7aa22f013e
This enables performing several reductions in parallel, each smaller than the size of the subgroup. One potential application is flash attention with subgroup-wide matrix multiplication and reduction combined in one kernel. The multiplication operation requires a 2D matrix to be distributed over the lanes of the subgroup, which then constrains the shape the following reduction can have if we want to keep data in registers.
625 lines
26 KiB
C++
625 lines
26 KiB
C++
//===- GPUToSPIRV.cpp - GPU to SPIR-V Patterns ----------------------------===//
|
|
//
|
|
// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
|
|
// See https://llvm.org/LICENSE.txt for license information.
|
|
// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
//
|
|
// This file implements patterns to convert GPU dialect to SPIR-V dialect.
|
|
//
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
#include "mlir/Conversion/GPUToSPIRV/GPUToSPIRV.h"
|
|
#include "mlir/Dialect/GPU/IR/GPUDialect.h"
|
|
#include "mlir/Dialect/SPIRV/IR/SPIRVAttributes.h"
|
|
#include "mlir/Dialect/SPIRV/IR/SPIRVDialect.h"
|
|
#include "mlir/Dialect/SPIRV/IR/SPIRVEnums.h"
|
|
#include "mlir/Dialect/SPIRV/IR/SPIRVOps.h"
|
|
#include "mlir/Dialect/SPIRV/IR/SPIRVTypes.h"
|
|
#include "mlir/Dialect/SPIRV/IR/TargetAndABI.h"
|
|
#include "mlir/Dialect/SPIRV/Transforms/SPIRVConversion.h"
|
|
#include "mlir/IR/BuiltinOps.h"
|
|
#include "mlir/IR/Matchers.h"
|
|
#include "mlir/Transforms/DialectConversion.h"
|
|
#include <optional>
|
|
|
|
using namespace mlir;
|
|
|
|
static constexpr const char kSPIRVModule[] = "__spv__";
|
|
|
|
namespace {
|
|
/// Pattern lowering GPU block/thread size/id to loading SPIR-V invocation
|
|
/// builtin variables.
|
|
template <typename SourceOp, spirv::BuiltIn builtin>
|
|
class LaunchConfigConversion : public OpConversionPattern<SourceOp> {
|
|
public:
|
|
using OpConversionPattern<SourceOp>::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(SourceOp op, typename SourceOp::Adaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override;
|
|
};
|
|
|
|
/// Pattern lowering subgroup size/id to loading SPIR-V invocation
|
|
/// builtin variables.
|
|
template <typename SourceOp, spirv::BuiltIn builtin>
|
|
class SingleDimLaunchConfigConversion : public OpConversionPattern<SourceOp> {
|
|
public:
|
|
using OpConversionPattern<SourceOp>::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(SourceOp op, typename SourceOp::Adaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override;
|
|
};
|
|
|
|
/// This is separate because in Vulkan workgroup size is exposed to shaders via
|
|
/// a constant with WorkgroupSize decoration. So here we cannot generate a
|
|
/// builtin variable; instead the information in the `spirv.entry_point_abi`
|
|
/// attribute on the surrounding FuncOp is used to replace the gpu::BlockDimOp.
|
|
class WorkGroupSizeConversion : public OpConversionPattern<gpu::BlockDimOp> {
|
|
public:
|
|
WorkGroupSizeConversion(TypeConverter &typeConverter, MLIRContext *context)
|
|
: OpConversionPattern(typeConverter, context, /*benefit*/ 10) {}
|
|
|
|
LogicalResult
|
|
matchAndRewrite(gpu::BlockDimOp op, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override;
|
|
};
|
|
|
|
/// Pattern to convert a kernel function in GPU dialect within a spirv.module.
|
|
class GPUFuncOpConversion final : public OpConversionPattern<gpu::GPUFuncOp> {
|
|
public:
|
|
using OpConversionPattern<gpu::GPUFuncOp>::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(gpu::GPUFuncOp funcOp, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override;
|
|
|
|
private:
|
|
SmallVector<int32_t, 3> workGroupSizeAsInt32;
|
|
};
|
|
|
|
/// Pattern to convert a gpu.module to a spirv.module.
|
|
class GPUModuleConversion final : public OpConversionPattern<gpu::GPUModuleOp> {
|
|
public:
|
|
using OpConversionPattern<gpu::GPUModuleOp>::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(gpu::GPUModuleOp moduleOp, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override;
|
|
};
|
|
|
|
/// Pattern to convert a gpu.return into a SPIR-V return.
