1// RUN: mlir-cuda-runner %s --shared-libs=%cuda_wrapper_library_dir/libcuda-runtime-wrappers%shlibext,%linalg_test_lib_dir/libmlir_runner_utils%shlibext --entry-point-result=void | FileCheck %s
2
3func @main() {
4  %data = alloc() : memref<2x6xf32>
5  %sum = alloc() : memref<2xf32>
6  %mul = alloc() : memref<2xf32>
7  %cst0 = constant 0.0 : f32
8  %cst1 = constant 1.0 : f32
9  %cst2 = constant 2.0 : f32
10  %cst4 = constant 4.0 : f32
11  %cst8 = constant 8.0 : f32
12  %cst16 = constant 16.0 : f32
13
14  %cst3 = constant 3.0 : f32
15  %cst6 = constant 6.0 : f32
16  %cst7 = constant 7.0 : f32
17  %cst10 = constant 10.0 : f32
18  %cst11 = constant 11.0 : f32
19
20  %c0 = constant 0 : index
21  %c1 = constant 1 : index
22  %c2 = constant 2 : index
23  %c3 = constant 3 : index
24  %c4 = constant 4 : index
25  %c5 = constant 5 : index
26  %c6 = constant 6 : index
27
28  %cast_data = memref_cast %data : memref<2x6xf32> to memref<*xf32>
29  gpu.host_register %cast_data : memref<*xf32>
30  %cast_sum = memref_cast %sum : memref<2xf32> to memref<*xf32>
31  gpu.host_register %cast_sum : memref<*xf32>
32  %cast_mul = memref_cast %mul : memref<2xf32> to memref<*xf32>
33  gpu.host_register %cast_mul : memref<*xf32>
34
35  store %cst0, %data[%c0, %c0] : memref<2x6xf32>
36  store %cst1, %data[%c0, %c1] : memref<2x6xf32>
37  store %cst2, %data[%c0, %c2] : memref<2x6xf32>
38  store %cst4, %data[%c0, %c3] : memref<2x6xf32>
39  store %cst8, %data[%c0, %c4] : memref<2x6xf32>
40  store %cst16, %data[%c0, %c5] : memref<2x6xf32>
41
42  store %cst2, %data[%c1, %c0] : memref<2x6xf32>
43  store %cst3, %data[%c1, %c1] : memref<2x6xf32>
44  store %cst6, %data[%c1, %c2] : memref<2x6xf32>
45  store %cst7, %data[%c1, %c3] : memref<2x6xf32>
46  store %cst10, %data[%c1, %c4] : memref<2x6xf32>
47  store %cst11, %data[%c1, %c5] : memref<2x6xf32>
48
49  // ADD + MUL
50  gpu.launch blocks(%bx, %by, %bz) in (%grid_x = %c2, %grid_y = %c1, %grid_z = %c1)
51             threads(%tx, %ty, %tz) in (%block_x = %c6, %block_y = %c1, %block_z = %c1) {
52    %val = load %data[%bx, %tx] : memref<2x6xf32>
53    %reduced0 = "gpu.all_reduce"(%val) ({}) { op = "add" } : (f32) -> (f32)
54    store %reduced0, %sum[%bx] : memref<2xf32>
55    %reduced1 = "gpu.all_reduce"(%val) ({}) { op = "mul" } : (f32) -> (f32)
56    store %reduced1, %mul[%bx] : memref<2xf32>
57    gpu.terminator
58  }
59
60  call @print_memref_f32(%cast_sum) : (memref<*xf32>) -> ()
61  // CHECK: [31, 39]
62
63  call @print_memref_f32(%cast_mul) : (memref<*xf32>) -> ()
64  // CHECK: [0, 27720]
65
66  return
67}
68
69func private @print_memref_f32(memref<*xf32>)
70