Skip to content

[mlir][GPUToNVVM] Fix memref function args/results #96392

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
Jun 23, 2024
Merged

[mlir][GPUToNVVM] Fix memref function args/results #96392

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
56 changes: 56 additions & 0 deletions mlir/lib/Conversion/GPUCommon/GPUOpsLowering.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -684,6 +684,62 @@ LogicalResult GPUDynamicSharedMemoryOpLowering::matchAndRewrite(
return success();
}

LogicalResult GPUReturnOpLowering::matchAndRewrite(
gpu::ReturnOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const {
Location loc = op.getLoc();
unsigned numArguments = op.getNumOperands();
SmallVector<Value, 4> updatedOperands;

bool useBarePtrCallConv = getTypeConverter()->getOptions().useBarePtrCallConv;
if (useBarePtrCallConv) {
// For the bare-ptr calling convention, extract the aligned pointer to
// be returned from the memref descriptor.
for (auto it : llvm::zip(op->getOperands(), adaptor.getOperands())) {
Type oldTy = std::get<0>(it).getType();
Value newOperand = std::get<1>(it);
if (isa<MemRefType>(oldTy) && getTypeConverter()->canConvertToBarePtr(
cast<BaseMemRefType>(oldTy))) {
MemRefDescriptor memrefDesc(newOperand);
newOperand = memrefDesc.allocatedPtr(rewriter, loc);
} else if (isa<UnrankedMemRefType>(oldTy)) {
// Unranked memref is not supported in the bare pointer calling
// convention.
return failure();
}
updatedOperands.push_back(newOperand);
}
} else {
updatedOperands = llvm::to_vector<4>(adaptor.getOperands());
(void)copyUnrankedDescriptors(rewriter, loc, op.getOperands().getTypes(),
updatedOperands,
/*toDynamic=*/true);
}

// If ReturnOp has 0 or 1 operand, create it and return immediately.
if (numArguments <= 1) {
rewriter.replaceOpWithNewOp<LLVM::ReturnOp>(
op, TypeRange(), updatedOperands, op->getAttrs());
return success();
}

// Otherwise, we need to pack the arguments into an LLVM struct type before
// returning.
auto packedType = getTypeConverter()->packFunctionResults(
op.getOperandTypes(), useBarePtrCallConv);
if (!packedType) {
return rewriter.notifyMatchFailure(op, "could not convert result types");
}

Value packed = rewriter.create<LLVM::UndefOp>(loc, packedType);
for (auto [idx, operand] : llvm::enumerate(updatedOperands)) {
packed = rewriter.create<LLVM::InsertValueOp>(loc, packed, operand, idx);
}
rewriter.replaceOpWithNewOp<LLVM::ReturnOp>(op, TypeRange(), packed,
op->getAttrs());
return success();
}

void mlir::populateGpuMemorySpaceAttributeConversions(
TypeConverter &typeConverter, const MemorySpaceMapping &mapping) {
typeConverter.addTypeAttributeConversion(
Expand Down
5 changes: 1 addition & 4 deletions mlir/lib/Conversion/GPUCommon/GPUOpsLowering.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -112,10 +112,7 @@ struct GPUReturnOpLowering : public ConvertOpToLLVMPattern<gpu::ReturnOp> {

LogicalResult
matchAndRewrite(gpu::ReturnOp op, OpAdaptor adaptor,
ConversionPatternRewriter &rewriter) const override {
rewriter.replaceOpWithNewOp<LLVM::ReturnOp>(op, adaptor.getOperands());
return success();
}
ConversionPatternRewriter &rewriter) const override;
};

namespace impl {
Expand Down
29 changes: 25 additions & 4 deletions mlir/test/Conversion/GPUToNVVM/gpu-to-nvvm.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1,4 +1,5 @@
// RUN: mlir-opt %s -convert-gpu-to-nvvm='has-redux=1' -split-input-file | FileCheck %s
// RUN: mlir-opt %s -convert-gpu-to-nvvm='has-redux=1 use-bare-ptr-memref-call-conv=1' -split-input-file | FileCheck %s --check-prefix=CHECK-BARE
// RUN: mlir-opt %s -transform-interpreter | FileCheck %s

gpu.module @test_module_0 {
Expand Down Expand Up @@ -670,7 +671,7 @@ gpu.module @test_module_32 {
}
}

gpu.module @gpumodule {
gpu.module @test_module_33 {
// CHECK-LABEL: func @kernel_with_block_size()
// CHECK: attributes {gpu.kernel, gpu.known_block_size = array<i32: 128, 1, 1>, nvvm.kernel, nvvm.maxntid = array<i32: 128, 1, 1>}
gpu.func @kernel_with_block_size() kernel attributes {known_block_size = array<i32: 128, 1, 1>} {
Expand All @@ -679,6 +680,28 @@ gpu.module @gpumodule {
}


gpu.module @test_module_34 {
// CHECK-LABEL: llvm.func @memref_signature(
// CHECK-SAME: %{{.*}}: !llvm.ptr, %{{.*}}: !llvm.ptr, %{{.*}}: i64, %{{.*}}: i64, %{{.*}}: i64, %{{.*}}: f32) -> !llvm.struct<(struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>, f32)>
// CHECK: llvm.mlir.undef
// CHECK: llvm.insertvalue
// CHECK: llvm.insertvalue
// CHECK: llvm.insertvalue
// CHECK: llvm.insertvalue
// CHECK: llvm.insertvalue
// CHECK: llvm.mlir.undef
// CHECK: llvm.insertvalue
// CHECK: llvm.insertvalue
// CHECK: llvm.return

// CHECK-BARE-LABEL: llvm.func @memref_signature(
// CHECK-BARE-SAME: %{{.*}}: !llvm.ptr, %{{.*}}: f32) -> !llvm.struct<(ptr, f32)>
gpu.func @memref_signature(%m: memref<2xf32>, %f: f32) -> (memref<2xf32>, f32) {
gpu.return %m, %f : memref<2xf32>, f32
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This example is nice. I think here gpu.func is a regular device function, so it can return. However, a GPU kernel cannot return anything. For nvidia, we differentiate kernel when nvvm.kernel attribute is present. I'm not sure how does it work for other models.
Does it make sense to add a verifier for that? What do you think?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

That makes sense. We already verify that for gpu.func but not for llvm.func. See #96399.

}
}


module attributes {transform.with_named_sequence} {
transform.named_sequence @__transform_main(%toplevel_module: !transform.any_op {transform.readonly}) {
%gpu_module = transform.structured.match ops{["gpu.module"]} in %toplevel_module
Expand All @@ -701,9 +724,7 @@ module attributes {transform.with_named_sequence} {
} with type_converter {
transform.apply_conversion_patterns.memref.memref_to_llvm_type_converter
{index_bitwidth = 64,
use_bare_ptr = true,
Copy link
Member Author

@matthias-springer matthias-springer Jun 22, 2024 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Note: These attributes do not exist.

use_bare_ptr_memref_call_conv = true,
use_opaque_pointers = true}
use_bare_ptr_call_conv = false}
} {
legal_dialects = ["llvm", "memref", "nvvm", "test"],
legal_ops = ["func.func", "gpu.module", "gpu.module_end", "gpu.yield"],
Expand Down
Loading