|
|
// TODO: This can go to DRR when GPU return has operands.
|
|
class GPUReturnOpConversion final : public OpConversionPattern<gpu::ReturnOp> {
|
|
public:
|
|
using OpConversionPattern<gpu::ReturnOp>::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(gpu::ReturnOp returnOp, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override;
|
|
};
|
|
|
|
/// Pattern to convert a gpu.barrier op into a spirv.ControlBarrier op.
|
|
class GPUBarrierConversion final : public OpConversionPattern<gpu::BarrierOp> {
|
|
public:
|
|
using OpConversionPattern::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(gpu::BarrierOp barrierOp, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override;
|
|
};
|
|
|
|
/// Pattern to convert a gpu.shuffle op into a spirv.GroupNonUniformShuffle op.
|
|
class GPUShuffleConversion final : public OpConversionPattern<gpu::ShuffleOp> {
|
|
public:
|
|
using OpConversionPattern::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(gpu::ShuffleOp shuffleOp, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override;
|
|
};
|
|
|
|
} // namespace
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// Builtins.
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
template <typename SourceOp, spirv::BuiltIn builtin>
|
|
LogicalResult LaunchConfigConversion<SourceOp, builtin>::matchAndRewrite(
|
|
SourceOp op, typename SourceOp::Adaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const {
|
|
auto *typeConverter = this->template getTypeConverter<SPIRVTypeConverter>();
|
|
Type indexType = typeConverter->getIndexType();
|
|
|
|
// For Vulkan, these SPIR-V builtin variables are required to be a vector of
|
|
// type <3xi32> by the spec:
|
|
// https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/NumWorkgroups.html
|
|
// https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/WorkgroupId.html
|
|
// https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/WorkgroupSize.html
|
|
// https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/LocalInvocationId.html
|
|
// https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/LocalInvocationId.html
|
|
// https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/GlobalInvocationId.html
|
|
//
|
|
// For OpenCL, it depends on the Physical32/Physical64 addressing model:
|
|
// https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_Env.html#_built_in_variables
|
|
bool forShader =
|
|
typeConverter->getTargetEnv().allows(spirv::Capability::Shader);
|
|
Type builtinType = forShader ? rewriter.getIntegerType(32) : indexType;
|
|
|
|
Value vector =
|
|
spirv::getBuiltinVariableValue(op, builtin, builtinType, rewriter);
|
|
Value dim = rewriter.create<spirv::CompositeExtractOp>(
|
|
op.getLoc(), builtinType, vector,
|
|
rewriter.getI32ArrayAttr({static_cast<int32_t>(op.getDimension())}));
|
|
if (forShader && builtinType != indexType)
|
|
dim = rewriter.create<spirv::UConvertOp>(op.getLoc(), indexType, dim);
|
|
rewriter.replaceOp(op, dim);
|
|
return success();
|
|
}
|
|
|
|
template <typename SourceOp, spirv::BuiltIn builtin>
|
|
LogicalResult
|
|
SingleDimLaunchConfigConversion<SourceOp, builtin>::matchAndRewrite(
|
|
SourceOp op, typename SourceOp::Adaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const {
|
|
auto *typeConverter = this->template getTypeConverter<SPIRVTypeConverter>();
|
|
Type indexType = typeConverter->getIndexType();
|
|
Type i32Type = rewriter.getIntegerType(32);
|
|
|
|
// For Vulkan, these SPIR-V builtin variables are required to be a vector of
|
|
// type i32 by the spec:
|
|
// https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/NumSubgroups.html
|
|
// https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/SubgroupId.html
|
|
// https://registry.khronos.org/vulkan/specs/1.3-extensions/man/html/SubgroupSize.html
|
|
//
|
|
// For OpenCL, they are also required to be i32:
|
|
// https://registry.khronos.org/OpenCL/specs/3.0-unified/html/OpenCL_Env.html#_built_in_variables
|
|
Value builtinValue =
|
|
spirv::getBuiltinVariableValue(op, builtin, i32Type, rewriter);
|
|
if (i32Type != indexType)
|
|
builtinValue = rewriter.create<spirv::UConvertOp>(op.getLoc(), indexType,
|
|
builtinValue);
|
|
rewriter.replaceOp(op, builtinValue);
|
|
return success();
|
|
}
|
|
|
|
LogicalResult WorkGroupSizeConversion::matchAndRewrite(
|
|
gpu::BlockDimOp op, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const {
|
|
DenseI32ArrayAttr workGroupSizeAttr = spirv::lookupLocalWorkGroupSize(op);
|
|
if (!workGroupSizeAttr)
|
|
return failure();
|
|
|
|
int val =
|
|
workGroupSizeAttr.asArrayRef()[static_cast<int32_t>(op.getDimension())];
|
|
auto convertedType =
|
|
getTypeConverter()->convertType(op.getResult().getType());
|
|
if (!convertedType)
|
|
return failure();
|
|
rewriter.replaceOpWithNewOp<spirv::ConstantOp>(
|
|
op, convertedType, IntegerAttr::get(convertedType, val));
|
|
return success();
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// GPUFuncOp
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
// Legalizes a GPU function as an entry SPIR-V function.
