-
Notifications
You must be signed in to change notification settings - Fork 87
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.
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
AmgxPgm cuda/hip/omp implementation and matrix/array utils #695
Conversation
Codecov Report
@@ Coverage Diff @@
## develop #695 +/- ##
===========================================
+ Coverage 94.05% 94.16% +0.10%
===========================================
Files 392 399 +7
Lines 30647 30829 +182
===========================================
+ Hits 28825 29029 +204
+ Misses 1822 1800 -22
Continue to review full report at Codecov.
|
785fa8f
to
3edc90d
Compare
3edc90d
to
be869fe
Compare
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM. I only have a few comments.
omp/multigrid/amgx_pgm_kernels.cpp
Outdated
agg_vals[i] = group; | ||
agg_vals[neighbor] = group; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This function gets me a bit confused for now. I didn't look too deep into it, but couldn't there be some concurrent access issues on these data as both agg_vals[i] and agg_vals[neighbor] could be accessed by different threads in at least two different places?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
it should be okay on the results aspect.
they always set the smaller index, so they should write the same result from two points.
I can do it with condition but I need to check what impact is.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It depends entirely on what the neighbors set look like. If it's symmetric (neighbor[i] = j and neighbor[j] = i), then there should be no issue as you say. But if you have cases where neighbor[i] = j, neighbor[j] = k and neighbor[k] = i (or anything else), then you could get different results depending on what gets executed first, I think?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This case you mentioned does not work in the previous one, either.
Those points with circular strong neighbor will be assigned in the last step.
The original one: the both side of strongest connection will set the smaller index as aggregation group.
Current one: only the smaller one proceed settin the aggregation of both side.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think especially considering the weaker memory model for C++ on ARM processors, we should give this some thought. There is a potential RAW race condition between the write in agg_vals[neighbor] = i;
and the read in agg_vals[i] == -1
. This might be especially problematic if the reads and writes are torn, i.e. implemented as two distinct memory transactions, and potentially even reordered.
If any (partial) write to agg_vals[neighbor]
become visible to the thread where i == neighbor
, the comparison will become false. I would argue though that this cannot cause any issues, since in the writing thread, i < neighbor
, so the reading thread would have i > neighbor
and thus would not write to memory itself. Does that sound right to you?
This will definitely produce ThreadSanitizer warnings, though, not sure how to silence those.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Sorry for late reply.
Yes, read may be after or before the write from the smaller threads and it will not cause any issue as the reason you said.
I am not sure how ThreadSanitizer warning shows on this part.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM! Only some minor things.
be869fe
to
b9f8993
Compare
format! |
Kudos, SonarCloud Quality Gate passed! |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
First part of my review, there are a few questions and suggestions.
I still need to take a closer look at the GPU implementation of the coarse matrix computation
// This is the undeterminstic implementation which is the same implementation of | ||
// the previous one but agg_val == agg_const_val. |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Would it be possible to remove the __restrict__
modifier from the two agg_val pointers and reused the same implementation?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes, it could be, but it will lose a little possibility such that compiler to optimize it
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
You could also go with a macro approach to reduce duplication, but I am not sure that makes the code very readable.
omp/multigrid/amgx_pgm_kernels.cpp
Outdated
vector<map<IndexType, ValueType>> row_list( | ||
source_nrows, map<IndexType, ValueType>{exec}, exec); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This looks a bit inefficient and hard to parallelize. You are doing something SpGEMM-like here, right? Do you think we could instead turn it into a two-pass algorithm (first compute row pointers, allocate, then fill the entries?) Do you have any preconditions on the data we could use? Are the column indices sorted?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
It should be almost the reference implementation.
I do not spend too much time for optimizing them (including CUDA generation)
the column index is not sorted and maybe repeated
cb8ca7e
to
9a0f58b
Compare
It can be reviewed again |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
LGTM!
