81 lines · plain
1// RUN: mlir-opt -convert-gpu-to-spirv %s -o - | FileCheck %s2 3module attributes {4 gpu.container_module,5 spirv.target_env = #spirv.target_env<6 #spirv.vce<v1.0, [Shader], [SPV_KHR_storage_buffer_storage_class]>, #spirv.resource_limits<>>7} {8 func.func @load_store(%arg0: memref<12x4xf32, #spirv.storage_class<StorageBuffer>>, %arg1: memref<12x4xf32, #spirv.storage_class<StorageBuffer>>, %arg2: memref<12x4xf32, #spirv.storage_class<StorageBuffer>>) {9 %c0 = arith.constant 0 : index10 %c12 = arith.constant 12 : index11 %0 = arith.subi %c12, %c0 : index12 %c1 = arith.constant 1 : index13 %c0_0 = arith.constant 0 : index14 %c4 = arith.constant 4 : index15 %1 = arith.subi %c4, %c0_0 : index16 %c1_1 = arith.constant 1 : index17 %c1_2 = arith.constant 1 : index18 gpu.launch_func @kernels::@load_store_kernel19 blocks in (%0, %c1_2, %c1_2) threads in (%1, %c1_2, %c1_2)20 args(%arg0 : memref<12x4xf32, #spirv.storage_class<StorageBuffer>>, %arg1 : memref<12x4xf32, #spirv.storage_class<StorageBuffer>>, %arg2 : memref<12x4xf32, #spirv.storage_class<StorageBuffer>>,21 %c0 : index, %c0_0 : index, %c1 : index, %c1_1 : index)22 return23 }24 25 // CHECK-LABEL: spirv.module @{{.*}} Logical GLSL45026 gpu.module @kernels {27 // CHECK-DAG: spirv.GlobalVariable @[[NUMWORKGROUPSVAR:.*]] built_in("NumWorkgroups") : !spirv.ptr<vector<3xi32>, Input>28 // CHECK-DAG: spirv.GlobalVariable @[[$LOCALINVOCATIONIDVAR:.*]] built_in("LocalInvocationId") : !spirv.ptr<vector<3xi32>, Input>29 // CHECK-DAG: spirv.GlobalVariable @[[$WORKGROUPIDVAR:.*]] built_in("WorkgroupId") : !spirv.ptr<vector<3xi32>, Input>30 // CHECK-LABEL: spirv.func @load_store_kernel31 // CHECK-SAME: %[[ARG0:.*]]: !spirv.ptr<!spirv.struct<(!spirv.array<48 x f32, stride=4> [0])>, StorageBuffer> {spirv.interface_var_abi = #spirv.interface_var_abi<(0, 0)>}32 // CHECK-SAME: %[[ARG1:.*]]: !spirv.ptr<!spirv.struct<(!spirv.array<48 x f32, stride=4> [0])>, StorageBuffer> {spirv.interface_var_abi = #spirv.interface_var_abi<(0, 1)>}33 // CHECK-SAME: %[[ARG2:.*]]: !spirv.ptr<!spirv.struct<(!spirv.array<48 x f32, stride=4> [0])>, StorageBuffer> {spirv.interface_var_abi = #spirv.interface_var_abi<(0, 2)>}34 // CHECK-SAME: %[[ARG3:.*]]: i32 {spirv.interface_var_abi = #spirv.interface_var_abi<(0, 3), StorageBuffer>}35 // CHECK-SAME: %[[ARG4:.*]]: i32 {spirv.interface_var_abi = #spirv.interface_var_abi<(0, 4), StorageBuffer>}36 // CHECK-SAME: %[[ARG5:.*]]: i32 {spirv.interface_var_abi = #spirv.interface_var_abi<(0, 5), StorageBuffer>}37 // CHECK-SAME: %[[ARG6:.