|
|
static spirv::FuncOp
|
|
lowerAsEntryFunction(gpu::GPUFuncOp funcOp, const TypeConverter &typeConverter,
|
|
ConversionPatternRewriter &rewriter,
|
|
spirv::EntryPointABIAttr entryPointInfo,
|
|
ArrayRef<spirv::InterfaceVarABIAttr> argABIInfo) {
|
|
auto fnType = funcOp.getFunctionType();
|
|
if (fnType.getNumResults()) {
|
|
funcOp.emitError("SPIR-V lowering only supports entry functions"
|
|
"with no return values right now");
|
|
return nullptr;
|
|
}
|
|
if (!argABIInfo.empty() && fnType.getNumInputs() != argABIInfo.size()) {
|
|
funcOp.emitError(
|
|
"lowering as entry functions requires ABI info for all arguments "
|
|
"or none of them");
|
|
return nullptr;
|
|
}
|
|
// Update the signature to valid SPIR-V types and add the ABI
|
|
// attributes. These will be "materialized" by using the
|
|
// LowerABIAttributesPass.
|
|
TypeConverter::SignatureConversion signatureConverter(fnType.getNumInputs());
|
|
{
|
|
for (const auto &argType :
|
|
enumerate(funcOp.getFunctionType().getInputs())) {
|
|
auto convertedType = typeConverter.convertType(argType.value());
|
|
if (!convertedType)
|
|
return nullptr;
|
|
signatureConverter.addInputs(argType.index(), convertedType);
|
|
}
|
|
}
|
|
auto newFuncOp = rewriter.create<spirv::FuncOp>(
|
|
funcOp.getLoc(), funcOp.getName(),
|
|
rewriter.getFunctionType(signatureConverter.getConvertedTypes(),
|
|
std::nullopt));
|
|
for (const auto &namedAttr : funcOp->getAttrs()) {
|
|
if (namedAttr.getName() == funcOp.getFunctionTypeAttrName() ||
|
|
namedAttr.getName() == SymbolTable::getSymbolAttrName())
|
|
continue;
|
|
newFuncOp->setAttr(namedAttr.getName(), namedAttr.getValue());
|
|
}
|
|
|
|
rewriter.inlineRegionBefore(funcOp.getBody(), newFuncOp.getBody(),
|
|
newFuncOp.end());
|
|
if (failed(rewriter.convertRegionTypes(&newFuncOp.getBody(), typeConverter,
|
|
&signatureConverter)))
|
|
return nullptr;
|
|
rewriter.eraseOp(funcOp);
|
|
|
|
// Set the attributes for argument and the function.
|
|
StringRef argABIAttrName = spirv::getInterfaceVarABIAttrName();
|
|
for (auto argIndex : llvm::seq<unsigned>(0, argABIInfo.size())) {
|
|
newFuncOp.setArgAttr(argIndex, argABIAttrName, argABIInfo[argIndex]);
|
|
}
|
|
newFuncOp->setAttr(spirv::getEntryPointABIAttrName(), entryPointInfo);
|
|
|
|
return newFuncOp;
|
|
}
|
|
|
|
/// Populates `argABI` with spirv.interface_var_abi attributes for lowering
|
|
/// gpu.func to spirv.func if no arguments have the attributes set
|
|
/// already. Returns failure if any argument has the ABI attribute set already.
