Use unified memory in CUDA debug builds #621
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
upsj
added
is:bug
upsj
changed the title
upsj
force-pushed
the
executor_debug_uvm
branch
from
e1d1f8f
to
6d90ba0
Compare
Codecov Report
@@ Coverage Diff @@
## develop #621 +/- ##
===========================================
- Coverage 92.86% 92.86% -0.01%
===========================================
Files 303 303
Lines 21115 21115
===========================================
- Hits 19609 19608 -1
- Misses 1506 1507 +1
Continue to review full report at Codecov.
|
upsj
changed the title
|
I definitely like the idea of using UVM in debug builds, but I am not so sure if we should be implementing two different implementations for Release and Debug. To me debug builds are meant to have the same implementation but maybe with more possible checks and information, but in this case the implementation (std::memcpy vs CudaMemcpyPeer) is different which kind of violates that principle. Nice job finding that bug for the multi-GPU raw_copy_to. I guess that was the issue with my CudaUVM memspace invalid device ordinal errors. Also you dont seem to be doing this for HIP device -->HIP device raw_copy_to. |
|
That's a good point, it might also be that you can just use cudaMemcpy as usual with UVM, I will check. |
upsj
force-pushed
the
executor_debug_uvm
branch
2 times, most recently
from
2709bba
to
50a6db6
Compare
also fix the incorrect peer memcpy device ID order
upsj
force-pushed
the
executor_debug_uvm
branch
from
aef52b1
to
15faee3
Compare
upsj
added
1:ST:ready-to-merge
|
Kudos, SonarCloud Quality Gate passed!
|
tcojean
added a commit
that referenced
this pull request
Aug 26, 2020
Release 1.3.0 of Ginkgo. The Ginkgo team is proud to announce the new minor release of Ginkgo version 1.3.0. This release brings CUDA 11 support, changes the default C++ standard to be C++14 instead of C++11, adds a new Diagonal matrix format and capacity for diagonal extraction, significantly improves the CMake configuration output format, adds the Ginkgo paper which got accepted into the Journal of Open Source Software (JOSS), and fixes multiple issues. Supported systems and requirements: + For all platforms, cmake 3.9+ + Linux and MacOS + gcc: 5.3+, 6.3+, 7.3+, all versions after 8.1+ + clang: 3.9+ + Intel compiler: 2017+ + Apple LLVM: 8.0+ + CUDA module: CUDA 9.0+ + HIP module: ROCm 2.8+ + Windows + MinGW and Cygwin: gcc 5.3+, 6.3+, 7.3+, all versions after 8.1+ + Microsoft Visual Studio: VS 2017 15.7+ + CUDA module: CUDA 9.0+, Microsoft Visual Studio + OpenMP module: MinGW or Cygwin. The current known issues can be found in the [known issues page](https://github.com/ginkgo-project/ginkgo/wiki/Known-Issues). Additions: + Add paper for Journal of Open Source Software (JOSS). [#479](#479) + Add a DiagonalExtractable interface. [#563](#563) + Add a new diagonal Matrix Format. [#580](#580) + Add Cuda11 support. [#603](#603) + Add information output after CMake configuration. [#610](#610) + Add a new preconditioner export example. [#595](#595) + Add a new cuda-memcheck CI job. [#592](#592) Changes: + Use unified memory in CUDA debug builds. [#621](#621) + Improve `BENCHMARKING.md` with more detailed info. [#619](#619) + Use C++14 standard instead of C++11. [#611](#611) + Update the Ampere sm information and CudaArchitectureSelector. [#588](#588) Fixes: + Fix documentation warnings and errors. [#624](#624) + Fix warnings for diagonal matrix format. [#622](#622) + Fix criterion factory parameters in CUDA. [#586](#586) + Fix the norm-type in the examples. [#612](#612) + Fix the WAW race in OpenMP is_sorted_by_column_index. [#617](#617) + Fix the example's exec_map by creating the executor only if requested. [#602](#602) + Fix some CMake warnings. [#614](#614) + Fix Windows building documentation. [#601](#601) + Warn when CXX and CUDA host compiler do not match. [#607](#607) + Fix reduce_add, prefix_sum, and doc-build. [#593](#593) + Fix find_library(cublas) issue on machines installing multiple cuda. [#591](#591) + Fix allocator in sellp read. [#589](#589) + Fix the CAS with HIP and NVIDIA backends. [#585](#585) Deletions: + Remove unused preconditioner parameter in LowerTrs. [#587](#587) Related PR: #625
tcojean
added a commit
that referenced
this pull request
Aug 27, 2020
The Ginkgo team is proud to announce the new minor release of Ginkgo version 1.3.0. This release brings CUDA 11 support, changes the default C++ standard to be C++14 instead of C++11, adds a new Diagonal matrix format and capacity for diagonal extraction, significantly improves the CMake configuration output format, adds the Ginkgo paper which got accepted into the Journal of Open Source Software (JOSS), and fixes multiple issues. Supported systems and requirements: + For all platforms, cmake 3.9+ + Linux and MacOS + gcc: 5.3+, 6.3+, 7.3+, all versions after 8.1+ + clang: 3.9+ + Intel compiler: 2017+ + Apple LLVM: 8.0+ + CUDA module: CUDA 9.0+ + HIP module: ROCm 2.8+ + Windows + MinGW and Cygwin: gcc 5.3+, 6.3+, 7.3+, all versions after 8.1+ + Microsoft Visual Studio: VS 2017 15.7+ + CUDA module: CUDA 9.0+, Microsoft Visual Studio + OpenMP module: MinGW or Cygwin. The current known issues can be found in the [known issues page](https://github.com/ginkgo-project/ginkgo/wiki/Known-Issues). Additions: + Add paper for Journal of Open Source Software (JOSS). [#479](#479) + Add a DiagonalExtractable interface. [#563](#563) + Add a new diagonal Matrix Format. [#580](#580) + Add Cuda11 support. [#603](#603) + Add information output after CMake configuration. [#610](#610) + Add a new preconditioner export example. [#595](#595) + Add a new cuda-memcheck CI job. [#592](#592) Changes: + Use unified memory in CUDA debug builds. [#621](#621) + Improve `BENCHMARKING.md` with more detailed info. [#619](#619) + Use C++14 standard instead of C++11. [#611](#611) + Update the Ampere sm information and CudaArchitectureSelector. [#588](#588) Fixes: + Fix documentation warnings and errors. [#624](#624) + Fix warnings for diagonal matrix format. [#622](#622) + Fix criterion factory parameters in CUDA. [#586](#586) + Fix the norm-type in the examples. [#612](#612) + Fix the WAW race in OpenMP is_sorted_by_column_index. [#617](#617) + Fix the example's exec_map by creating the executor only if requested. [#602](#602) + Fix some CMake warnings. [#614](#614) + Fix Windows building documentation. [#601](#601) + Warn when CXX and CUDA host compiler do not match. [#607](#607) + Fix reduce_add, prefix_sum, and doc-build. [#593](#593) + Fix find_library(cublas) issue on machines installing multiple cuda. [#591](#591) + Fix allocator in sellp read. [#589](#589) + Fix the CAS with HIP and NVIDIA backends. [#585](#585) Deletions: + Remove unused preconditioner parameter in LowerTrs. [#587](#587) Related PR: #627
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
This PR uses unified memory instead of device memory in CUDA debug builds, allowing you to directly access device memory in
cuda-gdb.It additionally fixes a bug in multi-GPU raw_copy_to, which used the wrong device ordinals due to a wrongly named parameter (src -> dest)