*]]: i32 {spirv.interface_var_abi = #spirv.interface_var_abi<(0, 6), StorageBuffer>}38 gpu.func @load_store_kernel(%arg0: memref<12x4xf32, #spirv.storage_class<StorageBuffer>>, %arg1: memref<12x4xf32, #spirv.storage_class<StorageBuffer>>, %arg2: memref<12x4xf32, #spirv.storage_class<StorageBuffer>>, %arg3: index, %arg4: index, %arg5: index, %arg6: index) kernel39 attributes {spirv.entry_point_abi = #spirv.entry_point_abi<workgroup_size = [16, 1, 1]>} {40 // CHECK: %[[ADDRESSWORKGROUPID:.*]] = spirv.mlir.addressof @[[$WORKGROUPIDVAR]]41 // CHECK: %[[WORKGROUPID:.*]] = spirv.Load "Input" %[[ADDRESSWORKGROUPID]]42 // CHECK: %[[WORKGROUPIDX:.*]] = spirv.CompositeExtract %[[WORKGROUPID]]{{\[}}0 : i32{{\]}}43 // CHECK: %[[ADDRESSLOCALINVOCATIONID:.*]] = spirv.mlir.addressof @[[$LOCALINVOCATIONIDVAR]]44 // CHECK: %[[LOCALINVOCATIONID:.*]] = spirv.Load "Input" %[[ADDRESSLOCALINVOCATIONID]]45 // CHECK: %[[LOCALINVOCATIONIDX:.*]] = spirv.CompositeExtract %[[LOCALINVOCATIONID]]{{\[}}0 : i32{{\]}}46 %0 = gpu.block_id x47 %1 = gpu.block_id y48 %2 = gpu.block_id z49 %3 = gpu.thread_id x50 %4 = gpu.thread_id y51 %5 = gpu.thread_id z52 %6 = gpu.grid_dim x53 %7 = gpu.grid_dim y54 %8 = gpu.grid_dim z55 %9 = gpu.block_dim x56 %10 = gpu.block_dim y57 %11 = gpu.block_dim z58 // CHECK: %[[INDEX1:.*]] = spirv.IAdd %[[ARG3]], %[[WORKGROUPIDX]]59 %12 = arith.addi %arg3, %0 : index60 // CHECK: %[[INDEX2:.*]] = spirv.IAdd %[[ARG4]], %[[LOCALINVOCATIONIDX]]61 %13 = arith.addi %arg4, %3 : index62 // CHECK: %[[ZERO:.*]] = spirv.Constant 0 : i3263 // CHECK: %[[STRIDE1_1:.*]] = spirv.Constant 4 : i3264 // CHECK: %[[UPDATE1_1:.*]] = spirv.IMul %[[INDEX1]], %[[STRIDE1_1]] : i3265 // CHECK: %[[OFFSET1_2:.*]] = spirv.IAdd %[[INDEX2]], %[[UPDATE1_1]] : i3266 // CHECK: %[[PTR1:.*]] = spirv.AccessChain %[[ARG0]]{{\[}}%[[ZERO]], %[[OFFSET1_2]]{{\]}}67 // CHECK-NEXT: %[[VAL1:.*]] = spirv.Load "StorageBuffer" %[[PTR1]]68 %14 = memref.load %arg0[%12, %13] : memref<12x4xf32, #spirv.storage_class<StorageBuffer>>69 // CHECK: %[[PTR2:.*]] = spirv.AccessChain %[[ARG1]]{{\[}}{{%.*}}, {{%.*}}{{\]}}70 // CHECK-NEXT: %[[VAL2:.*]] = spirv.Load "StorageBuffer" %[[PTR2]]71 %15 = memref.load %arg1[%12, %13] : memref<12x4xf32, #spirv.storage_class<StorageBuffer>>72 // CHECK: %[[VAL3:.*]] = spirv.FAdd %[[VAL1]], %[[VAL2]]73 %16 = arith.addf %14, %15 : f3274 // CHECK: %[[PTR3:.*]] = spirv.AccessChain %[[ARG2]]{{\[}}{{%.*}}, {{%.*}}{{\]}}75 // CHECK-NEXT: spirv.Store "StorageBuffer" %[[PTR3]], %[[VAL3]]76 memref.store %16, %arg2[%12, %13] : memref<12x4xf32, #spirv.storage_class<StorageBuffer>>77 gpu.return78 }79 }80}81