|
|
static LogicalResult
|
|
getDefaultABIAttrs(const spirv::TargetEnv &targetEnv, gpu::GPUFuncOp funcOp,
|
|
SmallVectorImpl<spirv::InterfaceVarABIAttr> &argABI) {
|
|
if (!spirv::needsInterfaceVarABIAttrs(targetEnv))
|
|
return success();
|
|
|
|
for (auto argIndex : llvm::seq<unsigned>(0, funcOp.getNumArguments())) {
|
|
if (funcOp.getArgAttrOfType<spirv::InterfaceVarABIAttr>(
|
|
argIndex, spirv::getInterfaceVarABIAttrName()))
|
|
return failure();
|
|
// Vulkan's interface variable requirements needs scalars to be wrapped in a
|
|
// struct. The struct held in storage buffer.
|
|
std::optional<spirv::StorageClass> sc;
|
|
if (funcOp.getArgument(argIndex).getType().isIntOrIndexOrFloat())
|
|
sc = spirv::StorageClass::StorageBuffer;
|
|
argABI.push_back(
|
|
spirv::getInterfaceVarABIAttr(0, argIndex, sc, funcOp.getContext()));
|
|
}
|
|
return success();
|
|
}
|
|
|
|
LogicalResult GPUFuncOpConversion::matchAndRewrite(
|
|
gpu::GPUFuncOp funcOp, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const {
|
|
if (!gpu::GPUDialect::isKernel(funcOp))
|
|
return failure();
|
|
|
|
auto *typeConverter = getTypeConverter<SPIRVTypeConverter>();
|
|
SmallVector<spirv::InterfaceVarABIAttr, 4> argABI;
|
|
if (failed(
|
|
getDefaultABIAttrs(typeConverter->getTargetEnv(), funcOp, argABI))) {
|
|
argABI.clear();
|
|
for (auto argIndex : llvm::seq<unsigned>(0, funcOp.getNumArguments())) {
|
|
// If the ABI is already specified, use it.
|
|
auto abiAttr = funcOp.getArgAttrOfType<spirv::InterfaceVarABIAttr>(
|
|
argIndex, spirv::getInterfaceVarABIAttrName());
|
|
if (!abiAttr) {
|
|
funcOp.emitRemark(
|
|
"match failure: missing 'spirv.interface_var_abi' attribute at "
|
|
"argument ")
|
|
<< argIndex;
|
|
return failure();
|
|
}
|
|
argABI.push_back(abiAttr);
|
|
}
|
|
}
|
|
|
|
auto entryPointAttr = spirv::lookupEntryPointABI(funcOp);
|
|
if (!entryPointAttr) {
|
|
funcOp.emitRemark(
|
|
"match failure: missing 'spirv.entry_point_abi' attribute");
|
|
return failure();
|
|
}
|
|
spirv::FuncOp newFuncOp = lowerAsEntryFunction(
|
|
funcOp, *getTypeConverter(), rewriter, entryPointAttr, argABI);
|
|
if (!newFuncOp)
|
|
return failure();
|
|
newFuncOp->removeAttr(
|
|
rewriter.getStringAttr(gpu::GPUDialect::getKernelFuncAttrName()));
|
|
return success();
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// ModuleOp with gpu.module.
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult GPUModuleConversion::matchAndRewrite(
|
|
gpu::GPUModuleOp moduleOp, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const {
|
|
auto *typeConverter = getTypeConverter<SPIRVTypeConverter>();
|
|
const spirv::TargetEnv &targetEnv = typeConverter->getTargetEnv();
|
|
spirv::AddressingModel addressingModel = spirv::getAddressingModel(
|
|
targetEnv, typeConverter->getOptions().use64bitIndex);
|
|
FailureOr<spirv::MemoryModel> memoryModel = spirv::getMemoryModel(targetEnv);
|
|
if (failed(memoryModel))
|
|
return moduleOp.emitRemark(
|
|
"cannot deduce memory model from 'spirv.target_env'");
|
|
|
|
// Add a keyword to the module name to avoid symbolic conflict.
|
|
std::string spvModuleName = (kSPIRVModule + moduleOp.getName()).str();
|
|
auto spvModule = rewriter.create<spirv::ModuleOp>(
|
|
moduleOp.getLoc(), addressingModel, *memoryModel, std::nullopt,
|
|
StringRef(spvModuleName));
|
|
|
|
// Move the region from the module op into the SPIR-V module.
|
|
Region &spvModuleRegion = spvModule.getRegion();
|
|
rewriter.inlineRegionBefore(moduleOp.getBodyRegion(), spvModuleRegion,
|
|
spvModuleRegion.begin());
|
|
// The spirv.module build method adds a block. Remove that.
