[Mlir-commits] [mlir] [MLIR][GPU-LLVM] Add in-pass signature update option for opencl kernels (PR #105664)
Petr Kurapov
llvmlistbot at llvm.org
Thu Aug 22 07:13:55 PDT 2024
https://github.com/kurapov-peter created https://github.com/llvm/llvm-project/pull/105664
This PR introduces a new option `force-opencl-address-spaces` to the `gpu-to-llvm-spv` conversion pass. It addresses a problem with signature incompatibility between `gpu.launch_func` and the kernel arguments when using a unified shared memory with OpenCL runtime. The discrepancy comes from the fact that the global address space is defined as `1` in OpenCL spec.
Note: this is a draft showing an alternative solution to the one proposed in https://github.com/llvm/llvm-project/pull/102925. Currently, it lacks the differentiation between kernels and non-kernel functions (generic address space should be attached to the latter one's arguments). I'll add it if we decide the approach is acceptable.
>From 2e8465cc42eb3cec6b4b94e5a01fc8bae0fb62c0 Mon Sep 17 00:00:00 2001
From: Petr Kurapov <petr.a.kurapov at intel.com>
Date: Mon, 19 Aug 2024 14:03:44 +0000
Subject: [PATCH] [MLIR][GPU-LLVM] Add in-pass signature update option for
opencl kernels
---
mlir/include/mlir/Conversion/Passes.td | 3 +
.../Conversion/GPUToLLVMSPV/GPUToLLVMSPV.cpp | 59 +++++++++++++++++++
.../GPUToLLVMSPV/gpu-to-llvm-spv.mlir | 14 +++++
3 files changed, 76 insertions(+)
diff --git a/mlir/include/mlir/Conversion/Passes.td b/mlir/include/mlir/Conversion/Passes.td
index 7bde9e490e4f4e..05f07421b8f526 100644
--- a/mlir/include/mlir/Conversion/Passes.td
+++ b/mlir/include/mlir/Conversion/Passes.td
@@ -542,6 +542,9 @@ def ConvertGpuOpsToLLVMSPVOps : Pass<"convert-gpu-to-llvm-spv", "gpu::GPUModuleO
Option<"indexBitwidth", "index-bitwidth", "unsigned",
/*default=kDeriveIndexBitwidthFromDataLayout*/"0",
"Bitwidth of the index type, 0 to use size of machine word">,
+ Option<"forceOpenclAddressSpaces", "force-opencl-address-spaces",
+ "bool", /*default=*/"false",
+ "Force kernel argument pointers to have address space global.">,
];
}
diff --git a/mlir/lib/Conversion/GPUToLLVMSPV/GPUToLLVMSPV.cpp b/mlir/lib/Conversion/GPUToLLVMSPV/GPUToLLVMSPV.cpp
index ced4236402923a..25cf6560257978 100644
--- a/mlir/lib/Conversion/GPUToLLVMSPV/GPUToLLVMSPV.cpp
+++ b/mlir/lib/Conversion/GPUToLLVMSPV/GPUToLLVMSPV.cpp
@@ -306,6 +306,51 @@ struct GPUShuffleConversion final : ConvertOpToLLVMPattern<gpu::ShuffleOp> {
}
};
+class MemorySpaceToOpenCLMemorySpaceConverter : public TypeConverter {
+public:
+ explicit MemorySpaceToOpenCLMemorySpaceConverter() {
+ addConversion([](Type t) { return t; });
+ addConversion(
+ [this](BaseMemRefType memRefType) -> std::optional<Type> {
+ std::optional<gpu::AddressSpace> addrSpace =
+ memorySpaceMap(memRefType.getMemorySpace());
+ if (!addrSpace) {
+ LLVM_DEBUG(
+ llvm::dbgs()
+ << "cannot convert " << memRefType
+ << " due to being unable to find address space in the map\n");
+ return std::nullopt;
+ }
+ auto addrSpaceAttr =
+ gpu::AddressSpaceAttr::get(memRefType.getContext(), *addrSpace);
+ if (auto rankedType = dyn_cast<MemRefType>(memRefType)) {
+ return MemRefType::get(memRefType.getShape(),
+ memRefType.getElementType(),
+ rankedType.getLayout(), addrSpaceAttr);
+ }
+ return UnrankedMemRefType::get(memRefType.getElementType(),
+ addrSpaceAttr);
+ });
+ addConversion([this](FunctionType type) {
+ auto inputs = llvm::map_to_vector(
+ type.getInputs(), [this](Type ty) { return convertType(ty); });
+ auto results = llvm::map_to_vector(
+ type.getResults(), [this](Type ty) { return convertType(ty); });
+ return FunctionType::get(type.getContext(), inputs, results);
+ });
+ }
+
+private:
+ std::optional<gpu::AddressSpace> memorySpaceMap(Attribute memSpaceAttr) {
+ if (!memSpaceAttr)
+ return gpu::AddressSpace::Global;
+ auto gpuAddrSpace = dyn_cast<gpu::AddressSpaceAttr>(memSpaceAttr);
+ if (!gpuAddrSpace)
+ return std::nullopt;
+ return gpuAddrSpace.getValue();
+ }
+};
+
//===----------------------------------------------------------------------===//
// GPU To LLVM-SPV Pass.
