1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
|
// RUN: mlir-opt %s -test-linalg-tensor-fusion-transform-patterns -resolve-shaped-type-result-dims -canonicalize -cse --split-input-file | FileCheck %s
// RUN: mlir-opt %s -test-linalg-tiled-loop-fusion-transform-patterns -resolve-shaped-type-result-dims -canonicalize -cse --split-input-file | FileCheck %s --check-prefix=TLOOP
module {
func @matmul_fusion(%A: tensor<?x?xf32>, %B: tensor<?x?xf32>,
%AB_init: tensor<?x?xf32>, %C: tensor<?x?xf32>,
%ABC_init: tensor<?x?xf32>) -> tensor<?x?xf32> {
%AB = linalg.matmul ins(%A, %B : tensor<?x?xf32>, tensor<?x?xf32>)
outs(%AB_init : tensor<?x?xf32>) -> tensor<?x?xf32> // <MxN1> <N1xN2>
%ABC = linalg.matmul {__internal_linalg_transform__ = "lhs_fusion"}
ins(%AB, %C : tensor<?x?xf32>, tensor<?x?xf32>)
outs(%ABC_init : tensor<?x?xf32>) -> tensor<?x?xf32> // <MxN2> <N2xN3>
return %ABC : tensor<?x?xf32>
}
}
// CHECK-DAG: #[[MAP1:.+]] = affine_map<(d0, d1) -> (32, d0 - d1)>
// CHECK-DAG: #[[MAP2:.+]] = affine_map<(d0)[s0] -> (16, -d0 + s0)>
// CHECK-DAG: #[[MAP3:.+]] = affine_map<(d0)[s0] -> (64, -d0 + s0)>
// CHECK-DAG: #[[MAP4:.+]] = affine_map<(d0, d1) -> (64, d0 - d1)>
// CHECK-DAG: #[[MAP5:.+]] = affine_map<(d0)[s0, s1] -> (-d0 + s0, 32, -d0 + s1)>
// CHECK: func @matmul_fusion
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]+]]: tensor<?x?xf32>
// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]+]]: tensor<?x?xf32>
// CHECK-SAME: %[[ARG2:[a-zA-Z0-9_]+]]: tensor<?x?xf32>
// CHECK-SAME: %[[ARG3:[a-zA-Z0-9_]+]]: tensor<?x?xf32>
// CHECK-SAME: %[[ARG4:[a-zA-Z0-9_]+]]: tensor<?x?xf32>
// CHECK-DAG: %[[C0:.+]] = constant 0 : index
// CHECK-DAG: %[[C1:.+]] = constant 1 : index
// CHECK-DAG: %[[C32:.+]] = constant 32 : index
// CHECK-DAG: %[[C64:.+]] = constant 64 : index
// CHECK-DAG: %[[C16:.+]] = constant 16 : index
// CHECK-DAG: %[[M:.+]] = tensor.dim %[[ARG0]], %[[C0]]
// CHECK: %[[RESULT:.+]] = scf.for %[[IV0:[a-zA-Z0-9]+]] =
// CHECK-SAME: %[[C0]] to %[[M]] step %[[C32]]
// CHECK-SAME: iter_args(%[[ARG6:.+]] = %[[ARG4]]) -> (tensor<?x?xf32>) {
// CHECK: %[[M_2:.+]] = tensor.dim %[[ARG6]], %[[C0]]
// CHECK: %[[TILE_M_2:.+]] = affine.min #[[MAP1]](%[[M_2]], %[[IV0]])
// CHECK: %[[N3:.+]] = tensor.dim %[[ARG6]], %[[C1]]
// CHECK: %[[ST_ARG6:.+]] = tensor.extract_slice %[[ARG6]][%[[IV0]], 0]
