Skip to content

[mlir] AMDGPUToROCDL: RawBufferOpLowering fixes #120642

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
Dec 20, 2024
Merged

[mlir] AMDGPUToROCDL: RawBufferOpLowering fixes #120642

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
20 changes: 14 additions & 6 deletions mlir/lib/Conversion/AMDGPUToROCDL/AMDGPUToROCDL.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -91,6 +91,13 @@ struct RawBufferOpLowering : public ConvertOpToLLVMPattern<GpuOp> {
Type llvmI32 = this->typeConverter->convertType(i32);
Type llvmI16 = this->typeConverter->convertType(rewriter.getI16Type());

auto toI32 = [&](Value val) -> Value {
if (val.getType() == llvmI32)
return val;

return rewriter.create<LLVM::TruncOp>(loc, llvmI32, val);
};

int64_t elementByteWidth = memrefType.getElementTypeBitWidth() / 8;
Value byteWidthConst = createI32Constant(rewriter, loc, elementByteWidth);

Expand Down Expand Up @@ -166,22 +173,22 @@ struct RawBufferOpLowering : public ConvertOpToLLVMPattern<GpuOp> {
Value stride = rewriter.create<LLVM::ConstantOp>(
loc, llvmI16, rewriter.getI16IntegerAttr(0));
Value numRecords;
if (memrefType.hasStaticShape()) {
if (memrefType.hasStaticShape() && memrefType.getLayout().isIdentity()) {
numRecords = createI32Constant(
rewriter, loc,
static_cast<int32_t>(memrefType.getNumElements() * elementByteWidth));
} else {
Value maxIndex;
for (uint32_t i = 0, e = memrefType.getRank(); i < e; ++i) {
Value size = memrefDescriptor.size(rewriter, loc, i);
Value stride = memrefDescriptor.stride(rewriter, loc, i);
Value size = toI32(memrefDescriptor.size(rewriter, loc, i));
Value stride = toI32(memrefDescriptor.stride(rewriter, loc, i));
stride = rewriter.create<LLVM::MulOp>(loc, stride, byteWidthConst);
Value maxThisDim = rewriter.create<LLVM::MulOp>(loc, size, stride);
maxIndex = maxIndex ? rewriter.create<LLVM::MaximumOp>(loc, maxIndex,
maxThisDim)
: maxThisDim;
}
numRecords = rewriter.create<LLVM::TruncOp>(loc, llvmI32, maxIndex);
numRecords = maxIndex;
}

// Flag word:
Expand Down Expand Up @@ -218,7 +225,8 @@ struct RawBufferOpLowering : public ConvertOpToLLVMPattern<GpuOp> {
Value strideOp;
if (ShapedType::isDynamic(strides[i])) {
strideOp = rewriter.create<LLVM::MulOp>(
loc, memrefDescriptor.stride(rewriter, loc, i), byteWidthConst);
loc, toI32(memrefDescriptor.stride(rewriter, loc, i)),
byteWidthConst);
} else {
strideOp =
createI32Constant(rewriter, loc, strides[i] * elementByteWidth);
Expand All @@ -240,7 +248,7 @@ struct RawBufferOpLowering : public ConvertOpToLLVMPattern<GpuOp> {
sgprOffset = createI32Constant(rewriter, loc, 0);
if (ShapedType::isDynamic(offset))
sgprOffset = rewriter.create<LLVM::AddOp>(
loc, memrefDescriptor.offset(rewriter, loc), sgprOffset);
loc, toI32(memrefDescriptor.offset(rewriter, loc)), sgprOffset);
else if (offset > 0)
sgprOffset = rewriter.create<LLVM::AddOp>(
loc, sgprOffset, createI32Constant(rewriter, loc, offset));
Expand Down
19 changes: 19 additions & 0 deletions mlir/test/Conversion/AMDGPUToROCDL/amdgpu-to-rocdl.mlir
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -30,6 +30,25 @@ func.func @gpu_gcn_raw_buffer_load_i32(%buf: memref<64xi32>, %idx: i32) -> i32 {
func.return %0 : i32
}

// CHECK-LABEL: func @gpu_gcn_raw_buffer_load_i32_strided
func.func @gpu_gcn_raw_buffer_load_i32_strided(%buf: memref<64xi32, strided<[?], offset: ?>>, %idx: i32) -> i32 {
// CHECK-DAG: %[[rstride:.*]] = llvm.mlir.constant(0 : i16)
// CHECK-DAG: %[[elem_size:.*]] = llvm.mlir.constant(4 : i32)
// CHECK: %[[size:.*]] = llvm.extractvalue %{{.*}}[3, 0] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>
// CHECK: %[[size32:.*]] = llvm.trunc %[[size]] : i64 to i32
// CHECK: %[[stride:.*]] = llvm.extractvalue %{{.*}}[4, 0] : !llvm.struct<(ptr, ptr, i64, array<1 x i64>, array<1 x i64>)>
// CHECK: %[[stride32:.*]] = llvm.trunc %[[stride]] : i64 to i32
// CHECK: %[[tmp:.*]] = llvm.mul %[[stride32]], %[[elem_size]] : i32
// CHECK: %[[numRecords:.*]] = llvm.mul %[[size32]], %[[tmp]] : i32
// GFX9: %[[flags:.*]] = llvm.mlir.constant(159744 : i32)
// RDNA: %[[flags:.*]] = llvm.mlir.constant(822243328 : i32)
// CHECK: %[[resource:.*]] = rocdl.make.buffer.rsrc %{{.*}}, %[[rstride]], %[[numRecords]], %[[flags]] : !llvm.ptr to <8>
// CHECK: %[[ret:.*]] = rocdl.raw.ptr.buffer.load %[[resource]], %{{.*}}, %{{.*}}, %{{.*}} : i32
// CHECK: return %[[ret]]
%0 = amdgpu.raw_buffer_load {boundsCheck = true} %buf[%idx] : memref<64xi32, strided<[?], offset: ?>>, i32 -> i32
func.return %0 : i32
}

// CHECK-LABEL: func @gpu_gcn_raw_buffer_load_i32_oob_off
func.func @gpu_gcn_raw_buffer_load_i32_oob_off(%buf: memref<64xi32>, %idx: i32) -> i32 {
// GFX9: %[[flags:.*]] = llvm.mlir.constant(159744 : i32)
Expand Down
Loading