1 // RUN: mlir-opt --transform-interpreter --cse -split-input-file %s | FileCheck %s
3 func.func @simple_matmul(%arg0 : tensor<?x?xf32>, %arg1 : tensor<?x?xf32>,
4 %arg2 : tensor<?x?xf32>) -> tensor<?x?xf32> {
5 %0 = linalg.matmul ins(%arg0, %arg1 : tensor<?x?xf32>, tensor<?x?xf32>)
6 outs(%arg2 : tensor<?x?xf32>) -> tensor<?x?xf32>
7 return %0 : tensor<?x?xf32>
10 module attributes {transform.with_named_sequence} {
11 transform.named_sequence @__transform_main(%arg1 : !transform.any_op {transform.readonly}) {
12 %matmul = transform.structured.match ops{["linalg.matmul"]} in %arg1
13 : (!transform.any_op) -> !transform.any_op
14 %a, %b, %c = transform.structured.tile_using_for %matmul tile_sizes [10, 20]
15 : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op)
19 // CHECK-DAG: #[[$MAP0:.+]] = affine_map<(d0)[s0] -> (-d0 + s0, 10)>
20 // CHECK-DAG: #[[$MAP1:.+]] = affine_map<(d0)[s0] -> (-d0 + s0, 20)>
21 // CHECK-LABEL: func.func @simple_matmul(
22 // CHECK-SAME: %[[ARG0:[a-zA-Z0-9]+]]: tensor<?x?xf32>
23 // CHECK-SAME: %[[ARG1:[a-zA-Z0-9]+]]: tensor<?x?xf32>
24 // CHECK-SAME: %[[ARG2:[a-zA-Z0-9]+]]: tensor<?x?xf32>
25 // CHECK-DAG: %[[C0:.+]] = arith.constant 0 : index
26 // CHECK-DAG: %[[C1:.+]] = arith.constant 1 : index
27 // CHECK-DAG: %[[M:.+]] = tensor.dim %[[ARG0]], %[[C0]]
28 // CHECK-DAG: %[[K:.+]] = tensor.dim %[[ARG0]], %[[C1]]
29 // CHECK-DAG: %[[N:.+]] = tensor.dim %[[ARG1]], %[[C1]]
30 // CHECK-DAG: %[[C10:.+]] = arith.constant 10 : index
31 // CHECK-DAG: %[[C20:.+]] = arith.constant 20 : index
32 // CHECK: %[[OUTER:[a-zA-Z0-9]+]] = scf.for %[[IV0:[a-zA-Z0-9]+]] = %[[C0]] to %[[M]] step %[[C10]]
33 // CHECK-SAME: iter_args(%[[INIT0:.+]] = %[[ARG2]])
34 // CHECK: %[[INNER:[a-zA-Z0-9]+]] = scf.for %[[IV1:[a-zA-Z0-9]+]] = %[[C0]] to %[[N]] step %[[C20]]
35 // CHECK-SAME: iter_args(%[[INIT1:.+]] = %[[INIT0]])
36 // CHECK-DAG: %[[TS_Y:.+]] = affine.min #[[$MAP0]](%[[IV0]])[%[[M]]]
37 // CHECK: %[[TS_X:.+]] = affine.min #[[$MAP1]](%[[IV1]])[%[[N]]]
38 // CHECK-DAG: %[[LHS_TILE:.+]] = tensor.extract_slice %[[ARG0]]
39 // CHECK-SAME: [%[[IV0]], 0] [%[[TS_Y]], %[[K]]] [1, 1]
40 // CHECK-DAG: %[[RHS_TILE:.+]] = tensor.extract_slice %[[ARG1]]
41 // CHECK-SAME: [0, %[[IV1]]] [%[[K]], %[[TS_X]]] [1, 1]
42 // CHECK-DAG: %[[INIT_TILE:.+]] = tensor.extract_slice %[[INIT1]]
43 // CHECK-SAME: [%[[IV0]], %[[IV1]]] [%[[TS_Y]], %[[TS_X]]] [1, 1]
44 // CHECK: %[[GEMM_TILE:.+]] = linalg.matmul
45 // CHECK-SAME: ins(%[[LHS_TILE]], %[[RHS_TILE]] :
