1// RUN: mlir-opt %s \ 2// RUN: | mlir-opt -gpu-kernel-outlining \ 3// RUN: | mlir-opt -pass-pipeline='builtin.module(gpu.module(strip-debuginfo,convert-gpu-to-nvvm),nvvm-attach-target)' \ 4// RUN: | mlir-opt -gpu-async-region -gpu-to-llvm -reconcile-unrealized-casts -gpu-module-to-binary="format=%gpu_compilation_format" \ 5// RUN: | mlir-opt -async-to-async-runtime -async-runtime-ref-counting \ 6// RUN: | mlir-opt -convert-async-to-llvm -convert-func-to-llvm -convert-arith-to-llvm -convert-cf-to-llvm -reconcile-unrealized-casts \ 7// RUN: | mlir-runner \ 8// RUN: --shared-libs=%mlir_cuda_runtime \ 9// RUN: --shared-libs=%mlir_async_runtime \ 10// RUN: --shared-libs=%mlir_runner_utils \ 11// RUN: --entry-point-result=void -O0 \ 12// RUN: | FileCheck %s 13 14func.func @main() { 15 %c0 = arith.constant 0 : index 16 %c1 = arith.constant 1 : index 17 %count = arith.constant 2 : index 18 19 // initialize h0 on host 20 %h0 = memref.alloc(%count) : memref<?xi32> 21 %h0_unranked = memref.cast %h0 : memref<?xi32> to memref<*xi32> 22 gpu.host_register %h0_unranked : memref<*xi32> 23 24 %v0 = arith.constant 42 : i32 25 memref.store %v0, %h0[%c0] : memref<?xi32> 26 memref.store %v0, %h0[%c1] : memref<?xi32> 27 28 // copy h0 to b0 on device. 29 %t0, %f0 = async.execute () -> !async.value<memref<?xi32>> { 30 %b0 = gpu.alloc(%count) : memref<?xi32> 31 gpu.memcpy %b0, %h0 : memref<?xi32>, memref<?xi32> 32 async.yield %b0 : memref<?xi32> 33 } 34 35 // copy h0 to b1 and b2 (fork) 36 %t1, %f1 = async.execute [%t0] ( 37 %f0 as %b0 : !async.value<memref<?xi32>> 38 ) -> !async.value<memref<?xi32>> { 39 %b1 = gpu.alloc(%count) : memref<?xi32> 40 gpu.memcpy %b1, %b0 : memref<?xi32>, memref<?xi32> 41 async.yield %b1 : memref<?xi32> 42 } 43 %t2, %f2 = async.execute [%t0] ( 44 %f0 as %b0 : !async.value<memref<?xi32>> 45 ) -> !async.value<memref<?xi32>> { 46 %b2 = gpu.alloc(%count) : memref<?xi32> 47 gpu.memcpy %b2, %b0 : memref<?xi32>, memref<?xi32> 48 async.yield %b2 : memref<?xi32> 49 } 50 51 // h0 = b1 + b2 (join). 52 %t3 = async.execute [%t1, %t2] ( 53 %f1 as %b1 : !async.value<memref<?xi32>>, 54 %f2 as %b2 : !async.value<memref<?xi32>> 55 ) { 56 gpu.launch blocks(%bx, %by, %bz) in (%grid_x = %c1, %grid_y = %c1, %grid_z = %c1) 57 threads(%tx, %ty, %tz) in (%block_x = %count, %block_y = %c1, %block_z = %c1) { 58 %v1 = memref.load %b1[%tx] : memref<?xi32> 59 %v2 = memref.load %b2[%tx] : memref<?xi32> 60 %sum = arith.addi %v1, %v2 : i32 61 memref.store %sum, %h0[%tx] : memref<?xi32> 62 gpu.terminator 63 } 64 async.yield 65 } 66 67 async.await %t3 : !async.token 68 // CHECK: [84, 84] 69 call @printMemrefI32(%h0_unranked) : (memref<*xi32>) -> () 70 return 71} 72 73func.func private @printMemrefI32(memref<*xi32>) 74