LibDDLA v0.0.5

LibDDLA v0.0.5

C++17 distributed dense linear algebra library for GPUs (CUDA/HIP) and CPU, ScaLAPACK-style 2D block-cyclic APIs over MPI. This release covers 56 commits since v0.0.4 (2026-07-31 → 2026-09-04).

Highlights

• New status-returning C API — every public entry point now returns ddlaStatus_t, with DDLA_CHECK macros and ddlaSetStream for stream ownership on the handle; ddla.h + ddla_handle_t.h are now the only public headers, and all others were moved into src/. (c633490, 352077f)
• CPU backend support with unified DdlaBackend dispatch — the backend layer is now a template (gemm<Backend, T>, comm_traits.h) covering CPU (OpenBLAS), CUDA, and HIP, including an optional dual CPU+GPU build selectable per handle at runtime. (41c7ace, 83f844b, 691f1b1, c4ce301)
• Leading-block sub-matrix support for the factor/solve/diag routines, plus expanded solve variants. (2c76bc3, d8d1181)
• Automatic CI on two HPC clusters — every push now runs the CPU suite on HPC3 and a HIP/DCU (gfx906) 40-case suite on KSSG via GitHub Actions; the HPC3 GPU stage runs as parallel 4-GPU + 6-GPU jobs. (f9ae8f5, ffd0f6e)

New Features

• ptrtrs gained side='R' (X*op(A)=B) support, and the solve family gained side/trans variants. (49f3164, d8d1181)
• plapiv gained row/col pivoting direction support; head-location solves added. (d8d1181)
• CPU random-number generation in random_generator (backend-agnostic matrix generation). (13deeab)
• Benchmark suite scoped to POTRF paths. (300794b)

Bug Fixes

• pposv was missing the row permutation of B when the pivot location is not -1. (a397609)
• pgetrf_bpiv is now a standard LU factorization, and pgetrs_bpiv applies the column pivot in the correct direction. (829202f)
• Comm/plapiv/pivot correctness fixes and general hardening. (6dd4bb3)
• gemmVbatched/gemmVbatched2s now accept lowercase transpose modes. (79cf31c)
• Public headers are self-contained via DDLA_VENDOR_* macros; private headers include the public header first and no longer redeclare default template/function arguments. (dfdaa6e, 0391b81, 165981d)
• Stream sync before reading last_value in the ppotrf head-correction path.
• MPI_CHECK in the CCL+TUNNEL build branches; int/status comparison fixes across tests. (33c29e0, c7c43cb)

Build System / Tooling

• Unified backend dispatch with DdlaBackend templates; CPU/GPU backend reduction. (41c7ace, 691f1b1)
• Unified comm interface in comm_traits.h; dead comm code removed. (c4ce301, a82ce35)
• ROCm include dir added to the HIP build for CXX-language TUs. (9cf0cdf)
• README updated to the current API and comm layer, including the dual CPU+GPU build documentation. (7852315)

CI

• New GitHub Actions pipeline: HPC3 CPU validation + KSSG HIP/DCU validation (gfx906, DTK 25.04.4, RCCL) on every push to main/develop. (f9ae8f5)
• HPC3 GPU validation split into 4-GPU (v100g32fat) + 6-GPU (v100m3) parallel jobs after chronic v100g32 congestion. (ffd0f6e)
• SAI Apptainer GPU validation retired in favor of the KSSG HIP pipeline. (f9ae8f5)

Tests & Benchmarks

• Test suite reorganized under ctest with cpu/gpu labels; 42 GPU cases + 5 CPU cases run per CI cycle; rank registration standardized (DDLA_TEST_NP). (22377a2, 6f3d566)

---

Full changelog: v0.0.4...develop (56 commits, 182 files changed, +11740/-3620 lines)

v0.0.4

LibDDLA v0.0.4

C++17 distributed dense linear algebra library for GPUs (CUDA/HIP), ScaLAPACK-style 2D block-cyclic APIs over MPI. This release covers 57 commits since v0.0.3 (2026-05-21 → 2026-07-27).

Highlights

• Unified CUDA/HIP batched GEMM backend (gemmVbatched / gemmVbatched2s) — a new vendor-neutral vbatched GEMM implementation shared across CUDA and HIP, supporting all four scalar types (float, double, std::complex<float>, std::complex<double>) with the public std::complex ABI preserved and thrust::complex used only internally (explicit FMA, no MAGMA dependency at runtime). (7aa2c1c)
• License changed from GPL to LGPL v3, with MAGMA's BSD-3-Clause license included under LICENSES/ for the vbatched-derived kernel configs. (717e8db, 7aa2c1c)
• README rewritten with accurate prerequisites, CMake option table, and build instructions for both CUDA and HIP backends. (e3eaaff, 3cbb91d, e8156f0, 9c404ac)

