Skip to content

[mlir][GPU] Expand LLVM function attribute copies #76755

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
Jan 3, 2024
Merged

[mlir][GPU] Expand LLVM function attribute copies #76755

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
68 changes: 42 additions & 26 deletions mlir/lib/Conversion/GPUCommon/GPUOpsLowering.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -26,9 +26,8 @@ GPUFuncOpLowering::matchAndRewrite(gpu::GPUFuncOp gpuFuncOp, OpAdaptor adaptor,

SmallVector<LLVM::GlobalOp, 3> workgroupBuffers;
workgroupBuffers.reserve(gpuFuncOp.getNumWorkgroupAttributions());
for (const auto &en : llvm::enumerate(gpuFuncOp.getWorkgroupAttributions())) {
BlockArgument attribution = en.value();

for (const auto [idx, attribution] :
llvm::enumerate(gpuFuncOp.getWorkgroupAttributions())) {
auto type = dyn_cast<MemRefType>(attribution.getType());
assert(type && type.hasStaticShape() && "unexpected type in attribution");

Expand All @@ -37,12 +36,12 @@ GPUFuncOpLowering::matchAndRewrite(gpu::GPUFuncOp gpuFuncOp, OpAdaptor adaptor,
auto elementType =
cast<Type>(typeConverter->convertType(type.getElementType()));
auto arrayType = LLVM::LLVMArrayType::get(elementType, numElements);
std::string name = std::string(
llvm::formatv("__wg_{0}_{1}", gpuFuncOp.getName(), en.index()));
std::string name =
std::string(llvm::formatv("__wg_{0}_{1}", gpuFuncOp.getName(), idx));
uint64_t alignment = 0;
if (auto alignAttr =
dyn_cast_or_null<IntegerAttr>(gpuFuncOp.getWorkgroupAttributionAttr(
en.index(), LLVM::LLVMDialect::getAlignAttrName())))
idx, LLVM::LLVMDialect::getAlignAttrName())))
alignment = alignAttr.getInt();
auto globalOp = rewriter.create<LLVM::GlobalOp>(
gpuFuncOp.getLoc(), arrayType, /*isConstant=*/false,
Expand Down Expand Up @@ -105,8 +104,7 @@ GPUFuncOpLowering::matchAndRewrite(gpu::GPUFuncOp gpuFuncOp, OpAdaptor adaptor,
rewriter.setInsertionPointToStart(&gpuFuncOp.front());
unsigned numProperArguments = gpuFuncOp.getNumArguments();

for (const auto &en : llvm::enumerate(workgroupBuffers)) {
LLVM::GlobalOp global = en.value();
for (const auto [idx, global] : llvm::enumerate(workgroupBuffers)) {
auto ptrType = LLVM::LLVMPointerType::get(rewriter.getContext(),
global.getAddrSpace());
Value address = rewriter.create<LLVM::AddressOfOp>(
Expand All @@ -119,18 +117,18 @@ GPUFuncOpLowering::matchAndRewrite(gpu::GPUFuncOp gpuFuncOp, OpAdaptor adaptor,
// existing memref infrastructure. This may use more registers than
// otherwise necessary given that memref sizes are fixed, but we can try
// and canonicalize that away later.
Value attribution = gpuFuncOp.getWorkgroupAttributions()[en.index()];
Value attribution = gpuFuncOp.getWorkgroupAttributions()[idx];
auto type = cast<MemRefType>(attribution.getType());
auto descr = MemRefDescriptor::fromStaticShape(
rewriter, loc, *getTypeConverter(), type, memory);
signatureConversion.remapInput(numProperArguments + en.index(), descr);
signatureConversion.remapInput(numProperArguments + idx, descr);
}

// Rewrite private memory attributions to alloca'ed buffers.
unsigned numWorkgroupAttributions = gpuFuncOp.getNumWorkgroupAttributions();
auto int64Ty = IntegerType::get(rewriter.getContext(), 64);
for (const auto &en : llvm::enumerate(gpuFuncOp.getPrivateAttributions())) {
Value attribution = en.value();
for (const auto [idx, attribution] :
llvm::enumerate(gpuFuncOp.getPrivateAttributions())) {
auto type = cast<MemRefType>(attribution.getType());
assert(type && type.hasStaticShape() && "unexpected type in attribution");

Expand All @@ -145,14 +143,14 @@ GPUFuncOpLowering::matchAndRewrite(gpu::GPUFuncOp gpuFuncOp, OpAdaptor adaptor,
uint64_t alignment = 0;
if (auto alignAttr =
dyn_cast_or_null<IntegerAttr>(gpuFuncOp.getPrivateAttributionAttr(
en.index(), LLVM::LLVMDialect::getAlignAttrName())))
idx, LLVM::LLVMDialect::getAlignAttrName())))
alignment = alignAttr.getInt();
Value allocated = rewriter.create<LLVM::AllocaOp>(
gpuFuncOp.getLoc(), ptrType, elementType, numElements, alignment);
auto descr = MemRefDescriptor::fromStaticShape(
rewriter, loc, *getTypeConverter(), type, allocated);
signatureConversion.remapInput(
numProperArguments + numWorkgroupAttributions + en.index(), descr);
numProperArguments + numWorkgroupAttributions + idx, descr);
}
}

