dla-future - DLA-Future 0.10.0

Changes

• Added inverse of a triangular matrix. (#1324)
• Added inverse from Cholesky factor and its C API. (#1324 and #1326)
• Improved Panel transposed broadcast to be more flexible. (#1325)
  • Note: previous usages without changes provide correct result but they perform an additional tile broadcast. This extra communication can be avoided by shrinking the transposed panel.
• Refactored Matrix API. (#1321)
  • Constructors changed.
  • Member functions have been updated. Old member functions are still available, but will be deprecated soon.

Bug fixes

• Fixed behaviour when CUDA archs is not specified for a CUDA build. (#1318)

- C++
Published by rasolca 8 months ago

dla-future - DLA-Future 0.9.0

Changes

• Automatically free all DLA-Future grids when finalizing the C API. (#1308)
• Tau factors are stored on GPU memory. (#1304)

Bug fixes

• Avoid potential stack overflows by transferring work to a new task on the default thread pool before launching GPU work. (#1292)

- C++
Published by RMeli 9 months ago

dla-future - DLA-Future 0.8.0

Changes

• Renamed tune parameters. (#1270)
  • dlaf:red2band-panel-nworkers becomes dlaf:red2band-panel-num-threads.
  • dlaf:tridiag-rank1-nworkers becomes dlaf:tridiag-rank1-num-threads.

Performance improvements

• Many improvements to reduction_to_band and its backtransformation. (#1214, #1256, #1263)

Bug fixes

• Fixed a bug in correctness checking of eigensolver miniapps when computing partial eigenspectrum. (#1283)

- C++
Published by rasolca 10 months ago

dla-future - DLA-Future 0.7.3

Bug fixes

• Changed C ScaLAPACK API indexing convention to 1-based for partial eigenspectrum. (#1248)

- C++
Published by RMeli about 1 year ago

dla-future - DLA-Future 0.7.1

Bug fixes

• Fixed a compilation issue with ROCm. (#1241)
• Fixed missing includes of <complex>. (#1243)

- C++
Published by msimberg about 1 year ago

dla-future - DLA-Future 0.7.0

Changes

• Added (generalized) eigensolver which computes only a part of the eigenspectrum. (#1194)
• Norm is now fully asynchronous. (#1221)

Performance improvements

• Refactored communication to use pika's transform_mpi and polling support. (#1125)
• Use custom coalescing heuristic for memory pools. (#1183)
• Added configuration option for number of CUDA streams and cuBLAS/SOLVER handles. (#1222, #1182)
• Some algorithmic clean-ups and improvements. (#1213, #1219, #1232)

Bug fixes

• Fixed builds with CUDA and stdexec enabled. (#1188)
• Work around buggy HIP complex operator overloads. (#1195)
• Namespace (SCA)LAPACK CMake modules with DLAF to avoid conflicts with other packages. (#1178)

- C++
Published by rasolca about 1 year ago

dla-future - DLA-Future 0.6.0

Changes

• Renamed ScaLAPACK-like generalized eigensolvers pXsygvx/pXhegvx to pXsygvd/pXhegvd. (#1168)
• Introduced generalized eigensolver where the matrix B is already factorized. (#1167)

Performance improvements

• Local eigenvector permutations in the distributed tridiagonal eigensolver are executed directly in GPU memory. (#1118)

Bug fixes

• Fixed ScaLAPACK detection in CMake for specific uenv cases. (#1159)

- C++
Published by rasolca over 1 year ago

dla-future - DLA-Future 0.5.0

Changes

• Introduced an option (*) for forcing contiguous GPU communication buffers. (#1096)
• Introduced an option (*) for enabling GPU aware MPI communication. (#1102)
• Removed special handling of Intel MKL, as it could lead to broken installations. (#1149)
  • Spack installations: spack will set the correct variables.
  • Manual installations: the user is responsible to correctly set variables (see BUILD.md).

(*) These options are available as spack variants.

Performance improvements

• Don't communicate in algorithms when using single rank communicators. (#1097)
• Fixed slow performance of local version of bt_band_to_tridiagonal (#1144)

Bug fixes

• Implemented a workaround for hipMemcpyDefault 2D memcpys, due to bugs in HIP. (#1106)
• Miniapps initialize HIP before MPI, as on older Cray MPICH versions initializing HIP after MPI leads to HIP not seeing any devices. (#1090)

- C++
Published by msimberg over 1 year ago

dla-future - DLA-Future 0.4.1

Bug fixes

• Update project version and export it in CMake. (#1121)

- C++
Published by msimberg over 1 year ago

dla-future - DLA-Future 0.4.0

Changes:

• Modified CommunicatorGrid to avoid blocking calls to MPI_Comm_dup. It now returns communicator pipelines. (#993)
• Added support for Intel oneMKL and the intel-oneapi-mkl spack package. (#1073) (*)

Performance improvements:

• Reduced the size of the matrix-matrix multiplications in the tridiagonal eigensolver to cover only the non deflated part of the eigenvectors. (#951 #967 #996 #997 #998)
• Introduced stackless threads where appropriate. (#1037)

Bug fixes:

• Use drop_operation_state to avoid stack overflows. (#1004)

Notes:

(*) At the time of the release the spack spec blaspp~openmp ^intel-oneapi-mkl threads=openmp doesn't build. If you rely on multithreaded BLAS we suggest to use blaspp+openmp ^intel-oneapi-mkl threads=openmp until https://github.com/spack/spack/pull/42087 gets merged.

- C++
Published by rasolca almost 2 years ago

dla-future - DLA-Future 0.3.1

Bugfix:

• Fixed compilation with gcc 9.3
• Fixed compilation with CUDA 11.2
• Improved eigensolver tests

- C++
Published by rasolca about 2 years ago

dla-future - DLA-Future 0.3.0

Changes:

• added C and ScaLAPACK API (generalized eigensolver) (#992)
• removed pika-algorithm dependency (#945)

Performance improvements:

• Fixed Cholesky priorities (#999)

- C++
Published by rasolca about 2 years ago

dla-future - DLA-Future 0.2.1

Bugfix:

• Fixed a problem in reduction_to_band that could have produced results filled with NaNs for certain corner cases. (E.g. input matrix with all off-band elements set to 0).

- C++
Published by rasolca over 2 years ago

dla-future - DLA-Future 0.2.0

Changes:

• renamed algorithms using snake case (#942)
• added C and ScaLAPACK API (cholesky and eigensolver) (#886)
• Matrix API:
  • initial support for matrices with different tile/block-size (#909)
  • initial support for matrix subpipelines (#898)
  • initial support for submatrices (#934)
  • initial support for matrix redistribution (#933) ## Bugfixes:
• fixed a problem in tridiagonal_eigensolver which produced wrong results for some classes of matrices (#960) ## Performance improvements:
• introduced busy barriers in reduction_to_band (#864)
• new band_to_tridiagonal algorithm implementation (#938, #946)
• improved the rank1 problem solution in tridiagonal_eigensolver (#904, #936)

- C++
Published by rasolca over 2 years ago

dla-future - DLA-Future 0.1.0

The first release of DLA-Future.

- C++
Published by rasolca over 2 years ago