Remove the GSL dependency from sss and xlqc by zjin-lcf · Pull Request #321 · ORNL/HeCBench

zjin-lcf · 2026-08-13T15:13:56Z

Resolves the GPL/CUDA-EULA license clash reported in #319 by taking option 1: both sss and xlqc no longer depend on GSL, so they stay in the suite and keep building by default.

sss

GSL was used only for a uniform random number generator (MT19937) and adaptive Gauss-Kronrod quadrature. Both are now implemented in a self-contained, dependency-free gsl_compat.h under src/sss-cuda/, so the benchmark needs nothing beyond a C++17 compiler. The gsl_* entry point names are kept so the surrounding sampler code is unchanged.

While in here:

The HIP and SYCL variants now share the host sources with the CUDA variant instead of carrying byte-identical copies (about 4.3k duplicated lines removed).
Fixed data races in the kernels that were previously masked.

xlqc

The Hartree-Fock SCF driver needs a dense symmetric eigensolver, matrix products and a linear solve for DIIS. These now come from Eigen (header-only, MPL2) behind a thin gsl_compat.h wrapper, replacing the GSL BLAS/eigen calls.

Eigen is discovered rather than hardcoded: the Makefiles use pkg-config --cflags eigen3 and CMake uses find_package(Eigen3), so a distro-installed Eigen (libeigen3-dev / eigen3-devel) needs no configuration. A non-standard prefix can be pointed at with PKG_CONFIG_PATH or make EIGEN_INC=-I/path/to/eigen3. If Eigen is absent, CMake skips xlqc with a message instead of failing.

Build system and docs

find_package(GSL) and the DEPEND_ON_GSL list are replaced by Eigen3 / DEPEND_ON_EIGEN (now xlqc only, since sss has no external dependency at all). The top-level README dependency table and the per-benchmark READMEs are updated accordingly.

Testing

All four programming-model variants of both benchmarks were built and run, and their output was checked against the pre-change GSL-based versions.

…g its path GSL is GPL, so a redistributed binary that links it together with the CUDA or HIP runtime cannot satisfy both licenses. Reported in #319 The SCF driver only needs a handful of GSL entry points: matrix and vector allocation and element access, dgemm, the symmetric eigensolver, and an LU solve for DIIS. gsl_compat.h maps exactly those onto Eigen (MPL2, header-only), keeping the existing gsl_* call sites in main, scf.cc and basis.cc unchanged. The header lives in xlqc-cuda and is shared by the hip, omp and sycl variants, which already compile those sources through -I../xlqc-cuda. Eigen's SelfAdjointEigenSolver returns ascending eigenvalues like gsl_eigen_symmv_sort, and PartialPivLU matches the pivoting of gsl_linalg_LU_decomp, so the SCF trajectory is preserved. The Makefiles carried the same "/path/to/..." placeholder that #305 replaces for GSL, so apply that fix to Eigen: EIGEN_INC defaults to `pkg-config --cflags eigen3` and a missing Eigen stops the build with an actionable message instead of a bare missing-header error. Eigen is header-only, so there is no library counterpart to GSL_LIB and LDFLAGS becomes empty. The check is skipped for `clean` so the tree can be cleaned without Eigen installed. CMake gains find_package(Eigen3) and guards xlqc on the Eigen3::Eigen imported target rather than a _FOUND variable, since that target is what carries the include path. xlqc-omp additionally never compiled the shared scf.cc, basis.cc and int_lib sources or unpacked its example data, so it could not link; both are fixed here. Co-authored-by: Cursor <cursoragent@cursor.com>