6fc8f67
to
e095dcb
Compare
- symmetric/hermitian/diag_dominant/hpd - array generator - add type test and documents for all in core/test/utils Remind to myself: class member is initialized by declaring order
Co-authored-by: Pratik Nayak <pratikvn@protonmail.com> Co-authored-by: Tobias Ribizel <ribizel@kit.edu>
Co-authored-by: Tobias Ribizel <ribizel@kit.edu>
need to improve it with less memory footprint
Kudos, SonarCloud Quality Gate passed! |
Ginkgo release 1.4.0 The Ginkgo team is proud to announce the new Ginkgo minor release 1.4.0. This release brings most of the Ginkgo functionality to the Intel DPC++ ecosystem which enables Intel-GPU and CPU execution. The only Ginkgo features which have not been ported yet are some preconditioners. Ginkgo's mixed-precision support is greatly enhanced thanks to: 1. The new Accessor concept, which allows writing kernels featuring on-the-fly memory compression, among other features. The accessor can be used as header-only, see the [accessor BLAS benchmarks repository](https://github.com/ginkgo-project/accessor-BLAS/tree/develop) as a usage example. 2. All LinOps now transparently support mixed-precision execution. By default, this is done through a temporary copy which may have a performance impact but already allows mixed-precision research. Native mixed-precision ELL kernels are implemented which do not see this cost. The accessor is also leveraged in a new CB-GMRES solver which allows for performance improvements by compressing the Krylov basis vectors. Many other features have been added to Ginkgo, such as reordering support, a new IDR solver, Incomplete Cholesky preconditioner, matrix assembly support (only CPU for now), machine topology information, and more! Supported systems and requirements: + For all platforms, cmake 3.13+ + C++14 compliant compiler + Linux and MacOS + gcc: 5.3+, 6.3+, 7.3+, all versions after 8.1+ + clang: 3.9+ + Intel compiler: 2018+ + Apple LLVM: 8.0+ + CUDA module: CUDA 9.0+ + HIP module: ROCm 3.5+ + DPC++ module: Intel OneAPI 2021.3. Set the CXX compiler to `dpcpp`. + Windows + MinGW and Cygwin: gcc 5.3+, 6.3+, 7.3+, all versions after 8.1+ + Microsoft Visual Studio: VS 2019 + CUDA module: CUDA 9.0+, Microsoft Visual Studio + OpenMP module: MinGW or Cygwin. Algorithm and important feature additions: + Add a new DPC++ Executor for SYCL execution and other base utilities [#648](#648), [#661](#661), [#757](#757), [#832](#832) + Port matrix formats, solvers and related kernels to DPC++. For some kernels, also make use of a shared kernel implementation for all executors (except Reference). [#710](#710), [#799](#799), [#779](#779), [#733](#733), [#844](#844), [#843](#843), [#789](#789), [#845](#845), [#849](#849), [#855](#855), [#856](#856) + Add accessors which allow multi-precision kernels, among other things. [#643](#643), [#708](#708) + Add support for mixed precision operations through apply in all LinOps. [#677](#677) + Add incomplete Cholesky factorizations and preconditioners as well as some improvements to ILU. [#672](#672), [#837](#837), [#846](#846) + Add an AMGX implementation and kernels on all devices but DPC++. [#528](#528), [#695](#695), [#860](#860) + Add a new mixed-precision capability solver, Compressed Basis GMRES (CB-GMRES). [#693](#693), [#763](#763) + Add the IDR(s) solver. [#620](#620) + Add a new fixed-size block CSR matrix format (for the Reference executor). [#671](#671), [#730](#730) + Add native mixed-precision support to the ELL format. [#717](#717), [#780](#780) + Add Reverse Cuthill-McKee reordering [#500](#500), [#649](#649) + Add matrix assembly support on CPUs. [#644](#644) + Extends ISAI from triangular to general and spd matrices. [#690](#690) Other additions: + Add the possibility to apply real matrices to complex vectors. [#655](#655), [#658](#658) + Add functions to compute the absolute of a matrix format. [#636](#636) + Add symmetric