41 lines · plain
1// RUN: mlir-opt %s --gpu-to-llvm | FileCheck %s2 3module attributes {gpu.container_module} {4 // CHECK-LABEL: llvm.func @main5 // CHECK-SAME: %[[size:.*]]: i646 func.func @main(%size : index) {7 // CHECK: %[[stream:.*]] = llvm.call @mgpuStreamCreate()8 %0 = gpu.wait async9 // CHECK: %[[gep:.*]] = llvm.getelementptr {{.*}}[%[[size]]]10 // CHECK: %[[size_bytes:.*]] = llvm.ptrtoint %[[gep]]11 // CHECK: %[[isHostShared:.*]] = llvm.mlir.constant 12 // CHECK: llvm.call @mgpuMemAlloc(%[[size_bytes]], %[[stream]], %[[isHostShared]])13 %1, %2 = gpu.alloc async [%0] (%size) : memref<?xf32>14 // CHECK: %[[float_ptr:.*]] = llvm.extractvalue {{.*}}[0]15 // CHECK: llvm.call @mgpuMemFree(%[[float_ptr]], %[[stream]])16 %3 = gpu.dealloc async [%2] %1 : memref<?xf32>17 // CHECK: llvm.call @mgpuStreamSynchronize(%[[stream]])18 // CHECK: llvm.call @mgpuStreamDestroy(%[[stream]])19 gpu.wait [%3]20 return21 }22 23 // CHECK-LABEL: llvm.func @alloc_sync24 // CHECK-SAME: %[[size:.*]]: i6425 func.func @alloc_sync(%size : index) {26 // CHECK: %[[gep:.*]] = llvm.getelementptr {{.*}}[%[[size]]]27 // CHECK: %[[size_bytes:.*]] = llvm.ptrtoint %[[gep]]28 // CHECK: %[[nullptr:.*]] = llvm.mlir.zero29 // CHECK: %[[isHostShared:.*]] = llvm.mlir.constant 30 // CHECK: llvm.call @mgpuMemAlloc(%[[size_bytes]], %[[nullptr]], %[[isHostShared]])31 %0 = gpu.alloc host_shared (%size) : memref<?xf32>32 // CHECK: %[[stream:.*]] = llvm.call @mgpuStreamCreate()33 %1 = gpu.wait async34 %2 = gpu.dealloc async [%1] %0 : memref<?xf32>35 // CHECK: llvm.call @mgpuStreamSynchronize(%[[stream]])36 // CHECK: llvm.call @mgpuStreamDestroy(%[[stream]])37 gpu.wait [%2]38 return39 }40}41