Release Notes#

This section lists significant changes, new features, performance improvements, and various issues for each release of cuSolverDx. Unless noted, listed issues should not impact functionality. When functionality is impacted, we offer a work-around to avoid the issue (if available).

0.6.0#

New Features#

  • Experimental cluster execution for GEQRF on Hopper and later GPUs (SM 900 or later)

  • Support for Rubin architecture sm_107 (SM<1070>).

  • TransposeMode conj_transposed is now supported for real data types in addition to complex. For real data types it is equivalent to transposed.

Deprecations#

  • Support for GCC 7 as the host compiler is deprecated and will be removed in a future release. A warning is emitted when compiling with GCC 7. We recommend GCC 10 or newer.

  • The LeadingDimension operator is deprecated and will be removed in a future release. If the matrix storage is not padded, use the default strides Solver::default_lda, Solver::default_ldb, and Solver::default_ldc. If the storage is padded, pass the leading dimensions as runtime arguments to execute. * The leading_dimension_of_v, leading_dimension_of_v_a, leading_dimension_of_v_b, and leading_dimension_of_v_c traits are deprecated.

  • Solver::shared_memory_size is deprecated. Use get_shared_memory_size with the runtime leading dimensions.

Resolved Issues#

  • A bug in the ::b_size trait for hegst was fixed.

Breaking Changes#

  • The accuracy of the get_shared_memory_size function was improved, reducing its value in many cases.

Known Issues#

Compiler / static assertions (cluster execution). cuSolverDx 0.6.0 triggers static assertions at compile time for cluster configurations known to be affected by specific compiler bugs. If your build fails with one of these assertions, you can define the corresponding macro to ignore the check and verify correctness yourself, or use an unaffected BlockDim or problem size (see each bullet below).

  • NVBUG 6700081 — CTK 13.0 and newer, SM90, cluster potrf / geqrf with BlocksPerCluster > 1, complex double, BlockDim < 128. May produce wrong results from a remote-DSMEM complex-double GEMM. Define CUSOLVERDX_IGNORE_NVBUG_6700081_ASSERT to suppress the assertion, or use BlockDim of at least 128, or run on Blackwell or newer GPUs to avoid the bug.

  • NVBUG 6750508 — CTK 13.0 and newer, SM90, cluster geqrf with BlocksPerCluster > 1. The assert covers complex double with BlockDim of at least 128, and complex float size 1928 x 49 with BlockDim 256 and BlocksPerCluster 8. Device LTO can miscompile the kernel for some test cases and silently produce wrong results. Define CUSOLVERDX_IGNORE_NVBUG_6750508_ASSERT to suppress the assertion, or use a real type, or run on Blackwell or newer GPUs to avoid the bug.

  • Cluster GEQRF SM100 — CTK 13.2, SM100, cluster geqrf, complex float, size 1899 x 44, BlockDim 128, BlocksPerCluster 4. Define CUSOLVERDX_IGNORE_GEQRF_CLUSTER_SM100_KNOWN_FAILURE_ASSERT to suppress the assertion, or use a different size, type, or BlockDim, or use a CTK other than 13.2.

  • Cluster GEQRF illegal instruction — CTK 13.0 and 13.1, SM90, cluster geqrf with BlocksPerCluster > 1, complex double, size 1153 x 37, BlockDim 128. May fail at runtime with an illegal instruction. Define CUSOLVERDX_IGNORE_GEQRF_CLUSTER_ILLEGAL_INSTRUCTION_ASSERT to suppress the assertion, use a different size or BlockDim, or use CTK 13.2 or newer to avoid the bug.

To highlight these issues cuSolverDx will hard fail with a static assertion if these conditions are met.

If your build fails with one of these assertions:

  • Prefer using a CUDA Toolkit version where the bug is fixed, or an unaffected BlockDim / size (see each bullet above).

  • Alternatively, define the corresponding ignore macro to suppress the assertion and verify correctness of results yourself.

