1// RUN: mlir-opt %s \ 2// RUN: | mlir-opt -convert-scf-to-cf \ 3// RUN: | mlir-opt -gpu-kernel-outlining \ 4// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-rocdl{use-bare-ptr-memref-call-conv=true}),rocdl-attach-target{chip=%chip})' \ 5// RUN: | mlir-opt -gpu-to-llvm=use-bare-pointers-for-kernels=true -reconcile-unrealized-casts -gpu-module-to-binary \ 6// RUN: | mlir-runner \ 7// RUN: --shared-libs=%mlir_rocm_runtime \ 8// RUN: --shared-libs=%mlir_runner_utils \ 9// RUN: --entry-point-result=void \ 10// RUN: | FileCheck %s 11 12func.func @vecadd(%arg0 : memref<5xf32>, %arg1 : memref<5xf32>, %arg2 : memref<5xf32>) { 13 %c0 = arith.constant 0 : index 14 %c1 = arith.constant 1 : index 15 %block_dim = arith.constant 5 : index 16 gpu.launch blocks(%bx, %by, %bz) in (%grid_x = %c1, %grid_y = %c1, %grid_z = %c1) 17 threads(%tx, %ty, %tz) in (%block_x = %block_dim, %block_y = %c1, %block_z = %c1) { 18 %a = memref.load %arg0[%tx] : memref<5xf32> 19 %b = memref.load %arg1[%tx] : memref<5xf32> 20 %c = arith.addf %a, %b : f32 21 memref.store %c, %arg2[%tx] : memref<5xf32> 22 gpu.terminator 23 } 24 return 25} 26 27// CHECK: [2.46, 2.46, 2.46, 2.46, 2.46] 28func.func @main() { 29 %c0 = arith.constant 0 : index 30 %c1 = arith.constant 1 : index 31 %c5 = arith.constant 5 : index 32 %cf1dot23 = arith.constant 1.23 : f32 33 %0 = memref.alloc() : memref<5xf32> 34 %1 = memref.alloc() : memref<5xf32> 35 %2 = memref.alloc() : memref<5xf32> 36 %3 = memref.cast %0 : memref<5xf32> to memref<?xf32> 37 %4 = memref.cast %1 : memref<5xf32> to memref<?xf32> 38 %5 = memref.cast %2 : memref<5xf32> to memref<?xf32> 39 scf.for %i = %c0 to %c5 step %c1 { 40 memref.store %cf1dot23, %3[%i] : memref<?xf32> 41 memref.store %cf1dot23, %4[%i] : memref<?xf32> 42 } 43 %6 = memref.cast %3 : memref<?xf32> to memref<*xf32> 44 %7 = memref.cast %4 : memref<?xf32> to memref<*xf32> 45 %8 = memref.cast %5 : memref<?xf32> to memref<*xf32> 46 gpu.host_register %6 : memref<*xf32> 47 gpu.host_register %7 : memref<*xf32> 48 gpu.host_register %8 : memref<*xf32> 49 %9 = call @mgpuMemGetDeviceMemRef1dFloat(%3) : (memref<?xf32>) -> (memref<?xf32>) 50 %10 = call @mgpuMemGetDeviceMemRef1dFloat(%4) : (memref<?xf32>) -> (memref<?xf32>) 51 %11 = call @mgpuMemGetDeviceMemRef1dFloat(%5) : (memref<?xf32>) -> (memref<?xf32>) 52 %12 = memref.cast %9 : memref<?xf32> to memref<5xf32> 53 %13 = memref.cast %10 : memref<?xf32> to memref<5xf32> 54 %14 = memref.cast %11 : memref<?xf32> to memref<5xf32> 55 56 call @vecadd(%12, %13, %14) : (memref<5xf32>, memref<5xf32>, memref<5xf32>) -> () 57 call @printMemrefF32(%8) : (memref<*xf32>) -> () 58 return 59} 60 61func.func private @mgpuMemGetDeviceMemRef1dFloat(%ptr : memref<?xf32>) -> (memref<?xf32>) 62func.func private @printMemrefF32(%ptr : memref<*xf32>) 63