xref: /llvm-project/mlir/test/Dialect/Affine/loop-tiling.mlir (revision 96ff0255f2ecd0bd5d1aba650a99e046a4bd5ec6)
1// RUN: mlir-opt %s -split-input-file -affine-loop-tile="tile-size=32" | FileCheck %s
2// RUN: mlir-opt %s -split-input-file -affine-loop-tile="cache-size=512" | FileCheck %s --check-prefix=MODEL
3// RUN: mlir-opt %s -split-input-file -affine-loop-tile="tile-size=32 separate" | FileCheck %s --check-prefix=SEPARATE
4
5// -----
6
7// CHECK-DAG: [[$UB:#map[0-9]*]] = affine_map<(d0) -> (d0 + 32)>
8// CHECK-DAG: [[$UB_MIN:#map[0-9]*]] = affine_map<(d0) -> (d0 + 32, 50)>
9// CHECK-DAG: [[$ID:#map[0-9]*]] = affine_map<(d0) -> (d0)>
10// CHECK-DAG: [[$ID_PLUS_21:#map[0-9]*]] = affine_map<(d0) -> (d0 + 21)>
11
12// CHECK-LABEL: func @loop_tiling()
13// CHECK-NEXT:   affine.for %{{.*}} = 0 to 256 step 32 {
14// CHECK-NEXT:     affine.for %{{.*}} = 0 to 512 step 32 {
15// CHECK-NEXT:       affine.for %{{.*}} = 0 to 1024 step 32 {
16// CHECK-NEXT:         affine.for %[[I:.*]] = [[$ID]](%{{.*}}) to [[$UB]](%{{.*}}) {
17// CHECK-NEXT:           affine.for %[[J:.*]] = [[$ID]](%{{.*}}) to [[$UB]](%{{.*}}) {
18// CHECK-NEXT:             affine.for %[[K:.*]] = [[$ID]](%{{.*}}) to [[$UB]](%{{.*}}) {
19// CHECK-NEXT:               "test.foo"(%[[I]], %[[J]], %[[K]])
20// CHECK-NEXT:             }
21// CHECK-NEXT:           }
22// CHECK-NEXT:         }
23// CHECK-NEXT:       }
24// CHECK-NEXT:     }
25// CHECK-NEXT:   }
26// CHECK-NEXT:   affine.for %{{.*}} = 0 to 50 step 32 {
27// CHECK-NEXT:     affine.for %[[X:.*]] = [[$ID]](%{{.*}}) to min [[$UB_MIN]](%{{.*}}) {
28// CHECK-NEXT:       "test.bar"(%[[X]], %[[X]])
29// CHECK-NEXT:     }
30// CHECK-NEXT:   }
31// CHECK-NEXT: affine.for %[[I:.*]] = 0 to 21 step 32 {
32// CHECK-NEXT:   affine.for %[[Y:.*]] = [[$ID]](%[[I]]) to [[$ID_PLUS_21]](%[[I]])  {
33// CHECK-NEXT:     "test.foobar"(%[[Y]])
34// CHECK-NEXT:   }
35// CHECK-NEXT: }
36// CHECK-NEXT:  return
37func.func @loop_tiling() {
38  affine.for %i = 0 to 256 {
39    affine.for %j = 0 to 512 {
40      affine.for %k = 0 to 1024 {
41        "test.foo"(%i, %j, %k) : (index, index, index) -> ()
42      }
43    }
44  }
45
46  affine.for %x = 0 to 50 {
47    "test.bar"(%x, %x) : (index, index) -> ()
48  }
49
50  // Intra-tile loop won't need a min expression.