46 // CHECK-SAME: outs(%[[INIT_TILE]] :
47 // CHECK: %[[UPDATE:.+]] = tensor.insert_slice %[[GEMM_TILE]] into %[[INIT1]]
48 // CHECK-SAME: [%[[IV0]], %[[IV1]]] [%[[TS_Y]], %[[TS_X]]] [1, 1]
49 // CHECK: scf.yield %[[UPDATE]]
50 // CHECK: scf.yield %[[INNER]]
51 // CHECK: return %[[OUTER]]
55 func.func @simple_matmul_memref(%arg0 : memref<?x?xf32>, %arg1 : memref<?x?xf32>,
56 %arg2 : memref<?x?xf32>) {
57 linalg.matmul ins(%arg0, %arg1 : memref<?x?xf32>, memref<?x?xf32>)
58 outs(%arg2 : memref<?x?xf32>)
62 module attributes {transform.with_named_sequence} {
63 transform.named_sequence @__transform_main(%arg1 : !transform.any_op {transform.readonly}) {
64 %matmul = transform.structured.match ops{["linalg.matmul"]} in %arg1
65 : (!transform.any_op) -> !transform.any_op
66 %a, %b, %c, %d = transform.structured.tile_using_for %matmul tile_sizes [10, 20, 30]
67 : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op, !transform.any_op)
71 // CHECK-DAG: #[[$MAP0:.+]] = affine_map<(d0)[s0] -> (-d0 + s0, 10)>
72 // CHECK-DAG: #[[$MAP1:.+]] = affine_map<(d0)[s0] -> (-d0 + s0, 20)>
73 // CHECK-DAG: #[[$MAP2:.+]] = affine_map<(d0)[s0] -> (-d0 + s0, 30)>
74 // CHECK-LABEL: func.func @simple_matmul_memref(
75 // CHECK-SAME: %[[ARG0:[a-zA-Z0-9]+]]: memref<?x?xf32>
76 // CHECK-SAME: %[[ARG1:[a-zA-Z0-9]+]]: memref<?x?xf32>
77 // CHECK-SAME: %[[ARG2:[a-zA-Z0-9]+]]: memref<?x?xf32>
78 // CHECK-DAG: %[[C0:.+]] = arith.constant 0 : index
79 // CHECK-DAG: %[[C1:.+]] = arith.constant 1 : index
80 // CHECK-DAG: %[[M:.+]] = memref.dim %[[ARG0]], %[[C0]]
81 // CHECK-DAG: %[[K:.+]] = memref.dim %[[ARG0]], %[[C1]]
82 // CHECK-DAG: %[[N:.+]] = memref.dim %[[ARG1]], %[[C1]]
83 // CHECK-DAG: %[[C10:.+]] = arith.constant 10 : index
84 // CHECK-DAG: %[[C20:.+]] = arith.constant 20 : index
85 // CHECK-DAG: %[[C30:.+]] = arith.constant 30 : index
86 // CHECK: scf.for %[[IV0:[a-zA-Z0-9]+]] = %[[C0]] to %[[M]] step %[[C10]]
87 // CHECK: scf.for %[[IV1:[a-zA-Z0-9]+]] = %[[C0]] to %[[N]] step %[[C20]]
88 // CHECK: scf.for %[[IV2:[a-zA-Z0-9]+]] = %[[C0]] to %[[K]] step %[[C30]]
89 // CHECK-DAG: %[[TS_M:.+]] = affine.min #[[$MAP0]](%[[IV0]])[%[[M]]]
90 // CHECK-DAG: %[[TS_N:.+]] = affine.min #[[$MAP1]](%[[IV1]])[%[[N]]]
91 // CHECK-DAG: %[[TS_K:.+]] = affine.min #[[$MAP2]](%[[IV2]])[%[[K]]]
92 // CHECK-DAG: %[[LHS_TILE:.+]] = memref.subview %[[ARG0]]
93 // CHECK-SAME: [%[[IV0]], %[[IV2]]] [%[[TS_M]], %[[TS_K]]] [1, 1]
94 // CHECK-DAG: %[[RHS_TILE:.+]] = memref.subview %[[ARG1]]
