BLAS interface(s)

Kokkos Kernels has multiple adapters for BLAS routines. For the sake of simplicity, this page will look only at {S,D,C,Z}GEMM with op(B) = B.

• When KokkosKernels_ENABLE_TPL_BLAS is set to ON,

  • the struct GEMM uses the TPL implementation when the Kokkos::View objects A, B, and C have the same layout.

• When KokkosKernels_ENABLE_TPL_CUBLAS is set to ON,

  • the struct GEMM uses the TPL implementation when the Kokkos::View objects A, B, and C have the same layout.

• The function KokkosBlas::gemm implements the cases op(A) = A, op(A) = A^T, and op(A) = A^H.

• The function KokkosBlas::gemm uses multiple variants:

  • when A has a LayoutRight, op(A) = A^T, and MN < 1600, then we use DotBasedGEMM.
  • else the structure KokkosBlas::Impl::GEMMImpl is used.
    • the structure allocates 3 scratch matrices and deep-copies blocks of A and B.