|
|
rewriter.eraseBlock(&spvModuleRegion.back());
|
|
|
|
// Some of the patterns call `lookupTargetEnv` during conversion and they
|
|
// will fail if called after GPUModuleConversion and we don't preserve
|
|
// `TargetEnv` attribute.
|
|
// Copy TargetEnvAttr only if it is attached directly to the GPUModuleOp.
|
|
if (auto attr = moduleOp->getAttrOfType<spirv::TargetEnvAttr>(
|
|
spirv::getTargetEnvAttrName()))
|
|
spvModule->setAttr(spirv::getTargetEnvAttrName(), attr);
|
|
|
|
rewriter.eraseOp(moduleOp);
|
|
return success();
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// GPU return inside kernel functions to SPIR-V return.
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult GPUReturnOpConversion::matchAndRewrite(
|
|
gpu::ReturnOp returnOp, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const {
|
|
if (!adaptor.getOperands().empty())
|
|
return failure();
|
|
|
|
rewriter.replaceOpWithNewOp<spirv::ReturnOp>(returnOp);
|
|
return success();
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// Barrier.
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult GPUBarrierConversion::matchAndRewrite(
|
|
gpu::BarrierOp barrierOp, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const {
|
|
MLIRContext *context = getContext();
|
|
// Both execution and memory scope should be workgroup.
|
|
auto scope = spirv::ScopeAttr::get(context, spirv::Scope::Workgroup);
|
|
// Require acquire and release memory semantics for workgroup memory.
|
|
auto memorySemantics = spirv::MemorySemanticsAttr::get(
|
|
context, spirv::MemorySemantics::WorkgroupMemory |
|
|
spirv::MemorySemantics::AcquireRelease);
|
|
rewriter.replaceOpWithNewOp<spirv::ControlBarrierOp>(barrierOp, scope, scope,
|
|
memorySemantics);
|
|
return success();
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// Shuffle
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
LogicalResult GPUShuffleConversion::matchAndRewrite(
|
|
gpu::ShuffleOp shuffleOp, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const {
|
|
// Require the shuffle width to be the same as the target's subgroup size,
|
|
// given that for SPIR-V non-uniform subgroup ops, we cannot select
|
|
// participating invocations.
|
|
auto targetEnv = getTypeConverter<SPIRVTypeConverter>()->getTargetEnv();
|
|
unsigned subgroupSize =
|
|
targetEnv.getAttr().getResourceLimits().getSubgroupSize();
|
|
IntegerAttr widthAttr;
|
|
if (!matchPattern(shuffleOp.getWidth(), m_Constant(&widthAttr)) ||
|
|
widthAttr.getValue().getZExtValue() != subgroupSize)
|
|
return rewriter.notifyMatchFailure(
|
|
shuffleOp, "shuffle width and target subgroup size mismatch");
|
|
|
|
Location loc = shuffleOp.getLoc();
|
|
Value trueVal = spirv::ConstantOp::getOne(rewriter.getI1Type(),
|
|
shuffleOp.getLoc(), rewriter);
|
|
auto scope = rewriter.getAttr<spirv::ScopeAttr>(spirv::Scope::Subgroup);
|
|
Value result;
|
|
|
|
switch (shuffleOp.getMode()) {
|
|
case gpu::ShuffleMode::XOR:
|
|
result = rewriter.create<spirv::GroupNonUniformShuffleXorOp>(
|
|
loc, scope, adaptor.getValue(), adaptor.getOffset());
|
|
break;
|
|
case gpu::ShuffleMode::IDX:
|
|
result = rewriter.create<spirv::GroupNonUniformShuffleOp>(
|
|
loc, scope, adaptor.getValue(), adaptor.getOffset());
|
|
break;
|
|
default:
|
|
return rewriter.notifyMatchFailure(shuffleOp, "unimplemented shuffle mode");
|
|
}
|
|
|
|
rewriter.replaceOp(shuffleOp, {result, trueVal});
|
|
return success();
|
|
}
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// Group ops
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