New Features

• Non-pivoting LU factorization/solve family: pgetrf_nopiv, pgetrs_nopiv, pgesv_nopiv, and a MAGMA-derived getrf_nopiv. (32c54aa, 568c88b)
• Banded-pivot LU factorization: pgetrf_bpiv. (46cdab2)
• ppotrf_bottom_right / potrf_bottom_right — Cholesky factorization variants operating on the bottom-right submatrix. (f95af64)
• pdam — distributed matrix scale/absolute-value-max style operation on a submatrix. (a629247)
• axpy interface. (7d8bdb4)
• pgemm transpose support (transa/transb ≠ 'N') plus general submatrix and non-square (n×m) MPI process grid support; ppotrf gained an upper-triangle (uplo == 'U') implementation. (1603864, dff3d79)
• Template-based random matrix generator, replacing the old non-generic implementation. (fa36e4f)
• deviceGetDeviceCount interface added to the device connector layer. (5e7a751)

Performance

• Improved pgetrf performance by optimizing the underlying pgetf2 panel factorization. (159da24, fdc6ce2, 0aac709)
• pgemm performance improvements alongside the transpose/submatrix support work. (589e3f6)
• Removed a redundant world-communicator barrier from the triangular-factorization (*trf) code paths. (f1f3295)

Bug Fixes

• Fixed pgemm correctness bug affecting asynchronous communication. (0d576d4)
• Fixed pgeadd bugs when transa/transb ≠ 'N', and a separate pgeadd/pswap bug fix with accompanying tests. (a8fc255, fc5fdaf)
• Fixed a rectangular block-size bug in the transport-block code path, and a related transport-block bug. (79c666e, 09938c0)
• Fixed a laswp interface bug caused by CUDA/HIP differences, and a broader CUDA/HIP interface-difference bug on GPU/DCU platforms. (a0297b4, e15c96f)
• Fixed HIP-specific bug. (7b5c888)
• Fixed *_CHECK macros not reporting the correct file/line location. (e4066eb)
• Added zero-size malloc and null-pointer free protection across allocation call sites. (0a42afc)

Build System / Tooling

• Renamed the ENABLE_* CMake/compile-definition family to DDLA_USE_* for clarity. (2870801, 5b8fbb5)
• Switched the DCU (HIP) compiler from hipcc to g++. (b8f8340)
• Updated HIP install/build scripts and CUDA/HIP install scripts. (91f4100, 2a6252b, ac6c76c, 2a118b9)
• Removed bench_pzgemm.cpp from src/ (moved/superseded by other benchmark updates) and cleaned up a stray tmp folder. (422ce7d, 1ff5713)
• Updated .gitignore. (c1a78dc)
• Added, then later removed, AGENTS.md contributor-agent instructions (no longer tracked in this repo). (8c55e6c, 8c5aa17, 3ddde74, 7b07283)

Tests & Benchmarks

• Added a large batch of new API-grid regression tests covering pgeadd, pgemm, pgesv/pgesv_nopiv, pgetf2(+panel), pgetrf(+_bpiv/_nopiv), pgetrs(+_nopiv), plapiv, pposv, ppotrf(+_bottom_right), ppotrs, pswap, ptran, ptrtrs, a subcommunicator LU test, and transport-block tests. (a5227e8, and the feature commits above)
• Added test_gemm_vbatched.cpp, test_pdam.cpp, test_getrf_nopiv.cpp, test_pgesv_nopiv.cpp, test_pgetrf_bpiv.cpp, test_pgetrf_nopiv.cpp, test_random_generate.cpp, and Hermitian-positive pgemm tests.
• Updated pgetrf/ptrtrs/ppotrf_bottom_right benchmark drivers, including a cusolverMp-based pzgetrf benchmark for CUDA 11.8. (5fe65ff, 5530d58, abb8250)
• Removed stray printf debug output from the LU factorization files. (acaa3f7)

---

Full changelog: v0.0.3...main (57 commits, 165 files changed, +26453/-3759 lines)

Note on tag naming: existing tags are v0.0.0, 0.0.2 (no v prefix), v0.0.3. This release recommends tagging as v0.0.4 to stay consistent with the majority convention.

v0.0.3

we give some interface for science computation
PGEMM
PZGEMM

PZGETF2
PZGETRF
PZGETRS
PZGESV
PZLAPIV

PTRTRS
PZTRTRS

PPOTRF
PPOTRS
PPOSV

v0.0.2

we give some interface for science computation
PGEMM
PZGETF2
PZGETRF
PZGEMM
PZLAPIV
PZGETRS
PZGESV

LibDDLA

v0.0.0 the package have supported PZGETRF, PZTRTRS, PZGEMM, PZGETRS, PZLAPIV