %[[V]], {{.*}} : vector<8x16xf32>, memref<8x16xf32>
434 linalg.fill ins(%arg0 : f32) outs(%A : memref<8x16xf32>)
435 return
436 }
437
438 module attributes {transform.with_named_sequence} {
439 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
440 %0 = transform.structured.match ops{["linalg.fill"]} in %arg1 : (!transform.any_op) -> !transform.any_op
441 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
442 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
443 transform.yield
444 }
445 }
446
447 // -----
448
449 // CHECK-LABEL: func @test_vectorize_fill
450 func.func @test_vectorize_fill_0d(%A : memref<f32>, %arg0 : f32) {
451 // CHECK-SAME: (%[[M:.*]]: memref<f32>, %[[val:.*]]: f32)
452 // CHECK: %[[VEC:.*]] = vector.broadcast %[[val]] : f32 to vector<f32>
453 // CHECK: vector.transfer_write %[[VEC]], %[[M]][] : vector<f32>, memref<f32>
454 linalg.fill ins(%arg0 : f32) outs(%A : memref<f32>)
455 return
456 }
457
458 module attributes {transform.with_named_sequence} {
459 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
460 %0 = transform.structured.match ops{["linalg.fill"]} in %arg1 : (!transform.any_op) -> !transform.any_op
461 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
462 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
463 transform.yield
464 }
465 }
466
467 // -----
468
469 // CHECK-LABEL: func @test_vectorize_copy
470 func.func @test_vectorize_copy(%A : memref<8x16xf32>, %B : memref<8x16xf32>) {
471 // CHECK: %[[V:.*]] = vector.transfer_read {{.*}} : memref<8x16xf32>, vector<8x16xf32>
472 // CHECK: vector.transfer_write %[[V]], {{.*}} : vector<8x16xf32>, memref<8x16xf32>
473 memref.copy %A, %B : memref<8x16xf32> to memref<8x16xf32>
474 return
475 }
476
477 module attributes {transform.with_named_sequence} {
478 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
479 %0 = transform.structured.match ops{["memref.copy"]} in %arg1 : (!transform.any_op) -> !transform.any_op
480 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
481 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
482 transform.yield
483 }
484 }
485
486 // -----
487
488 // CHECK-LABEL: func @test_vectorize_copy_0d
489 func.func @test_vectorize_copy_0d(%A : memref<f32>, %B : memref<f32>) {
490 // CHECK-SAME: (%[[A:.*]]: memref<f32>, %[[B:.*]]: memref<f32>)
491 // CHECK: %[[V:.*]] = vector.transfer_read %[[A]][]{{.*}} : memref<f32>, vector<f32>
492 // CHECK: %[[val:.*]] = vector.extract %[[V]][] : f32 from vector<f32>
493 // CHECK: %[[VV:.*]] = vector.broadcast %[[val]] : f32 to vector<f32>
494 // CHECK: vector.transfer_write %[[VV]], %[[B]][] : vector<f32>, memref<f32>
495 memref.copy %A, %B : memref<f32> to memref<f32>
496 return
497 }
498
499 module attributes {transform.with_named_sequence} {
500 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
501 %0 = transform.structured.match ops{["memref.copy"]} in %arg1 : (!transform.any_op) -> !transform.any_op
502 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
503 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
504 transform.yield
505 }
506 }
507
508 // -----
509
510 // CHECK-LABEL: func @test_vectorize_copy_complex
511 // CHECK-NOT: vector<
512 func.func @test_vectorize_copy_complex(%A : memref<8x16xcomplex<f32>>, %B : memref<8x16xcomplex<f32>>) {
513 memref.copy %A, %B : memref<8x16xcomplex<f32>> to memref<8x16xcomplex<f32>>
514 return
515 }
516
517 module attributes {transform.with_named_sequence} {
518 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
519 %0 = transform.structured.match ops{["memref.copy"]} in %arg1 : (!transform.any_op) -> !transform.any_op
520 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
521 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
522 transform.yield
523 }
524 }
525
526 // -----
527
528 // CHECK-LABEL: func @test_vectorize_trailing_index
529 // CHECK-SAME: (%[[ARG0:.*]]: memref<1x2x4x8xindex>)
530 func.func @test_vectorize_trailing_index(%arg0: memref<1x2x4x8xindex>) {
531 // CHECK-DAG: %[[CST0:.*]] = arith.constant dense<[0, 1, 2, 3, 4, 5, 6, 7]> : vector<8xindex>
532 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index
533 linalg.generic {
534 indexing_maps = [
535 affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>],
536 iterator_types = ["parallel", "parallel", "parallel", "parallel"]}
537 outs(%arg0: memref<1x2x4x8xindex>) {
538 ^bb0(%arg1: index):
539 // CHECK: %[[BCST:.*]] = vector.broadcast %[[CST0]] : vector<8xindex> to vector<1x2x4x8xindex>
540 // CHECK: vector.transfer_write %[[BCST]], %[[ARG0]][%[[C0]], %[[C0]], %[[C0]], %[[C0]]] {{.*}} : vector<1x2x4x8xindex>, memref<1x2x4x8xindex>
541 %0 = linalg.index 3 : index
542 linalg.yield %0 : index
543 }
544 return
545 }
546
547 module attributes {transform.with_named_sequence} {
548 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
549 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
550 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
551 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
552 transform.yield
553 }
554 }
555
556 // -----
557
558 // CHECK-LABEL: func @test_vectorize_inner_index
559 // CHECK-SAME: (%[[ARG0:.*]]: memref<1x2x4x8xindex>)
560 func.func @test_vectorize_inner_index(%arg0: memref<1x2x4x8xindex>) {
561 // CHECK-DAG: %[[CST0:.*]] = arith.constant dense<[0, 1]> : vector<2xindex>
562 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index
563 linalg.generic {
564 indexing_maps = [
565 affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>],
566 iterator_types = ["parallel", "parallel", "parallel", "parallel"]}
567 outs(%arg0: memref<1x2x4x8xindex>) {
568 ^bb0(%arg1: index):
569 // CHECK: %[[BCST:.*]] = vector.broadcast %[[CST0]] : vector<2xindex> to vector<1x8x4x2xindex>
570 // CHECK: %[[TRAN:.*]] = vector.transpose %[[BCST]], [0, 3, 2, 1] : vector<1x8x4x2xindex> to vector<1x2x4x8xindex>
571 // CHECK: vector.transfer_write %[[TRAN]], %[[ARG0]][%[[C0]], %[[C0]], %[[C0]], %[[C0]]] {{.*}} : vector<1x2x4x8xindex>, memref<1x2x4x8xindex>
572 %0 = linalg.index 1 : index
573 linalg.yield %0 : index
574 }
575 return
576 }
577
578 module attributes {transform.with_named_sequence} {
579 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
580 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
581 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
582 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
583 transform.yield
584 }
585 }
586
587 // -----
588
589 // CHECK-LABEL: func @generic_vectorize
590 // CHECK-SAME: (%[[ARG0:.*]]: memref<4x256xf32>, %[[ARG1:.*]]: memref<4x256xf32>,
591 // CHECK-SAME: %[[ARG2:.*]]: memref<256xf32>, %[[ARG3:.*]]: f32)
592 func.func @generic_vectorize(%arg0: memref<4x256xf32>,
593 %arg1: memref<4x256xf32>,
594 %arg2: memref<256xf32>, %i: f32) {
595 // CHECK-DAG: %[[CST0:.*]] = arith.constant dense<2.000000e+00> : vector<4x256xf32>
596 // CHECK-DAG: %[[CST1:.*]] = arith.constant dense<1.000000e+00> : vector<4x256xf32>
597 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index
598 %c1_f32 = arith.constant 1.0 : f32
599 linalg.generic {
600 indexing_maps = [
601 affine_map<(d0, d1) -> (d0, d1)>,
602 affine_map<(d0, d1) -> (d1)>,
603 affine_map<(d0, d1) -> (d0, d1)>,
604 affine_map<(d0, d1) -> (d0, d1)>,
605 affine_map<(d0, d1) -> (d0, d1)>,
606 affine_map<(d0, d1) -> (d0, d1)>,
607 affine_map<(d0, d1) -> (d0, d1)>,
608 affine_map<(d0, d1) -> (d0, d1)>,
609 affine_map<(d0, d1) -> (d0, d1)>,
610 affine_map<(d0, d1) -> (d0, d1)>,
611 affine_map<(d0, d1) -> (d0, d1)>,
612 affine_map<(d0, d1) -> (d0, d1)>],