// CHECK-SAME: [%[[TILE_M_2]], %[[N3]]]
// CHECK: %[[TILE_M_3:.+]] = affine.min #[[MAP5]](%[[IV0]])[%[[M]], %[[M]]]
// CHECK: %[[N1:.+]] = tensor.dim %[[ARG0]], %[[C1]]
// CHECK: %[[ST_ARG0:.+]] = tensor.extract_slice %[[ARG0]][%[[IV0]], 0]
// CHECK-SAME: [%[[TILE_M_3]], %[[N1]]]
// CHECK: %[[M_3:.+]] = tensor.dim %[[ARG2]], %[[C0]]
// CHECK: %[[TILE_M_4:.+]] = affine.min #[[MAP5]](%[[IV0]])[%[[M_3]], %[[M]]]
// CHECK: %[[N2_2:.+]] = tensor.dim %[[ARG2]], %[[C1]]
// CHECK: %[[ST_ARG2:.+]] = tensor.extract_slice %[[ARG2]][%[[IV0]], 0]
// CHECK-SAME: [%[[TILE_M_4]], %[[N2_2]]]
// CHECK: %[[LHS:.+]] = linalg.matmul
// CHECK-SAME: __internal_linalg_transform__ = "after_lhs_fusion_producer"
// CHECK-SAME: ins(%[[ST_ARG0]], %[[ARG1]] : tensor<?x?xf32>, tensor<?x?xf32>)
// CHECK-SAME: outs(%[[ST_ARG2]] : tensor<?x?xf32>)
// CHECK: %[[N2:.+]] = tensor.dim %[[ARG1]], %[[C1]]
// CHECK: %[[N3_2:.+]] = tensor.dim %[[ARG3]], %[[C1]]
// CHECK: %[[YIELD0:.+]] = scf.for %[[IV1:[a-zA-Z0-9]+]] =
// CHECK-SAME: %[[C0]] to %[[N3_2]] step %[[C64]]
// CHECK-SAME: iter_args(%[[ARG8:.+]] = %[[ST_ARG6]]) -> (tensor<?x?xf32>) {
// CHECK: %[[YIELD1:.+]] = scf.for %[[IV2:[a-zA-Z0-9]+]] =
// CHECK-SAME: %[[C0]] to %[[N2]] step %[[C16]]
// CHECK-SAME: iter_args(%[[ARG10:.+]] = %[[ARG8]]) -> (tensor<?x?xf32>) {
// CHECK: %[[TILE_N2:.+]] = affine.min #[[MAP2]](%[[IV2]])[%[[N2]]]
// CHECK: %[[ST_LHS:.+]] = tensor.extract_slice %[[LHS]][0, %[[IV2]]]
// CHECK-SAME: [%[[TILE_M_3]], %[[TILE_N2]]]
// CHECK: %[[N2_3:.+]] = tensor.dim %[[ARG3]], %[[C0]]
// CHECK: %[[TILE_N2_2:.+]] = affine.min #[[MAP2]](%[[IV2]])[%[[N2_3]]]
// CHECK: %[[TILE_N3:.+]] = affine.min #[[MAP3]](%[[IV1]])[%[[N3_2]]]
// CHECK: %[[ST_ARG3:.+]] = tensor.extract_slice %[[ARG3]][%[[IV2]], %[[IV1]]]
// CHECK-SAME: [%[[TILE_N2_2]], %[[TILE_N3]]]
// CHECK: %[[M_4:.+]] = tensor.dim %[[ARG10]], %[[C0]]
// CHECK: %[[N3_3:.+]] = tensor.dim %[[ARG10]], %[[C1]]
// CHECK: %[[TILE_N3_2:.+]] = affine.min #[[MAP4]](%[[N3_3]], %[[IV1]])
// CHECK: %[[ST_ARG4:.+]] = tensor.extract_slice %[[ARG10]][0, %[[IV1]]]
// CHECK-SAME: [%[[M_4]], %[[TILE_N3_2]]]
// CHECK: %[[ST_RESULT:.+]] = linalg.matmul
// CHECK-SAME: __internal_linalg_transform__ = "after_lhs_fusion"
// CHECK-SAME: ins(%[[ST_LHS]], %[[ST_ARG3]]
// CHECK-SAME: : tensor<?x?xf32>, tensor<?x?xf32>)
