/* //@HEADER // ***************************************************************************** // Adelus // // Copyright 2020 NTESS and the Adelus contributors. // SPDX-License-Identifier: BSD-3-Clause // ***************************************************************************** //@HEADER */ #ifndef __ADELUS_PERMRHS_HPP__ #define __ADELUS_PERMRHS_HPP__ #include #include #include #include #include "Adelus_defines.h" #include "Adelus_macros.h" #include "Adelus_mytime.hpp" #include "Kokkos_Core.hpp" namespace Adelus { template inline void permute_rhs(HandleType& ahandle, ZViewType& RHS, PViewType& permute) { using value_type = typename ZViewType::value_type; using execution_space = typename ZViewType::device_type::execution_space ; using memory_space = typename ZViewType::device_type::memory_space ; using ViewVectorType = Kokkos::View; #ifdef ADELUS_HOST_PINNED_MEM_MPI #if defined(KOKKOS_ENABLE_CUDA) using ViewVectorHostPinnType = Kokkos::View;//CudaHostPinnedSpace #elif defined(KOKKOS_ENABLE_HIP) using ViewVectorHostPinnType = Kokkos::View;//HIPHostPinnedSpace #endif #endif MPI_Comm col_comm = ahandle.get_col_comm(); int myrow = ahandle.get_myrow(); int nprocs_col = ahandle.get_nprocs_col(); int nrows_matrix = ahandle.get_nrows_matrix(); int pivot_row, k_row; ViewVectorType tmpr( "tmpr", RHS.extent(1) ); ViewVectorType tmps( "tmps", RHS.extent(1) ); #if defined(ADELUS_HOST_PINNED_MEM_MPI) && (defined(KOKKOS_ENABLE_CUDA) || defined(KOKKOS_ENABLE_HIP)) ViewVectorHostPinnType h_tmpr( "h_tmpr", RHS.extent(1) ); ViewVectorHostPinnType h_tmps( "h_tmps", RHS.extent(1) ); #endif MPI_Status msgstatus; //TODO: try this later //MPI_Datatype strided_vec_type; //int strided_vec_nblocks = RHS.extent(1); //int strided_vec_blocklen = 1; //int strided_vec_stride = RHS.extent(0); //MPI_Type_vector( strided_vec_nblocks, strided_vec_blocklen, strided_vec_stride, // ADELUS_MPI_DATA_TYPE, &strided_vec_type); //MPI_Type_commit(&strided_vec_type); #ifdef GET_TIMING double permuterhstime,t1; t1 = MPI_Wtime(); #endif for (int k=0;k<=nrows_matrix-2;k++) { k_row=k%nprocs_col; if (ahandle.get_my_rhs() > 0) { if (myrow==k_row) pivot_row = static_cast(permute(k/nprocs_col)); MPI_Bcast(&pivot_row,1,MPI_INT,k_row,col_comm); int pivot_row_pid = pivot_row%nprocs_col; if (k != pivot_row) { if (k_row == pivot_row_pid) {//pivot row is in the same rank if (myrow == k_row) { int curr_lrid = k/nprocs_col; int piv_lrid = pivot_row/nprocs_col; Kokkos::parallel_for(Kokkos::RangePolicy(0,RHS.extent(1)), KOKKOS_LAMBDA (const int i) { value_type tmp = RHS(curr_lrid,i); RHS(curr_lrid,i) = RHS(piv_lrid,i); RHS(piv_lrid,i) = tmp; }); Kokkos::fence(); } } else {//pivot row is is a different rank if (myrow == k_row) { int curr_lrid = k/nprocs_col; Kokkos::parallel_for(Kokkos::RangePolicy(0,RHS.extent(1)), KOKKOS_LAMBDA (const int i) { tmps(i) = RHS(curr_lrid,i); }); #if defined(ADELUS_HOST_PINNED_MEM_MPI) && (defined(KOKKOS_ENABLE_CUDA) || defined(KOKKOS_ENABLE_HIP)) Kokkos::deep_copy(h_tmps,tmps); MPI_Send(reinterpret_cast(h_tmps.data()),RHS.extent(1)*sizeof(value_type),MPI_CHAR,pivot_row_pid,2,col_comm); MPI_Recv(reinterpret_cast(h_tmpr.data()),RHS.extent(1)*sizeof(value_type),MPI_CHAR,pivot_row_pid,3,col_comm,&msgstatus); Kokkos::deep_copy(tmpr,h_tmpr); #else //GPU-aware MPI Kokkos::fence(); MPI_Send(reinterpret_cast(tmps.data()),RHS.extent(1)*sizeof(value_type),MPI_CHAR,pivot_row_pid,2,col_comm); MPI_Recv(reinterpret_cast(tmpr.data()),RHS.extent(1)*sizeof(value_type),MPI_CHAR,pivot_row_pid,3,col_comm,&msgstatus); #endif Kokkos::parallel_for(Kokkos::RangePolicy(0,RHS.extent(1)), KOKKOS_LAMBDA (const int i) { RHS(curr_lrid,i) = tmpr(i); }); Kokkos::fence(); } if (myrow == pivot_row_pid) { int piv_lrid = pivot_row/nprocs_col; Kokkos::parallel_for(Kokkos::RangePolicy(0,RHS.extent(1)), KOKKOS_LAMBDA (const int i) { tmps(i) = RHS(piv_lrid,i); }); #if defined(ADELUS_HOST_PINNED_MEM_MPI) && (defined(KOKKOS_ENABLE_CUDA) || defined(KOKKOS_ENABLE_HIP)) Kokkos::deep_copy(h_tmps,tmps); MPI_Recv(reinterpret_cast(h_tmpr.data()),RHS.extent(1)*sizeof(value_type),MPI_CHAR,k_row,2,col_comm,&msgstatus); MPI_Send(reinterpret_cast(h_tmps.data()),RHS.extent(1)*sizeof(value_type),MPI_CHAR,k_row,3,col_comm); Kokkos::deep_copy(tmpr,h_tmpr); #else // GPU-aware MPI Kokkos::fence(); MPI_Recv(reinterpret_cast(tmpr.data()),RHS.extent(1)*sizeof(value_type),MPI_CHAR,k_row,2,col_comm,&msgstatus); MPI_Send(reinterpret_cast(tmps.data()),RHS.extent(1)*sizeof(value_type),MPI_CHAR,k_row,3,col_comm); #endif Kokkos::parallel_for(Kokkos::RangePolicy(0,RHS.extent(1)), KOKKOS_LAMBDA (const int i) { RHS(piv_lrid,i) = tmpr(i); }); Kokkos::fence(); } }//End of pivot row is is a different rank }// End of if (k != pivot_row) }// End of if (my_num_rhs > 0) }// End of for (k=0;k<=nrows_matrix-2;k++) #ifdef GET_TIMING permuterhstime = MPI_Wtime()-t1; showtime(ahandle.get_comm_id(), ahandle.get_comm(), ahandle.get_myrank(), ahandle.get_nprocs_cube(), "Time to permute rhs", &permuterhstime); #endif }// End of function permute_rhs }//namespace Adelus #endif