95 // CHECK-SAME: [%[[IV2]], %[[IV1]]] [%[[TS_K]], %[[TS_N]]] [1, 1]
96 // CHECK-DAG: %[[OUT_TILE:.+]] = memref.subview %[[ARG2]]
97 // CHECK-SAME: [%[[IV0]], %[[IV1]]] [%[[TS_M]], %[[TS_N]]] [1, 1]
98 // CHECK: linalg.matmul
99 // CHECK-SAME: ins(%[[LHS_TILE]], %[[RHS_TILE]] :
100 // CHECK-SAME: outs(%[[OUT_TILE]] :
104 #map0 = affine_map<(d0, d1, d2) -> (d0, d1, d2)>
105 #map1 = affine_map<(d0, d1, d2) -> (d0, d2, d1)>
106 #map2 = affine_map<(d0, d1, d2) -> (d2, d0, d1)>
107 func.func @multi_result(%arg0 : tensor<128x200x300xf32>) -> (tensor<128x300x200xf32>, tensor<300x128x200xf32>) {
108 %init0 = tensor.empty() : tensor<128x300x200xf32>
109 %init1 = tensor.empty() : tensor<300x128x200xf32>
110 %0:2 = linalg.generic {
111 indexing_maps = [#map0, #map1, #map2],
112 iterator_types = ["parallel", "parallel", "parallel"]}
113 ins(%arg0 : tensor<128x200x300xf32>)
114 outs(%init0, %init1 : tensor<128x300x200xf32>, tensor<300x128x200xf32>) {
115 ^bb0(%b0 : f32, %b1 : f32, %b2 : f32):
116 linalg.yield %b0, %b0 : f32, f32
117 } -> (tensor<128x300x200xf32>, tensor<300x128x200xf32>)
118 return %0#0, %0#1 : tensor<128x300x200xf32>, tensor<300x128x200xf32>
121 module attributes {transform.with_named_sequence} {
122 transform.named_sequence @__transform_main(%arg1 : !transform.any_op {transform.readonly}) {
123 %generic = transform.structured.match ops{["linalg.generic"]} in %arg1
124 : (!transform.any_op) -> !transform.any_op
125 %a, %b, %c = transform.structured.tile_using_for %generic tile_sizes [10, 0, 20]
126 : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op)
130 // CHECK-DAG: #[[$MAP0:.+]] = affine_map<(d0) -> (-d0 + 128, 10)>
131 // CHECK-LABEL: func.func @multi_result(
132 // CHECK-SAME: %[[ARG0:[a-zA-Z0-9]+]]: tensor<128x200x300xf32>)
133 // CHECK-DAG: %[[INIT0:.+]] = tensor.empty()
134 // CHECK-DAG: %[[INIT1:.+]] = tensor.empty()
135 // CHECK-DAG: %[[C0:.+]] = arith.constant 0 : index
136 // CHECK-DAG: %[[C128:.+]] = arith.constant 128 : index
137 // CHECK-DAG: %[[C300:.+]] = arith.constant 300 : index
138 // CHECK-DAG: %[[C10:.+]] = arith.constant 10 : index
139 // CHECK-DAG: %[[C20:.+]] = arith.constant 20 : index
140 // CHECK: %[[OUTER:[a-zA-Z0-9]+]]:2 = scf.for %[[IV0:[a-zA-Z0-9]+]] = %[[C0]] to %[[C128]] step %[[C10]]
141 // CHECK-SAME: iter_args(%[[ARG1:[a-zA-Z0-9]+]] = %[[INIT0]], %[[ARG2:[a-zA-Z0-9]+]] = %[[INIT1]])
142 // CHECK: %[[INNER:[a-zA-Z0-9]+]]:2 = scf.for %[[IV1:[a-zA-Z0-9]+]] = %[[C0]] to %[[C300]] step %[[C20]]