// CHECK-SAME: outs(%[[ST_ARG4]] : tensor<?x?xf32>)
// CHECK: %[[UPDATE1:.+]] = tensor.insert_slice %[[ST_RESULT]]
// CHECK-SAME: into %[[ARG10]][0, %[[IV1]]] [%[[M_4]], %[[TILE_N3_2]]]
// CHECK: scf.yield %[[UPDATE1]]
// CHECK: }
// CHECK: scf.yield %[[YIELD1]]
// CHECK: }
// CHECK: %[[UPDATE0:.+]] = tensor.insert_slice %[[YIELD0]] into
// CHECK-SAME: %[[ARG6]][%[[IV0]], 0] [%[[TILE_M_2]], %[[N3]]]
// CHECK: scf.yield %[[UPDATE0]]
// CHECK: }
// CHECK: return %[[RESULT]]
// TLOOP-LABEL: func @matmul_fusion(
// TLOOP-SAME: %[[A:[a-zA-Z0-9_]+]]: tensor<?x?xf32>,
// TLOOP-SAME: %[[B:[a-zA-Z0-9_]+]]: tensor<?x?xf32>,
// TLOOP-SAME: %[[AB_INIT:[a-zA-Z0-9_]+]]: tensor<?x?xf32>,
// TLOOP-SAME: %[[C:[a-zA-Z0-9_]+]]: tensor<?x?xf32>,
// TLOOP-SAME: %[[ABC_INIT:[a-zA-Z0-9_]+]]: tensor<?x?xf32>) -> tensor<?x?xf32> {
// TLOOP-DAG: %[[C32:.*]] = constant 32 : index
// TLOOP-DAG: %[[C64:.*]] = constant 64 : index
// TLOOP-DAG: %[[C16:.*]] = constant 16 : index
// TLOOP-DAG: %[[C0:.*]] = constant 0 : index
// TLOOP-DAG: %[[C1:.*]] = constant 1 : index
// TLOOP: %[[DIM_A0:.*]] = tensor.dim %[[A]], %[[C0]] : [[TY:.*]]
// TLOOP: %[[ABC:.*]] = linalg.tiled_loop (%[[IV0:.*]]) = (%[[C0]])
// TLOOP-SAME: to (%[[DIM_A0]]) step (%[[C32]])
// TLOOP-SAME: ins (%[[C_:.*]] = %[[C]]: tensor<?x?xf32>,
// TLOOP-SAME: %[[A_:.*]] = %[[A]]: tensor<?x?xf32>,
// TLOOP-SAME: %[[B_:.*]] = %[[B]]: tensor<?x?xf32>,
// TLOOP-SAME: %[[AB_INIT_:.*]] = %[[AB_INIT]]: tensor<?x?xf32>)
// TLOOP-SAME: outs (%[[ABC_INIT_:.*]] = %[[ABC_INIT]]: tensor<?x?xf32>) {
// TLOOP: %[[ABC_INIT_SUB:.*]] = tensor.extract_slice %[[ABC_INIT_]][%[[IV0]], 0]
// TLOOP: %[[A_SUB:.*]] = tensor.extract_slice %[[A_]][%[[IV0]], 0]
// TLOOP: %[[AB_INIT_SUB:.*]] = tensor.extract_slice %[[AB_INIT_]][%[[IV0]], 0]
// TLOOP: %[[AB_SUB:.*]] = linalg.matmul
// TLOOP-SAME: ins(%[[A_SUB]], %[[B_]] : {{.*}}) outs(%[[AB_INIT_SUB]]
// TLOOP: %[[DIM_B_1:.*]] = tensor.dim %[[B_]], %[[C1]] : [[TY]]
// TLOOP: %[[DIM_C_1:.*]] = tensor.dim %[[C_]], %[[C1]] : [[TY]]
// TLOOP: %[[ABC_SUB_:.*]] = linalg.tiled_loop (%[[IV1:.*]], %[[IV2:.*]]) =
// TLOOP-SAME: (%[[C0]], %[[C0]]) to (%[[DIM_C_1]], %[[DIM_B_1]])
// TLOOP-SAME: step (%[[C64]], %[[C16]])
// TLOOP-SAME: ins (%[[AB_SUB_:.*]] = %[[AB_SUB]]: [[TY]],
// TLOOP-SAME: %[[C__:.*]] = %[[C_]]: [[TY]])