permutation and improve existing permutations. [#684](#684), [#657](#657), [#663](#663) + Add a MachineTopology class with HWLOC support [#554](#554), [#697](#697) + Add an implicit residual norm criterion. [#702](#702), [#818](#818), [#850](#850) + Row-major accessor is generalized to more than 2 dimensions and a new "block column-major" accessor has been added. [#707](#707) + Add an heat equation example. [#698](#698), [#706](#706) + Add ccache support in CMake and CI. [#725](#725), [#739](#739) + Allow tuning and benchmarking variables non intrusively. [#692](#692) + Add triangular solver benchmark [#664](#664) + Add benchmarks for BLAS operations [#772](#772), [#829](#829) + Add support for different precisions and consistent index types in benchmarks. [#675](#675), [#828](#828) + Add a Github bot system to facilitate development and PR management. [#667](#667), [#674](#674), [#689](#689), [#853](#853) + Add Intel (DPC++) CI support and enable CI on HPC systems. [#736](#736), [#751](#751), [#781](#781) + Add ssh debugging for Github Actions CI. [#749](#749) + Add pipeline segmentation for better CI speed. [#737](#737) Changes: + Add a Scalar Jacobi specialization and kernels. [#808](#808), [#834](#834), [#854](#854) + Add implicit residual log for solvers and benchmarks. [#714](#714) + Change handling of the conjugate in the dense dot product. [#755](#755) + Improved Dense stride handling. [#774](#774) + Multiple improvements to the OpenMP kernels performance, including COO, an exclusive prefix sum, and more. [#703](#703), [#765](#765), [#740](#740) + Allow specialization of submatrix and other dense creation functions in solvers. [#718](#718) + Improved Identity constructor and treatment of rectangular matrices. [#646](#646) + Allow CUDA/HIP executors to select allocation mode. [#758](#758) + Check if executors share the same memory. [#670](#670) + Improve test install and smoke testing support. [#721](#721) + Update the JOSS paper citation and add publications in the documentation. [#629](#629), [#724](#724) + Improve the version output. [#806](#806) + Add some utilities for dim and span. [#821](#821) + Improved solver and preconditioner benchmarks. [#660](#660) + Improve benchmark timing and output. [#669](#669), [#791](#791), [#801](#801), [#812](#812) Fixes: + Sorting fix for the Jacobi preconditioner. [#659](#659) + Also log the first residual norm in CGS [#735](#735) + Fix BiCG and HIP CSR to work with complex matrices. [#651](#651) + Fix Coo SpMV on strided vectors. [#807](#807) + Fix segfault of extract_diagonal, add short-and-fat test. [#769](#769) + Fix device_reset issue by moving counter/mutex to device. [#810](#810) + Fix `EnableLogging` superclass. [#841](#841) + Support ROCm 4.1.x and breaking HIP_PLATFORM changes. [#726](#726) + Decreased test size for a few device tests. [#742](#742) + Fix multiple issues with our CMake HIP and RPATH setup. [#712](#712), [#745](#745), [#709](#709) + Cleanup our CMake installation step. [#713](#713) + Various simplification and fixes to the Windows CMake setup. [#720](#720), [#785](#785) + Simplify third-party integration. [#786](#786) + Improve Ginkgo device arch flags management. [#696](#696) + Other fixes and improvements to the CMake setup. [#685](#685), [#792](#792), [#705](#705), [#836](#836) + Clarification of dense norm documentation [#784](#784) + Various development tools fixes and improvements [#738](#738), [#830](#830), [#840](#840) + Make multiple operators/constructors explicit. [#650](#650), [#761](#761) + Fix some issues, memory leaks and warnings found by MSVC. [#666](#666), [#731](#731) + Improved solver memory estimates and consistent iteration counts [#691](#691) + Various logger improvements and fixes [#728](#728), [#743](#743), [#754](#754) + Fix for ForwardIterator requirements in iterator_factory. [#665](#665) + Various benchmark fixes. [#647](#647), [#673](#673), [#722](#722) + Various CI fixes and improvements. [#642](#642), [#641](#641), [#795](#795), [#783](#783), [#793](#793), [#852](#852) Related PR: #857
