review comments

- clarify description comment
- add integration test for the vulkan runner

Author: inbelic

mlir/lib/Dialect/SPIRV/Transforms/SPIRVWebGPUTransforms.cpp

@@ -190,8 +190,7 @@ struct ExpandAddCarryPattern final : OpRewritePattern<IAddCarryOp> {
     Value zero =
         rewriter.create<ConstantOp>(loc, argTy, getScalarOrSplatAttr(argTy, 0));
 
-    // Emulate 64-bit unsigned addition by allowing our addition to overflow,
-    // and then set the carry accordingly.
+    // Calculate the carry by checking if the addition resulted in an overflow.
     Value out = rewriter.create<IAddOp>(loc, lhs, rhs);
     Value cmp = rewriter.create<ULessThanOp>(loc, out, lhs);
     Value carry = rewriter.create<SelectOp>(loc, cmp, one, zero);

mlir/test/mlir-vulkan-runner/iaddcarry_extended.mlir

@@ -0,0 +1,74 @@
+// Make sure that addition with carry produces expected results
+// with and without expansion to primitive add/cmp ops for WebGPU.
+
+// RUN: mlir-vulkan-runner %s \
+// RUN:  --shared-libs=%vulkan-runtime-wrappers,%mlir_runner_utils \
+// RUN:  --entry-point-result=void | FileCheck %s
+
+// RUN: mlir-vulkan-runner %s --vulkan-runner-spirv-webgpu-prepare \
+// RUN:  --shared-libs=%vulkan-runtime-wrappers,%mlir_runner_utils \
+// RUN:  --entry-point-result=void | FileCheck %s
+
+// CHECK: [0, 1, 0, 42]
+// CHECK: [0, 0, 1, 1]
+module attributes {
+  gpu.container_module,
+  spirv.target_env = #spirv.target_env<
+    #spirv.vce<v1.4, [Shader], [SPV_KHR_storage_buffer_storage_class]>, #spirv.resource_limits<>>
+} {
+  gpu.module @kernels {
+    gpu.func @kernel_add(%arg0 : memref<4xi32>, %arg1 : memref<4xi32>, %arg2 : memref<4xi32>, %arg3 : memref<4xi8>)
+      kernel attributes { spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [1, 1, 1]>} {
+      %0 = gpu.block_id x
+      %lhs = memref.load %arg0[%0] : memref<4xi32>
+      %rhs = memref.load %arg1[%0] : memref<4xi32>
+      %sum, %carry = arith.addui_extended %lhs, %rhs : i32
+
+      // We will convert to i8 as this is the smallest value we can use for
+      // fill/print in the runner.
+      %carry_i8 = arith.extui %carry : i1 to i8
+
+      memref.store %sum, %arg2[%0] : memref<4xi32>
+      memref.store %carry_i8, %arg3[%0] : memref<4xi8>
+      gpu.return
+    }
+  }
+
+  func.func @main() {
+    %buf0 = memref.alloc() : memref<4xi32>
+    %buf1 = memref.alloc() : memref<4xi32>
+    %buf2 = memref.alloc() : memref<4xi32>
+    %buf3 = memref.alloc() : memref<4xi8>
+    %i32_0 = arith.constant 0 : i32
+    %i8_0 = arith.constant 0 : i32
+
+    // Initialize output buffers.
+    %buf4 = memref.cast %buf2 : memref<4xi32> to memref<?xi32>
+    %buf5 = memref.cast %buf3 : memref<4xi8> to memref<?xi8>
+    call @fillResource1DInt(%buf4, %i32_0) : (memref<?xi32>, i32) -> ()
+    call @fillResource1DInt8(%buf5, %i8_0) : (memref<?xi32>, i32) -> ()
+
+    %idx_0 = arith.constant 0 : index
+    %idx_1 = arith.constant 1 : index
+    %idx_4 = arith.constant 4 : index
+
+    // Initialize input buffers.
+    %lhs_vals = arith.constant dense<[0, 0, -2147483647, 43]> : vector<4xi32>
+    %rhs_vals = arith.constant dense<[0, 1, 1, -2147483647]> : vector<4xi32>
+    vector.store %lhs_vals, %buf0[%idx_0] : memref<4xi32>, vector<4xi32>
+    vector.store %rhs_vals, %buf1[%idx_0] : memref<4xi32>, vector<4xi32>
+
+    gpu.launch_func @kernels::@kernel_add
+        blocks in (%idx_4, %idx_1, %idx_1) threads in (%idx_1, %idx_1, %idx_1)
+        args(%buf0 : memref<4xi32>, %buf1 : memref<4xi32>, %buf2 : memref<4xi32>, %buf3 : memref<4xi8>)
+    %buf_sum = memref.cast %buf4 : memref<?xi32> to memref<*xi32>
+    %buf_carry = memref.cast %buf5 : memref<?xi32> to memref<*xi8>
+    call @printMemrefI32(%buf_sum) : (memref<*xi32>) -> ()
+    call @printMemrefI32(%buf_carry) : (memref<*xi8>) -> ()
+    return
+  }
+  func.func private @fillResource1DInt8(%0 : memref<?xi8>, %1 : i8)
+  func.func private @fillResource1DInt(%0 : memref<?xi32>, %1 : i32)
+  func.func private @printMemrefI8(%ptr : memref<*xi8>)
+  func.func private @printMemrefI32(%ptr : memref<*xi32>)
+}