Change dot, dotc, vector_norm2, and vector_abs_sum to imitate reduce, so that they return their r... more Change dot, dotc, vector_norm2, and vector_abs_sum to imitate reduce, so that they return their result, instead of taking an output parameter. Users may set the result type via optional init parameter. Minor changes to "expression template" classes, based on implementation experience Briefly address LEWGI request of exploring concepts for input arguments. Lazy ranges style API was NOT explored. blas_interface.md 4/14/2021 2 / 141 Revision 2 (pre-Cologne) to be submitted 2020-01-13 Add "Future work" section. Remove "Options and votes" section (which were addressed in SG6, SG14, and LEWGI). Remove basic_mdarray overloads. Remove batched linear algebra operations. Remove over-and underflow requirement for vector_norm2. Mandate any extent compatibility checks that can be done at compile time. Add missing functions {symmetric,hermitian}_matrix_rank_k_update and triangular_matrix_{left,right}_product. Remove packed_view function. Fix wording for {conjugate,transpose,conjugate_transpose}_view, so that implementations may optimize the return type. Make sure that transpose_view of a layout_blas_packed matrix returns a layout_blas_packed matrix with opposite Triangle and StorageOrder. Remove second template parameter T from accessor_conjugate. Make scaled_scalar and conjugated_scalar exposition only. Add in-place overloads of triangular_matrix_matrix_{left,right}_solve, triangular_matrix_{left,right}_product, and triangular_matrix_vector_solve. Add alpha overloads to {symmetric,hermitian}_matrix_rank_{1,k}_update. Add Cholesky factorization and solve examples. Revision 3 (electronic) to be submitted 2021-04-15 Per LEWG request, add a section on our investigation of constraining template parameters with concepts, in the manner of P1813R0 with the numeric algorithms. We concluded that we disagree with the approach of P1813R0, and that the Standard's current GENERALIZED_SUM approach better expresses numeric algorithms' behavior. Update references to the current revision of P0009 (mdspan). Per LEWG request, introduce std::linalg namespace and put everything in there. Per LEWG request, replace the linalg_ prefix with the aforementioned namespace. We renamed linalg_add to add, linalg_copy to copy, and linalg_swap to swap_elements. Per LEWG request, do not use _view as a suffix, to avoid confusion with "views" in the sense of Ranges. We renamed conjugate_view to conjugated, conjugate_transpose_view to conjugate_transposed, scaled_view to scaled, and transpose_view to transposed. blas_interface.md 4/14/2021 3 / 141 Change wording from "then implementations will use T's precision or greater for intermediate terms in the sum," to "then intermediate terms in the sum use T's precision or greater." Thanks to Jens Maurer for this suggestion (and many others!). Before, a Note on vector_norm2 said, "We recommend that implementers document their guarantees regarding overflow and underflow of vector_norm2 for floating-point return types." Implementations always document "implementation-defined behavior" per [defs.impl.defined]. (Thanks to Jens Maurer for pointing out that "We recommend..." does not belong in the Standard.) Thus, we changed this from a Note to normative wording in Remarks: "If either in_vector_t::element_type or T are floating-point types or complex versions thereof, then any guarantees regarding overflow and underflow of vector_norm2 are implementationdefined." Define return types of the dot, dotc, vector_norm2, and vector_abs_sum overloads with auto return type. Remove the explicitly stated constraint on add and copy that the rank of the array arguments be no more than 2. This is redundant, because we already impose this via the existing constraints on template parameters named in_object*_t, inout_object*_t, or out_object*_t. If we later wish to relax this restriction, then we only have to do so in one place. Add vector_sum_of_squares. First, this gives implementers a path to implementing vector_norm2 in a way that achieves the over/underflow guarantees intended by the BLAS Standard. Second, this is a useful algorithm in itself for parallelizing vector 2-norm computation. Add matrix_frob_norm, matrix_one_norm, and matrix_inf_norm (thanks to coauthor Piotr Luszczek). Address LEWG request for us to investigate support for GPU memory. See section "Explicit support for asynchronous return of scalar values."
In this report we describe the logic and interface we develop for the MAGMA-sparse library to all... more In this report we describe the logic and interface we develop for the MAGMA-sparse library to allow for easy integration as third-party library into a top-level software ecosystem. The design choices are based on extensive consultation with other software library developers, in particular the Trilinos software development team. The interface documentation is at this point not exhaustive, but a first proposal for setting a standard. Although the interface description targets the MAGMA-sparse software module, we hope that the design choices carry beyond this specific library, and are attractive for adoption in other packages. This report is not intended as static document, but will be updated over time to reflect the agile software development in the ECP 1.3.3.11 STMS11-PEEKS project.
This is the definitive user manual for the IFPACK2 package in the Trilinos project. IFPACK2 provi... more This is the definitive user manual for the IFPACK2 package in the Trilinos project. IFPACK2 provides implementations of iterative algorithms (e.g., Jacobi, SOR, additive Schwarz) and processorbased incomplete factorizations. IFPACK2 is part of the Trilinos TPETRA solver stack, is templated on index, scalar, and node types, and leverages node-level parallelism indirectly through its use of TPETRA kernels. IFPACK2 can be used to solve to matrix systems with greater than 2 billion rows (using 64-bit indices). Any options not documented in this manual should be considered strictly experimental.