// TLOOP-SAME: outs (%[[ABC_INIT_SUB_:.*]] = %[[ABC_INIT_SUB]]: [[TY]])
// TLOOP-SAME: iterators["parallel", "reduction"] {
// TLOOP: %[[AB_SUB_SUB:.*]] = tensor.extract_slice %[[AB_SUB_]][0, %[[IV2]]]
// TLOOP: %[[C__SUB:.*]] = tensor.extract_slice %[[C__]][%[[IV2]], %[[IV1]]]
// TLOOP: %[[ABS_INIT_SUB_SUB:.*]] = tensor.extract_slice %[[ABC_INIT_SUB_]][0, %[[IV1]]]
// TLOOP: %[[ABC_SUB_SUB:.*]] = linalg.matmul
// TLOOP-SAME: ins(%[[AB_SUB_SUB]], %[[C__SUB]] : [[TY]], [[TY]])
// TLOOP-SAME: outs(%[[ABS_INIT_SUB_SUB]] : [[TY]]) -> [[TY]]
// TLOOP: %[[RES0:.*]] = tensor.insert_slice %[[ABC_SUB_SUB]]
// TLOOP-SAME: into %[[ABC_INIT_SUB_]][0, %[[IV1]]]
// TLOOP: linalg.yield %[[RES0]] : [[TY]]
// TLOOP: }
// TLOOP: %[[RES1:.*]] = tensor.insert_slice %[[ABC_SUB_]] into %[[ABC_INIT_]][%[[IV0]], 0]
// TLOOP: linalg.yield %[[RES1]] : [[TY]]
// TLOOP: }
// TLOOP: return %[[ABC]] : [[TY]]
// -----
module {
func @matmul_plus_matmul(%arg0: tensor<?x?xf32>, %arg1: tensor<?x?xf32>,
%arg2: tensor<?x?xf32>) -> tensor<?x?xf32>{
%c0 = constant 0 : index
%c1 = constant 1 : index
%0 = tensor.dim %arg2, %c0 : tensor<?x?xf32>
%1 = tensor.dim %arg2, %c1 : tensor<?x?xf32>
%2 = linalg.matmul ins(%arg0, %arg1 : tensor<?x?xf32>, tensor<?x?xf32>)
outs(%arg2 : tensor<?x?xf32>) -> tensor<?x?xf32>
%3 = tensor.dim %2, %c0 : tensor<?x?xf32>
%4 = tensor.dim %2, %c1 : tensor<?x?xf32>
%5 = linalg.init_tensor [%3, %4] : tensor<?x?xf32>
%6 = linalg.generic
{indexing_maps = [affine_map<(d0, d1) -> (d0, d1)>,
affine_map<(d0, d1) -> (d0, d1)>,
affine_map<(d0, d1) -> (d0, d1)>],
iterator_types = ["parallel", "parallel"],
__internal_linalg_transform__ = "transpose_fusion"}
ins(%2, %2 : tensor<?x?xf32>, tensor<?x?xf32>)
outs(%5 : tensor<?x?xf32>) {
^bb0(%arg3 : f32, %arg4 : f32, %arg5 : f32) :
%7 = addf %arg3, %arg4 : f32
linalg.yield %7 : f32
} -> tensor<?x?xf32>
return %6 : tensor<?x?xf32>
}
}
// CHECK: func @matmul_plus_matmul
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]+]]: tensor<?x?xf32>
// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]+]]: tensor<?x?xf32>
// CHECK-SAME: %[[ARG2:[a-zA-Z0-9_]+]]: tensor<?x?xf32>
// CHECK: %[[RESULT:.+]] = scf.for %[[IV0:[a-zA-Z0-9_]+]]
// CHECK-SAME: iter_args(%[[ARG4:.+]] = %{{[a-zA-Z0-9_]+}})
// CHECK: %[[YIELD:.+]] = scf.for %[[IV1:[a-zA-Z0-9_]+]]