613 iterator_types = ["parallel", "parallel"]}
614 ins(%arg1, %arg2: memref<4x256xf32>, memref<256xf32>)
615 outs(
616 %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0 :
617 memref<4x256xf32>, memref<4x256xf32>, memref<4x256xf32>, memref<4x256xf32>,
618 memref<4x256xf32>, memref<4x256xf32>, memref<4x256xf32>, memref<4x256xf32>,
619 memref<4x256xf32>, memref<4x256xf32>) {
620 ^bb0(%arg3 : f32, %arg4 : f32, %arg5: f32, %arg6: f32, %arg7: f32, %arg8: f32,
621 // CHECK: %[[V2:.*]] = vector.transfer_read %[[ARG1]][%[[C0]], %[[C0]]], {{.*}} : memref<4x256xf32>, vector<4x256xf32>
622 // CHECK: %[[V0:.*]] = vector.transfer_read %[[ARG2]][%[[C0]]], {{.*}} : memref<256xf32>, vector<4x256xf32>
623 // CHECK: %[[V3:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], {{.*}} : memref<4x256xf32>, vector<4x256xf32>
624 // CHECK: %[[V1:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], {{.*}} : memref<4x256xf32>, vector<4x256xf32>
625 %arg9 : f32, %arg10 : f32, %arg11 : f32, %arg12 : f32, %arg13 : f32,
626 %arg14 : f32):
627 // CHECK: %[[ADD:.*]] = arith.addf %[[V0]], %[[V1]] : vector<4x256xf32>
628 %6 = arith.addf %arg4, %arg6 : f32
629 // CHECK: %[[CMP:.*]] = arith.cmpf ogt, %[[V2]], %[[V1]] : vector<4x256xf32>
630 %7 = arith.cmpf ogt, %arg3, %arg6 : f32
631 // CHECK: %[[ARG3B:.*]] = vector.broadcast %[[ARG3]] : f32 to vector<4x256xf32>
632 %8 = arith.constant 2.0 : f32
633 // CHECK: %[[DIV:.*]] = arith.divf %[[V3]], %[[ARG3B]] : vector<4x256xf32>
634 %9 = arith.divf %arg5, %i : f32
635 // CHECK: %[[EXP:.*]] = math.exp2 %[[V3]] : vector<4x256xf32>
636 %10 = math.exp2 %arg5 : f32
637 // CHECK: %[[MUL:.*]] = arith.mulf %[[V3]], %[[CST0]] : vector<4x256xf32>
638 %11 = arith.mulf %arg5, %8 : f32
639 // CHECK: %[[RSQRT:.*]] = math.rsqrt %[[V3]] : vector<4x256xf32>
640 %12 = math.rsqrt %arg5 : f32
641 // CHECK: %[[SEL:.*]] = arith.select %[[CMP]], %[[V3]], %[[V1]] : vector<4x256xi1>, vector<4x256xf32>
642 %13 = arith.select %7, %arg5, %arg6 : f32
643 // CHECK: %[[SUB:.*]] = arith.subf %[[V3]], %[[V0]] : vector<4x256xf32>
644 %14 = arith.subf %arg5, %arg4 : f32
645 // CHECK: %[[TAN:.*]] = math.tanh %[[V3]] : vector<4x256xf32>
646 %15 = math.tanh %arg5 : f32
647 // CHECK: vector.transfer_write %[[ADD]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>
648 // CHECK: vector.transfer_write %[[CST0]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>
649 // CHECK: vector.transfer_write %[[CST1]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>
650 // CHECK: vector.transfer_write %[[DIV]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>
651 // CHECK: vector.transfer_write %[[EXP]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>
652 // CHECK: vector.transfer_write %[[MUL]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>
653 // CHECK: vector.transfer_write %[[RSQRT]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>
654 // CHECK: vector.transfer_write %[[SEL]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>
655 // CHECK: vector.transfer_write %[[SUB]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>
656 // CHECK: vector.transfer_write %[[TAN]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, memref<4x256xf32>
657 linalg.yield %6, %8, %c1_f32, %9, %10, %11, %12, %13, %14, %15 : f32, f32,
658 f32, f32, f32, f32, f32, f32, f32, f32
659 }
660 return
661 }
662
663 module attributes {transform.with_named_sequence} {
664 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
665 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
666 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
667 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op
668 transform.yield
669 }
670 }
671
672 // -----
673
674 // CHECK-LABEL: func @generic_vectorize_tensor
675 // CHECK-SAME: (%[[ARG0:.*]]: tensor<4x256xf32>, %[[ARG1:.*]]: tensor<4x256xf32>,
676 // CHECK-SAME: %[[ARG2:.*]]: tensor<256xf32>, %[[ARG3:.*]]: f32)
677 func.func @generic_vectorize_tensor(%arg0: tensor<4x256xf32>,
678 %arg1: tensor<4x256xf32>, %arg2: tensor<256xf32>,
679 %i: f32) -> (tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>,
680 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>,
681 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>) {
682 %c1_f32 = arith.constant 1.0 : f32
683 %r:10 = linalg.generic {
684 indexing_maps = [
685 affine_map<(d0, d1) -> (d0, d1)>,
686 affine_map<(d0, d1) -> (d1)>,
687 affine_map<(d0, d1) -> (d0, d1)>,
688 affine_map<(d0, d1) -> (d0, d1)>,
689 affine_map<(d0, d1) -> (d0, d1)>,
690 affine_map<(d0, d1) -> (d0, d1)>,
691 affine_map<(d0, d1) -> (d0, d1)>,
692 affine_map<(d0, d1) -> (d0, d1)>,
693 affine_map<(d0, d1) -> (d0, d1)>,
694 affine_map<(d0, d1) -> (d0, d1)>,
695 affine_map<(d0, d1) -> (d0, d1)>,
696 affine_map<(d0, d1) -> (d0, d1)>],
697 iterator_types = ["parallel", "parallel"]}
698 ins(%arg1, %arg2: tensor<4x256xf32>, tensor<256xf32>)
699 outs(
700 %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0, %arg0 :
701 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>,
702 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>,
703 tensor<4x256xf32>, tensor<4x256xf32>) {
704 ^bb0(%arg3 : f32, %arg4 : f32, %arg5: f32, %arg6: f32, %arg7: f32, %arg8: f32,
705 %arg9 : f32, %arg10 : f32, %arg11 : f32, %arg12 : f32, %arg13 : f32,
706 %arg14 : f32):
707 // CHECK-DAG: %[[CST0:.*]] = arith.constant dense<2.000000e+00> : vector<4x256xf32>
708 // CHECK-DAG: %[[CST1:.*]] = arith.constant dense<1.000000e+00> : vector<4x256xf32>
709 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index
710 // CHECK: %[[V2:.*]] = vector.transfer_read %[[ARG1]][%[[C0]], %[[C0]]], {{.*}} : tensor<4x256xf32>, vector<4x256xf32>
711 // CHECK: %[[V0:.*]] = vector.transfer_read %[[ARG2]][%[[C0]]], {{.*}} : tensor<256xf32>, vector<4x256xf32>
712 // CHECK: %[[V3:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], {{.*}} : tensor<4x256xf32>, vector<4x256xf32>
713 // CHECK: %[[V1:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], {{.*}} : tensor<4x256xf32>, vector<4x256xf32>
714 // CHECK: %[[ADD:.*]] = arith.addf %[[V0]], %[[V1]] : vector<4x256xf32>
715 %6 = arith.addf %arg4, %arg6 : f32
716 // CHECK: %[[CMP:.*]] = arith.cmpf ogt, %[[V2]], %[[V1]] : vector<4x256xf32>
717 %7 = arith.cmpf ogt, %arg3, %arg6 : f32
718 // CHECK: %[[ARG3B:.*]] = vector.broadcast %[[ARG3]] : f32 to vector<4x256xf32>
719 %8 = arith.constant 2.0 : f32
720 // CHECK: %[[DIV:.*]] = arith.divf %[[V3]], %[[ARG3B]] : vector<4x256xf32>
721 %9 = arith.divf %arg5, %i : f32
722 // CHECK: %[[EXP:.*]] = math.exp2 %[[V3]] : vector<4x256xf32>
723 %10 = math.exp2 %arg5 : f32
724 // CHECK: %[[MUL:.*]] = arith.mulf %[[V3]], %[[CST0]] : vector<4x256xf32>
725 %11 = arith.mulf %arg5, %8 : f32
726 // CHECK: %[[RSQRT:.*]] = math.rsqrt %[[V3]] : vector<4x256xf32>
727 %12 = math.rsqrt %arg5 : f32
728 // CHECK: %[[SEL:.*]] = arith.select %[[CMP]], %[[V3]], %[[V1]] : vector<4x256xi1>, vector<4x256xf32>
729 %13 = arith.select %7, %arg5, %arg6 : f32
730 // CHECK: %[[SUB:.*]] = arith.subf %[[V3]], %[[V0]] : vector<4x256xf32>
731 %14 = arith.subf %arg5, %arg4 : f32
732 // CHECK: %[[TAN:.*]] = math.tanh %[[V3]] : vector<4x256xf32>
733 %15 = math.tanh %arg5 : f32
734 // CHECK: %[[R0:.*]] = vector.transfer_write %[[ADD]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>
735 // CHECK: %[[R1:.*]] = vector.transfer_write %[[CST0]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>
736 // CHECK: %[[R2:.*]] = vector.transfer_write %[[CST1]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>
737 // CHECK: %[[R3:.*]] = vector.transfer_write %[[DIV]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>