Expand All @@ -169,15 +167,16 @@ GPUFuncOpLowering::matchAndRewrite(gpu::GPUFuncOp gpuFuncOp, OpAdaptor adaptor,
if (getTypeConverter()->getOptions().useBarePtrCallConv) {
OpBuilder::InsertionGuard guard(rewriter);
rewriter.setInsertionPointToStart(&llvmFuncOp.getBody().front());
for (const auto &en : llvm::enumerate(gpuFuncOp.getArgumentTypes())) {
auto memrefTy = dyn_cast<MemRefType>(en.value());
for (const auto [idx, argTy] :
llvm::enumerate(gpuFuncOp.getArgumentTypes())) {
auto memrefTy = dyn_cast<MemRefType>(argTy);
if (!memrefTy)
continue;
assert(memrefTy.hasStaticShape() &&
"Bare pointer convertion used with dynamically-shaped memrefs");
// Use a placeholder when replacing uses of the memref argument to prevent
// circular replacements.
auto remapping = signatureConversion.getInputMapping(en.index());
auto remapping = signatureConversion.getInputMapping(idx);
assert(remapping && remapping->size == 1 &&
"Type converter should produce 1-to-1 mapping for bare memrefs");
BlockArgument newArg =
Expand All @@ -193,19 +192,23 @@ GPUFuncOpLowering::matchAndRewrite(gpu::GPUFuncOp gpuFuncOp, OpAdaptor adaptor,

// Get memref type from function arguments and set the noalias to
// pointer arguments.
for (const auto &en : llvm::enumerate(gpuFuncOp.getArgumentTypes())) {
auto memrefTy = en.value().dyn_cast<MemRefType>();
NamedAttrList argAttr = argAttrs
? argAttrs[en.index()].cast<DictionaryAttr>()
: NamedAttrList();

for (const auto [idx, argTy] :
llvm::enumerate(gpuFuncOp.getArgumentTypes())) {
auto remapping = signatureConversion.getInputMapping(idx);
NamedAttrList argAttr =
argAttrs ? argAttrs[idx].cast<DictionaryAttr>() : NamedAttrList();
auto copyAttribute = [&](StringRef attrName) {
Attribute attr = argAttr.erase(attrName);
if (!attr)
return;
for (size_t i = 0, e = remapping->size; i < e; ++i)
llvmFuncOp.setArgAttr(remapping->inputNo + i, attrName, attr);
};
auto copyPointerAttribute = [&](StringRef attrName) {
Attribute attr = argAttr.erase(attrName);

// This is a proxy for the bare pointer calling convention.
if (!attr)
return;
auto remapping = signatureConversion.getInputMapping(en.index());
if (remapping->size > 1 &&
attrName == LLVM::LLVMDialect::getNoAliasAttrName()) {
emitWarning(llvmFuncOp.getLoc(),
Expand All @@ -224,10 +227,23 @@ GPUFuncOpLowering::matchAndRewrite(gpu::GPUFuncOp gpuFuncOp, OpAdaptor adaptor,
if (argAttr.empty())
continue;

if (memrefTy) {
copyAttribute(LLVM::LLVMDialect::getReturnedAttrName());
copyAttribute(LLVM::LLVMDialect::getNoUndefAttrName());
copyAttribute(LLVM::LLVMDialect::getInRegAttrName());
bool lowersToPointer = false;
for (size_t i = 0, e = remapping->size; i < e; ++i) {
lowersToPointer |= isa<LLVM::LLVMPointerType>(
llvmFuncOp.getArgument(remapping->inputNo + i).getType());
}

if (lowersToPointer) {
copyPointerAttribute(LLVM::LLVMDialect::getNoAliasAttrName());
copyPointerAttribute(LLVM::LLVMDialect::getNoCaptureAttrName());
copyPointerAttribute(LLVM::LLVMDialect::getNoFreeAttrName());
copyPointerAttribute(LLVM::LLVMDialect::getAlignAttrName());
copyPointerAttribute(LLVM::LLVMDialect::getReadonlyAttrName());
copyPointerAttribute(LLVM::LLVMDialect::getWriteOnlyAttrName());
copyPointerAttribute(LLVM::LLVMDialect::getReadnoneAttrName());
copyPointerAttribute(LLVM::LLVMDialect::getNonNullAttrName());
copyPointerAttribute(LLVM::LLVMDialect::getDereferenceableAttrName());
copyPointerAttribute(
Expand Down
25 changes: 25 additions & 0 deletions mlir/test/Conversion/GPUCommon/memref-arg-attrs.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -24,6 +24,17 @@ gpu.module @kernel {
// ROCDL-SAME: !llvm.ptr {llvm.writeonly}
// NVVM-SAME: !llvm.ptr {llvm.writeonly}

// -----

gpu.module @kernel {
gpu.func @test_func_readonly_ptr(%arg0 : !llvm.ptr {llvm.readonly} ) {
gpu.return
}
}

// CHECK-LABEL: llvm.func @test_func_readonly_ptr
// ROCDL-SAME: !llvm.ptr {llvm.readonly}
// NVVM-SAME: !llvm.ptr {llvm.readonly}

// -----

Expand Down Expand Up @@ -62,3 +73,17 @@ gpu.module @kernel {
// CHECK-LABEL: llvm.func @test_func_dereferenceable_or_null
// ROCDL-SAME: !llvm.ptr {llvm.dereferenceable_or_null = 4 : i64}
// NVVM-SAME: !llvm.ptr {llvm.dereferenceable_or_null = 4 : i64}

// -----

gpu.module @kernel {
gpu.func @test_func_noundef(%arg0 : memref<f32> {llvm.noundef} ) {
gpu.return
}
}

// CHECK-LABEL: llvm.func @test_func_noundef
// ROCDL-SAME: !llvm.ptr {llvm.noundef}
// ROCDL-SAME: i64 {llvm.noundef}
// NVVM-SAME: !llvm.ptr {llvm.noundef}
// NVVM-SAME: i64 {llvm.noundef}