template <typename UniformOp, typename NonUniformOp>
|
|
static Value createGroupReduceOpImpl(OpBuilder &builder, Location loc,
|
|
Value arg, bool isGroup, bool isUniform) {
|
|
Type type = arg.getType();
|
|
auto scope = mlir::spirv::ScopeAttr::get(builder.getContext(),
|
|
isGroup ? spirv::Scope::Workgroup
|
|
: spirv::Scope::Subgroup);
|
|
auto groupOp = spirv::GroupOperationAttr::get(builder.getContext(),
|
|
spirv::GroupOperation::Reduce);
|
|
if (isUniform) {
|
|
return builder.create<UniformOp>(loc, type, scope, groupOp, arg)
|
|
.getResult();
|
|
}
|
|
return builder.create<NonUniformOp>(loc, type, scope, groupOp, arg, Value{})
|
|
.getResult();
|
|
}
|
|
|
|
static std::optional<Value> createGroupReduceOp(OpBuilder &builder,
|
|
Location loc, Value arg,
|
|
gpu::AllReduceOperation opType,
|
|
bool isGroup, bool isUniform) {
|
|
enum class ElemType { Float, Boolean, Integer };
|
|
using FuncT = Value (*)(OpBuilder &, Location, Value, bool, bool);
|
|
struct OpHandler {
|
|
gpu::AllReduceOperation kind;
|
|
ElemType elemType;
|
|
FuncT func;
|
|
};
|
|
|
|
Type type = arg.getType();
|
|
ElemType elementType;
|
|
if (isa<FloatType>(type)) {
|
|
elementType = ElemType::Float;
|
|
} else if (auto intTy = dyn_cast<IntegerType>(type)) {
|
|
elementType = (intTy.getIntOrFloatBitWidth() == 1) ? ElemType::Boolean
|
|
: ElemType::Integer;
|
|
} else {
|
|
return std::nullopt;
|
|
}
|
|
|
|
// TODO(https://github.com/llvm/llvm-project/issues/73459): The SPIR-V spec
|
|
// does not specify how -0.0 / +0.0 and NaN values are handled in *FMin/*FMax
|
|
// reduction ops. We should account possible precision requirements in this
|
|
// conversion.
|
|
|
|
using ReduceType = gpu::AllReduceOperation;
|
|
const OpHandler handlers[] = {
|
|
{ReduceType::ADD, ElemType::Integer,
|
|
&createGroupReduceOpImpl<spirv::GroupIAddOp,
|
|
spirv::GroupNonUniformIAddOp>},
|
|
{ReduceType::ADD, ElemType::Float,
|
|
&createGroupReduceOpImpl<spirv::GroupFAddOp,
|
|
spirv::GroupNonUniformFAddOp>},
|
|
{ReduceType::MUL, ElemType::Integer,
|
|
&createGroupReduceOpImpl<spirv::GroupIMulKHROp,
|
|
spirv::GroupNonUniformIMulOp>},
|
|
{ReduceType::MUL, ElemType::Float,
|
|
&createGroupReduceOpImpl<spirv::GroupFMulKHROp,
|
|
spirv::GroupNonUniformFMulOp>},
|
|
{ReduceType::MINUI, ElemType::Integer,
|
|
&createGroupReduceOpImpl<spirv::GroupUMinOp,
|
|
spirv::GroupNonUniformUMinOp>},
|
|
{ReduceType::MINSI, ElemType::Integer,
|
|
&createGroupReduceOpImpl<spirv::GroupSMinOp,
|
|
spirv::GroupNonUniformSMinOp>},
|
|
{ReduceType::MINNUMF, ElemType::Float,
|
|
&createGroupReduceOpImpl<spirv::GroupFMinOp,
|
|
spirv::GroupNonUniformFMinOp>},
|
|
{ReduceType::MAXUI, ElemType::Integer,
|
|
&createGroupReduceOpImpl<spirv::GroupUMaxOp,
|
|
spirv::GroupNonUniformUMaxOp>},
|
|
{ReduceType::MAXSI, ElemType::Integer,
|
|
&createGroupReduceOpImpl<spirv::GroupSMaxOp,
|
|
spirv::GroupNonUniformSMaxOp>},
|
|
{ReduceType::MAXNUMF, ElemType::Float,
|
|
&createGroupReduceOpImpl<spirv::GroupFMaxOp,
|
|
spirv::GroupNonUniformFMaxOp>},
|
|
{ReduceType::MINIMUMF, ElemType::Float,
|
|
&createGroupReduceOpImpl<spirv::GroupFMinOp,
|
|
spirv::GroupNonUniformFMinOp>},
|
|
{ReduceType::MAXIMUMF, ElemType::Float,
|
|
&createGroupReduceOpImpl<spirv::GroupFMaxOp,
|
|
spirv::GroupNonUniformFMaxOp>}};
|
|
|
|
for (const OpHandler &handler : handlers)
|
|
if (handler.kind == opType && elementType == handler.elemType)
|
|
return handler.func(builder, loc, arg, isGroup, isUniform);
|
|
|
|
return std::nullopt;
|
|
}
|
|
|
|
/// Pattern to convert a gpu.all_reduce op into a SPIR-V group op.
