Skip to content

Fixed-size reductions #1061

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
cliffburdick merged 2 commits into from
Sep 30, 2025
Merged

Fixed-size reductions #1061

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
f68a62f
Use cub's new fixed size segmented reduce
cliffburdick Sep 12, 2025
a06a7c3
Added fixed-size segmented reduce
cliffburdick Sep 29, 2025
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
38 changes: 37 additions & 1 deletion include/matx/transforms/cub.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -659,12 +659,30 @@ inline void ExecSort(OutputTensor &a_out,
// type of reduction where there's not a single output, since any type of reduction can be generalized
// to a segmented type
if constexpr (OutputTensor::Rank() > 0) {
#if CUB_MAJOR_VERSION >= 3 && CUB_MINOR_VERSION >= 2
[[maybe_unused]] cudaError_t err;
if (is_tensor_view_v<InputOperator> && a.IsContiguous() && a_out.IsContiguous()) {
const int seg_size = static_cast<int>(TotalSize(a) / TotalSize(out_base));
err = cub::DeviceSegmentedReduce::Reduce(d_temp, temp_storage_bytes, in_base.Data(), out_base.Data(), static_cast<cuda::std::int64_t>(TotalSize(out_base)), seg_size, cparams_.reduce_op,
cparams_.init, stream);
}
else {
auto ft = [&](auto &&in, auto &&out, auto &&begin, auto &&end) {
return cub::DeviceSegmentedReduce::Reduce(d_temp, temp_storage_bytes, in, out, static_cast<int>(TotalSize(out_base)), begin, end, cparams_.reduce_op,
cparams_.init, stream);
};
err = ReduceInput(ft, out_base, in_base);
}

MATX_ASSERT_STR_EXP(err, cudaSuccess, matxCudaError, "Error in cub::DeviceSegmentedReduce::Reduce");
#else
auto ft = [&](auto &&in, auto &&out, auto &&begin, auto &&end) {
return cub::DeviceSegmentedReduce::Reduce(d_temp, temp_storage_bytes, in, out, static_cast<int>(TotalSize(out_base)), begin, end, cparams_.reduce_op,
cparams_.init, stream);
};
[[maybe_unused]] auto rv = ReduceInput(ft, out_base, in_base);
MATX_ASSERT_STR_EXP(rv, cudaSuccess, matxCudaError, "Error in cub::DeviceSegmentedReduce::Reduce");
#endif
}
else {
auto ft = [&](auto &&in, auto &&out, [[maybe_unused]] auto &&unused1, [[maybe_unused]] auto &&unused2) {
Expand Down Expand Up @@ -710,11 +728,29 @@ inline void ExecSort(OutputTensor &a_out,
// type of reduction where there's not a single output, since any type of reduction can be generalized
// to a segmented type
if constexpr (OutputTensor::Rank() > 0) {
// Check if fixed-size reductions are supported
#if CUB_MAJOR_VERSION >= 3 && CUB_MINOR_VERSION >= 2
[[maybe_unused]] cudaError_t err;
if (is_tensor_view_v<InputOperator> && a.IsContiguous() && a_out.IsContiguous()) {
const int seg_size = static_cast<int>(TotalSize(a) / TotalSize(out_base));
err = cub::DeviceSegmentedReduce::Sum(d_temp, temp_storage_bytes, in_base.Data(), out_base.Data(), static_cast<cuda::std::int64_t>(TotalSize(out_base)), seg_size, stream);
}
else {
auto ft = [&](auto &&in, auto &&out, auto &&begin, auto &&end) {
return cub::DeviceSegmentedReduce::Sum(d_temp, temp_storage_bytes, in, out, static_cast<int>(TotalSize(out_base)), begin, end, stream);
};

err = ReduceInput(ft, out_base, in_base);
}

MATX_ASSERT_STR_EXP(err, cudaSuccess, matxCudaError, "Error in cub::DeviceSegmentedReduce::Sum");
#else
auto ft = [&](auto &&in, auto &&out, auto &&begin, auto &&end) {
return cub::DeviceSegmentedReduce::Sum(d_temp, temp_storage_bytes, in, out, static_cast<int>(TotalSize(out_base)), begin, end, stream);
return cub::DeviceSegmentedReduce::Sum(d_temp, temp_storage_bytes, in, out, static_cast<int>(TotalSize(out_base)), begin, end, stream);
};
[[maybe_unused]] auto rv = ReduceInput(ft, out_base, in_base);
MATX_ASSERT_STR_EXP(rv, cudaSuccess, matxCudaError, "Error in cub::DeviceSegmentedReduce::Sum");
#endif
}
else {
auto ft = [&](auto &&in, auto &&out, [[maybe_unused]] auto &&unused1, [[maybe_unused]] auto &&unused2) {
Expand Down