738 // CHECK: %[[R4:.*]] = vector.transfer_write %[[EXP]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>
739 // CHECK: %[[R5:.*]] = vector.transfer_write %[[MUL]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>
740 // CHECK: %[[R6:.*]] = vector.transfer_write %[[RSQRT]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>
741 // CHECK: %[[R7:.*]] = vector.transfer_write %[[SEL]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>
742 // CHECK: %[[R8:.*]] = vector.transfer_write %[[SUB]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>
743 // CHECK: %[[R9:.*]] = vector.transfer_write %[[TAN]], %[[ARG0]][%[[C0]], %[[C0]]] {{.*}} : vector<4x256xf32>, tensor<4x256xf32>
744 linalg.yield %6, %8, %c1_f32, %9, %10, %11, %12, %13, %14, %15 : f32, f32,
745 f32, f32, f32, f32, f32, f32, f32, f32
746 } -> (tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>,
747 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>,
748 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>)
749 // CHECK: return %[[R0]], %[[R1]], %[[R2]], %[[R3]], %[[R4]], %[[R5]], %[[R6]], %[[R7]], %[[R8]], %[[R9]] : tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>
750 return %r#0, %r#1, %r#2, %r#3, %r#4, %r#5, %r#6, %r#7, %r#8, %r#9:
751 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>,
752 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>,
753 tensor<4x256xf32>, tensor<4x256xf32>, tensor<4x256xf32>
754 }
755
756 module attributes {transform.with_named_sequence} {
757 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
758 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
759 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
760 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op
761 transform.yield
762 }
763 }
764
765 // -----
766
767 // CHECK-DAG: #[[$MAP0:.*]] = affine_map<(d0, d1) -> (d0, 0, 0, d1)>
768 // CHECK-DAG: #[[$MAP1:.*]] = affine_map<(d0) -> (d0, 0, 0, 0)>
769 // CHECK-DAG: #[[$MAP2:.*]] = affine_map<(d0) -> (0, 0, d0, 0)>
770 // CHECK-DAG: #[[$MAP3:.*]] = affine_map<(d0, d1) -> (d1, 0, d0, 0)>
771 // CHECK: func @generic_vectorize_broadcast_transpose
772 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index
773 // CHECK-DAG: %[[PV:.*]] = ub.poison : f32
774 // CHECK: %[[V0:.*]] = vector.transfer_read %{{.*}}[%[[C0]], %[[C0]]], %[[PV]] {in_bounds = [true, true, true, true], permutation_map = #[[$MAP0]]} : memref<4x4xf32>, vector<4x4x4x4xf32>
775 // CHECK: %[[V1:.*]] = vector.transfer_read %{{.*}}[%[[C0]]], %[[PV]] {in_bounds = [true, true, true, true], permutation_map = #[[$MAP1]]} : memref<4xf32>, vector<4x4x4x4xf32>
776 // CHECK: %[[V2:.*]] = vector.transfer_read %{{.*}}[%[[C0]]], %[[PV]] {in_bounds = [true, true, true, true], permutation_map = #[[$MAP2]]} : memref<4xf32>, vector<4x4x4x4xf32>
777 // CHECK: %[[V3:.*]] = vector.transfer_read %{{.*}}[%[[C0]], %[[C0]]], %[[PV]] {in_bounds = [true, true, true, true], permutation_map = #[[$MAP3]]} : memref<4x4xf32>, vector<4x4x4x4xf32>
778 // CHECK: %[[SUB:.*]] = arith.subf %[[V0]], %[[V1]] : vector<4x4x4x4xf32>
779 // CHECK: %[[ADD0:.*]] = arith.addf %[[V2]], %[[SUB]] : vector<4x4x4x4xf32>
780 // CHECK: %[[ADD1:.*]] = arith.addf %[[V3]], %[[ADD0]] : vector<4x4x4x4xf32>
781 // CHECK: vector.transfer_write %[[ADD1]], {{.*}} : vector<4x4x4x4xf32>, memref<4x4x4x4xf32>
782 func.func @generic_vectorize_broadcast_transpose(
783 %A: memref<4xf32>, %B: memref<4x4xf32>, %C: memref<4x4x4x4xf32>) {
784 linalg.generic {
785 indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0, d3)>,
786 affine_map<(d0, d1, d2, d3) -> (d0)>,
787 affine_map<(d0, d1, d2, d3) -> (d2)>,
788 affine_map<(d0, d1, d2, d3) -> (d2, d0)>,
789 affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>],
790 iterator_types = ["parallel", "parallel", "parallel", "parallel"]}
791 ins(%B, %A, %A, %B: memref<4x4xf32>, memref<4xf32>, memref<4xf32>, memref<4x4xf32>)
792 outs(%C : memref<4x4x4x4xf32>) {
793 ^bb0(%arg0: f32, %arg1: f32, %arg2: f32, %arg3: f32, %arg4: f32):
794 %s = arith.subf %arg0, %arg1 : f32
795 %a = arith.addf %arg2, %s : f32
796 %b = arith.addf %arg3, %a : f32
797 linalg.yield %b : f32
798 }
799 return
800 }
801
802 module attributes {transform.with_named_sequence} {
803 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
804 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
805 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
806 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op
807 transform.yield
808 }
809 }
810
811 // -----
812
813 // Test different input maps.
814 #matmul_trait = {
815 indexing_maps = [
816 affine_map<(d0, d1, d2, d3) -> (d1, d0)>,
817 affine_map<(d0, d1, d2, d3) -> (d3, d1)>,
818 affine_map<(d0, d1, d2, d3) -> (d3, d1, d0, d2)>,
819 affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
820 ],
821 iterator_types = ["parallel", "parallel", "parallel", "parallel"]
822 }
823
824 // CHECK-DAG: #[[MAP0:.*]] = affine_map<(d0, d1) -> (d1, d0, 0, 0)>
825 // CHECK-DAG: #[[MAP1:.*]] = affine_map<(d0, d1) -> (0, d1, 0, d0)>
826 // CHECK-DAG: #[[MAP2:.*]] = affine_map<(d0, d1, d2, d3) -> (d2, d1, d3, d0)>
827 // CHECK: func @vectorization_transpose
828 // CHECK: vector.transfer_read {{.*}}{in_bounds = [true, true, true, true], permutation_map = #[[MAP0]]} : memref<14x7xf32>, vector<7x14x8x16xf32>
829 // CHECK: vector.transfer_read {{.*}}{in_bounds = [true, true, true, true], permutation_map = #[[MAP1]]} : memref<16x14xf32>, vector<7x14x8x16xf32>
830 // CHECK: vector.transfer_read {{.*}}{in_bounds = [true, true, true, true], permutation_map = #[[MAP2]]} : memref<16x14x7x8xf32>, vector<7x14x8x16xf32>
831 // CHECK: arith.addf {{.*}} : vector<7x14x8x16xf32>
832 // CHECK: arith.addf {{.*}} : vector<7x14x8x16xf32>
833 // CHECK: vector.transfer_write {{.*}} : vector<7x14x8x16xf32>, memref<7x14x8x16xf32>
834 func.func @vectorization_transpose(%A: memref<14x7xf32>, %B: memref<16x14xf32>,
835 %C: memref<16x14x7x8xf32>, %D: memref<7x14x8x16xf32>) {
836 linalg.generic #matmul_trait
837 ins(%A, %B, %C : memref<14x7xf32>, memref<16x14xf32>, memref<16x14x7x8xf32>)
838 outs(%D : memref<7x14x8x16xf32>) {
839 ^bb(%a: f32, %b: f32, %c: f32, %d: f32) :
840 %e = arith.addf %a, %b: f32
841 %f = arith.addf %e, %c: f32
842 linalg.yield %f : f32
843 }
844 return
845 }
846
847 module attributes {transform.with_named_sequence} {
848 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
849 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
850 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
851 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op
852 transform.yield
853 }
854 }
855
856 // -----
857
858 // CHECK-LABEL: func @matmul_tensors
859 // CHECK-SAME: (%[[ARG0:.*]]: tensor<8x4xf32>, %[[ARG1:.*]]: tensor<4x12xf32>,
860 // CHECK-SAME: %[[ARG2:.*]]: tensor<8x12xf32>) -> tensor<8x12xf32>
861 func.func @matmul_tensors(
862 %arg0: tensor<8x4xf32>, %arg1: tensor<4x12xf32>, %arg2: tensor<8x12xf32>)
863 -> tensor<8x12xf32> {
864 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index
865 // CHECK-DAG: %[[V0:.*]] = vector.transfer_read %[[ARG0]][%[[C0]], %[[C0]]], {{.*}} : tensor<8x4xf32>, vector<8x12x4xf32>
866 // CHECK-DAG: %[[V1:.*]] = vector.transfer_read %[[ARG1]][%[[C0]], %[[C0]]], {{.*}} : tensor<4x12xf32>, vector<8x12x4xf32>
867 // CHECK-DAG: %[[V2:.*]] = vector.transfer_read %[[ARG2]][%[[C0]], %[[C0]]], {{.*}} : tensor<8x12xf32>, vector<8x12xf32>
868 //
869 // linalg matmul lowers gets expanded to a 3D reduction, canonicalization later
870 // convert it to a 2D contract.
