Skip to content
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

Jump to bottom

[REVIEW] mdspan integration. #437

Merged
merged 15 commits into from
Feb 22, 2022
+1,254 −2
Merged

[REVIEW] mdspan integration. #437

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
15 commits
Select commit Hold shift + click to select a range
073d576
Implement mdarray and integrate mdspan.
trivialfis Jul 14, 2020
04c5b39
Move into detail.
trivialfis Jan 25, 2022
7b69dd0
Rename header to hpp.
trivialfis Jan 25, 2022
1b01e68
Factory methods.
trivialfis Jan 25, 2022
511fafa
pragma
trivialfis Jan 26, 2022
debcaba
Remove postfix in naming.
trivialfis Jan 26, 2022
297f9d4
checks.
trivialfis Jan 26, 2022
07be704
Optional matrix layout.
trivialfis Jan 28, 2022
e4fdf32
clang format.
trivialfis Jan 28, 2022
c28a8b3
Add alias for c/f contiguous.
trivialfis Jan 29, 2022
db9ec8a
scalar & rename.
trivialfis Jan 29, 2022
c02fbaf
Update cmake module.
trivialfis Feb 14, 2022
daf9955
Merge branch 'branch-22.04' into fea-mdspan
trivialfis Feb 14, 2022
96f3484
Add license.
trivialfis Feb 14, 2022
94dc609
Adding copyright to detail/mdarray.hpp
cjnolet Feb 17, 2022
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
6 changes: 4 additions & 2 deletions cpp/CMakeLists.txt
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -111,6 +111,7 @@ include(cmake/thirdparty/get_rmm.cmake)
include(cmake/thirdparty/get_cuco.cmake)
include(cmake/thirdparty/get_libcudacxx.cmake)
include(cmake/thirdparty/get_faiss.cmake)
include(cmake/thirdparty/get_mdspan.cmake)

if(BUILD_TESTS)
include(cmake/thirdparty/get_gtest.cmake)
Expand All @@ -137,7 +138,8 @@ target_link_libraries(raft INTERFACE
CUDA::cusparse
$<$<BOOL:${NVTX}>:CUDA::nvToolsExt>
rmm::rmm
cuco::cuco)
cuco::cuco
std::mdspan)

target_compile_definitions(raft INTERFACE $<$<BOOL:${NVTX}>:NVTX_ENABLED>)
target_compile_features(raft INTERFACE cxx_std_17 $<BUILD_INTERFACE:cuda_std_17>)
Expand Down Expand Up @@ -248,7 +250,7 @@ install(DIRECTORY include/raft/

# Temporary install of raft.hpp while the file is removed
install(FILES include/raft.hpp
DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/raft)
DESTINATION ${CMAKE_INSTALL_INCLUDEDIR}/raft)

##############################################################################
# - install export -----------------------------------------------------------
Expand Down
15 changes: 15 additions & 0 deletions cpp/cmake/thirdparty/get_mdspan.cmake
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,15 @@
function(find_and_configure_mdspan VERSION)
rapids_cpm_find(
mdspan ${VERSION}
GLOBAL_TARGETS std::mdspan
BUILD_EXPORT_SET raft-exports
INSTALL_EXPORT_SET raft-exports
CPM_ARGS
GIT_REPOSITORY https://github.com/trivialfis/mdspan
Copy link
Member

Choose a reason for hiding this comment

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

I guess we need to think about how to handle this: trivialfis/mdspan@f18e694 and whether those changes are absolutely necessary based on kokkos/mdspan#109 (comment).

cc @teju85 @seunghwak @achirkin @jrhemstad and @brycelelbach for thoughts here. We're trying to avoid forcing 64-bit indexing when we know we can safely use 32-bit (or smaller).

Copy link
Member

@teju85 teju85 Jan 26, 2022 •
edited
Loading

Choose a reason for hiding this comment

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

+1 for what Corey said.

There are now a few handful of important kernels in RAFT that are having register pressure. So, increased usage of registers can cause occupancy cliffs. Based on some crude numbers, perf loss due to this was somewhere around ~3-10%! Anything we can do to NOT affect register usage for these kernels will help. So, I'd certainly want the ability to be able to use 32b for indexing in such kernels, when possible.

Copy link
Member

Choose a reason for hiding this comment

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

We might also want to consider whether the mdspan fork should be moved into an official rapidsai repository for the time being.

GIT_TAG 0193f075e977cc5f3c957425fd899e53d598f524
OPTIONS "MDSPAN_ENABLE_CUDA ON"
"MDSPAN_CXX_STANDARD ON"
)
endfunction()

find_and_configure_mdspan(0.2.0)
Loading