|
|
class GPUAllReduceConversion final
|
|
: public OpConversionPattern<gpu::AllReduceOp> {
|
|
public:
|
|
using OpConversionPattern::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(gpu::AllReduceOp op, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
auto opType = op.getOp();
|
|
|
|
// gpu.all_reduce can have either reduction op attribute or reduction
|
|
// region. Only attribute version is supported.
|
|
if (!opType)
|
|
return failure();
|
|
|
|
auto result =
|
|
createGroupReduceOp(rewriter, op.getLoc(), adaptor.getValue(), *opType,
|
|
/*isGroup*/ true, op.getUniform());
|
|
if (!result)
|
|
return failure();
|
|
|
|
rewriter.replaceOp(op, *result);
|
|
return success();
|
|
}
|
|
};
|
|
|
|
/// Pattern to convert a gpu.subgroup_reduce op into a SPIR-V group op.
|
|
class GPUSubgroupReduceConversion final
|
|
: public OpConversionPattern<gpu::SubgroupReduceOp> {
|
|
public:
|
|
using OpConversionPattern::OpConversionPattern;
|
|
|
|
LogicalResult
|
|
matchAndRewrite(gpu::SubgroupReduceOp op, OpAdaptor adaptor,
|
|
ConversionPatternRewriter &rewriter) const override {
|
|
if (op.getClusterSize())
|
|
return rewriter.notifyMatchFailure(
|
|
op, "lowering for clustered reduce not implemented");
|
|
|
|
if (!isa<spirv::ScalarType>(adaptor.getValue().getType()))
|
|
return rewriter.notifyMatchFailure(op, "reduction type is not a scalar");
|
|
|
|
auto result = createGroupReduceOp(rewriter, op.getLoc(), adaptor.getValue(),
|
|
adaptor.getOp(),
|
|
/*isGroup=*/false, adaptor.getUniform());
|
|
if (!result)
|
|
return failure();
|
|
|
|
rewriter.replaceOp(op, *result);
|
|
return success();
|
|
}
|
|
};
|
|
|
|
//===----------------------------------------------------------------------===//
|
|
// GPU To SPIRV Patterns.
|
|
//===----------------------------------------------------------------------===//
|
|
|
|
void mlir::populateGPUToSPIRVPatterns(SPIRVTypeConverter &typeConverter,
|
|
RewritePatternSet &patterns) {
|
|
patterns.add<
|
|
GPUBarrierConversion, GPUFuncOpConversion, GPUModuleConversion,
|
|
GPUReturnOpConversion, GPUShuffleConversion,
|
|
LaunchConfigConversion<gpu::BlockIdOp, spirv::BuiltIn::WorkgroupId>,
|
|
LaunchConfigConversion<gpu::GridDimOp, spirv::BuiltIn::NumWorkgroups>,
|
|
LaunchConfigConversion<gpu::BlockDimOp, spirv::BuiltIn::WorkgroupSize>,
|
|
LaunchConfigConversion<gpu::ThreadIdOp,
|
|
spirv::BuiltIn::LocalInvocationId>,
|
|
LaunchConfigConversion<gpu::GlobalIdOp,
|
|
spirv::BuiltIn::GlobalInvocationId>,
|
|
SingleDimLaunchConfigConversion<gpu::SubgroupIdOp,
|
|
spirv::BuiltIn::SubgroupId>,
|
|
SingleDimLaunchConfigConversion<gpu::NumSubgroupsOp,
|
|
spirv::BuiltIn::NumSubgroups>,
|
|
SingleDimLaunchConfigConversion<gpu::SubgroupSizeOp,
|
|
spirv::BuiltIn::SubgroupSize>,
|
|
WorkGroupSizeConversion, GPUAllReduceConversion,
|
|
GPUSubgroupReduceConversion>(typeConverter, patterns.getContext());
|
|
}
|