//===----------------------------------------------------------------------===//
@@ -325,6 +370,20 @@ struct GPUToLLVMSPVConversionPass final
LLVMTypeConverter converter(context, options);
LLVMConversionTarget target(*context);
+ if (forceOpenclAddressSpaces) {
+ MemorySpaceToOpenCLMemorySpaceConverter converter;
+ AttrTypeReplacer replacer;
+ replacer.addReplacement([&converter](BaseMemRefType origType)
+ -> std::optional<BaseMemRefType> {
+ return converter.convertType<BaseMemRefType>(origType);
+ });
+
+ replacer.recursivelyReplaceElementsIn(getOperation(),
+ /*replaceAttrs=*/true,
+ /*replaceLocs=*/false,
+ /*replaceTypes=*/true);
+ }
+
target.addIllegalOp<gpu::BarrierOp, gpu::BlockDimOp, gpu::BlockIdOp,
gpu::GPUFuncOp, gpu::GlobalIdOp, gpu::GridDimOp,
gpu::ReturnOp, gpu::ShuffleOp, gpu::ThreadIdOp>();
diff --git a/mlir/test/Conversion/GPUToLLVMSPV/gpu-to-llvm-spv.mlir b/mlir/test/Conversion/GPUToLLVMSPV/gpu-to-llvm-spv.mlir
index ec4f4a304d5073..d100f36ae42521 100644
--- a/mlir/test/Conversion/GPUToLLVMSPV/gpu-to-llvm-spv.mlir
+++ b/mlir/test/Conversion/GPUToLLVMSPV/gpu-to-llvm-spv.mlir
@@ -2,6 +2,8 @@
// RUN: | FileCheck --check-prefixes=CHECK-64,CHECK %s
// RUN: mlir-opt -pass-pipeline="builtin.module(gpu.module(convert-gpu-to-llvm-spv{index-bitwidth=32}))" -split-input-file -verify-diagnostics %s \
// RUN: | FileCheck --check-prefixes=CHECK-32,CHECK %s
+// RUN: mlir-opt -pass-pipeline="builtin.module(gpu.module(convert-gpu-to-llvm-spv{force-opencl-address-spaces}))" -split-input-file -verify-diagnostics %s \
+// RUN: | FileCheck --check-prefixes=OPENCL %s
gpu.module @builtins {
// CHECK-64: llvm.func spir_funccc @_Z14get_num_groupsj(i32) -> i64 attributes {
@@ -515,3 +517,15 @@ gpu.module @kernels {
gpu.return
}
}
+
+// -----
+
+gpu.module @kernels {
+// OPENCL-LABEL: llvm.func spir_funccc @no_address_spaces(
+// OPENCL-SAME: %{{[a-zA-Z_][a-zA-Z0-9_]*}}: !llvm.ptr<1>
+// OPENCL-SAME: %{{[a-zA-Z_][a-zA-Z0-9_]*}}: !llvm.ptr<1>
+// OPENCL-SAME: %{{[a-zA-Z_][a-zA-Z0-9_]*}}: !llvm.ptr<1>
+ gpu.func @no_address_spaces(%arg0: memref<f32>, %arg1: memref<f32, #gpu.address_space<global>>, %arg2: memref<f32>) {
+ gpu.return
+ }
+}
More information about the Mlir-commits
mailing list