871 // CHECK: %[[MUL:.*]] = arith.mulf %[[V0]], %[[V1]] : vector<8x12x4xf32>
872 // CHECK: %[[R:.*]] = vector.multi_reduction <add>, %[[MUL]], %[[V2]] [2] : vector<8x12x4xf32> to vector<8x12xf32>
873 // CHECK: %[[W:.*]] = vector.transfer_write %[[R]], %[[ARG2]][%[[C0]], %[[C0]]] {in_bounds = [true, true]} : vector<8x12xf32>, tensor<8x12xf32>
874 %0 = linalg.matmul ins(%arg0, %arg1: tensor<8x4xf32>, tensor<4x12xf32>)
875 outs(%arg2: tensor<8x12xf32>)
876 -> tensor<8x12xf32>
877 // CHECK: return %[[W]] : tensor<8x12xf32>
878 return %0 : tensor<8x12xf32>
879 }
880
881 module attributes {transform.with_named_sequence} {
882 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
883 %0 = transform.structured.match ops{["linalg.matmul"]} in %arg1 : (!transform.any_op) -> !transform.any_op
884 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
885 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns, disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op
886 transform.yield
887 }
888 }
889
890 // -----
891
892 // CHECK-LABEL: func @sum_exp
893 func.func @sum_exp(%input: tensor<4x16x8xf32>, %output: tensor<4x16xf32>)
894 -> tensor<4x16xf32>
895 {
896 // CHECK: vector.transfer_read {{.*}} : tensor<4x16x8xf32>, vector<4x16x8xf32>
897 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true]} : tensor<4x16xf32>, vector<4x16xf32>
898 // CHECK: math.exp {{.*}} : vector<4x16x8xf32>
899 // CHECK: vector.multi_reduction <add>, %{{.*}}, %{{.*}} [2] : vector<4x16x8xf32> to vector<4x16xf32>
900 // CHECK: vector.transfer_write {{.*}} : vector<4x16xf32>, tensor<4x16xf32>
901 // CHECK: return {{.*}} : tensor<4x16xf32>
902 %0 = linalg.generic {
903 indexing_maps = [
904 affine_map<(d0, d1, d2) -> (d0, d1, d2)>,
905 affine_map<(d0, d1, d2) -> (d0, d1)>
906 ],
907 iterator_types = ["parallel", "parallel", "reduction"]
908 } ins(%input : tensor<4x16x8xf32>) outs(%output : tensor<4x16xf32>) {
909 ^bb0(%arg0: f32, %arg1: f32):
910 %1 = math.exp %arg0 : f32
911 %2 = arith.addf %1, %arg1 : f32
912 linalg.yield %2 : f32
913 } -> tensor<4x16xf32>
914 return %0 : tensor<4x16xf32>
915 }
916
917
918 module attributes {transform.with_named_sequence} {
919 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
920 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
921 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
922 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op
923 transform.yield
924 }
925 }
926
927 // -----
928
929 // CHECK-DAG: #[[$M1:.*]] = affine_map<(d0, d1) -> (d1, d0, 0, 0)>
930 // CHECK-DAG: #[[$M2:.*]] = affine_map<(d0, d1) -> (0, 0, d1, d0)>
931 // CHECK-DAG: #[[$M3:.*]] = affine_map<(d0, d1) -> (d1, d0)>
932
933 // CHECK-LABEL: func @sum_exp_2
934 func.func @sum_exp_2(%input: tensor<3x2xf32>, %input_2: tensor<5x4xf32>, %output: tensor<5x2xf32>)
935 -> tensor<5x2xf32>
936 {
937 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true, true, true], permutation_map = #[[$M1]]} : tensor<3x2xf32>, vector<2x3x4x5xf32>
938 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true, true, true], permutation_map = #[[$M2]]} : tensor<5x4xf32>, vector<2x3x4x5xf32>
939 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true], permutation_map = #[[$M3]]} : tensor<5x2xf32>, vector<2x5xf32>
940 // CHECK: math.exp {{.*}} : vector<2x3x4x5xf32>
941 // CHECK: math.exp {{.*}} : vector<2x3x4x5xf32>
942 // CHECK: addf {{.*}} : vector<2x3x4x5xf32>
943 // CHECK: vector.multi_reduction <add>, {{.*}}, %{{.*}} [1, 2] : vector<2x3x4x5xf32> to vector<2x5xf32>
944 // CHECK: vector.transfer_write {{.*}} {in_bounds = [true, true], permutation_map = #[[$M3]]} : vector<2x5xf32>, tensor<5x2xf32>
945 // CHECK: return {{.*}} : tensor<5x2xf32>
946 %0 = linalg.generic {
947 indexing_maps = [
948 affine_map<(d0, d1, d2, d3) -> (d1, d0)>,
949 affine_map<(d0, d1, d2, d3) -> (d3, d2)>,
950 affine_map<(d0, d1, d2, d3) -> (d3, d0)>
951 ],
952 iterator_types = ["parallel", "reduction", "reduction", "parallel"]
953 } ins(%input, %input_2 : tensor<3x2xf32>, tensor<5x4xf32>) outs(%output : tensor<5x2xf32>) {
954 ^bb0(%arg0: f32, %arg1: f32, %arg2: f32):
955 %1 = math.exp %arg0 : f32
956 %2 = math.exp %arg1 : f32
957 %3 = arith.addf %1, %2 : f32
958 %4 = arith.addf %3, %arg2 : f32
959 linalg.yield %4 : f32
960 } -> tensor<5x2xf32>
961 return %0 : tensor<5x2xf32>
962 }
963
964
965 module attributes {transform.with_named_sequence} {
966 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
967 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
968 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
969 %5 = transform.structured.vectorize_children_and_apply_patterns %4 { disable_multi_reduction_to_contract_patterns, disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op
970 transform.yield
971 }
972 }
973
974 // -----
975
976 // CHECK-LABEL: func @red_maximumf_2d(
977 func.func @red_maximumf_2d(%arg0: tensor<4x4xf32>) -> tensor<4xf32> {
978 // CHECK: %[[CMINF:.+]] = arith.constant dense<-3.402820e+38> : vector<4xf32>
979 // CHECK: tensor.empty() : tensor<4xf32>
980 // CHECK: vector.multi_reduction <maximumf>, {{.*}}, %[[CMINF]] [1] : vector<4x4xf32> to vector<4xf32>
981 // CHECK: vector.transfer_write {{.*}} : vector<4xf32>, tensor<4xf32>
982 %ident = arith.constant -3.40282e+38 : f32
983 %init = tensor.empty() : tensor<4xf32>
984 %fill = linalg.fill ins(%ident : f32) outs(%init : tensor<4xf32>) -> tensor<4xf32>
985 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,
986 affine_map<(d0, d1) -> (d0)>],
987 iterator_types = ["parallel", "reduction"]}
988 ins(%arg0 : tensor<4x4xf32>) outs(%fill : tensor<4xf32>) {
989 ^bb0(%in0: f32, %out0: f32):
990 %max = arith.maximumf %in0, %out0 : f32
991 linalg.yield %max : f32
992 } -> tensor<4xf32>
993 return %red : tensor<4xf32>
994 }
995
996
997 module attributes {transform.with_named_sequence} {
998 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
999 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1000 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1001 %5 = transform.structured.vectorize_children_and_apply_patterns %4 { vectorize_padding } : (!transform.any_op) -> !transform.any_op
1002 transform.yield
1003 }
1004 }
1005
1006 // -----
1007
1008 // CHECK-LABEL: func @red_maxnumf_2d(
1009 func.func @red_maxnumf_2d(%arg0: tensor<4x4xf32>) -> tensor<4xf32> {
1010 // CHECK: %[[CMINF:.+]] = arith.constant dense<-3.402820e+38> : vector<4xf32>
1011 // CHECK: tensor.empty() : tensor<4xf32>
1012 // CHECK: vector.multi_reduction <maxnumf>, {{.*}}, %[[CMINF]] [1] : vector<4x4xf32> to vector<4xf32>
1013 // CHECK: vector.transfer_write {{.*}} : vector<4xf32>, tensor<4xf32>
1014 %ident = arith.constant -3.40282e+38 : f32
1015 %init = tensor.empty() : tensor<4xf32>
1016 %fill = linalg.fill ins(%ident : f32) outs(%init : tensor<4xf32>) -> tensor<4xf32>
1017 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,
1018 affine_map<(d0, d1) -> (d0)>],
1019 iterator_types = ["parallel", "reduction"]}
1020 ins(%arg0 : tensor<4x4xf32>) outs(%fill : tensor<4xf32>) {
1021 ^bb0(%in0: f32, %out0: f32):
1022 %max = arith.maxnumf %in0, %out0 : f32
1023 linalg.yield %max : f32
1024 } -> tensor<4xf32>
1025 return %red : tensor<4xf32>
1026 }
1027
1028
1029 module attributes {transform.with_named_sequence} {
1030 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
1031 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1032 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1033 %5 = transform.structured.vectorize_children_and_apply_patterns %4 { vectorize_padding } : (!transform.any_op) -> !transform.any_op
1034 transform.yield
1035 }
1036 }
1037
1038 // -----
1039
1040 // CHECK-LABEL: func @red_minimumf_2d(
1041 func.func @red_minimumf_2d(%arg0: tensor<4x4xf32>) -> tensor<4xf32> {
1042 // CHECK: %[[CMAXF:.+]] = arith.constant dense<3.402820e+38> : vector<4xf32>
1043 // CHECK: tensor.empty() : tensor<4xf32>
1044 // CHECK: vector.transfer_read {{.*}} : tensor<4x4xf32>, vector<4x4xf32>
1045 // CHECK: vector.multi_reduction <minimumf>, {{.*}}, %[[CMAXF]] [1] : vector<4x4xf32> to vector<4xf32>
1046 // CHECK: vector.transfer_write {{.*}} : vector<4xf32>, tensor<4xf32>
1047 %maxf32 = arith.constant 3.40282e+38 : f32
1048 %init = tensor.empty() : tensor<4xf32>
1049 %fill = linalg.fill ins(%maxf32 : f32) outs(%init : tensor<4xf32>) -> tensor<4xf32>
1050 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,
1051 affine_map<(d0, d1) -> (d0)>],
1052 iterator_types = ["parallel", "reduction"]}
1053 ins(%arg0 : tensor<4x4xf32>) outs(%fill : tensor<4xf32>) {
1054 ^bb0(%in0: f32, %out0: f32):
1055 %min = arith.minimumf %out0, %in0 : f32
1056 linalg.yield %min : f32
1057 } -> tensor<4xf32>
1058 return %red : tensor<4xf32>
1059 }
1060
1061
1062 module attributes {transform.with_named_sequence} {
1063 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
1064 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1065 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1066 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op
1067 transform.yield
1068 }
1069 }
1070
1071 // -----
1072
1073 // CHECK-LABEL: func @red_minnumf_2d(
1074 func.func @red_minnumf_2d(%arg0: tensor<4x4xf32>) -> tensor<4xf32> {
1075 // CHECK: %[[CMAXF:.+]] = arith.constant dense<3.402820e+38> : vector<4xf32>
1076 // CHECK: tensor.empty() : tensor<4xf32>
1077 // CHECK: vector.transfer_read {{.*}} : tensor<4x4xf32>, vector<4x4xf32>
1078 // CHECK: vector.multi_reduction <minnumf>, {{.*}}, %[[CMAXF]] [1] : vector<4x4xf32> to vector<4xf32>
1079 // CHECK: vector.transfer_write {{.*}} : vector<4xf32>, tensor<4xf32>
1080 %maxf32 = arith.constant 3.40282e+38 : f32
1081 %init = tensor.empty() : tensor<4xf32>
1082 %fill = linalg.fill ins(%maxf32 : f32) outs(%init : tensor<4xf32>) -> tensor<4xf32>
1083 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,
1084 affine_map<(d0, d1) -> (d0)>],
1085 iterator_types = ["parallel", "reduction"]}
1086 ins(%arg0 : tensor<4x4xf32>) outs(%fill : tensor<4xf32>) {
1087 ^bb0(%in0: f32, %out0: f32):
1088 %min = arith.minnumf %out0, %in0 : f32
1089 linalg.yield %min : f32
1090 } -> tensor<4xf32>
1091 return %red : tensor<4xf32>
1092 }
1093
1094
1095 module attributes {transform.with_named_sequence} {
1096 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
1097 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1098 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1099 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op
1100 transform.yield
1101 }
1102 }
1103
1104 // -----
1105
1106 // CHECK-LABEL: func @red_mul_2d(
1107 func.func @red_mul_2d(%arg0: tensor<4x4xf32>) -> tensor<4xf32> {