// CHECK-SAME: iter_args(%[[ARG6:.+]] = %[[ARG4]])
// CHECK: %[[ST_ARG6:.+]] = tensor.extract_slice %[[ARG6]][%[[IV0]], %[[IV1]]]
// CHECK: %[[ST_ARG0:.+]] = tensor.extract_slice %[[ARG0]][%[[IV0]], 0]
// CHECK: %[[ST_ARG1:.+]] = tensor.extract_slice %[[ARG1]][0, %[[IV1]]]
// CHECK: %[[ST_ARG2:.+]] = tensor.extract_slice %[[ARG2]][%[[IV0]], %[[IV1]]]
// CHECK: %[[LHS:.+]] = linalg.matmul
// CHECK-SAME: ins(%[[ST_ARG0]], %[[ST_ARG1]]
// CHECK-SAME: : tensor<?x?xf32>, tensor<?x?xf32>)
// CHECK-SAME: outs(%[[ST_ARG2]] : tensor<?x?xf32>)
// CHECK: %[[ST_RESULT:.+]] = linalg.generic
// CHECK-SAME: ins(%[[LHS]] : tensor<?x?xf32>)
// CHECK-SAME: outs(%[[ST_ARG6]] : tensor<?x?xf32>)
// CHECK: %[[UPDATE:.+]] = tensor.insert_slice %[[ST_RESULT]]
// CHECK-SAME: into %[[ARG6]][%[[IV0]], %[[IV1]]]
// CHECK: scf.yield %[[UPDATE]]
// CHECK: scf.yield %[[YIELD]]
// CHECK: return %[[RESULT]]
// TLOOP-LABEL: func @matmul_plus_matmul
// TLOOP-SAME: %[[A:[a-zA-Z0-9_]+]]: tensor<?x?xf32>,
// TLOOP-SAME: %[[B:[a-zA-Z0-9_]+]]: tensor<?x?xf32>,
// TLOOP-SAME: %[[AB:[a-zA-Z0-9_]+]]: tensor<?x?xf32>
// TLOOP-DAG: %[[C32:.*]] = constant 32 : index
// TLOOP-DAG: %[[C64:.*]] = constant 64 : index
// TLOOP-DAG: %[[C0:.*]] = constant 0 : index
// TLOOP-DAG: %[[C1:.*]] = constant 1 : index
// TLOOP: %[[DIM_A_0:.*]] = tensor.dim %[[A]], %[[C0]] : [[TY:.*]]
// TLOOP: %[[DIM_B_1:.*]] = tensor.dim %[[B]], %[[C1]] : [[TY]]
// TLOOP: %[[INIT:.*]] = linalg.init_tensor [%[[DIM_A_0]], %[[DIM_B_1]]]
// TLOOP: %[[RESULT:.*]] = linalg.tiled_loop (%[[IV0:.*]], %[[IV1:.*]]) =
// TLOOP-SAME: (%[[C0]], %[[C0]]) to (%[[DIM_A_0]], %[[DIM_B_1]])
// TLOOP-SAME: step (%[[C32]], %[[C64]])
// TLOOP-SAME: ins (%[[A_:.*]] = %[[A]]: [[TY]],
// TLOOP-SAME: %[[B_:.*]] = %[[B]]: [[TY]],
// TLOOP-SAME: %[[AB_:.*]] = %[[AB]]: [[TY]])
// TLOOP-SAME: outs (%[[INIT_:.*]] = %[[INIT]]: [[TY]]) {
// TLOOP: %[[INIT_SUB:.*]] = tensor.extract_slice %[[INIT_]][%[[IV0]], %[[IV1]]]
// TLOOP: %[[A_SUB:.*]] = tensor.extract_slice %[[A_]][%[[IV0]], 0]
// TLOOP: %[[B_SUB:.*]] = tensor.extract_slice %[[B_]][0, %[[IV1]]]
// TLOOP: %[[AB_SUB_INIT:.*]] = tensor.extract_slice %[[AB_]][%[[IV0]], %[[IV1]]]
// TLOOP: %[[AB_SUB:.*]] = linalg.matmul
// TLOOP-SAME: ins(%[[A_SUB]], %[[B_SUB]] : [[TY]], [[TY]])
// TLOOP-SAME: outs(%[[AB_SUB_INIT]] : [[TY]])
// TLOOP: %[[DOUBLE_AB:.*]] = linalg.generic
