Skip to content

[mlir][nvgpu] Use the strides of the memref descriptor to construct the TMA descriptor #85403

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

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open

[mlir][nvgpu] Use the strides of the memref descriptor to construct the TMA descriptor #85403

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
32 changes: 19 additions & 13 deletions mlir/lib/ExecutionEngine/CudaRuntimeWrappers.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -427,13 +427,21 @@ namespace {

template <int rank>
void mgpuGetMemRefDataAndShape(void *raw_descriptor, char **addr,
uint64_t *globalDim) {
uint64_t *globalDim, uint64_t *globalStrides,
const CUtensorMapDataType tensorDataType) {
auto descriptor =
reinterpret_cast<StridedMemRefType<char, rank> *>(raw_descriptor);
*addr = descriptor->data;
for (int i = 0; i < rank; ++i) {
globalDim[i] = static_cast<uint64_t>(descriptor->sizes[rank - i - 1]);
}
static constexpr int elementSizeInBytes[] = {1, 2, 4, 4, 8, 8, 2,
4, 8, 2, 4, 4, 4};
// TODO(grypp): Check that the minormost stride is equal to the element size.
Copy link
Member

Choose a reason for hiding this comment

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

LLVM doesn't use TODO with name. Let's just keep this as TODO

for (int i = 0; i < rank - 1; ++i) {
globalStrides[i] = static_cast<uint64_t>(
descriptor->strides[rank - i - 2] * elementSizeInBytes[tensorDataType]);
}
}

} // namespace
Expand All @@ -457,19 +465,24 @@ extern "C" MLIR_CUDA_WRAPPERS_EXPORT void *mgpuTensorMapEncodeTiledMemref(
char *globalAddress = nullptr;
switch (tensorRank) {
case 1:
mgpuGetMemRefDataAndShape<1>(ranked_descriptor, &globalAddress, globalDim);
mgpuGetMemRefDataAndShape<1>(ranked_descriptor, &globalAddress, globalDim,
globalStrides, tensorDataType);
break;
case 2:
mgpuGetMemRefDataAndShape<2>(ranked_descriptor, &globalAddress, globalDim);
mgpuGetMemRefDataAndShape<2>(ranked_descriptor, &globalAddress, globalDim,
globalStrides, tensorDataType);
break;
case 3:
mgpuGetMemRefDataAndShape<3>(ranked_descriptor, &globalAddress, globalDim);
mgpuGetMemRefDataAndShape<3>(ranked_descriptor, &globalAddress, globalDim,
globalStrides, tensorDataType);
break;
case 4:
mgpuGetMemRefDataAndShape<4>(ranked_descriptor, &globalAddress, globalDim);
mgpuGetMemRefDataAndShape<4>(ranked_descriptor, &globalAddress, globalDim,
globalStrides, tensorDataType);
break;
case 5:
mgpuGetMemRefDataAndShape<5>(ranked_descriptor, &globalAddress, globalDim);
mgpuGetMemRefDataAndShape<5>(ranked_descriptor, &globalAddress, globalDim,
globalStrides, tensorDataType);
break;
default:
fprintf(
Expand All @@ -478,17 +491,10 @@ extern "C" MLIR_CUDA_WRAPPERS_EXPORT void *mgpuTensorMapEncodeTiledMemref(
return NULL;
}

static const int elementSizeInBytes[] = {1, 2, 4, 4, 8, 8, 2,
4, 8, 2, 4, 4, 4};
for (int64_t r = 0; r < tensorRank; ++r) {
elementStrides[r] = uint32_t(1);
boxDim[r] = static_cast<uint32_t>(inputBoxDims[tensorRank - r - 1]);
}

globalStrides[0] = globalDim[0] * elementSizeInBytes[tensorDataType];
for (int r = 1; r < tensorRank - 1; r++)
globalStrides[r] = globalStrides[r - 1] * globalDim[r];

ScopedContext scopedContext;
mgpuTensorMapEncodeTiled(&tensorMap, tensorDataType, tensorRank32,
globalAddress, globalDim, globalStrides, boxDim,
Expand Down