1108 // CHECK: tensor.empty() : tensor<4xf32>
1109 // CHECK: vector.transfer_read {{.*}} : tensor<4x4xf32>, vector<4x4xf32>
1110 // CHECK: vector.multi_reduction <mul>, {{.*}}, {{.*}} [1] : vector<4x4xf32> to vector<4xf32>
1111 // CHECK: vector.transfer_write {{.*}} : vector<4xf32>, tensor<4xf32>
1112 %ident = arith.constant 1.0 : f32
1113 %init = tensor.empty() : tensor<4xf32>
1114 %fill = linalg.fill ins(%ident : f32) outs(%init : tensor<4xf32>) -> tensor<4xf32>
1115 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,
1116 affine_map<(d0, d1) -> (d0)>],
1117 iterator_types = ["parallel", "reduction"]}
1118 ins(%arg0 : tensor<4x4xf32>) outs(%fill : tensor<4xf32>) {
1119 ^bb0(%in0: f32, %out0: f32):
1120 %mul = arith.mulf %in0, %out0 : f32
1121 linalg.yield %mul : f32
1122 } -> tensor<4xf32>
1123 return %red : tensor<4xf32>
1124 }
1125
1126
1127 module attributes {transform.with_named_sequence} {
1128 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
1129 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1130 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1131 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op
1132 transform.yield
1133 }
1134 }
1135
1136 // -----
1137
1138 // CHECK-LABEL: func @red_or_2d(
1139 func.func @red_or_2d(%arg0: tensor<4x4xi1>) -> tensor<4xi1> {
1140 // CHECK: tensor.empty() : tensor<4xi1>
1141 // CHECK: vector.transfer_read {{.*}} : tensor<4x4xi1>, vector<4x4xi1>
1142 // CHECK: vector.multi_reduction <or>, {{.*}}, {{.*}} [1] : vector<4x4xi1> to vector<4xi1>
1143 // CHECK: vector.transfer_write {{.*}} : vector<4xi1>, tensor<4xi1>
1144 %ident = arith.constant false
1145 %init = tensor.empty() : tensor<4xi1>
1146 %fill = linalg.fill ins(%ident : i1) outs(%init : tensor<4xi1>) -> tensor<4xi1>
1147 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,
1148 affine_map<(d0, d1) -> (d0)>],
1149 iterator_types = ["parallel", "reduction"]}
1150 ins(%arg0 : tensor<4x4xi1>) outs(%fill : tensor<4xi1>) {
1151 ^bb0(%in0: i1, %out0: i1):
1152 %or = arith.ori %in0, %out0 : i1
1153 linalg.yield %or : i1
1154 } -> tensor<4xi1>
1155 return %red : tensor<4xi1>
1156 }
1157
1158
1159 module attributes {transform.with_named_sequence} {
1160 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
1161 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1162 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1163 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op
1164 transform.yield
1165 }
1166 }
1167
1168 // -----
1169
1170 // CHECK-LABEL: func @red_and_2d(
1171 func.func @red_and_2d(%arg0: tensor<4x4xi1>) -> tensor<4xi1> {
1172 // CHECK: tensor.empty() : tensor<4xi1>
1173 // CHECK: vector.transfer_read {{.*}} : tensor<4x4xi1>, vector<4x4xi1>
1174 // CHECK: vector.multi_reduction <and>, {{.*}}, {{.*}} [1] : vector<4x4xi1> to vector<4xi1>
1175 // CHECK: vector.transfer_write {{.*}} : vector<4xi1>, tensor<4xi1>
1176 %ident = arith.constant true
1177 %init = tensor.empty() : tensor<4xi1>
1178 %fill = linalg.fill ins(%ident : i1) outs(%init : tensor<4xi1>) -> tensor<4xi1>
1179 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,
1180 affine_map<(d0, d1) -> (d0)>],
1181 iterator_types = ["parallel", "reduction"]}
1182 ins(%arg0 : tensor<4x4xi1>) outs(%fill : tensor<4xi1>) {
1183 ^bb0(%in0: i1, %out0: i1):
1184 %and = arith.andi %in0, %out0 : i1
1185 linalg.yield %and : i1
1186 } -> tensor<4xi1>
1187 return %red : tensor<4xi1>
1188 }
1189
1190
1191 module attributes {transform.with_named_sequence} {
1192 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
1193 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1194 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1195 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op
1196 transform.yield
1197 }
1198 }
1199
1200 // -----
1201
1202 // CHECK-LABEL: func @red_xor_2d(
1203 func.func @red_xor_2d(%arg0: tensor<4x4xi1>) -> tensor<4xi1> {
1204 // CHECK: tensor.empty() : tensor<4xi1>
1205 // CHECK: vector.transfer_read {{.*}} : tensor<4x4xi1>, vector<4x4xi1>
1206 // CHECK: vector.multi_reduction <xor>, {{.*}}, {{.*}} [1] : vector<4x4xi1> to vector<4xi1>
1207 // CHECK: vector.transfer_write {{.*}} : vector<4xi1>, tensor<4xi1>
1208 %ident = arith.constant false
1209 %init = tensor.empty() : tensor<4xi1>
1210 %fill = linalg.fill ins(%ident : i1) outs(%init : tensor<4xi1>) -> tensor<4xi1>
1211 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,
1212 affine_map<(d0, d1) -> (d0)>],
1213 iterator_types = ["parallel", "reduction"]}
1214 ins(%arg0 : tensor<4x4xi1>) outs(%fill : tensor<4xi1>) {
1215 ^bb0(%in0: i1, %out0: i1):
1216 %xor = arith.xori %in0, %out0 : i1
1217 linalg.yield %xor : i1
1218 } -> tensor<4xi1>
1219 return %red : tensor<4xi1>
1220 }
1221
1222
1223 module attributes {transform.with_named_sequence} {
1224 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
1225 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1226 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1227 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op
1228 transform.yield
1229 }
1230 }
1231
1232 // -----
1233
1234 // CHECK-DAG: #[[$M5:.*]] = affine_map<(d0, d1) -> (d0, 0)>
1235
1236 // CHECK-LABEL: func @explicit_broadcast(
1237 func.func @explicit_broadcast(%arg0: tensor<4x4xf32>, %arg1: tensor<4x1xf32>) -> tensor<4x4xf32> {
1238 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true]} : tensor<4x4xf32>, vector<4x4xf32>
1239 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true], permutation_map = #[[$M5]]} : tensor<4x1xf32>, vector<4x4xf32>
1240 // CHECK: subf {{.*}} : vector<4x4xf32>
1241 // CHECK: vector.transfer_write {{.*}} {in_bounds = [true, true]} : vector<4x4xf32>, tensor<4x4xf32>
1242 %c0 = arith.constant 0.0 : f32
1243 %init = tensor.empty() : tensor<4x4xf32>
1244 %fill = linalg.fill ins(%c0 : f32) outs(%init : tensor<4x4xf32>) -> tensor<4x4xf32>
1245 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,
1246 affine_map<(d0, d1) -> (d0, 0)>,
1247 affine_map<(d0, d1) -> (d0, d1)>],
1248 iterator_types = ["parallel", "parallel"]}
1249 ins(%arg0, %arg1 : tensor<4x4xf32>, tensor<4x1xf32>)
1250 outs(%fill : tensor<4x4xf32>) {
1251 ^bb0(%arg7: f32, %arg8: f32, %arg9: f32):
1252 %40 = arith.subf %arg7, %arg8 : f32
1253 linalg.yield %40 : f32
1254 } -> tensor<4x4xf32>
1255 return %red : tensor<4x4xf32>
1256 }
1257
1258
1259 module attributes {transform.with_named_sequence} {
1260 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
1261 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1262 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1263 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op
1264 transform.yield
1265 }
1266 }
1267
1268 // -----
1269
1270 // CHECK-DAG: #[[$M6:.*]] = affine_map<(d0, d1) -> (d0, 0)>
1271
1272 // CHECK-LABEL: func @fused_broadcast_red_2d
1273 func.func @fused_broadcast_red_2d(%arg0: tensor<4x4xf32>, %arg1: tensor<4x1xf32>) -> tensor<4xf32> {
1274 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true]} : tensor<4x4xf32>, vector<4x4xf32>
1275 // CHECK: vector.transfer_read {{.*}} {in_bounds = [true, true], permutation_map = #[[$M6]]} : tensor<4x1xf32>, vector<4x4xf32>
1276 // CHECK: subf {{.*}} : vector<4x4xf32>
1277 // CHECK: math.exp {{.*}} : vector<4x4xf32>
1278 // CHECK: vector.multi_reduction <add>, {{.*}}, {{.*}} : vector<4x4xf32> to vector<4xf32>
1279 // CHECK: vector.transfer_write {{.*}} {in_bounds = [true]} : vector<4xf32>, tensor<4xf32>
1280 %c0 = arith.constant 0.0 : f32
1281 %init = tensor.empty() : tensor<4xf32>
1282 %fill = linalg.fill ins(%c0 : f32) outs(%init : tensor<4xf32>) -> tensor<4xf32>
1283 %red = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,
1284 affine_map<(d0, d1) -> (d0, 0)>,
1285 affine_map<(d0, d1) -> (d0)>],
1286 iterator_types = ["parallel", "reduction"]}
1287 ins(%arg0, %arg1 : tensor<4x4xf32>, tensor<4x1xf32>)
1288 outs(%fill : tensor<4xf32>) {
1289 ^bb0(%arg7: f32, %arg8: f32, %arg9: f32):
1290 %40 = arith.subf %arg7, %arg8 : f32
1291 %41 = math.exp %40 : f32
1292 %42 = arith.addf %41, %arg9 : f32
1293 linalg.yield %42 : f32
1294 } -> tensor<4xf32>
1295 return %red : tensor<4xf32>
1296 }
1297
1298
1299 module attributes {transform.with_named_sequence} {
1300 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
1301 %0 = transform.structured.match ops{["linalg.fill"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1302 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1303 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
1304
1305 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1306 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1307 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op
1308 transform.yield
1309 }
1310 }
1311
1312 // -----
1313
1314 // CHECK-LABEL: func @reduce_1d(
1315 // CHECK-SAME: %[[A:.*]]: tensor<32xf32>
1316 func.func @reduce_1d(%arg0: tensor<32xf32>) -> tensor<f32> {
1317 // CHECK-DAG: %[[F0:.*]] = arith.constant 0.000000e+00 : f32
1318 // CHECK-DAG: %[[C0:.*]] = arith.constant 0 : index
1319 %f0 = arith.constant 0.000000e+00 : f32
1320
1321 // CHECK: %[[init:.*]] = tensor.empty() : tensor<f32>
1322 %0 = tensor.empty() : tensor<f32>
1323
1324 %1 = linalg.fill ins(%f0 : f32) outs(%0 : tensor<f32>) -> tensor<f32>
1325 // CHECK: %[[r:.*]] = vector.transfer_read %[[A]][%[[C0]]]
1326 // CHECK-SAME: : tensor<32xf32>, vector<32xf32>
1327 // CHECK: %[[red:.*]] = vector.multi_reduction <add>, %[[r]], %[[F0]] [0]
1328 // CHECK-SAME: : vector<32xf32> to f32
1329 // CHECK: %[[red_v1:.*]] = vector.broadcast %[[red]] : f32 to vector<f32>
1330 // CHECK: %[[res:.*]] = vector.transfer_write %[[red_v1]], %[[init]][]
1331 // CHECK-SAME: : vector<f32>, tensor<f32>
1332 %2 = linalg.generic {
1333 indexing_maps = [affine_map<(d0) -> (d0)>,
1334 affine_map<(d0) -> ()>],
1335 iterator_types = ["reduction"]}
1336 ins(%arg0 : tensor<32xf32>)
1337 outs(%1 : tensor<f32>) {
1338 ^bb0(%a: f32, %b: f32):
1339 %3 = arith.addf %a, %b : f32
1340 linalg.yield %3 : f32
1341 } -> tensor<f32>
1342
1343 return %2 : tensor<f32>
1344 }
1345
1346 module attributes {transform.with_named_sequence} {
1347 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
1348 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1349 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1350 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
1351 transform.yield
1352 }
1353 }
1354
1355
1356 // -----
1357
1358 // This test checks that vectorization does not occur when an input indexing map
1359 // is not a projected permutation. In the future, this can be converted to a
1360 // positive test when support is added.