Release 1.4.0 to master The Ginkgo team is proud to announce the new Ginkgo minor release 1.4.0. This release brings most of the Ginkgo functionality to the Intel DPC++ ecosystem which enables Intel-GPU and CPU execution. The only Ginkgo features which have not been ported yet are some preconditioners. Ginkgo's mixed-precision support is greatly enhanced thanks to: 1. The new Accessor concept, which allows writing kernels featuring on-the-fly memory compression, among other features. The accessor can be used as header-only, see the [accessor BLAS benchmarks repository](https://github.com/ginkgo-project/accessor-BLAS/tree/develop) as a usage example. 2. All LinOps now transparently support mixed-precision execution. By default, this is done through a temporary copy which may have a performance impact but already allows mixed-precision research. Native mixed-precision ELL kernels are implemented which do not see this cost. The accessor is also leveraged in a new CB-GMRES solver which allows for performance improvements by compressing the Krylov basis vectors. Many other features have been added to Ginkgo, such as reordering support, a new IDR solver, Incomplete Cholesky preconditioner, matrix assembly support (only CPU for now), machine topology information, and more! Supported systems and requirements: + For all platforms, cmake 3.13+ + C++14 compliant compiler + Linux and MacOS + gcc: 5.3+, 6.3+, 7.3+, all versions after 8.1+ + clang: 3.9+ + Intel compiler: 2018+ + Apple LLVM: 8.0+ + CUDA module: CUDA 9.0+ + HIP module: ROCm 3.5+ + DPC++ module: Intel OneAPI 2021.3. Set the CXX compiler to `dpcpp`. + Windows + MinGW and Cygwin: gcc 5.3+, 6.3+, 7.3+, all versions after 8.1+ + Microsoft Visual Studio: VS 2019 + CUDA module: CUDA 9.0+, Microsoft Visual Studio + OpenMP module: MinGW or Cygwin. Algorithm and important feature additions: + Add a new DPC++ Executor for SYCL execution and other base utilities [#648](#648), [#661](#661), [#757](#757), [#832](#832) + Port matrix formats, solvers and related kernels to DPC++. For some kernels, also make use of a shared kernel implementation for all executors (except Reference). [#710](#710), [#799](#799), [#779](#779), [#733](#733), [#844](#844), [#843](#843), [#789](#789), [#845](#845), [#849](#849), [#855](#855), [#856](#856) + Add accessors which allow multi-precision kernels, among other things. [#643](#643), [#708](#708) + Add support for mixed precision operations through apply in all LinOps. [#677](#677) + Add incomplete Cholesky factorizations and preconditioners as well as some improvements to ILU. [#672](#672), [#837](#837), [#846](#846) + Add an AMGX implementation and kernels on all devices but DPC++. [#528](#528), [#695](#695), [#860](#860) + Add a new mixed-precision capability solver, Compressed Basis GMRES (CB-GMRES). [#693](#693), [#763](#763) + Add the IDR(s) solver. [#620](#620) + Add a new fixed-size block CSR matrix format (for the Reference executor). [#671](#671), [#730](#730) + Add native mixed-precision support to the ELL format. [#717](#717), [#780](#780) + Add Reverse Cuthill-McKee reordering [#500](#500), [#649](#649) + Add matrix assembly support on CPUs. [#644](#644) + Extends ISAI from triangular to general and spd matrices. [#690](#690) Other additions: + Add the possibility to apply real matrices to complex vectors. [#655](#655), [#658](#658) + Add functions to compute the absolute of a matrix format. [#636](#636) + Add symmetric permutation and improve existing permutations. [#684](#684), [#657](#657), [#663](#663) + Add a MachineTopology class with HWLOC support [#554](#554), [#697](#697) + Add an implicit residual norm criterion. [#702](#702), [#818](#818), [#850](#850) + Row-major accessor is generalized to more