…x kernel data races GSL is GPL, so distributing sss binaries built from CUDA/HIP sources conflicts with the CUDA EULA (#319). sss only used GSL for a Mersenne Twister stream and one semi-infinite integral, so replace it with a self-contained gsl_compat.h that reproduces mt19937 with GSL's 2002 Knuth seeding and implements adaptive Gauss-Kronrod (QK15) quadrature from the public domain QUADPACK algorithm. The RNG stream is bit-identical to GSL 2.7.1 and the integral agrees to ~1e-12, so results are unchanged. sss now builds with no external dependency, so drop the GSL discovery and the guards that skipped the benchmark when GSL was absent. Deduplicate the variants: the files that carry no GPU API calls now live only in sss-cuda and are pulled in through an include path, removing about 4500 lines of copies. Only main, kernels, and DPmixGGM_SSSmoves stay per-variant, and sss-cuda/kernels.cu is now shared with sss-hip since the two were identical and it uses no CUDA-only names. Add the missing barriers in CanDeleteEdge and CanAddEdge. Both kernels let every thread write shared state and then had thread 0 read all of it without a barrier, and CanDeleteEdge tested a thread-0-only counter for its loop exit, so threads could leave the loop divergently. This was benign on NVIDIA, where a 32-thread block is one warp in lockstep, but on gfx908 thread 0 read stale values: sss-hip scored -5721.97 and reported num_allModels=2731275 against -5675.31 and 12504300 elsewhere, converging on a worse optimum after 4.6x fewer iterations and so also reporting a misleadingly short run time. All three variants now reproduce the CPU reference path iteration for iteration and write identical output. Also fix two sss-sycl reporting bugs: the stopwatch was restarted on every search restart, so wall_time only covered the final restart (0.0036s instead of 35s), and the initial progress line printed k before it was assigned. Co-authored-by: Cursor <cursoragent@cursor.com>

Copilot

Pull request overview

Removes GSL from sss and xlqc, resolving the license conflict while retaining default builds.

Changes:

Adds self-contained GSL-compatible RNG/quadrature for sss.
Replaces xlqc linear algebra with an Eigen-backed compatibility layer.
Consolidates duplicated sources and updates dependency discovery and documentation.

Reviewed changes

Copilot reviewed 54 out of 54 changed files in this pull request and generated 3 comments.

Show a summary per file

File	Description
`src/xlqc-sycl/README.md`	Documents Eigen dependency.
`src/xlqc-sycl/Makefile`	Replaces GSL flags with Eigen discovery.
`src/xlqc-sycl/main.cpp`	Uses compatibility layer and preserves cleanup on failure.
`src/xlqc-sycl/CMakeLists.txt`	Links Eigen target.
`src/xlqc-omp/README.md`	Documents Eigen dependency.
`src/xlqc-omp/Makefile.nvc`	Configures Eigen for NVHPC.
`src/xlqc-omp/Makefile.aomp`	Configures Eigen for AOMP.
`src/xlqc-omp/Makefile`	Configures Eigen for OpenMP.
`src/xlqc-omp/main.cpp`	Uses compatibility layer and returns validation status.
`src/xlqc-omp/CMakeLists.txt`	Builds shared XLQC sources with Eigen.
`src/xlqc-hip/README.md`	Documents Eigen dependency.
`src/xlqc-hip/Makefile`	Replaces GSL and adjusts HIP compilation.
`src/xlqc-hip/main.cu`	Uses compatibility layer and preserves cleanup.
`src/xlqc-hip/CMakeLists.txt`	Links Eigen for HIP.
`src/xlqc-cuda/scf.cc`	Routes SCF operations through compatibility APIs.
`src/xlqc-cuda/README.md`	Documents Eigen setup.
`src/xlqc-cuda/Makefile`	Replaces GSL flags with Eigen flags.
`src/xlqc-cuda/main.cu`	Uses compatibility layer and returns validation status.
`src/xlqc-cuda/gsl_compat.h`	Implements Eigen-backed GSL subset.
`src/xlqc-cuda/CMakeLists.txt`	Links Eigen for CUDA.
`src/xlqc-cuda/basis.cc`	Removes obsolete GSL includes.
`src/sss-sycl/utilities.cpp`	Removes duplicated utility implementation.
`src/sss-sycl/README.md`	Documents dependency-free compatibility layer.
`src/sss-sycl/Makefile`	Uses shared CUDA host sources.
`src/sss-sycl/main.cpp`	Uses local GSL replacement and fixes initialization timing.
`src/sss-sycl/kernels.cpp`	Adds work-group synchronization.
`src/sss-sycl/gwish.cpp`	Removes duplicated implementation.
`src/sss-sycl/graph.h`	Removes duplicated declarations.
`src/sss-sycl/graph.cpp`	Removes duplicated graph implementation.
`src/sss-sycl/DPmixGGM.cpp`	Removes duplicated sampler implementation.
`src/sss-sycl/DPmixGGM_Lists.cpp`	Removes duplicated list implementation.
`src/sss-sycl/CMakeLists.txt`	Removes GSL and includes shared sources.
`src/sss-hip/utilities.cpp`	Removes duplicated utility implementation.
`src/sss-hip/README.md`	Documents dependency-free compatibility layer.
`src/sss-hip/Makefile`	Uses shared CUDA host sources.
`src/sss-hip/main.cu`	Uses local GSL replacement.
`src/sss-hip/kernels.cu`	Removes duplicated kernel implementation.
`src/sss-hip/gwish.cpp`	Removes duplicated implementation.
`src/sss-hip/graph.h`	Removes duplicated declarations.
`src/sss-hip/graph.cpp`	Removes duplicated graph implementation.
`src/sss-hip/DPmixGGM.cpp`	Removes duplicated sampler implementation.
`src/sss-hip/DPmixGGM_Lists.cpp`	Removes duplicated list implementation.
`src/sss-hip/CMakeLists.txt`	Removes GSL and includes shared sources.
`src/sss-cuda/README.md`	Documents self-contained dependencies.
`src/sss-cuda/Makefile`	Removes GSL build flags.
`src/sss-cuda/main.cu`	Uses local GSL replacement.
`src/sss-cuda/kernels.cu`	Adds block synchronization.
`src/sss-cuda/gwish.cpp`	Adds required math declarations.
`src/sss-cuda/gsl_compat.h`	Implements MT19937 and adaptive quadrature.
`src/sss-cuda/DPmixGGM.cpp`	Adds required math declarations.
`src/sss-cuda/CMakeLists.txt`	Removes GSL dependency.
`README.md`	Replaces GSL dependency listing with Eigen.
`CMakeLists.txt`	Discovers Eigen instead of GSL.
`cmake/modules/BenchmarkMacros.cmake`	Skips XLQC when Eigen is unavailable.