1361
1362 // CHECK-LABEL: func @not_projected_permutation
1363 func.func @not_projected_permutation(%arg0: tensor<8x8xf32>) -> tensor<6x6x3x3xf32> {
1364 %c0 = arith.constant 0.0 : f32
1365 %init = tensor.empty() : tensor<6x6x3x3xf32>
1366 %fill = linalg.fill ins(%c0 : f32) outs(%init : tensor<6x6x3x3xf32>) -> tensor<6x6x3x3xf32>
1367 // CHECK: linalg.generic
1368 %result = linalg.generic {indexing_maps = [affine_map<(d0, d1, d2, d3) -> (d0 + d2, d1 + d3)>,
1369 affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>],
1370 iterator_types = ["parallel", "parallel", "parallel", "parallel"]}
1371 ins(%arg0 : tensor<8x8xf32>)
1372 outs(%fill : tensor<6x6x3x3xf32>) {
1373 ^bb0(%arg7: f32, %arg9: f32):
1374 linalg.yield %arg7 : f32
1375 } -> tensor<6x6x3x3xf32>
1376 return %result : tensor<6x6x3x3xf32>
1377 }
1378
1379 module attributes {transform.with_named_sequence} {
1380 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
1381 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1382 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1383 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
1384 transform.yield
1385 }
1386 }
1387
1388 // -----
1389
1390 // Check vectorization can handle cases where outputs are a mix of reduced and non-reduced values.
1391 func.func @mixed_parallel_reduced_results(%arg0 : tensor<2x4x8xf32>,
1392 %arg1 : tensor<2x4xf32>, %arg2 : tensor<2x4x8xf32>, %arg3 : tensor<2x4xf32>) ->
1393 (tensor<2x4x8xf32>, tensor<2x4xf32>) {
1394 %0:2 = linalg.generic {
1395 indexing_maps = [affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>,
1396 affine_map<(d0, d1, d2) -> (d0, d1, d2)>, affine_map<(d0, d1, d2) -> (d0, d1)>],
1397 iterator_types = ["parallel", "parallel", "reduction"]}
1398 ins(%arg0, %arg1 : tensor<2x4x8xf32>, tensor<2x4xf32>)
1399 outs(%arg2, %arg3 : tensor<2x4x8xf32>, tensor<2x4xf32>) {
1400 ^bb0(%b0 : f32, %b1 : f32, %b2 : f32, %b3 : f32):
1401 %1 = arith.mulf %b0, %b1 : f32
1402 %2 = arith.addf %1, %b3 : f32
1403 linalg.yield %1, %2 : f32, f32
1404 } -> (tensor<2x4x8xf32>, tensor<2x4xf32>)
1405 return %0#0, %0#1 : tensor<2x4x8xf32>, tensor<2x4xf32>
1406 }
1407 // CHECK-LABEL: func @mixed_parallel_reduced_results(
1408 // CHECK-SAME: %[[ARG0:[a-zA-Z0-9]+]]: tensor<2x4x8xf32>
1409 // CHECK-SAME: %[[ARG1:[a-zA-Z0-9]+]]: tensor<2x4xf32>
1410 // CHECK-SAME: %[[ARG2:[a-zA-Z0-9]+]]: tensor<2x4x8xf32>
1411 // CHECK-SAME: %[[ARG3:[a-zA-Z0-9]+]]: tensor<2x4xf32>
1412 // CHECK-DAG: %[[V0:.+]] = vector.transfer_read %[[ARG0]]
1413 // CHECK-DAG: %[[V1:.+]] = vector.transfer_read %[[ARG1]]
1414 // CHECK-DAG: %[[V2:.+]] = vector.transfer_read %[[ARG3]]
1415 // CHECK-DAG: %[[MUL:.+]] = arith.mulf %[[V0]], %[[V1]]
1416 // CHECK-DAG: %[[ADD:.+]] = vector.multi_reduction <add>, %[[MUL]], %[[V2]]
1417 // CHECK-DAG: vector.transfer_write %[[MUL]], %[[ARG2]]
1418 // CHECK-DAG: vector.transfer_write %[[ADD]], %[[ARG3]]
1419
1420 module attributes {transform.with_named_sequence} {
1421 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
1422 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1423 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1424 %2 = transform.structured.vectorize_children_and_apply_patterns %1 { disable_multi_reduction_to_contract_patterns, disable_transfer_permutation_map_lowering_patterns } : (!transform.any_op) -> !transform.any_op
1425 transform.yield
1426 }
1427 }
1428
1429 // -----
1430
1431 func.func @vectorize_map(%arg0: memref<64xf32>,
1432 %arg1: memref<64xf32>, %arg2: memref<64xf32>) {
1433 linalg.map ins(%arg0, %arg1 : memref<64xf32>, memref<64xf32>)
1434 outs(%arg2 : memref<64xf32>)
1435 (%in: f32, %in_0: f32) {
1436 %0 = arith.addf %in, %in_0 : f32
1437 linalg.yield %0 : f32
1438 }
1439 return
1440 }
1441 // CHECK-LABEL: func @vectorize_map
1442 // CHECK: %[[LHS:.*]] = vector.transfer_read
1443 // CHECK-NEXT: %[[RHS:.*]] = vector.transfer_read
1444 // CHECK-NEXT: arith.addf %[[LHS]], %[[RHS]] : vector<64xf32>
1445
1446 module attributes {transform.with_named_sequence} {
1447 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
1448 %0 = transform.structured.match ops{["linalg.map"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1449 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1450 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
1451 transform.yield
1452 }
1453 }
1454
1455 // -----
1456
1457 func.func @vectorize_transpose(%arg0: memref<16x32x64xf32>,
1458 %arg1: memref<32x64x16xf32>) {
1459 linalg.transpose ins(%arg0 : memref<16x32x64xf32>)
1460 outs(%arg1 : memref<32x64x16xf32>) permutation = [1, 2, 0]
1461 return
1462 }
1463 // CHECK-LABEL: func @vectorize_transpose
1464 // CHECK: vector.transpose
1465 // CHECK-SAME: [1, 2, 0] : vector<16x32x64xf32> to vector<32x64x16xf32>
1466
1467 module attributes {transform.with_named_sequence} {
1468 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
1469 %0 = transform.structured.match ops{["linalg.transpose"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1470 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1471 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
1472 transform.yield
1473 }
1474 }
1475
1476 // -----
1477
1478 func.func @vectorize_reduce(%arg0: memref<16x32x64xf32>,
1479 %arg1: memref<16x64xf32>) {
1480 linalg.reduce ins(%arg0 : memref<16x32x64xf32>)
1481 outs(%arg1 : memref<16x64xf32>) dimensions = [1]
1482 (%in: f32, %init: f32) {
1483 %0 = arith.addf %in, %init : f32
1484 linalg.yield %0 : f32
1485 }
1486 return
1487 }
1488 // CHECK-LABEL: func @vectorize_reduce
1489 // CHECK: vector.multi_reduction <add>
1490 // CHECK-SAME: : vector<16x32x64xf32> to vector<16x64xf32>
1491
1492 module attributes {transform.with_named_sequence} {
1493 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
1494 %0 = transform.structured.match ops{["linalg.reduce"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1495 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1496 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
1497 transform.yield
1498 }
1499 }
1500
1501 // -----
1502
1503 // This is a regression test. This IR cannot be vectorized, but
1504 // structured.vectorize_children_and_apply_patterns should nevertheless succeed.