143 // CHECK-SAME: iter_args(%[[ARG3:[a-zA-Z0-9]+]] = %[[ARG1]], %[[ARG4:[a-zA-Z0-9]+]] = %[[ARG2]])
144 // CHECK-DAG: %[[TS_Y:.+]] = affine.min #[[$MAP0]](%[[IV0]])
145 // CHECK-DAG: %[[ARG_TILE:.+]] = tensor.extract_slice %[[ARG0]]
146 // CHECK-SAME: [%[[IV0]], 0, %[[IV1]]] [%[[TS_Y]], 200, 20] [1, 1, 1]
147 // CHECK-DAG: %[[INIT0_TILE:.+]] = tensor.extract_slice %[[ARG3]]
148 // CHECK-SAME: [%[[IV0]], %[[IV1]], 0] [%[[TS_Y]], 20, 200] [1, 1, 1]
149 // CHECK-DAG: %[[INIT1_TILE:.+]] = tensor.extract_slice %[[ARG4]]
150 // CHECK-SAME: [%[[IV1]], %[[IV0]], 0] [20, %[[TS_Y]], 200] [1, 1, 1]
151 // CHECK: %[[RESULT_TILE:.+]]:2 = linalg.generic
152 // CHECK-SAME: ins(%[[ARG_TILE]] :
153 // CHECK-SAME: outs(%[[INIT0_TILE]], %[[INIT1_TILE]] :
154 // CHECK: %[[UPDATE0:.+]] = tensor.insert_slice %[[RESULT_TILE]]#0 into %[[ARG3]]
155 // CHECK-SAME: [%[[IV0]], %[[IV1]], 0] [%[[TS_Y]], 20, 200] [1, 1, 1]
156 // CHECK: %[[UPDATE1:.+]] = tensor.insert_slice %[[RESULT_TILE]]#1 into %[[ARG4]]
157 // CHECK-SAME: [%[[IV1]], %[[IV0]], 0] [20, %[[TS_Y]], 200] [1, 1, 1]
158 // CHECK: scf.yield %[[UPDATE0]], %[[UPDATE1]]
159 // CHECK: scf.yield %[[INNER]]#0, %[[INNER]]#1
160 // CHECK: return %[[OUTER]]#0, %[[OUTER]]#1
164 func.func @conv2D(%arg0 : tensor<?x?x?x?xf32>, %arg1 : tensor<?x?x?x?xf32>,
165 %arg2 : tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32> {
166 %0 = linalg.conv_2d_nhwc_hwcf {
167 strides = dense<[2, 3]> : tensor<2xi64>,
168 dilation = dense<[4, 5]> : tensor<2xi64>}
169 ins(%arg0, %arg1 : tensor<?x?x?x?xf32>, tensor<?x?x?x?xf32>)
170 outs(%arg2 : tensor<?x?x?x?xf32>) -> tensor<?x?x?x?xf32>
171 return %0 : tensor<?x?x?x?xf32>
174 module attributes {transform.with_named_sequence} {
175 transform.named_sequence @__transform_main(%arg1 : !transform.any_op {transform.readonly}) {
176 %conv = transform.structured.match ops{["linalg.conv_2d_nhwc_hwcf"]} in %arg1
177 : (!transform.any_op) -> !transform.any_op
178 %a, %b, %c, %d = transform.structured.tile_using_for %conv tile_sizes [0, 0, 0, 0, 10, 20, 30]
179 : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op, !transform.any_op)
183 // CHECK-DAG: #[[$MAP0:.+]] = affine_map<(d0)[s0] -> (-d0 + s0, 10)>
184 // CHECK-DAG: #[[$MAP1:.+]] = affine_map<(d0)[s0] -> (-d0 + s0, 20)>
185 // CHECK-DAG: #[[$MAP2:.+]] = affine_map<(d0)[s0] -> (-d0 + s0, 30)>
186 // CHECK-DAG: #[[$MAP3:.+]] = affine_map<(d0)[s0] -> (d0 + s0 * 2 - 2)>
187 // CHECK-DAG: #[[$MAP4:.+]] = affine_map<(d0)[s0] -> (d0 + s0 * 3 - 3)>
188 // CHECK-LABEL: func.func @conv2D(
