1// RUN: mlir-vulkan-runner %s --shared-libs=%vulkan_wrapper_library_dir/libvulkan-runtime-wrappers%shlibext,%linalg_test_lib_dir/libmlir_runner_utils%shlibext --entry-point-result=void | FileCheck %s
2
3// CHECK-COUNT-32: [2.2, 2.2, 2.2, 2.2]
4module attributes {
5  gpu.container_module,
6  spv.target_env = #spv.target_env<
7    #spv.vce<v1.0, [Shader], [SPV_KHR_storage_buffer_storage_class]>,
8    #spv.resource_limits<>>
9} {
10  gpu.module @kernels {
11    gpu.func @kernel_sub(%arg0 : memref<8x4x4xf32>, %arg1 : memref<4x4xf32>, %arg2 : memref<8x4x4xf32>)
12      kernel attributes { spv.entry_point_abi = #spv.entry_point_abi<local_size = dense<[1, 1, 1]>: vector<3xi32>>} {
13      %x = gpu.block_id x
14      %y = gpu.block_id y
15      %z = gpu.block_id z
16      %1 = memref.load %arg0[%x, %y, %z] : memref<8x4x4xf32>
17      %2 = memref.load %arg1[%y, %z] : memref<4x4xf32>
18      %3 = arith.subf %1, %2 : f32
19      memref.store %3, %arg2[%x, %y, %z] : memref<8x4x4xf32>
20      gpu.return
21    }
22  }
23
24  func.func @main() {
25    %arg0 = memref.alloc() : memref<8x4x4xf32>
26    %arg1 = memref.alloc() : memref<4x4xf32>
27    %arg2 = memref.alloc() : memref<8x4x4xf32>
28    %0 = arith.constant 0 : i32
29    %1 = arith.constant 1 : i32
30    %2 = arith.constant 2 : i32
31    %value0 = arith.constant 0.0 : f32
32    %value1 = arith.constant 3.3 : f32
33    %value2 = arith.constant 1.1 : f32
34    %arg3 = memref.cast %arg0 : memref<8x4x4xf32> to memref<?x?x?xf32>
35    %arg4 = memref.cast %arg1 : memref<4x4xf32> to memref<?x?xf32>
36    %arg5 = memref.cast %arg2 : memref<8x4x4xf32> to memref<?x?x?xf32>
37    call @fillResource3DFloat(%arg3, %value1) : (memref<?x?x?xf32>, f32) -> ()
38    call @fillResource2DFloat(%arg4, %value2) : (memref<?x?xf32>, f32) -> ()
39    call @fillResource3DFloat(%arg5, %value0) : (memref<?x?x?xf32>, f32) -> ()
40
41    %cst1 = arith.constant 1 : index
42    %cst4 = arith.constant 4 : index
43    %cst8 = arith.constant 8 : index
44    gpu.launch_func @kernels::@kernel_sub
45        blocks in (%cst8, %cst4, %cst4) threads in (%cst1, %cst1, %cst1)
46        args(%arg0 : memref<8x4x4xf32>, %arg1 : memref<4x4xf32>, %arg2 : memref<8x4x4xf32>)
47    %arg6 = memref.cast %arg5 : memref<?x?x?xf32> to memref<*xf32>
48    call @printMemrefF32(%arg6) : (memref<*xf32>) -> ()
49    return
50  }
51  func.func private @fillResource2DFloat(%0 : memref<?x?xf32>, %1 : f32)
52  func.func private @fillResource3DFloat(%0 : memref<?x?x?xf32>, %1 : f32)
53  func.func private @printMemrefF32(%ptr : memref<*xf32>)
54}
55