0.5.0#

New Features#

  • Experimental cluster execution for POTRF on Hopper and later GPUs (SM 900 or later)

  • Support for batched or non-batched reduction of Hermitian/symmetric generalized eigenproblem to standard form.

  • Support for batched or non-batched generalized Hermitian/symmetric eigenvalue solver (HEGV), including optional eigenvectors for all three LAPACK itype problem forms.

  • More examples, including new examples using the cluster operator, and advanced examples demonstrating hybrid memory strategies and block algorithms to process large size matrices.

  • The static library artifact of cuSolverDx (libcusolverdx.a) is being deprecated and will be removed in the next release.

  • Added CUSOLVERDX_SKIP_IF_NOT_APPLICABLE_SM() macro to skip kernel compilation for unintended device architectures, reducing compilation time and the final binary size. See SM for details.

0.4.0#

New Features#

  • BDSVD and GESVD functions now support computing left and/or right singular vectors in addition to singular values. See bdsvd for the bidiagonal SVD device API and job options (e.g. job options), and gesvd for the general matrix SVD device API and job options (e.g. U/VT storage).

  • Thread operators: cuSolverDx introduces the Thread operator, which runs the solver in a thread context so that each thread executes the operation independently. This is intended for small problem sizes that fit in registers; see Execution Operators and the Thread operator for usage and restrictions.

  • Support for batched and non-batched Modified LU factorization of unitary matrices, useful for constructing the explicit tall-skinny Householder reflector after Tall-Skinny QR (TSQR).

  • Performance improvements for heev function on Hopper.

  • More examples, including examples using the new thread operator, are included.

Breaking Changes#

  • cuSolverDx 0.4.0 requires CUDA Toolkit 13.0 or newer.

  • Support for SM70 and SM72 CUDA architectures is removed in cuSolverDx 0.4.0.

  • The suggested_leading_dimension trait has been removed. With thread operators and the expanded set of functions, a single recommended leading dimension is no longer meaningful, and the confusion it caused outweighed its usefulness. For matrices in shared memory, use the LeadingDimension operator to specify padding when necessary and tune for the best performance in your use case.

Known Issues#

Compiler / static assertions (block execution). cuSolverDx 0.4.0 triggers static assertions at compile time for configurations known to be affected by specific compiler bugs. If your build fails with one of these assertions, you can define the corresponding macro to ignore the check and verify correctness yourself, or use a CUDA Toolkit version where the bug is fixed.

  • NVBUG 5972531 — CTK 13.2+, SM100+, htev with vectors; a few specific sizes and block dimensions. May produce wrong results. Define CUSOLVERDX_IGNORE_NVBUG_5972531_ASSERT to suppress the assertion, or use CTK 13.0 or 13.1 to avoid the bug.

  • NVBUG 5986343 — CTK 13.0/13.1, SM80/86/89/90, bdsvd with vectors; a few specific sizes and block dimensions. May produce wrong results. Define CUSOLVERDX_IGNORE_NVBUG_5986343_ASSERT to suppress the assertion, or use CTK 13.2+ to avoid the bug.

  • NVBUG 5908003 — CTK 13.2+, SM100+, gesvd with vectors, complex; a few specific sizes (and block dimensions for V1). May produce wrong results. Define CUSOLVERDX_IGNORE_NVBUG_5908003_ASSERT to suppress the assertion, or use CTK 13.0 or 13.1 to avoid the bug.

  • NVBUG 5288270 — CTK 13.x, SM120, gesv_no_pivot, real; a few specific matrix sizes and block dimensions. May produce wrong results (release builds). Define CUSOLVERDX_IGNORE_NVBUG_5288270_ASSERT to suppress the assertion, or use a CUDA Toolkit build where the issue is fixed.

To highlight these issues cuSolverDx will hard fail with a static assertion if these conditions are met.

If your build fails with one of these assertions:

  • Prefer using a CUDA Toolkit version where the bug is fixed (see each bullet above).

