|
173
|
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-4: [6, 6, 6, 6]
|
|
|
4 module 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 {max_compute_workgroup_invocations = 128 : i32,
|
|
|
9 max_compute_workgroup_size = dense<[128, 128, 64]> : vector<3xi32>}>
|
|
|
10 } {
|
|
|
11 gpu.module @kernels {
|
|
|
12 gpu.func @kernel_mul(%arg0 : memref<4x4xf32>, %arg1 : memref<4x4xf32>, %arg2 : memref<4x4xf32>)
|
|
|
13 kernel attributes { spv.entry_point_abi = {local_size = dense<[1, 1, 1]>: vector<3xi32> }} {
|
|
|
14 %x = "gpu.block_id"() {dimension = "x"} : () -> index
|
|
|
15 %y = "gpu.block_id"() {dimension = "y"} : () -> index
|
|
|
16 %1 = load %arg0[%x, %y] : memref<4x4xf32>
|
|
|
17 %2 = load %arg1[%x, %y] : memref<4x4xf32>
|
|
|
18 %3 = mulf %1, %2 : f32
|
|
|
19 store %3, %arg2[%x, %y] : memref<4x4xf32>
|
|
|
20 gpu.return
|
|
|
21 }
|
|
|
22 }
|
|
|
23
|
|
|
24 func @main() {
|
|
|
25 %arg0 = alloc() : memref<4x4xf32>
|
|
|
26 %arg1 = alloc() : memref<4x4xf32>
|
|
|
27 %arg2 = alloc() : memref<4x4xf32>
|
|
|
28 %0 = constant 0 : i32
|
|
|
29 %1 = constant 1 : i32
|
|
|
30 %2 = constant 2 : i32
|
|
|
31 %value0 = constant 0.0 : f32
|
|
|
32 %value1 = constant 2.0 : f32
|
|
|
33 %value2 = constant 3.0 : f32
|
|
|
34 %arg3 = memref_cast %arg0 : memref<4x4xf32> to memref<?x?xf32>
|
|
|
35 %arg4 = memref_cast %arg1 : memref<4x4xf32> to memref<?x?xf32>
|
|
|
36 %arg5 = memref_cast %arg2 : memref<4x4xf32> to memref<?x?xf32>
|
|
|
37 call @fillResource2DFloat(%arg3, %value1) : (memref<?x?xf32>, f32) -> ()
|
|
|
38 call @fillResource2DFloat(%arg4, %value2) : (memref<?x?xf32>, f32) -> ()
|
|
|
39 call @fillResource2DFloat(%arg5, %value0) : (memref<?x?xf32>, f32) -> ()
|
|
|
40
|
|
|
41 %cst1 = constant 1 : index
|
|
|
42 %cst4 = constant 4 : index
|
|
|
43 "gpu.launch_func"(%cst4, %cst4, %cst1, %cst1, %cst1, %cst1, %arg0, %arg1, %arg2) { kernel = @kernels::@kernel_mul }
|
|
|
44 : (index, index, index, index, index, index, memref<4x4xf32>, memref<4x4xf32>, memref<4x4xf32>) -> ()
|
|
|
45 %arg6 = memref_cast %arg5 : memref<?x?xf32> to memref<*xf32>
|
|
|
46 call @print_memref_f32(%arg6) : (memref<*xf32>) -> ()
|
|
|
47 return
|
|
|
48 }
|
|
|
49 func @fillResource2DFloat(%0 : memref<?x?xf32>, %1 : f32)
|
|
|
50 func @print_memref_f32(%ptr : memref<*xf32>)
|
|
|
51 }
|
|
|
52
|