189 // CHECK-SAME: %[[INPUT:[a-zA-Z0-9]+]]: tensor<?x?x?x?xf32>
190 // CHECK-SAME: %[[FILTER:[a-zA-Z0-9]+]]: tensor<?x?x?x?xf32>
191 // CHECK-SAME: %[[INIT:[a-zA-Z0-9]+]]: tensor<?x?x?x?xf32>
192 // CHECK-DAG: %[[C0:.+]] = arith.constant 0 : index
193 // CHECK-DAG: %[[C1:.+]] = arith.constant 1 : index
194 // CHECK-DAG: %[[C2:.+]] = arith.constant 2 : index
195 // CHECK-DAG: %[[C3:.+]] = arith.constant 3 : index
196 // CHECK-DAG: %[[N:.+]] = tensor.dim %[[INPUT]], %[[C0]]
197 // CHECK-DAG: %[[C:.+]] = tensor.dim %[[INPUT]], %[[C3]]
198 // CHECK-DAG: %[[P:.+]] = tensor.dim %[[FILTER]], %[[C0]]
199 // CHECK-DAG: %[[Q:.+]] = tensor.dim %[[FILTER]], %[[C1]]
200 // CHECK-DAG: %[[F:.+]] = tensor.dim %[[FILTER]], %[[C3]]
201 // CHECK-DAG: %[[R:.+]] = tensor.dim %[[INIT]], %[[C1]]
202 // CHECK-DAG: %[[S:.+]] = tensor.dim %[[INIT]], %[[C2]]
203 // CHECK-DAG: %[[C10:.+]] = arith.constant 10 : index
204 // CHECK-DAG: %[[C20:.+]] = arith.constant 20 : index
205 // CHECK-DAG: %[[C30:.+]] = arith.constant 30 : index
206 // CHECK: scf.for %[[IV0:[a-zA-Z0-9]+]] = %[[C0]] to %[[P]] step %[[C10]]
207 // CHECK-SAME: iter_args(%[[INIT0:.+]] = %[[INIT]])
208 // CHECK: scf.for %[[IV1:[a-zA-Z0-9]+]] = %[[C0]] to %[[Q]] step %[[C20]]
209 // CHECK-SAME: iter_args(%[[INIT1:.+]] = %[[INIT0]])
210 // CHECK: scf.for %[[IV2:[a-zA-Z0-9]+]] = %[[C0]] to %[[C]] step %[[C30]]
211 // CHECK-SAME: iter_args(%[[INIT2:.+]] = %[[INIT1]])
212 // CHECK-DAG: %[[TS_P:.+]] = affine.min #[[$MAP0]](%[[IV0]])[%[[P]]]
213 // CHECK-DAG: %[[TS_Q:.+]] = affine.min #[[$MAP1]](%[[IV1]])[%[[Q]]]
214 // CHECK-DAG: %[[TS_C:.+]] = affine.min #[[$MAP2]](%[[IV2]])[%[[C]]]
215 // CHECK-DAG: %[[TS_H:.+]] = affine.apply #[[$MAP3]](%[[TS_P]])[%[[R]]]
216 // CHECK-DAG: %[[TS_W:.+]] = affine.apply #[[$MAP4]](%[[TS_Q]])[%[[S]]]
217 // CHECK-DAG: %[[INPUT_TILE:.+]] = tensor.extract_slice %[[INPUT]]
218 // CHECK-SAME: [0, %[[IV0]], %[[IV1]], %[[IV2]]] [%[[N]], %[[TS_H]], %[[TS_W]], %[[TS_C]]]
219 // CHECK-DAG: %[[FILTER_TILE:.+]] = tensor.extract_slice %[[FILTER]]
220 // CHECK-SAME: [%[[IV0]], %[[IV1]], %[[IV2]], 0] [%[[TS_P]], %[[TS_Q]], %[[TS_C]], %[[F]]]
221 // CHECK-DAG: %[[INIT_TILE:.+]] = tensor.extract_slice %[[INIT2]]
222 // CHECK-SAME: [0, 0, 0, 0] [%[[N]], %[[R]], %[[S]], %[[F]]]
223 // CHECK: %[[CONV_TILE:.+]] = linalg.conv_2d_nhwc_hwcf
224 // CHECK-SAME: dilation = dense<[4, 5]> : tensor<2xi64>, strides = dense<[2, 3]> : tensor<2xi64>
225 // CHECK-SAME: ins(%[[INPUT_TILE]], %[[FILTER_TILE]] :
226 // CHECK-SAME: outs(%[[INIT_TILE]] :