than 2 dimensions and a new "block column-major" accessor has been added. [#707](#707) + Add an heat equation example. [#698](#698), [#706](#706) + Add ccache support in CMake and CI. [#725](#725), [#739](#739) + Allow tuning and benchmarking variables non intrusively. [#692](#692) + Add triangular solver benchmark [#664](#664) + Add benchmarks for BLAS operations [#772](#772), [#829](#829) + Add support for different precisions and consistent index types in benchmarks. [#675](#675), [#828](#828) + Add a Github bot system to facilitate development and PR management. [#667](#667), [#674](#674), [#689](#689), [#853](#853) + Add Intel (DPC++) CI support and enable CI on HPC systems. [#736](#736), [#751](#751), [#781](#781) + Add ssh debugging for Github Actions CI. [#749](#749) + Add pipeline segmentation for better CI speed. [#737](#737) Changes: + Add a Scalar Jacobi specialization and kernels. [#808](#808), [#834](#834), [#854](#854) + Add implicit residual log for solvers and benchmarks. [#714](#714) + Change handling of the conjugate in the dense dot product. [#755](#755) + Improved Dense stride handling. [#774](#774) + Multiple improvements to the OpenMP kernels performance, including COO, an exclusive prefix sum, and more. [#703](#703), [#765](#765), [#740](#740) + Allow specialization of submatrix and other dense creation functions in solvers. [#718](#718) + Improved Identity constructor and treatment of rectangular matrices. [#646](#646) + Allow CUDA/HIP executors to select allocation mode. [#758](#758) + Check if executors share the same memory. [#670](#670) + Improve test install and smoke testing support. [#721](#721) + Update the JOSS paper citation and add publications in the documentation. [#629](#629), [#724](#724) + Improve the version output. [#806](#806) + Add some utilities for dim and span. [#821](#821) + Improved solver and preconditioner benchmarks. [#660](#660) + Improve benchmark timing and output. [#669](#669), [#791](#791), [#801](#801), [#812](#812) Fixes: + Sorting fix for the Jacobi preconditioner. [#659](#659) + Also log the first residual norm in CGS [#735](#735) + Fix BiCG and HIP CSR to work with complex matrices. [#651](#651) + Fix Coo SpMV on strided vectors. [#807](#807) + Fix segfault of extract_diagonal, add short-and-fat test. [#769](#769) + Fix device_reset issue by moving counter/mutex to device. [#810](#810) + Fix `EnableLogging` superclass. [#841](#841) + Support ROCm 4.1.x and breaking HIP_PLATFORM changes. [#726](#726) + Decreased test size for a few device tests. [#742](#742) + Fix multiple issues with our CMake HIP and RPATH setup. [#712](#712), [#745](#745), [#709](#709) + Cleanup our CMake installation step. [#713](#713) + Various simplification and fixes to the Windows CMake setup. [#720](#720), [#785](#785) + Simplify third-party integration. [#786](#786) + Improve Ginkgo device arch flags management. [#696](#696) + Other fixes and improvements to the CMake setup. [#685](#685), [#792](#792), [#705](#705), [#836](#836) + Clarification of dense norm documentation [#784](#784) + Various development tools fixes and improvements [#738](#738), [#830](#830), [#840](#840) + Make multiple operators/constructors explicit. [#650](#650), [#761](#761) + Fix some issues, memory leaks and warnings found by MSVC. [#666](#666), [#731](#731) + Improved solver memory estimates and consistent iteration counts [#691](#691) + Various logger improvements and fixes [#728](#728), [#743](#743), [#754](#754) + Fix for ForwardIterator requirements in iterator_factory. [#665](#665) + Various benchmark fixes. [#647](#647), [#673](#673), [#722](#722) + Various CI fixes and improvements. [#642](#642), [#641](#641), [#795](#795), [#783](#783), [#793](#793), [#852](#852) Related PR: #866
This PR implements the amgxpgm device code with aggregation size = 2 except for dpc++.
Also, collect those utils in core/test/utils
TODO: