1 // Make sure that addition with carry produces expected results
2 // with and without expansion to primitive add/cmp ops for WebGPU.
4 // RUN: mlir-vulkan-runner %s \
5 // RUN: --shared-libs=%vulkan-runtime-wrappers,%mlir_runner_utils \
6 // RUN: --entry-point-result=void | FileCheck %s
8 // RUN: mlir-vulkan-runner %s --vulkan-runner-spirv-webgpu-prepare \
9 // RUN: --shared-libs=%vulkan-runtime-wrappers,%mlir_runner_utils \
10 // RUN: --entry-point-result=void | FileCheck %s
12 // CHECK: [0, 42, 0, 42]
13 // CHECK: [1, 0, 1, 1]
16 spirv.target_env = #spirv.target_env<
17 #spirv.vce<v1.4, [Shader], [SPV_KHR_storage_buffer_storage_class]>, #spirv.resource_limits<>>
20 gpu.func @kernel_add(%arg0 : memref<4xi32>, %arg1 : memref<4xi32>, %arg2 : memref<4xi32>, %arg3 : memref<4xi32>)
21 kernel attributes { spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [1, 1, 1]>} {
23 %lhs = memref.load %arg0[%0] : memref<4xi32>
24 %rhs = memref.load %arg1[%0] : memref<4xi32>
25 %sum, %carry = arith.addui_extended %lhs, %rhs : i32, i1
27 %carry_i32 = arith.extui %carry : i1 to i32
29 memref.store %sum, %arg2[%0] : memref<4xi32> memref.store %carry_i32, %arg3[%0] : memref<4xi32>
35 %buf0 = memref.alloc() : memref<4xi32>
36 %buf1 = memref.alloc() : memref<4xi32>
37 %buf2 = memref.alloc() : memref<4xi32>
38 %buf3 = memref.alloc() : memref<4xi32>
39 %i32_0 = arith.constant 0 : i32
41 // Initialize output buffers.
42 %buf4 = memref.cast %buf2 : memref<4xi32> to memref<?xi32>
43 %buf5 = memref.cast %buf3 : memref<4xi32> to memref<?xi32>
44 call @fillResource1DInt(%buf4, %i32_0) : (memref<?xi32>, i32) -> ()
45 call @fillResource1DInt(%buf5, %i32_0) : (memref<?xi32>, i32) -> ()
47 %idx_0 = arith.constant 0 : index
48 %idx_1 = arith.constant 1 : index
49 %idx_4 = arith.constant 4 : index
51 // Initialize input buffers.
52 %lhs_vals = arith.constant dense<[-1, 24, 4294967295, 43]> : vector<4xi32>
53 %rhs_vals = arith.constant dense<[1, 18, 1, 4294967295]> : vector<4xi32>
54 vector.store %lhs_vals, %buf0[%idx_0] : memref<4xi32>, vector<4xi32>
55 vector.store %rhs_vals, %buf1[%idx_0] : memref<4xi32>, vector<4xi32>
57 gpu.launch_func @kernels::@kernel_add
58 blocks in (%idx_4, %idx_1, %idx_1) threads in (%idx_1, %idx_1, %idx_1)
59 args(%buf0 : memref<4xi32>, %buf1 : memref<4xi32>, %buf2 : memref<4xi32>, %buf3 : memref<4xi32>)
60 %buf_sum = memref.cast %buf4 : memref<?xi32> to memref<*xi32>
61 %buf_carry = memref.cast %buf5 : memref<?xi32> to memref<*xi32>
62 call @printMemrefI32(%buf_sum) : (memref<*xi32>) -> ()
63 call @printMemrefI32(%buf_carry) : (memref<*xi32>) -> ()
66 func.func private @fillResource1DInt(%0 : memref<?xi32>, %1 : i32)
67 func.func private @printMemrefI32(%ptr : memref<*xi32>)