diff --git a/blas/tpls/KokkosBlas1_rotg_tpl_spec_avail.hpp b/blas/tpls/KokkosBlas1_rotg_tpl_spec_avail.hpp index 4d1320c402..a17a5cdb05 100644 --- a/blas/tpls/KokkosBlas1_rotg_tpl_spec_avail.hpp +++ b/blas/tpls/KokkosBlas1_rotg_tpl_spec_avail.hpp @@ -139,7 +139,7 @@ KOKKOSBLAS1_ROTG_TPL_SPEC_AVAIL_CUBLAS(double, Kokkos::LayoutRight, Kokkos::Cuda, Kokkos::CudaSpace) KOKKOSBLAS1_ROTG_TPL_SPEC_AVAIL_CUBLAS(float, Kokkos::LayoutLeft, Kokkos::Cuda, Kokkos::CudaSpace) -KOKKOSBLAS1_ROTG_TPL_SPEC_AVAIL_CUBLAS(float, Kokkos::LayoutLeft, Kokkos::Cuda, +KOKKOSBLAS1_ROTG_TPL_SPEC_AVAIL_CUBLAS(float, Kokkos::LayoutRight, Kokkos::Cuda, Kokkos::CudaSpace) KOKKOSBLAS1_ROTG_TPL_SPEC_AVAIL_CUBLAS(Kokkos::complex, Kokkos::LayoutLeft, Kokkos::Cuda, @@ -160,7 +160,7 @@ KOKKOSBLAS1_ROTG_TPL_SPEC_AVAIL_CUBLAS(double, Kokkos::LayoutRight, Kokkos::Cuda, Kokkos::CudaUVMSpace) KOKKOSBLAS1_ROTG_TPL_SPEC_AVAIL_CUBLAS(float, Kokkos::LayoutLeft, Kokkos::Cuda, Kokkos::CudaUVMSpace) -KOKKOSBLAS1_ROTG_TPL_SPEC_AVAIL_CUBLAS(float, Kokkos::LayoutLeft, Kokkos::Cuda, +KOKKOSBLAS1_ROTG_TPL_SPEC_AVAIL_CUBLAS(float, Kokkos::LayoutRight, Kokkos::Cuda, Kokkos::CudaUVMSpace) KOKKOSBLAS1_ROTG_TPL_SPEC_AVAIL_CUBLAS(Kokkos::complex, Kokkos::LayoutLeft, Kokkos::Cuda, diff --git a/blas/tpls/KokkosBlas1_rotg_tpl_spec_decl.hpp b/blas/tpls/KokkosBlas1_rotg_tpl_spec_decl.hpp index 7847abe6ec..c53fd9f5a0 100644 --- a/blas/tpls/KokkosBlas1_rotg_tpl_spec_decl.hpp +++ b/blas/tpls/KokkosBlas1_rotg_tpl_spec_decl.hpp @@ -318,44 +318,44 @@ namespace Impl { } \ }; -#define KOKKOSBLAS1_ZROTG_TPL_SPEC_DECL_CUBLAS(LAYOUT, EXECSPACE, MEMSPACE, \ - ETI_SPEC_AVAIL) \ - template \ - struct Rotg< \ - EXECSPACE, \ - Kokkos::View, LAYOUT, \ - Kokkos::Device, \ - Kokkos::MemoryTraits>, \ - Kokkos::View, \ - Kokkos::MemoryTraits>, \ - true, ETI_SPEC_AVAIL> { \ - using SViewTypeKokkos::View, LAYOUT, \ - Kokkos::Device, \ - Kokkos::MemoryTraits>; \ - using MViewType = \ - Kokkos::View, \ - Kokkos::MemoryTraits>; \ - static void rotg(EXECSPACE const& space, SViewType const& a, \ - SViewType const& b, MViewType const& c, \ - SViewType const& s) { \ - Kokkos::Profiling::pushRegion( \ - "KokkosBlas::nrm1[TPL_CUBLAS,complex]"); \ - rotg_print_specialization, EXECSPACE>(); \ - KokkosBlas::Impl::CudaBlasSingleton& singleton = \ - KokkosBlas::Impl::CudaBlasSingleton::singleton(); \ - KOKKOS_CUBLAS_SAFE_CALL_IMPL( \ - cublasSetStream(singleton.handle, space.cuda_stream())); \ - cublasZrotg(singleton.handle, \ - reinterpret_cast(a.data()), \ - reinterpret_cast(b.data()), c.data(), \ - reinterpret_cast(s.data())); \ - Kokkos::Profiling::popRegion(); \ - } \ +#define KOKKOSBLAS1_ZROTG_TPL_SPEC_DECL_CUBLAS(LAYOUT, EXECSPACE, MEMSPACE, \ + ETI_SPEC_AVAIL) \ + template <> \ + struct Rotg< \ + EXECSPACE, \ + Kokkos::View, LAYOUT, \ + Kokkos::Device, \ + Kokkos::MemoryTraits>, \ + Kokkos::View, \ + Kokkos::MemoryTraits>, \ + true, ETI_SPEC_AVAIL> { \ + using SViewType = Kokkos::View, LAYOUT, \ + Kokkos::Device, \ + Kokkos::MemoryTraits>; \ + using MViewType = \ + Kokkos::View, \ + Kokkos::MemoryTraits>; \ + static void rotg(EXECSPACE const& space, SViewType const& a, \ + SViewType const& b, MViewType const& c, \ + SViewType const& s) { \ + Kokkos::Profiling::pushRegion( \ + "KokkosBlas::nrm1[TPL_CUBLAS,complex]"); \ + rotg_print_specialization, EXECSPACE>(); \ + KokkosBlas::Impl::CudaBlasSingleton& singleton = \ + KokkosBlas::Impl::CudaBlasSingleton::singleton(); \ + KOKKOS_CUBLAS_SAFE_CALL_IMPL( \ + cublasSetStream(singleton.handle, space.cuda_stream())); \ + cublasZrotg(singleton.handle, \ + reinterpret_cast(a.data()), \ + reinterpret_cast(b.data()), c.data(), \ + reinterpret_cast(s.data())); \ + Kokkos::Profiling::popRegion(); \ + } \ }; #define KOKKOSBLAS1_CROTG_TPL_SPEC_DECL_CUBLAS(LAYOUT, EXECSPACE, MEMSPACE, \ ETI_SPEC_AVAIL) \ - template \ + template <> \ struct Rotg, LAYOUT, \ Kokkos::Device, \ @@ -529,7 +529,7 @@ namespace Impl { #define KOKKOSBLAS1_ZROTG_TPL_SPEC_DECL_ROCBLAS(LAYOUT, EXECSPACE, MEMSPACE, \ ETI_SPEC_AVAIL) \ - template \ + template <> \ struct Rotg< \ EXECSPACE, \ Kokkos::View, LAYOUT, \ @@ -565,7 +565,7 @@ namespace Impl { #define KOKKOSBLAS1_CROTG_TPL_SPEC_DECL_ROCBLAS(LAYOUT, EXECSPACE, MEMSPACE, \ ETI_SPEC_AVAIL) \ - template \ + template <> \ struct Rotg, LAYOUT, \ Kokkos::Device, \