Change dot, dotc, vector_norm2, and vector_abs_sum to imitate reduce, so that they return their r... more Change dot, dotc, vector_norm2, and vector_abs_sum to imitate reduce, so that they return their result, instead of taking an output parameter. Users may set the result type via optional init parameter. Minor changes to "expression template" classes, based on implementation experience Briefly address LEWGI request of exploring concepts for input arguments. Lazy ranges style API was NOT explored. blas_interface.md 4/14/2021 2 / 141 Revision 2 (pre-Cologne) to be submitted 2020-01-13 Add "Future work" section. Remove "Options and votes" section (which were addressed in SG6, SG14, and LEWGI). Remove basic_mdarray overloads. Remove batched linear algebra operations. Remove over-and underflow requirement for vector_norm2. Mandate any extent compatibility checks that can be done at compile time. Add missing functions {symmetric,hermitian}_matrix_rank_k_update and triangular_matrix_{left,right}_product. Remove packed_view function. Fix wording for {conjugate,transpose,conjugate_transpose}_view, so that implementations may optimize the return type. Make sure that transpose_view of a layout_blas_packed matrix returns a layout_blas_packed matrix with opposite Triangle and StorageOrder. Remove second template parameter T from accessor_conjugate. Make scaled_scalar and conjugated_scalar exposition only. Add in-place overloads of triangular_matrix_matrix_{left,right}_solve, triangular_matrix_{left,right}_product, and triangular_matrix_vector_solve. Add alpha overloads to {symmetric,hermitian}_matrix_rank_{1,k}_update. Add Cholesky factorization and solve examples. Revision 3 (electronic) to be submitted 2021-04-15 Per LEWG request, add a section on our investigation of constraining template parameters with concepts, in the manner of P1813R0 with the numeric algorithms. We concluded that we disagree with the approach of P1813R0, and that the Standard's current GENERALIZED_SUM approach better expresses numeric algorithms' behavior. Update references to the current revision of P0009 (mdspan). Per LEWG request, introduce std::linalg namespace and put everything in there. Per LEWG request, replace the linalg_ prefix with the aforementioned namespace. We renamed linalg_add to add, linalg_copy to copy, and linalg_swap to swap_elements. Per LEWG request, do not use _view as a suffix, to avoid confusion with "views" in the sense of Ranges. We renamed conjugate_view to conjugated, conjugate_transpose_view to conjugate_transposed, scaled_view to scaled, and transpose_view to transposed. blas_interface.md 4/14/2021 3 / 141 Change wording from "then implementations will use T's precision or greater for intermediate terms in the sum," to "then intermediate terms in the sum use T's precision or greater." Thanks to Jens Maurer for this suggestion (and many others!). Before, a Note on vector_norm2 said, "We recommend that implementers document their guarantees regarding overflow and underflow of vector_norm2 for floating-point return types." Implementations always document "implementation-defined behavior" per [defs.impl.defined]. (Thanks to Jens Maurer for pointing out that "We recommend..." does not belong in the Standard.) Thus, we changed this from a Note to normative wording in Remarks: "If either in_vector_t::element_type or T are floating-point types or complex versions thereof, then any guarantees regarding overflow and underflow of vector_norm2 are implementationdefined." Define return types of the dot, dotc, vector_norm2, and vector_abs_sum overloads with auto return type. Remove the explicitly stated constraint on add and copy that the rank of the array arguments be no more than 2. This is redundant, because we already impose this via the existing constraints on template parameters named in_object*_t, inout_object*_t, or out_object*_t. If we later wish to relax this restriction, then we only have to do so in one place. Add vector_sum_of_squares. First, this gives implementers a path to implementing vector_norm2 in a way that achieves the over/underflow guarantees intended by the BLAS Standard. Second, this is a useful algorithm in itself for parallelizing vector 2-norm computation. Add matrix_frob_norm, matrix_one_norm, and matrix_inf_norm (thanks to coauthor Piotr Luszczek). Address LEWG request for us to investigate support for GPU memory. See section "Explicit support for asynchronous return of scalar values."
In this report we describe the logic and interface we develop for the MAGMA-sparse library to all... more In this report we describe the logic and interface we develop for the MAGMA-sparse library to allow for easy integration as third-party library into a top-level software ecosystem. The design choices are based on extensive consultation with other software library developers, in particular the Trilinos software development team. The interface documentation is at this point not exhaustive, but a first proposal for setting a standard. Although the interface description targets the MAGMA-sparse software module, we hope that the design choices carry beyond this specific library, and are attractive for adoption in other packages. This report is not intended as static document, but will be updated over time to reflect the agile software development in the ECP 1.3.3.11 STMS11-PEEKS project.
This is the definitive user manual for the IFPACK2 package in the Trilinos project. IFPACK2 provi... more This is the definitive user manual for the IFPACK2 package in the Trilinos project. IFPACK2 provides implementations of iterative algorithms (e.g., Jacobi, SOR, additive Schwarz) and processorbased incomplete factorizations. IFPACK2 is part of the Trilinos TPETRA solver stack, is templated on index, scalar, and node types, and leverages node-level parallelism indirectly through its use of TPETRA kernels. IFPACK2 can be used to solve to matrix systems with greater than 2 billion rows (using 64-bit indices). Any options not documented in this manual should be considered strictly experimental.
Uploads
Papers by Andrew Ponte