51  affine.for %y = 0 to 21 {
52    "test.foobar"(%y) : (index) -> ()
53  }
54
55  return
56}
57
58// -----
59
60// CHECK-DAG: [[$IDENTITY:#map[0-9]*]] = affine_map<(d0) -> (d0)>
61// CHECK-DAG: [[$LB:#map[0-9]*]] = affine_map<()[s0] -> (0, s0)>
62// CHECK-DAG: [[$UB:#map[0-9]*]] = affine_map<()[s0, s1] -> (s0, 4096 floordiv s1)>
63// CHECK-DAG: [[$UB_INTRA_TILE:#map[0-9]*]] = affine_map<(d0)[s0, s1] -> (d0 + 32, s0, 4096 floordiv s1)>
64
65#lb = affine_map<()[s0] -> (0, s0)>
66#ub = affine_map<()[s0, s1] -> (s0, 4096 floordiv s1)>
67// CHECK-LABEL: func @loop_max_min_bound(%{{.*}}: memref<?xi32>, %{{.*}}: index, %{{.*}}: index) {
68func.func @loop_max_min_bound(%A : memref<? x i32>, %L : index, %U : index) {
69  %c0 = arith.constant 0 : index
70  %M = memref.dim %A, %c0 : memref<? x i32>
71  affine.for %i = max #lb()[%L] to min #ub()[%M, %U] {
72    arith.addi %i, %i : index
73  }
74  return
75// CHECK:       affine.for %{{.*}} = max [[$LB]]()[%{{.*}}] to min [[$UB]]()[%{{.*}}, %{{.*}}] step 32 {
76// CHECK-NEXT:    affine.for %[[I:.*]] = [[$IDENTITY]](%{{.*}}) to min [[$UB_INTRA_TILE]](%{{.*}})[%{{.*}}, %{{.*}}] {
77// CHECK-NEXT:      arith.addi %[[I]], %[[I]]
78// CHECK-NEXT:    }
79// CHECK-NEXT:  }
80}
81
82// -----
83
84// Cache size is set to 512 KiB. This loop nest accesses about 49 MiB, and the
85// tile sizes chosen would be 6 x 6 x 6. However, to avoid min/max, which is
86// possible here, they are adjusted to 4 x 4 x 5.
87
88// MODEL-LABEL: func @simple_matmul
89func.func @simple_matmul(%arg0: memref<256x256xvector<64xf32>>, %arg1: memref<256x256xvector<64xf32>>, %arg2: memref<256x256xvector<64xf32>>) -> memref<256x256xvector<64xf32>> {
90  affine.for %i = 0 to 256 {
91    affine.for %j = 0 to 256 {
92      affine.for %k = 0 to 250 {
93        %l = affine.load %arg0[%i, %k] : memref<256x256xvector<64xf32>>
94        %r = affine.load %arg1[%k, %j] : memref<256x256xvector<64xf32>>
95        %o = affine.load %arg2[%i, %j] : memref<256x256xvector<64xf32>>
96        %m = arith.mulf %l, %r : vector<64xf32>
97        %a = arith.addf %o, %m : vector<64xf32>
98        affine.store %a, %arg2[%i, %j] : memref<256x256xvector<64xf32>>
99      }
100    }
101  }
102  return %arg2 : memref<256x256xvector<64xf32>>
103}
104// MODEL:       affine.for %{{.*}} = 0 to 256 step 4 {
105// MODEL-NEXT:    affine.for %{{.*}} = 0 to 256 step 4 {
106// MODEL-NEXT:      affine.for %{{.*}} = 0 to 250 step 5 {
107
108
109// -----
110
111// CHECK-DAG: [[$UBMAP:#map[0-9]*]] = affine_map<(d0)[s0] -> (d0 + 32, s0)>
112
113func.func @tile_using_symbolic_loop_upper_bounds(%arg0: memref<?x?xf32>, %arg1: memref<?x?xf32>, %arg2: memref<?x?xf32>) {
114  %cst = arith.constant 0.000000e+00 : f32
115  %c0 = arith.constant 0 : index
116  %0 = memref.dim %arg0, %c0 : memref<?x?xf32>
117  affine.for %i0 = 0 to %0 {
118    affine.for %i1 = 0 to %0 {
