Skip to content

Improves reduction code and removes kernels/boolean_reductions.hpp #1557

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 6 commits into from
Feb 22, 2024
Merged

Improves reduction code and removes kernels/boolean_reductions.hpp #1557

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
6 commits
Select commit Hold shift + click to select a range
40d57ef
Modularizes sequential reduction calls
ndgrigorian Feb 21, 2024
0c3633b
Moves reduction dispatching into individual reduction files
ndgrigorian Feb 21, 2024
dbe5746
py_boolean_reductions throws a value error when atomics are not suppo…
ndgrigorian Feb 22, 2024
ddb946a
Removes kernels/boolean_reductions.hpp
ndgrigorian Feb 22, 2024
5fa2dbf
Address PR feedback
ndgrigorian Feb 22, 2024
713a177
Adds `IsSycl[Op]` structs
ndgrigorian Feb 22, 2024
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
732 changes: 0 additions & 732 deletions dpctl/tensor/libtensor/include/kernels/boolean_reductions.hpp
View file
Open in desktop

This file was deleted.

253 changes: 127 additions & 126 deletions dpctl/tensor/libtensor/include/kernels/linalg_functions/dot_product.hpp
View file
Open in desktop

Large diffs are not rendered by default.

45 changes: 10 additions & 35 deletions dpctl/tensor/libtensor/include/kernels/linalg_functions/gemm.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2584,11 +2584,7 @@ gemm_batch_tree_nm_impl(sycl::queue &exec_q,
(reduction_nelems + preferred_reductions_per_wi * wg - 1) /
(preferred_reductions_per_wi * wg);

// max_max_wg prevents running out of resources on CPU
constexpr size_t max_max_wg = 2048;
size_t max_wg = std::min(
max_max_wg,
dev.get_info<sycl::info::device::max_work_group_size>() / 2);
size_t max_wg = reduction_detail::get_work_group_size(dev);

if (reduction_nelems <= preferred_reductions_per_wi * max_wg) {
resTy *tmp = sycl::malloc_device<resTy>(
Expand Down Expand Up @@ -2666,17 +2662,20 @@ gemm_batch_tree_nm_impl(sycl::queue &exec_q,
using OuterInnerDimsIndexerT =
dpctl::tensor::offset_utils::StridedIndexer;
using TmpIndexerT = dpctl::tensor::offset_utils::NoOpIndexer;

const OuterInnerDimsIndexerT lhs_indexer(
inner_nd + lhs_outer_nd, 0, lhs_outer_inner_shapes_strides);
const OuterInnerDimsIndexerT rhs_indexer(
inner_nd + rhs_outer_nd, 0, rhs_outer_inner_shapes_strides);
constexpr TmpIndexerT res_indexer{};

using dpctl::tensor::offset_utils::Strided1DIndexer;
using dpctl::tensor::offset_utils::StridedIndexer;
using dpctl::tensor::offset_utils::ThreeOffsets_CombinedIndexer;
using dpctl::tensor::offset_utils::UnpackedStridedIndexer;
using BatchDimsIndexerT = ThreeOffsets_CombinedIndexer<
StridedIndexer, UnpackedStridedIndexer, Strided1DIndexer>;

const StridedIndexer lhs_batch_indexer(batch_nd, lhs_batch_offset,
batch_shape_strides);
const UnpackedStridedIndexer rhs_batch_indexer(
Expand Down Expand Up @@ -2969,11 +2968,7 @@ gemm_batch_contig_tree_k_impl(sycl::queue &exec_q,
(reduction_nelems + preferred_reductions_per_wi * wg - 1) /
(preferred_reductions_per_wi * wg);

// max_max_wg prevents running out of resources on CPU
constexpr size_t max_max_wg = 2048;
size_t max_wg = std::min(
max_max_wg,
dev.get_info<sycl::info::device::max_work_group_size>() / 2);
size_t max_wg = reduction_detail::get_work_group_size(dev);

if (reduction_nelems <= preferred_reductions_per_wi * max_wg) {
resTy *tmp = sycl::malloc_device<resTy>(
Expand Down Expand Up @@ -3172,11 +3167,7 @@ gemm_batch_contig_tree_nm_impl(sycl::queue &exec_q,
(reduction_nelems + preferred_reductions_per_wi * wg - 1) /
(preferred_reductions_per_wi * wg);

// max_max_wg prevents running out of resources on CPU
constexpr size_t max_max_wg = 2048;
size_t max_wg = std::min(
max_max_wg,
dev.get_info<sycl::info::device::max_work_group_size>() / 2);
size_t max_wg = reduction_detail::get_work_group_size(dev);

if (reduction_nelems <= preferred_reductions_per_wi * max_wg) {
resTy *tmp = sycl::malloc_device<resTy>(
Expand Down Expand Up @@ -3558,11 +3549,7 @@ sycl::event gemm_tree_k_impl(sycl::queue &exec_q,
(reduction_nelems + preferred_reductions_per_wi * wg - 1) /
(preferred_reductions_per_wi * wg);

// max_max_wg prevents running out of resources on CPU
constexpr size_t max_max_wg = 2048;
size_t max_wg = std::min(
max_max_wg,
dev.get_info<sycl::info::device::max_work_group_size>() / 2);
size_t max_wg = reduction_detail::get_work_group_size(dev);

if (reduction_nelems <= preferred_reductions_per_wi * max_wg) {
resTy *tmp = sycl::malloc_device<resTy>(
Expand Down Expand Up @@ -3728,11 +3715,7 @@ sycl::event gemm_tree_nm_impl(sycl::queue &exec_q,
(reduction_nelems + preferred_reductions_per_wi * wg - 1) /
(preferred_reductions_per_wi * wg);

// max_max_wg prevents running out of resources on CPU
constexpr size_t max_max_wg = 2048;
size_t max_wg = std::min(
max_max_wg,
dev.get_info<sycl::info::device::max_work_group_size>() / 2);
size_t max_wg = reduction_detail::get_work_group_size(dev);

if (reduction_nelems <= preferred_reductions_per_wi * max_wg) {
resTy *tmp = sycl::malloc_device<resTy>(
Expand Down Expand Up @@ -3982,11 +3965,7 @@ sycl::event gemm_contig_tree_k_impl(sycl::queue &exec_q,
(reduction_nelems + preferred_reductions_per_wi * wg - 1) /
(preferred_reductions_per_wi * wg);

// max_max_wg prevents running out of resources on CPU
constexpr size_t max_max_wg = 2048;
size_t max_wg = std::min(
max_max_wg,
dev.get_info<sycl::info::device::max_work_group_size>() / 2);
size_t max_wg = reduction_detail::get_work_group_size(dev);

if (reduction_nelems <= preferred_reductions_per_wi * max_wg) {
resTy *tmp = sycl::malloc_device<resTy>(
Expand Down Expand Up @@ -4139,11 +4118,7 @@ sycl::event gemm_contig_tree_nm_impl(sycl::queue &exec_q,
(reduction_nelems + preferred_reductions_per_wi * wg - 1) /
(preferred_reductions_per_wi * wg);

// max_max_wg prevents running out of resources on CPU
constexpr size_t max_max_wg = 2048;
size_t max_wg = std::min(
max_max_wg,
dev.get_info<sycl::info::device::max_work_group_size>() / 2);
size_t max_wg = reduction_detail::get_work_group_size(dev);

if (reduction_nelems <= preferred_reductions_per_wi * max_wg) {
resTy *tmp = sycl::malloc_device<resTy>(
Expand Down
Loading