1505
1506 #map = affine_map<(d0) -> (d0)>
1507 // CHECK-LABEL: @not_vectorizable
1508 func.func @not_vectorizable(%arg0: tensor<1x?xf32>, %arg1: index, %arg2: index, %arg3: index) -> tensor<1x128xf32> {
1509 %c0 = arith.constant 0 : index
1510 %0 = tensor.empty() : tensor<1x128xf32>
1511 %1 = scf.for %arg5 = %arg2 to %arg1 step %arg3 iter_args(%arg6 = %0) -> (tensor<1x128xf32>) {
1512 %extracted_slice = tensor.extract_slice %arg6[0, 0] [1, %arg1] [1, 1] : tensor<1x128xf32> to tensor<?xf32>
1513 %sz0 = tensor.dim %extracted_slice, %c0 : tensor<?xf32>
1514 %expanded = tensor.expand_shape %extracted_slice [[0, 1]] output_shape [1, %sz0] : tensor<?xf32> into tensor<1x?xf32>
1515 %extracted_slice_0 = tensor.extract_slice %arg0[0, %arg3] [1, %arg2] [1, 1] : tensor<1x?xf32> to tensor<?xf32>
1516 %extracted_slice_1 = tensor.extract_slice %expanded[0, %arg3] [1, %arg2] [1, 1] : tensor<1x?xf32> to tensor<?xf32>
1517 %2 = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel"]} ins(%extracted_slice_0 : tensor<?xf32>) outs(%extracted_slice_1 : tensor<?xf32>) {
1518 ^bb0(%in: f32, %out: f32):
1519 %3 = arith.addf %in, %out : f32
1520 linalg.yield %3 : f32
1521 } -> tensor<?xf32>
1522 %inserted_slice = tensor.insert_slice %2 into %expanded[0, %arg3] [1, %arg2] [1, 1] : tensor<?xf32> into tensor<1x?xf32>
1523 %collapsed = tensor.collapse_shape %inserted_slice [[0, 1]] : tensor<1x?xf32> into tensor<?xf32>
1524 %inserted_slice_2 = tensor.insert_slice %collapsed into %arg6[0, 0] [1, %arg1] [1, 1] : tensor<?xf32> into tensor<1x128xf32>
1525 scf.yield %inserted_slice_2 : tensor<1x128xf32>
1526 }
1527 return %1 : tensor<1x128xf32>
1528 }
1529 module attributes {transform.with_named_sequence} {
1530 transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
1531 %0 = transform.structured.match ops{["func.func"]} in %arg0 : (!transform.any_op) -> !transform.any_op
1532 %1 = transform.structured.vectorize_children_and_apply_patterns %0 : (!transform.any_op) -> !transform.any_op
1533 transform.yield
1534 }
1535 }
1536
1537 // -----
1538
1539 // Regression test: %13 was incorrectly detected as a reduction and
1540 // vectorization failed.
1541
1542 func.func @wrong_reduction_detection(%input: tensor<120x64xf32>) -> tensor<120x64xf32> {
1543 %c0 = arith.constant 0 : index
1544 %c4 = arith.constant 4 : index
1545 %c64 = arith.constant 64 : index
1546 %cst_6 = arith.constant 4.000000e+00 : f32
1547 %1 = scf.for %arg0 = %c0 to %c64 step %c4 iter_args(%arg1 = %input) -> (tensor<120x64xf32>) {
1548 %extracted_slice = tensor.extract_slice %arg1[%c0, %arg0] [1, 4] [1, 1] : tensor<120x64xf32> to tensor<1x4xf32>
1549 %10 = linalg.fill {__internal_linalg_transform__ = "1"} ins(%cst_6 : f32) outs(%extracted_slice : tensor<1x4xf32>) -> tensor<1x4xf32>
1550 %11 = linalg.generic {indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>], iterator_types = ["parallel", "parallel"]} outs(%10 : tensor<1x4xf32>) {
1551 ^bb0(%out: f32):
1552 %12 = linalg.index 0 : index
1553 %13 = arith.addi %arg0, %12 : index
1554 %18 = arith.index_cast %13 : index to i32
1555 %20 = arith.uitofp %18 : i32 to f32
1556 %67 = arith.mulf %out, %20 : f32
1557 linalg.yield %67 : f32
1558 } -> tensor<1x4xf32>
1559 %inserted_slice = tensor.insert_slice %11 into %arg1[%c0, %arg0] [1, 4] [1, 1] : tensor<1x4xf32> into tensor<120x64xf32>
1560 scf.yield %inserted_slice : tensor<120x64xf32>
1561 }
1562 return %1 : tensor<120x64xf32>
1563 }
1564
1565 module attributes {transform.with_named_sequence} {
1566 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
1567 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1568 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1569 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
1570 transform.yield
1571 }
1572 }
1573
1574 // CHECK-LABEL: @wrong_reduction_detection
1575 // CHECK: vector.broadcast
1576 // CHECK: vector.transfer_write
1577
1578 // -----
1579
1580 // Don't vectorize tensor<0xf32> : (!transform.any_op) -> !transform.any_op
1581 // CHECK-LABEL: @tensor_size0
1582 // CHECK: linalg.generic
1583 func.func @tensor_size0(%arg0: tensor<0xf32>,
1584 %arg1: tensor<f32>) -> tensor<f32> {
1585 %0 = linalg.generic
1586 {indexing_maps = [affine_map<(d0) -> (d0)>, affine_map<(d0) -> ()>],
1587 iterator_types = ["reduction"]}
1588 ins(%arg0 : tensor<0xf32>) outs(%arg1 : tensor<f32>) {
1589 ^bb0(%in: f32, %out: f32):
1590 %12 = arith.addf %out, %in : f32
1591 linalg.yield %12 : f32
1592 } -> tensor<f32>
1593 return %0 : tensor<f32>
1594 }
1595
1596 module attributes {transform.with_named_sequence} {
1597 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
1598 %0 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1599 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1600 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
1601 transform.yield
1602 }
1603 }
1604
1605 // -----
1606
1607 func.func @zero_dim_tensor(%input: tensor<f32>, %output: tensor<f32>) -> tensor<f32>
1608 {
1609 %0 = linalg.generic { indexing_maps = [ affine_map<() -> ()>, affine_map<() -> ()> ],
1610 iterator_types = [] }
1611 ins(%input : tensor<f32>)
1612 outs(%output : tensor<f32>) {
1613 ^bb0(%arg0: f32, %arg1: f32):
1614 %2 = arith.addf %arg0, %arg1 : f32
1615 linalg.yield %2 : f32
1616 } -> tensor<f32>
1617 return %0 : tensor<f32>
1618 }
1619
1620 module attributes {transform.with_named_sequence} {
1621 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
1622 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1623 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1624 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op
1625 transform.yield
1626 }
1627 }
1628
1629 // CHECK-LABEL: func @zero_dim_tensor
1630 // CHECK: vector.transfer_read {{.*}} : tensor<f32>, vector<f32>
1631 // CHECK: vector.extract
1632 // CHECK: vector.transfer_read {{.*}} : tensor<f32>, vector<f32>
1633 // CHECK: vector.extract
1634 // CHECK: arith.addf {{.*}} : f32
1635 // CHECK: vector.broadcast %{{.*}} : f32 to vector<f32>
1636 // CHECK: vector.transfer_write {{.*}} : vector<f32>, tensor<f32>
1637
1638 // -----
1639
1640 // Make sure we generate the right transfer writes for multi-output generic ops
1641 // with different permutation maps.