227 // CHECK: tensor.insert_slice %[[CONV_TILE]] into %[[INIT2]]
228 // CHECK-SAME: [0, 0, 0, 0] [%[[N]], %[[R]], %[[S]], %[[F]]]
232 func.func @indexed_semantics(%arg0: tensor<?x?xf32>, %arg1: tensor<?x?xf32>) -> tensor<?x?xf32> {
233 // Check that we correctly amend "linalg.index" results.
235 %0 = linalg.generic {
236 indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,
237 affine_map<(d0, d1) -> (d0, d1)>],
238 iterator_types = ["parallel", "parallel"]}
239 ins(%arg0: tensor<?x?xf32>)
240 outs(%arg1: tensor<?x?xf32>) {
241 ^bb0(%arg2: f32, %arg3: f32):
242 %1 = linalg.index 0 : index
243 %2 = linalg.index 1 : index
244 %3 = arith.addi %1, %2 : index
245 %4 = arith.index_cast %3 : index to i64
246 %5 = arith.uitofp %4 : i64 to f32
247 %6 = arith.addf %5, %arg2 : f32
248 linalg.yield %6 : f32
249 } -> (tensor<?x?xf32>)
250 return %0 : tensor<?x?xf32>
253 module attributes {transform.with_named_sequence} {
254 transform.named_sequence @__transform_main(%arg1 : !transform.any_op {transform.readonly}) {
255 %generic = transform.structured.match ops{["linalg.generic"]} in %arg1
256 : (!transform.any_op) -> !transform.any_op
257 %a, %b, %c = transform.structured.tile_using_for %generic tile_sizes [10, 20]
258 : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op)
262 // CHECK: #[[$MAP_ADD:.+]] = affine_map<(d0, d1) -> (d0 + d1)>
263 // CHECK-LABEL: @indexed_semantics
264 // CHECK: scf.for %[[I0:.+]] = %{{.*}} to %{{.*}} step %{{.*}}
265 // CHECK: scf.for %[[I1:.+]] = %{{.*}} to %{{.*}} step %{{.*}}
266 // CHECK: %[[INDEX0:.+]] = linalg.index 0
267 // CHECK: %[[INDEX0_AMENDED:.+]] = affine.apply #[[$MAP_ADD]](%[[INDEX0]], %[[I0]])
268 // CHECK: %[[INDEX1:.+]] = linalg.index 1
269 // CHECK: %[[INDEX1_AMENDED:.+]] = affine.apply #[[$MAP_ADD]](%[[INDEX1]], %[[I1]])
270 // CHECK: arith.addi %[[INDEX0_AMENDED]], %[[INDEX1_AMENDED]]
274 func.func @interchange_matmul(%arg0 : tensor<?x?xf32>, %arg1 : tensor<?x?xf32>,
275 %arg2 : tensor<?x?xf32>) -> tensor<?x?xf32> {
276 %0 = linalg.matmul ins(%arg0, %arg1 : tensor<?x?xf32>, tensor<?x?xf32>)
277 outs(%arg2 : tensor<?x?xf32>) -> tensor<?x?xf32>
278 return %0 : tensor<?x?xf32>
281 module attributes {transform.with_named_sequence} {
282 transform.named_sequence @__transform_main(%arg1 : !transform.any_op {transform.readonly}) {
283 %matmul = transform.structured.match ops{["linalg.matmul"]} in %arg1
284 : (!transform.any_op) -> !transform.any_op
285 %a, %b, %c, %d = transform.structured.tile_using_for %matmul tile_sizes [10, 20, 30] interchange = [1, 2, 0]