// TLOOP-SAME: ins(%[[AB_SUB]] : [[TY]]) outs(%[[INIT_SUB]] : [[TY]])
// TLOOP: %[[RESULT_SUB:.*]] = tensor.insert_slice
// TLOOP-SAME: %[[DOUBLE_AB:.*]] into %[[INIT_]][%[[IV0]], %[[IV1]]]
// TLOOP: linalg.yield %[[RESULT_SUB]] : [[TY]]
// TLOOP: }
// TLOOP: return %[[RESULT]] : [[TY]]
// -----
module {
func @matmul_out_fusion(%arg0: tensor<?x?xf32>, %arg1: tensor<?x?xf32>,
%arg2: tensor<?x?xf32>) -> tensor<?x?xf32> {
%c0 = constant 0.0 : f32
%0 = linalg.fill(%c0, %arg0) : f32, tensor<?x?xf32> -> tensor<?x?xf32>
%1 = linalg.matmul {__internal_linalg_transform__ = "out_fusion"}
ins(%arg1, %arg2 : tensor<?x?xf32>, tensor<?x?xf32>)
outs(%0 : tensor<?x?xf32>) -> tensor<?x?xf32>
return %1 : tensor<?x?xf32>
}
}
// CHECK-LABEL: func @matmul_out_fusion(
// CHECK-SAME: %[[ARG0:[a-zA-Z0-9_]+]]: tensor<?x?xf32>
// CHECK-SAME: %[[ARG1:[a-zA-Z0-9_]+]]: tensor<?x?xf32>
// CHECK-SAME: %[[ARG2:[a-zA-Z0-9_]+]]: tensor<?x?xf32>
// CHECK: %[[C0:.*]] = constant 0.0{{.*}} : f32
// CHECK-NOT: fill
// CHECK: scf.for %[[I:.*]]{{.*}}iter_args(%{{.*}} = %[[ARG0]]) -> (tensor<?x?xf32>) {
// CHECK: scf.for %[[J:.*]]
// CHECK: %[[ST:.*]] = tensor.extract_slice %[[ARG0]]
// CHECK: %[[ST_FILL:.*]] = linalg.fill(%[[C0]], %[[ST]]) {__internal_linalg_transform__ = "after_out_fusion_producer"} : f32, tensor<?x?xf32> -> tensor<?x?xf32>
// CHECK: %[[ST_MM_RES:.*]] = scf.for %[[K:.*]]{{.*}}iter_args(%[[BB:.*]] = %[[ST_FILL]]) -> (tensor<?x?xf32>) {
// CHECK-NOT: fill
// CHECK: %[[ST_MM:.*]] = linalg.matmul {__internal_linalg_transform__ = "after_out_fusion"} ins(%{{.*}}, %{{.*}} : tensor<?x?xf32>, tensor<?x?xf32>) outs(%[[BB]] : tensor<?x?xf32>) -> tensor<?x?xf32>
// CHECK: scf.yield %[[ST_MM]] : tensor<?x?xf32>
// CHECK: %[[MM:.*]] = tensor.insert_slice %[[ST_MM_RES]] into {{.*}}
// CHECK: scf.yield %[[MM]] : tensor<?x?xf32>
// TLOOP-LABEL: func @matmul_out_fusion(
// TLOOP-SAME: %[[OUT:[a-zA-Z0-9_]+]]: tensor<?x?xf32>
// TLOOP-SAME: %[[A:[a-zA-Z0-9_]+]]: tensor<?x?xf32>
// TLOOP-SAME: %[[B:[a-zA-Z0-9_]+]]: tensor<?x?xf32>
// TLOOP-DAG: %[[C0_F32:.*]] = constant 0.0
// TLOOP-DAG: %[[C32:.*]] = constant 32 : index
// TLOOP-DAG: %[[C64:.*]] = constant 64 : index
// TLOOP-DAG: %[[C16:.*]] = constant 16 : index
// TLOOP-DAG: %[[C0:.*]] = constant 0 : index
// TLOOP-DAG: %[[C1:.*]] = constant 1 : index