  • Alternatively, define the corresponding CUSOLVERDX_IGNORE_NVBUG_<id>_ASSERT macro to suppress the assertion and verify correctness of results yourself.

  • Adding -Xptxas -O0 to the compile command will limit PTX optimization phase and produce correct binary, although potentially slower.

0.3.0#

New Features#

Breaking Changes#

  • cuSolverDx 0.3.0 updates the dimensions for function trsm, to be consistent with Lapack trsm and cublasXtrsm notation. Specifically, M is the number of rows in matrix B, and N is the number of columns in matrix B, with the triangular matrix A sized accordingly. K, if specified, is ignored.

  • To use the IO-related copy_2d methods or shared memory management tools, the header file cusolverdx_io.hpp needs to be included in your code.

0.2.1#

New Features#

  • Support for CUDA 13.0.

  • Support for Thor SM renaming from sm_101 to sm_110 starting from CUDA 13.0. * Note: For CUDA 12.9 and older releases, Thor stays labeled as sm_101.

Known Issues#

  • CUDA 12.8, 12.9 and 13.0 could miscompile kernels using gesv_no_pivot function with high register pressure when all of the following conditions are met:

    • SM is set to 1200 (SM120), and

    • Type is set to type::real.

    The code corruptions may manifest as intermittent incorrect results.

    To highlight this issue cuSolverDx 0.2.1 will overprotectively hard fail if these conditions are met.

    If you are using cuSolverDx, and this happens to your code, you can:

    • Define the CUSOLVERDX_IGNORE_NVBUG_5288270_ASSERT macro to ignore these assertions and verify correctness of the results

    • If the case is indeed affected by the bug, adding the -Xptxas -O0 flag to the compilation command will limit PTX optimization phase and produce correct binary, although potentially slower

0.2.0#

New Features#

Breaking Changes#

  • cuSolverDx 0.2.0 requires CUDA Toolkit 12.6.3 or newer.

  • cuSolverDx 0.2.0 requires FillMode operator to be specified explicitly for Cholesky functions (potrf, potrs, posv), instead of using the default value if not specified.

Known Issues#

  • CUDA 12.8.0 and 12.9.0 could miscompile kernels using gesv_no_pivot function with high register pressure when all of the following conditions are met:

    • SM is set to 1200 (SM120), and

    • Type is set to type::real.

    The code corruptions may manifest as intermittent incorrect results.

    To highlight this issue cuSolverDx 0.2.0 will overprotectively hard fail if these conditions are met.

    If you are using cuSolverDx, and this happens to your code, you can:

    • Define the CUSOLVERDX_IGNORE_NVBUG_5288270_ASSERT macro to ignore these assertions and verify correctness of the results

    • If the case is indeed affected by the bug, adding the -Xptxas -O0 flag to the compilation command will limit PTX optimization phase and produce correct binary, although potentially slower

0.1.0#

The first early access (EA) release of cuSolverDx library.

New Features#

Known Issues#

  • For CUDA Toolkits older than 12.1 Update 1 when using NVRTC and nvJitLink to perform runtime compilation of kernels that use cuSolverDx it is required to use fatbin file libcusolverdx.fatbin when linking instead of libcusolverdx.a. Path to the fatbin library file is defined in cusolverdx_FATBIN cmake variable.

  • NVCC compiler in CUDA Toolkit from 12.2 to 12.4 reports incorrect compilation error when value_type type of a solver description type is used. The affected code with workarounds are presented below. The issue is fixed in CTK 12.5 and newer.

    using Solver = decltype(Size<32>()
                + Precision<double>()
                + Type<type::complex>()
                + Function<function::potrf>()
                + SM<800>()
                + Block());
using type = typename Solver::value_type; // compilation error

// Workaround #1
using type = typename decltype(Solver())::value_type;

// Workaround #2 (used in cuSolverDx examples)
template <typename T>
using value_type_t = typename T::value_type;

using type = value_type_t<Solver>;