286 : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op, !transform.any_op)
290 // CHECK-DAG: #[[$MAP0:.+]] = affine_map<(d0)[s0] -> (-d0 + s0, 20)>
291 // CHECK-DAG: #[[$MAP1:.+]] = affine_map<(d0)[s0] -> (-d0 + s0, 30)>
292 // CHECK-DAG: #[[$MAP2:.+]] = affine_map<(d0)[s0] -> (-d0 + s0, 10)>
293 // CHECK-LABEL: func.func @interchange_matmul(
294 // CHECK-SAME: %[[ARG0:[a-zA-Z0-9]+]]: tensor<?x?xf32>
295 // CHECK-SAME: %[[ARG1:[a-zA-Z0-9]+]]: tensor<?x?xf32>
296 // CHECK-SAME: %[[ARG2:[a-zA-Z0-9]+]]: tensor<?x?xf32>
297 // CHECK-DAG: %[[C0:.+]] = arith.constant 0 : index
298 // CHECK-DAG: %[[C1:.+]] = arith.constant 1 : index
299 // CHECK-DAG: %[[M:.+]] = tensor.dim %[[ARG0]], %[[C0]]
300 // CHECK-DAG: %[[K:.+]] = tensor.dim %[[ARG0]], %[[C1]]
301 // CHECK-DAG: %[[N:.+]] = tensor.dim %[[ARG1]], %[[C1]]
302 // CHECK-DAG: %[[C10:.+]] = arith.constant 10 : index
303 // CHECK-DAG: %[[C20:.+]] = arith.constant 20 : index
304 // CHECK-DAG: %[[C30:.+]] = arith.constant 30 : index
305 // CHECK: %[[OUTER:[a-zA-Z0-9]+]] = scf.for %[[IV0:[a-zA-Z0-9]+]] = %[[C0]] to %[[N]] step %[[C20]]
306 // CHECK-SAME: iter_args(%[[INIT0:.+]] = %[[ARG2]])
307 // CHECK: %[[INNER1:[a-zA-Z0-9]+]] = scf.for %[[IV1:[a-zA-Z0-9]+]] = %[[C0]] to %[[K]] step %[[C30]]
308 // CHECK-SAME: iter_args(%[[INIT1:.+]] = %[[INIT0]])
309 // CHECK: %[[INNER2:[a-zA-Z0-9]+]] = scf.for %[[IV2:[a-zA-Z0-9]+]] = %[[C0]] to %[[M]] step %[[C10]]
310 // CHECK-SAME: iter_args(%[[INIT2:.+]] = %[[INIT1]])
311 // CHECK-DAG: %[[TS_N:.+]] = affine.min #[[$MAP0]](%[[IV0]])[%[[N]]]
312 // CHECK-DAG: %[[TS_K:.+]] = affine.min #[[$MAP1]](%[[IV1]])[%[[K]]]
313 // CHECK-DAG: %[[TS_M:.+]] = affine.min #[[$MAP2]](%[[IV2]])[%[[M]]]
314 // CHECK-DAG: %[[LHS_TILE:.+]] = tensor.extract_slice %[[ARG0]]
315 // CHECK-SAME: [%[[IV2]], %[[IV1]]] [%[[TS_M]], %[[TS_K]]] [1, 1]
316 // CHECK-DAG: %[[RHS_TILE:.+]] = tensor.extract_slice %[[ARG1]]
317 // CHECK-SAME: [%[[IV1]], %[[IV0]]] [%[[TS_K]], %[[TS_N]]] [1, 1]
318 // CHECK-DAG: %[[INIT_TILE:.+]] = tensor.extract_slice %[[INIT2]]
319 // CHECK-SAME: [%[[IV2]], %[[IV0]]] [%[[TS_M]], %[[TS_N]]] [1, 1]
320 // CHECK: %[[GEMM_TILE:.+]] = linalg.matmul
321 // CHECK-SAME: ins(%[[LHS_TILE]], %[[RHS_TILE]] :
322 // CHECK-SAME: outs(%[[INIT_TILE]] :
323 // CHECK: %[[UPDATE:.+]] = tensor.insert_slice %[[GEMM_TILE]] into %[[INIT2]]
324 // CHECK-SAME: [%[[IV2]], %[[IV0]]] [%[[TS_M]], %[[TS_N]]] [1, 1]
325 // CHECK: scf.yield %[[UPDATE]]