1642
1643 func.func @multi_output_generic_different_perm_maps(%in0: tensor<4x1xf32>,
1644 %out0: tensor<4x1xf32>,
1645 %out1: tensor<1x4xf32>) -> (tensor<4x1xf32>, tensor<1x4xf32>) {
1646 %13:2 = linalg.generic {indexing_maps = [ affine_map<(d0, d1) -> (d1, d0)>,
1647 affine_map<(d0, d1) -> (d1, d0)>,
1648 affine_map<(d0, d1) -> (d0, d1)> ],
1649 iterator_types = ["parallel", "parallel"]}
1650 ins(%in0 : tensor<4x1xf32>)
1651 outs(%out0, %out1 : tensor<4x1xf32>, tensor<1x4xf32>) {
1652 ^bb0(%in: f32, %out: f32, %out_2: f32):
1653 %16 = arith.addf %in, %in : f32
1654 linalg.yield %16, %16 : f32, f32
1655 } -> (tensor<4x1xf32>, tensor<1x4xf32>)
1656 return %13#0, %13#1 : tensor<4x1xf32>, tensor<1x4xf32>
1657 }
1658
1659 module attributes {transform.with_named_sequence} {
1660 transform.named_sequence @__transform_main(%arg1: !transform.any_op {transform.readonly}) {
1661 %3 = transform.structured.match ops{["linalg.generic"]} in %arg1 : (!transform.any_op) -> !transform.any_op
1662 %4 = transform.get_parent_op %3 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1663 %5 = transform.structured.vectorize_children_and_apply_patterns %4 : (!transform.any_op) -> !transform.any_op
1664 transform.yield
1665 }
1666 }
1667
1668 // CHECK-LABEL: func @multi_output_generic_different_perm_maps
1669 // CHECK: %[[VAL_5:.*]] = vector.transfer_read %{{.*}} {in_bounds = [true, true]} : tensor<4x1xf32>, vector<4x1xf32>
1670 // CHECK: %[[VAL_6:.*]] = arith.addf %[[VAL_5]], %[[VAL_5]] : vector<4x1xf32>
1671 // CHECK: %[[VAL_7:.*]] = vector.transpose %[[VAL_6]], [1, 0] : vector<4x1xf32> to vector<1x4xf32>
1672 // CHECK: %[[VAL_8:.*]] = vector.transpose %[[VAL_7]], [1, 0] : vector<1x4xf32> to vector<4x1xf32>
1673 // CHECK: vector.transfer_write %[[VAL_8]], %{{.*}} {in_bounds = [true, true]} : vector<4x1xf32>, tensor<4x1xf32>
1674 // CHECK: vector.transfer_write %[[VAL_7]], %{{.*}} {in_bounds = [true, true]} : vector<1x4xf32>, tensor<1x4xf32>
1675
1676 // -----
1677
1678 // Extracted from: https://github.com/llvm/llvm-project/issues/97247
1679
1680 #map = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
1681 #map1 = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, 0)>
1682
1683 func.func @generic_with_reduction_and_broadcast(%arg0: tensor<1x12x197x197xf32>) -> (tensor<1x12x197x1xf32>) {
1684 %0 = tensor.empty() : tensor<1x12x197x1xf32>
1685 %1 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "reduction"]} ins(%arg0 : tensor<1x12x197x197xf32>) outs(%0 : tensor<1x12x197x1xf32>) {
1686 ^bb0(%in: f32, %out: f32):
1687 %818 = arith.addf %in, %out : f32
1688 linalg.yield %818 : f32
1689 } -> tensor<1x12x197x1xf32>
1690 return %1 : tensor<1x12x197x1xf32>
1691 }
1692 module attributes {transform.with_named_sequence} {
1693 transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
1694 %0 = transform.structured.match ops{["linalg.generic"]} in %arg0 : (!transform.any_op) -> !transform.any_op
1695 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1696 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
1697 transform.yield
1698 }
1699 }
1700
1701 // CHECK: #[[$ATTR_32:.+]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2)>
1702
1703 // CHECK-LABEL: func.func @generic_with_reduction_and_broadcast(
1704 // CHECK-SAME: %[[VAL_0:.*]]: tensor<1x12x197x197xf32>) -> tensor<1x12x197x1xf32> {
1705 // CHECK: %[[VAL_1:.*]] = ub.poison : f32
1706 // CHECK: %[[VAL_2:.*]] = arith.constant 0 : index
1707 // CHECK: %[[VAL_3:.*]] = tensor.empty() : tensor<1x12x197x1xf32>
1708 // CHECK: %[[VAL_4:.*]] = vector.transfer_read %[[VAL_0]]{{\[}}%[[VAL_2]], %[[VAL_2]], %[[VAL_2]], %[[VAL_2]]], %[[VAL_1]] {in_bounds = [true, true, true, true]} : tensor<1x12x197x197xf32>, vector<1x12x197x197xf32>
1709 // CHECK: %[[VAL_5:.*]] = vector.transfer_read %[[VAL_3]]{{\[}}%[[VAL_2]], %[[VAL_2]], %[[VAL_2]], %[[VAL_2]]], %[[VAL_1]] {in_bounds = [true, true, true], permutation_map = #[[$ATTR_32]]} : tensor<1x12x197x1xf32>, vector<1x12x197xf32>
1710 // CHECK: %[[VAL_6:.*]] = vector.multi_reduction <add>, %[[VAL_4]], %[[VAL_5]] [3] : vector<1x12x197x197xf32> to vector<1x12x197xf32>
1711 // CHECK: %[[VAL_7:.*]] = vector.broadcast %[[VAL_6]] : vector<1x12x197xf32> to vector<1x1x12x197xf32>
1712 // CHECK: %[[VAL_8:.*]] = vector.transpose %[[VAL_7]], [1, 2, 3, 0] : vector<1x1x12x197xf32> to vector<1x12x197x1xf32>
1713 // CHECK: %[[VAL_9:.*]] = vector.transfer_write %[[VAL_8]], %[[VAL_3]]{{\[}}%[[VAL_2]], %[[VAL_2]], %[[VAL_2]], %[[VAL_2]]] {in_bounds = [true, true, true, true]} : vector<1x12x197x1xf32>, tensor<1x12x197x1xf32>
1714 // CHECK: return %[[VAL_9]] : tensor<1x12x197x1xf32>
1715
1716 // -----
1717
1718 // Input identical as the test in vectorization.mlir. Output is different -
1719 // vector sizes are inferred (rather than user-specified) and hence _no_
1720 // masking was used.
1721
1722 func.func @test_vectorize_pack(%arg0: tensor<32x8x16xf32>, %arg1: tensor<4x1x32x16x2xf32>) -> tensor<4x1x32x16x2xf32> {
1723 %pack = linalg.pack %arg0 outer_dims_perm = [1, 2, 0] inner_dims_pos = [2, 1] inner_tiles = [16, 2] into %arg1 : tensor<32x8x16xf32> -> tensor<4x1x32x16x2xf32>
1724 return %pack : tensor<4x1x32x16x2xf32>
1725 }
1726
1727 module attributes {transform.with_named_sequence} {
1728 transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
1729 %0 = transform.structured.match ops{["linalg.pack"]} in %arg0 : (!transform.any_op) -> !transform.any_op
1730 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1731 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
1732 transform.yield
1733 }
1734 }
1735
1736 // CHECK-LABEL: func.func @test_vectorize_pack(
1737 // CHECK-SAME: %[[VAL_0:.*]]: tensor<32x8x16xf32>,
1738 // CHECK-SAME: %[[VAL_1:.*]]: tensor<4x1x32x16x2xf32>) -> tensor<4x1x32x16x2xf32> {
1739 // CHECK: %[[VAL_2:.*]] = arith.constant 0.000000e+00 : f32
1740 // CHECK: %[[VAL_3:.*]] = arith.constant 0 : index
1741 // CHECK: %[[VAL_4:.*]] = vector.transfer_read %[[VAL_0]]{{\[}}%[[VAL_3]], %[[VAL_3]], %[[VAL_3]]], %[[VAL_2]] {in_bounds = [true, true, true]} : tensor<32x8x16xf32>, vector<32x8x16xf32>
1742 // CHECK: %[[VAL_5:.*]] = vector.shape_cast %[[VAL_4]] : vector<32x8x16xf32> to vector<32x4x2x1x16xf32>
1743 // CHECK: %[[VAL_6:.*]] = vector.transpose %[[VAL_5]], [1, 3, 0, 4, 2] : vector<32x4x2x1x16xf32> to vector<4x1x32x16x2xf32>
1744 // CHECK: %[[VAL_7:.*]] = tensor.empty() : tensor<4x1x32x16x2xf32>
1745 // CHECK: %[[VAL_8:.*]] = vector.transfer_write %[[VAL_6]], %[[VAL_7]]{{\[}}%[[VAL_3]], %[[VAL_3]], %[[VAL_3]], %[[VAL_3]], %[[VAL_3]]] {in_bounds = [true, true, true, true, true]} : vector<4x1x32x16x2xf32>, tensor<4x1x32x16x2xf32>
1746 // CHECK: return %[[VAL_8]] : tensor<4x1x32x16x2xf32>
1747
1748 // -----
1749
1750 // Input identical as the test in vectorization.mlir. Output is different -
1751 // vector sizes are inferred (rather than user-specified) and hence _no_
1752 // masking was used.
1753
1754 func.func @test_vectorize_padded_pack(%arg0: tensor<32x7x15xf32>, %arg1: tensor<32x4x1x16x2xf32>) -> tensor<32x4x1x16x2xf32> {
1755 %pad = arith.constant 0.000000e+00 : f32
1756 %pack = linalg.pack %arg0 padding_value(%pad : f32) inner_dims_pos = [2, 1] inner_tiles = [16, 2] into %arg1 : tensor<32x7x15xf32> -> tensor<32x4x1x16x2xf32>
1757 return %pack : tensor<32x4x1x16x2xf32>
1758 }
1759
1760 // CHECK-LABEL: func.func @test_vectorize_padded_pack(
1761 // CHECK-SAME: %[[VAL_0:.*]]: tensor<32x7x15xf32>,
1762 // CHECK-SAME: %[[VAL_1:.*]]: tensor<32x4x1x16x2xf32>) -> tensor<32x4x1x16x2xf32> {
1763 // CHECK: %[[VAL_2:.*]] = arith.constant 0.000000e+00 : f32
1764 // CHECK: %[[VAL_3:.*]] = arith.constant 0 : index
1765 // CHECK: %[[VAL_4:.*]] = vector.transfer_read %[[VAL_0]]{{\[}}%[[VAL_3]], %[[VAL_3]], %[[VAL_3]]], %[[VAL_2]] {in_bounds = [true, false, false]} : tensor<32x7x15xf32>, vector<32x8x16xf32>
1766 // CHECK: %[[VAL_5:.*]] = vector.shape_cast %[[VAL_4]] : vector<32x8x16xf32> to vector<32x4x2x1x16xf32>
1767 // CHECK: %[[VAL_6:.*]] = vector.transpose %[[VAL_5]], [0, 1, 3, 4, 2] : vector<32x4x2x1x16xf32> to vector<32x4x1x16x2xf32>
1768 // CHECK: %[[VAL_7:.*]] = tensor.empty() : tensor<32x4x1x16x2xf32>
1769 // CHECK: %[[VAL_8:.*]] = vector.transfer_write %[[VAL_6]], %[[VAL_7]]{{\[}}%[[VAL_3]], %[[VAL_3]], %[[VAL_3]], %[[VAL_3]], %[[VAL_3]]] {in_bounds = [true, true, true, true, true]} : vector<32x4x1x16x2xf32>, tensor<32x4x1x16x2xf32>
1770 // CHECK: return %[[VAL_8]] : tensor<32x4x1x16x2xf32>
1771
1772 module attributes {transform.with_named_sequence} {
1773 transform.named_sequence @__transform_main(%arg0: !transform.any_op {transform.readonly}) {
1774 %0 = transform.structured.match ops{["linalg.pack"]} in %arg0 : (!transform.any_op) -> !transform.any_op
1775 %1 = transform.get_parent_op %0 {isolated_from_above} : (!transform.any_op) -> !transform.any_op
1776 %2 = transform.structured.vectorize_children_and_apply_patterns %1 : (!transform.any_op) -> !transform.any_op
1777 transform.yield
1778 }
1779 }