119      affine.store %cst, %arg2[%i0, %i1] : memref<?x?xf32>
120      affine.for %i2 = 0 to %0 {
121        %1 = affine.load %arg0[%i0, %i2] : memref<?x?xf32>
122        %2 = affine.load %arg1[%i2, %i1] : memref<?x?xf32>
123        %3 = arith.mulf %1, %2 : f32
124        %4 = affine.load %arg2[%i0, %i1] : memref<?x?xf32>
125        %5 = arith.addf %4, %3 : f32
126        affine.store %5, %arg2[%i0, %i1] : memref<?x?xf32>
127      }
128    }
129  }
130  return
131}
132
133// CHECK:       memref.dim %{{.*}}, %c0 : memref<?x?xf32>
134// CHECK-NEXT:  affine.for %{{.*}} = 0 to %{{.*}} step 32 {
135// CHECK-NEXT:    affine.for %{{.*}} = 0 to %{{.*}} step 32 {
136// CHECK-NEXT:      affine.for %{{.*}} = #[[$MAP:.*]](%{{.*}}) to min [[$UBMAP]](%{{.*}})[%{{.*}}] {
137// CHECK-NEXT:        affine.for %{{.*}} = #[[$MAP]](%{{.*}}) to min [[$UBMAP]](%{{.*}})[%{{.*}}] {
138// CHECK-NEXT:          affine.store %{{.*}}, %{{.*}}[%{{.*}}, %{{.*}}] : memref<?x?xf32>
139// CHECK-NEXT:          affine.for %{{.*}} = 0 to %{{.*}} {
140// CHECK-NEXT:            affine.load
141// CHECK-NEXT:            affine.load
142// CHECK-NEXT:            arith.mulf
143// CHECK-NEXT:            affine.load
144// CHECK-NEXT:            arith.addf
145// CHECK-NEXT:            affine.store
146// CHECK-NEXT:          }
147// CHECK-NEXT:        }
148// CHECK-NEXT:      }
149// CHECK-NEXT:    }
150// CHECK-NEXT:  }
151// CHECK-NEXT:  return
152
153// -----
154
155// CHECK-DAG: [[MAP0:#map[0-9]*]] = affine_map<(d0) -> (d0)>
156// CHECK-DAG: [[MAP1:#map[0-9]*]] = affine_map<()[s0, s1] -> (s0 + s1)>
157// CHECK-DAG: [[$UBMAP:#map[0-9]*]] = affine_map<(d0)[s0, s1] -> (d0 + 32, s0 + s1)>
158
159func.func @tile_using_loop_upper_bounds_in_two_symbols(%arg0: memref<?xf32>, %limit: index) {
160  %c0 = arith.constant 0 : index
161  %dim0 = memref.dim %arg0, %c0 : memref<?xf32>
162  affine.for %i0 = 0 to affine_map<()[s0, s1] -> (s0 + s1)> ()[%dim0, %limit] {
163    %v0 = affine.load %arg0[%i0] : memref<?xf32>
164  }
165  return
166}
167
168// CHECK:       memref.dim %{{.*}}, %c0 : memref<?xf32>
169// CHECK-NEXT:  affine.for %{{.*}} = 0 to [[MAP1]]()[%{{.*}}, %{{.*}}] step 32 {
170// CHECK-NEXT:    affine.for %{{.*}} = [[MAP0]](%{{.*}}) to min [[$UBMAP]](%{{.*}})[%{{.*}}, %{{.*}}] {
171// CHECK-NEXT:      affine.load
172// CHECK-NEXT:    }
173// CHECK-NEXT:  }
174
175// -----
176
177// CHECK-DAG:  #[[$ID:.*]] = affine_map<(d0) -> (d0)>
178// CHECK-DAG:  [[$UBMAP:#map[0-9]*]] = affine_map<(d0)[s0] -> (d0 + 160, s0)>
179
180func.func @tile_loop_with_non_unit_step(%arg0 : memref<50xf32>, %arg1 : index) {
181  affine.for %i = 0 to %arg1 step 5 {
182    affine.load %arg0[%i] : memref<50xf32>
183  }
184  return
185}
186
187// CHECK-LABEL: func @tile_loop_with_non_unit_step(%arg{{.*}}: memref<50xf32>, %arg{{.*}}: index)
188// CHECK:     affine.for %[[I:.*]] = 0 to %[[N:.*]] step 160 {