326 // CHECK: scf.yield %[[INNER2]]
327 // CHECK: scf.yield %[[INNER1]]
328 // CHECK: return %[[OUTER]]
332 func.func @linalg_copy_matmul(%a: memref<?x?xf32>, %b: memref<?x?xf32>) {
333 linalg.copy ins(%a : memref<?x?xf32>) outs(%b : memref<?x?xf32>)
337 module attributes {transform.with_named_sequence} {
338 transform.named_sequence @__transform_main(%arg1 : !transform.any_op {transform.readonly}) {
339 %copy = transform.structured.match ops{["linalg.copy"]} in %arg1
340 : (!transform.any_op) -> !transform.any_op
341 %a, %b, %c = transform.structured.tile_using_for %copy tile_sizes [10, 20]
342 : (!transform.any_op) -> (!transform.any_op, !transform.any_op, !transform.any_op)
346 // CHECK-LABEL: func @linalg_copy_matmul(
349 // CHECK: memref.subview
350 // CHECK: memref.subview
351 // CHECK: linalg.copy
355 func.func @check_scalar_operation(%arg0 : tensor<f32>) -> tensor<f32> {
356 %init = tensor.empty() : tensor<f32>
357 %0 = linalg.generic {
358 indexing_maps = [affine_map<() -> ()>, affine_map<() -> ()>],
360 ins(%arg0 : tensor<f32>) outs(%init : tensor<f32>){
361 ^bb0(%b0 : f32, %b1 : f32):
362 %1 = arith.mulf %b0, %b0 : f32
363 linalg.yield %1 : f32
365 return %0 : tensor<f32>
368 module attributes {transform.with_named_sequence} {
369 transform.named_sequence @__transform_main(%arg1 : !transform.any_op {transform.readonly}) {
370 %generic = transform.structured.match ops{["linalg.generic"]} in %arg1
371 : (!transform.any_op) -> !transform.any_op
372 %a = transform.structured.tile_using_for %generic tile_sizes []
373 : (!transform.any_op) -> (!transform.any_op)
377 // CHECK-LABEL: func @check_scalar_operation
378 // CHECK-NOT: scf.for
379 // CHECK: linalg.generic
383 func.func @check_scalar_memref_operation(%arg0 : memref<f32>, %arg1 : memref<f32>){
385 indexing_maps = [affine_map<() -> ()>, affine_map<() -> ()>],
387 ins(%arg0 : memref<f32>) outs(%arg1 : memref<f32>){
388 ^bb0(%b0 : f32, %b1 : f32):
389 %1 = arith.mulf %b0, %b0 : f32
390 linalg.yield %1 : f32
395 module attributes {transform.with_named_sequence} {
396 transform.named_sequence @__transform_main(%arg1 : !transform.any_op {transform.readonly}) {
397 %generic = transform.structured.match ops{["linalg.generic"]} in %arg1
398 : (!transform.any_op) -> !transform.any_op
399 %a = transform.structured.tile_using_for %generic tile_sizes []
400 : (!transform.any_op) -> (!transform.any_op)
404 // CHECK-LABEL: func @check_scalar_memref_operation
405 // CHECK-NOT: scf.for
406 // CHECK: linalg.generic