1// RUN: mlir-opt -test-gpu-greedy-parallel-loop-mapping -split-input-file %s | FileCheck %s
2
3func.func @parallel_loop(%arg0 : index, %arg1 : index, %arg2 : index,
4                    %arg3 : index) {
5  %zero = arith.constant 0 : index
6  %one = arith.constant 1 : index
7  %four = arith.constant 4 : index
8  scf.parallel (%i0, %i1) = (%arg0, %arg1) to (%arg2, %arg3)
9                                          step (%four, %four)  {
10    scf.parallel (%si0, %si1) = (%zero, %zero) to (%four, %four)
11                                            step (%one, %one)  {
12    }
13  }
14  return
15}
16
17// CHECK: #[[$MAP:.*]] = affine_map<(d0) -> (d0)>
18// CHECK-LABEL:   func @parallel_loop(
19// CHECK:           scf.parallel
20// CHECK:             scf.parallel
21// CHECK:      {mapping = [{bound = #[[$MAP]], map = #[[$MAP]], processor = 3 : i64},
22// CHECK-SAME:             {bound = #[[$MAP]], map = #[[$MAP]], processor = 4 : i64}]}
23// CHECK:      {mapping = [{bound = #[[$MAP]], map = #[[$MAP]], processor = 0 : i64},
24// CHECK-SAME:             {bound = #[[$MAP]], map = #[[$MAP]], processor = 1 : i64}]}
25// CHECK-NOT: mapping
26
27// -----
28
29func.func @parallel_loop_4d(%arg0 : index, %arg1 : index, %arg2 : index,
30                       %arg3 : index) {
31  %zero = arith.constant 0 : index
32  %one = arith.constant 1 : index
33  %four = arith.constant 4 : index
34  scf.parallel (%i0, %i1, %i2, %i3) = (%zero, %zero, %zero, %zero) to (%arg0, %arg1, %arg2, %arg3)
35                                       step (%four, %four, %four, %four)  {
36    scf.parallel (%si0, %si1, %si2, %si3) = (%zero, %zero, %zero, %zero) to (%four, %four, %four, %four)
37                                             step (%one, %one, %one, %one)  {
38      scf.parallel (%ti0, %ti1, %ti2, %ti3) = (%zero, %zero, %zero, %zero) to (%four, %four, %four, %four)
39                                               step (%one, %one, %one, %one)  {
40      }
41    }
42  }
43  return
44}
45
46// CHECK: #[[$MAP:.*]] = affine_map<(d0) -> (d0)>
47// CHECK-LABEL:   func @parallel_loop_4d(
48// CHECK:           scf.parallel
49// CHECK:             scf.parallel
50// CHECK:               scf.parallel
51// CHECK:      {mapping = [{bound = #[[$MAP]], map = #[[$MAP]], processor = 6 : i64},
52// CHECK-SAME:             {bound = #[[$MAP]], map = #[[$MAP]], processor = 6 : i64},
53// CHECK-SAME:             {bound = #[[$MAP]], map = #[[$MAP]], processor = 6 : i64},
54// CHECK-SAME:             {bound = #[[$MAP]], map = #[[$MAP]], processor = 6 : i64}]}
55// CHECK:      {mapping = [{bound = #[[$MAP]], map = #[[$MAP]], processor = 3 : i64},
56// CHECK-SAME:             {bound = #[[$MAP]], map = #[[$MAP]], processor = 4 : i64},
57// CHECK-SAME:             {bound = #[[$MAP]], map = #[[$MAP]], processor = 5 : i64},
58// CHECK-SAME:             {bound = #[[$MAP]], map = #[[$MAP]], processor = 6 : i64}]}
59// CHECK:      {mapping = [{bound = #[[$MAP]], map = #[[$MAP]], processor = 0 : i64},
60// CHECK-SAME:             {bound = #[[$MAP]], map = #[[$MAP]], processor = 1 : i64},
61// CHECK-SAME:             {bound = #[[$MAP]], map = #[[$MAP]], processor = 2 : i64},
62// CHECK-SAME:             {bound = #[[$MAP]], map = #[[$MAP]], processor = 6 : i64}]}
63// CHECK-NOT: mapping
64