// TLOOP: %[[DIM_A_0:.*]] = tensor.dim %[[A]], %[[C0]] : [[TY:.*]]
// TLOOP: %[[DIM_B_1:.*]] = tensor.dim %[[B]], %[[C1]] : [[TY]]
// TLOOP: %[[AB:.*]] = linalg.tiled_loop (%[[I:.*]], %[[J:.*]]) =
// TLOOP-SAME: (%[[C0]], %[[C0]]) to (%[[DIM_A_0]], %[[DIM_B_1]])
// TLOOP-SAME: step (%[[C32]], %[[C64]])
// TLOOP-SAME: ins (%[[A_:.*]] = %[[A]]: [[TY]],
// TLOOP-SAME: %[[B_:.*]] = %[[B]]: [[TY]],
// TLOOP-SAME: %[[C0_F32_:.*]] = %[[C0_F32]]
// TLOOP-SAME: outs (%[[OUT_:.*]] = %[[OUT]]: [[TY]]) {
// TLOOP: %[[DIM_A__1:.*]] = tensor.dim %[[A_]], %[[C1]] : [[TY]]
// TLOOP: %[[A_SUB:.*]] = tensor.extract_slice %[[A_]][%[[I]], 0]
// TLOOP: %[[B_SUB:.*]] = tensor.extract_slice %[[B_]][0, %[[J]]]
// TLOOP: %[[OUT_SUB:.*]] = tensor.extract_slice %[[OUT_]][%[[I]], %[[J]]]
// TLOOP: %[[INIT_SUB:.*]] = linalg.fill(%[[C0_F32_]], %[[OUT_SUB]])
// TLOOP: %[[AB_SUB:.*]] = linalg.tiled_loop (%[[K:.*]]) = (%[[C0]])
// TLOOP-SAME: to (%[[DIM_A__1]]) step (%[[C16]])
// TLOOP-SAME: ins (%[[A_SUB_:.*]] = %[[A_SUB]]: [[TY]],
// TLOOP-SAME: %[[B_SUB_:.*]] = %[[B_SUB]]: [[TY]])
// TLOOP-SAME: outs (%[[INIT_SUB_:.*]] = %[[INIT_SUB]]: [[TY]])
// TLOOP-SAME: iterators["reduction"] {
// TLOOP: %[[A_SUB_SUB:.*]] = tensor.extract_slice %[[A_SUB_]][0, %[[K]]]
// TLOOP: %[[B_SUB_SUB:.*]] = tensor.extract_slice %[[B_SUB_]][%[[K]], 0]
// TLOOP: %[[AB_SUB_SUB:.*]] = linalg.matmul
// TLOOP-SAME: ins(%[[A_SUB_SUB]], %[[B_SUB_SUB]] : [[TY]], [[TY]])
// TLOOP-SAME: outs(%[[INIT_SUB_]] : [[TY]]) -> [[TY]]
// TLOOP: linalg.yield %[[AB_SUB_SUB]] : [[TY]]
// TLOOP: }
// TLOOP: %[[SUB_RESULT:.*]] = tensor.insert_slice %[[AB_SUB]]
// TLOOP-SAME: into %[[OUT_]][%[[I]], %[[J]]]
// TLOOP: linalg.yield %[[SUB_RESULT]] : [[TY]]
// TLOOP: }
// TLOOP: return %[[AB]] : [[TY]]
// -----
module {
func @generic_plus_matmul(%arg0: tensor<?x?xf32>, %arg1: tensor<?x?xf32>,
%arg2: tensor<?x?xf32>) -> tensor<?x?xf32> {
%c0 = constant 0.0 : f32
%0 = linalg.generic {
indexing_maps = [affine_map<(m, n) -> ()>, affine_map<(m, n) -> (m, n)>],
iterator_types = ["parallel", "parallel"]}
ins(%c0 : f32)
outs(%arg0: tensor<?x?xf32>) {
^bb(%0: f32, %1: f32) :
linalg.yield %0 : f32
} -> tensor<?x?xf32>
%1 = linalg.matmul {__internal_linalg_transform__ = "out_fusion"}
ins(%arg1, %arg2 : tensor<?x?xf32>, tensor<?x?xf32>)
outs(%0 : tensor<?x?xf32>) -> tensor<?x?xf32>