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

zjin-lcf · 2026-08-13T18:59:23Z

+------------
+None beyond a C++17 compiler. The uniform random number generator (MT19937) and
+the adaptive Gauss-Kronrod quadrature the sampler needs are implemented in
+`gsl_compat.h` in this folder.


The file was updated. Thanks.

zjin-lcf and others added 4 commits August 13, 2026 08:13

[xlqc] double check the source and build files

fc44fe2

[sss] double check gsl_compat.h

d2fc4a2

zjin-lcf requested a review from Geekdude August 13, 2026 15:48

Geekdude requested a balanced review from Copilot August 13, 2026 17:32

Copilot started reviewing on behalf of Geekdude August 13, 2026 17:33 View session

Geekdude approved these changes Aug 13, 2026

View reviewed changes

Copilot AI reviewed Aug 13, 2026

View reviewed changes

keitaTN merged commit 9fb2cc3 into master Aug 13, 2026
1 check passed

Geekdude deleted the gsl-license-319 branch August 13, 2026 18:05

Geekdude mentioned this pull request Aug 13, 2026

Combining CUDA binaries with GPL code violates the CUDA EULA. #319

Closed

Navigation Menu

Uh oh!

Remove the GSL dependency from sss and xlqc - #321

Remove the GSL dependency from sss and xlqc#321
keitaTN merged 4 commits into
masterfrom
gsl-license-319

zjin-lcf commented Aug 13, 2026

Uh oh!

Copilot AI left a comment

Uh oh!

Uh oh!

Uh oh!

zjin-lcf Aug 13, 2026

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants

Uh oh!

Conversation

zjin-lcf commented Aug 13, 2026

sss

xlqc

Build system and docs

Testing

Uh oh!

Copilot AI left a comment

Choose a reason for hiding this comment

Pull request overview

Reviewed changes

Uh oh!

Uh oh!

Uh oh!

zjin-lcf Aug 13, 2026

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

4 participants