189// CHECK-NEXT:       affine.for %[[II:.*]] = [[$ID:.*]](%[[I]]) to min
190// [[$UBMAP]](%[[I]])[%[[N]]] step 5 {
191// CHECK-NEXT:         affine.load %arg{{.*}}[%arg{{.*}}] : memref<50xf32>
192
193// -----
194
195func.func @tile_size_larger_than_trip_count_symbolic_bound(%M: index, %N :  index) {
196  affine.for %i = affine_map<(d0) -> (d0)>(%M) to affine_map<(d0) -> (d0 + 2)>(%M) {
197    affine.for %j = affine_map<(d0) -> (d0)>(%N) to affine_map<(d0) -> (d0 + 4)>(%N) {
198      "test.foo" () : () -> ()
199    }
200  }
201  return
202}
203
204// CHECK-DAG: #[[$ID:.*]] = affine_map<(d0) -> (d0)>
205// CHECK-DAG: #[[$ID_PLUS_2:.*]] = affine_map<(d0) -> (d0 + 2)>
206// CHECK-DAG: #[[$ID_PLUS_4:.*]] = affine_map<(d0) -> (d0 + 4)>
207// CHECK: %[[M:.*]]: index, %[[N:.*]]: index
208// CHECK:      affine.for %[[I:.*]] = #[[$ID]](%[[M]]) to #[[$ID_PLUS_2]](%[[M]]) step 32
209// CHECK-NEXT:   affine.for %[[J:.*]] = #[[$ID]](%[[N]]) to #[[$ID_PLUS_4]](%[[N]]) step 32
210// CHECK-NEXT:     affine.for %arg4 = #[[$ID]](%[[I]]) to #[[$ID_PLUS_2]](%[[I]])
211// CHECK-NEXT:       affine.for %arg5 = #[[$ID]](%[[J]]) to #[[$ID_PLUS_4]](%[[J]])
212// CHECK-NEXT:         "test.foo"
213
214// -----
215
216// CHECK-LABEL: func @trip_count_one
217// SEPARATE-LABEL: func @trip_count_one
218func.func @trip_count_one(%arg0: memref<196608x1xf32>, %arg1: memref<196608x1xf32>)
219    -> memref<196608x1xf32> {
220  affine.for %i1 = 0 to 196608 {
221    affine.for %i3 = 0 to 1 {
222      %4 = affine.load %arg0[%i1, %i3] : memref<196608x1xf32>
223      affine.store %4, %arg1[%i1, %i3] : memref<196608x1xf32>
224    }
225  }
226  // CHECK: affine.load %{{.*}}[%{{.*}}, %{{.*}}] : memref<196608x1xf32>
227  return %arg1 : memref<196608x1xf32>
228}
229// To make sure SEPARATE-DAGs further below do not match with something above.
230// SEPARATE: return
231
232// -----
233
234func.func @separate_full_tile_2d(%M : index, %N : index) {
235  affine.for %i = 0 to %M {
236    affine.for %j = 0 to %N {
237      "test.foo"() : () -> ()
238    }
239  }
240  return
241}
242
243// -----
244
245#ub = affine_map<(d0)[s0] -> (d0, s0)>
246// CHECK-LABEL: func @non_hyperrectangular_loop
247func.func @non_hyperrectangular_loop() {
248  %N = arith.constant 128 : index
249  affine.for %i = 0 to %N {
250    affine.for %j = 0 to min #ub(%i)[%N] {
251      "test.foo"() : () -> ()
252    }
253 }
254  // No tiling is performed here.
255  // CHECK:      arith.constant
256  // CHECK-NEXT: affine.for
257  // CHECK-NEXT:   affine.for
258  // CHECK-NEXT:     test.foo
259  return
260}
261
262// -----
263
264// No tiling supported on loops with yield values.
265
266// CHECK-LABEL: func @yield_values
267func.func @yield_values(%init : index) {
268  %r = affine.for %i = 0 to 10 iter_args(%s = %init) -> index {
269    "test.foo"() : () -> ()
270    affine.yield %s : index
271  }
272  // No tiling here.
273  // CHECK-NEXT: affine.for {{.*}} {
274  // CHECK-NEXT:   test.foo
275  return
276}
277
278// -----
279
280// SEPARATE-DAG: #[[$SEP_COND:.*]] = affine_set<(d0, d1)[s0, s1] : (-d0 + s0 - 32 >= 0, -d1 + s1 - 32 >= 0)>
281// SEPARATE-DAG: #[[$LB:.*]] = affine_map<(d0) -> (d0)>
282// SEPARATE-DAG: #[[$FULL_TILE_UB:.*]] = affine_map<(d0) -> (d0 + 32)>
283// SEPARATE-DAG: #[[$PART_TILE_UB:.*]] = affine_map<(d0)[s0] -> (d0 + 32, s0)>
284
285// SEPARATE-LABEL: func @separate_full_tile_2d(
286// SEPARATE: %[[M:.*]]: index, %[[N:.*]]: index
287
288// SEPARATE:       affine.for %[[I:.*]] =
289// SEPARATE-NEXT:    affine.for %[[J:.*]] =
290// SEPARATE-NEXT:      affine.if #[[$SEP_COND]](%arg2, %arg3)[%arg0, %arg1] {
291// SEPARATE-NEXT:        affine.for %{{.*}} = #[[$LB]](%[[I]]) to #[[$FULL_TILE_UB]](%[[I]]) {
292// SEPARATE-NEXT:          affine.for %{{.*}} = #[[$LB]](%[[J]]) to #[[$FULL_TILE_UB]](%[[J]]) {
293// SEPARATE-NEXT:           "test.foo"
294// SEPARATE-NEXT:          }
295// SEPARATE-NEXT:        }
296// SEPARATE-NEXT:      } else {
297// SEPARATE-NEXT:        affine.for %{{.*}} = #[[$LB]](%[[I]]) to min #[[$PART_TILE_UB]](%[[I]])[%[[M]]] {
298// SEPARATE-NEXT:          affine.for %{{.*}} = #[[$LB]](%[[J]]) to min #[[$PART_TILE_UB]](%[[J]])[%[[N]]] {
299// SEPARATE-NEXT:           "test.foo"
300// SEPARATE-NEXT:          }
301// SEPARATE-NEXT:        }
302// SEPARATE-NEXT:      }
303// SEPARATE-NEXT:    }
304// SEPARATE-NEXT:  }
305// SEPARATE-NEXT:  return
306
307// -----
308
309func.func @separate_full_tile_1d_max_min(%M : index, %N : index, %P : index, %Q : index) {
310  affine.for %i0 = max affine_map<(d0, d1) -> (d0, d1)>  (%M, %N) to min affine_map< (d0, d1) -> (d0, d1)> (%P, %Q) {
311  }
312  return
313}
314
315// SEPARATE-DAG: #[[$SEP_COND:.*]] = affine_set<(d0)[s0, s1] : (-d0 + s0 - 32 >= 0, -d0 + s1 - 32 >= 0)>
316// SEPARATE-DAG: #[[TILE_LB:.*]] = affine_map<(d0) -> (d0)>
317// SEPARATE-DAG: #[[$FULL_TILE_UB:.*]] = affine_map<(d0) -> (d0 + 32)>
318// SEPARATE-DAG: #[[PARTIAL_TILE_UB:.*]] = affine_map<(d0, d1, d2) -> (d2 + 32, d0, d1)>
319
320// SEPARATE:         affine.for %arg4
321// SEPARATE-NEXT:      affine.if #[[$SEP_COND]](%arg4)[%arg2, %arg3] {
322// SEPARATE-NEXT:        affine.for %arg5 = #[[TILE_LB]](%arg4) to #[[$FULL_TILE_UB]](%arg4) {
323// SEPARATE-NEXT:        }
324// SEPARATE-NEXT:      } else {
325// SEPARATE-NEXT:        affine.for %arg5 = #[[TILE_LB]](%arg4) to min #[[PARTIAL_TILE_UB]](%arg2, %arg3, %arg4) {
326// SEPARATE-NEXT:        }
327// SEPARATE-NEXT:      }
328// SEPARATE-NEXT:    }
329