return %1 : tensor<?x?xf32>
}
}
// TLOOP-LABEL: func @generic_plus_matmul(
// TLOOP-SAME: %[[OUT:[a-zA-Z0-9_]+]]: tensor<?x?xf32>
// TLOOP-SAME: %[[A:[a-zA-Z0-9_]+]]: tensor<?x?xf32>
// TLOOP-SAME: %[[B:[a-zA-Z0-9_]+]]: tensor<?x?xf32>
// TLOOP-DAG: %[[C0_F32:.*]] = constant 0.0
// TLOOP-DAG: %[[C32:.*]] = constant 32 : index
// TLOOP-DAG: %[[C64:.*]] = constant 64 : index
// TLOOP-DAG: %[[C16:.*]] = constant 16 : index
// TLOOP-DAG: %[[C0:.*]] = constant 0 : index
// TLOOP-DAG: %[[C1:.*]] = constant 1 : index
// TLOOP: %[[DIM_A_0:.*]] = tensor.dim %[[A]], %[[C0]] : [[TY:.*]]
// TLOOP: %[[DIM_B_1:.*]] = tensor.dim %[[B]], %[[C1]] : [[TY]]
// TLOOP: %[[AB:.*]] = linalg.tiled_loop (%[[I:.*]], %[[J:.*]]) =
// TLOOP-SAME: (%[[C0]], %[[C0]]) to (%[[DIM_A_0]], %[[DIM_B_1]])
// TLOOP-SAME: step (%[[C32]], %[[C64]])
// TLOOP-SAME: ins (%[[A_:.*]] = %[[A]]: [[TY]],
// TLOOP-SAME: %[[B_:.*]] = %[[B]]: [[TY]],
// TLOOP-SAME: %[[C0_F32_:.*]] = %[[C0_F32]]
// TLOOP-SAME: outs (%[[OUT_:.*]] = %[[OUT]]: [[TY]]) {
// TLOOP: %[[DIM_A__1:.*]] = tensor.dim %[[A_]], %[[C1]] : [[TY]]
// TLOOP: %[[A_SUB:.*]] = tensor.extract_slice %[[A_]][%[[I]], 0]
// TLOOP: %[[B_SUB:.*]] = tensor.extract_slice %[[B_]][0, %[[J]]]
// TLOOP: %[[OUT_SUB:.*]] = tensor.extract_slice %[[OUT_]][%[[I]], %[[J]]]
// TLOOP: %[[INIT_SUB:.*]] = linalg.generic
// TLOOP-SAME: ins(%[[C0_F32_]]
// TLOOP-SAME: outs(%[[OUT_SUB]]
// TLOOP: %[[AB_SUB:.*]] = linalg.tiled_loop (%[[K:.*]]) = (%[[C0]])
// TLOOP-SAME: to (%[[DIM_A__1]]) step (%[[C16]])
// TLOOP-SAME: ins (%[[A_SUB_:.*]] = %[[A_SUB]]: [[TY]],
// TLOOP-SAME: %[[B_SUB_:.*]] = %[[B_SUB]]: [[TY]])
// TLOOP-SAME: outs (%[[INIT_SUB_:.*]] = %[[INIT_SUB]]: [[TY]])
// TLOOP-SAME: iterators["reduction"] {
// TLOOP: %[[A_SUB_SUB:.*]] = tensor.extract_slice %[[A_SUB_]][0, %[[K]]]
// TLOOP: %[[B_SUB_SUB:.*]] = tensor.extract_slice %[[B_SUB_]][%[[K]], 0]
// TLOOP: %[[AB_SUB_SUB:.*]] = linalg.matmul
// TLOOP-SAME: ins(%[[A_SUB_SUB]], %[[B_SUB_SUB]] : [[TY]], [[TY]])
// TLOOP-SAME: outs(%[[INIT_SUB_]] : [[TY]]) -> [[TY]]
// TLOOP: linalg.yield %[[AB_SUB_SUB]] : [[TY]]
// TLOOP: }
// TLOOP: %[[SUB_RESULT:.*]] = tensor.insert_slice %[[AB_SUB]]
// TLOOP-SAME: into %[[OUT_]][%[[I]], %[[J]]]
// TLOOP: linalg.yield %[[SUB_RESULT]] : [[TY]]
// TLOOP: }
// TLOOP: return %[[AB]] : [[TY]]
|
