From 7f3f96b6604c7fff1cb46cd694bdcc1cb8f2d90a Mon Sep 17 00:00:00 2001 From: Bruno Turcksin Date: Thu, 2 Aug 2018 14:51:35 +0000 Subject: [PATCH] Add tests --- tests/cuda/parallel_vector_01.cu | 110 ++++++++ tests/cuda/parallel_vector_01.mpirun=2.output | 6 + tests/cuda/parallel_vector_02.cu | 121 ++++++++ tests/cuda/parallel_vector_02.mpirun=2.output | 6 + tests/cuda/parallel_vector_03.cu | 126 +++++++++ tests/cuda/parallel_vector_03.mpirun=2.output | 6 + tests/cuda/parallel_vector_03a.cu | 180 ++++++++++++ .../cuda/parallel_vector_03a.mpirun=2.output | 2 + tests/cuda/parallel_vector_04.cu | 181 ++++++++++++ tests/cuda/parallel_vector_04.mpirun=2.output | 3 + tests/cuda/parallel_vector_05.cu | 122 +++++++++ tests/cuda/parallel_vector_05.mpirun=2.output | 3 + tests/cuda/parallel_vector_06.cu | 168 ++++++++++++ tests/cuda/parallel_vector_06.mpirun=2.output | 11 + tests/cuda/parallel_vector_08.cu | 150 ++++++++++ tests/cuda/parallel_vector_08.mpirun=2.output | 3 + tests/cuda/parallel_vector_10.cu | 116 ++++++++ tests/cuda/parallel_vector_10.mpirun=2.output | 3 + tests/cuda/parallel_vector_11.cu | 258 ++++++++++++++++++ tests/cuda/parallel_vector_11.mpirun=2.output | 16 ++ tests/cuda/parallel_vector_12.cu | 241 ++++++++++++++++ tests/cuda/parallel_vector_12.mpirun=2.output | 8 + tests/cuda/parallel_vector_13.cu | 137 ++++++++++ tests/cuda/parallel_vector_13.mpirun=2.output | 8 + tests/cuda/parallel_vector_14.cu | 185 +++++++++++++ tests/cuda/parallel_vector_14.mpirun=2.output | 4 + tests/cuda/parallel_vector_15.cu | 135 +++++++++ tests/cuda/parallel_vector_15.mpirun=2.output | 10 + tests/cuda/parallel_vector_21.cu | 104 +++++++ tests/cuda/parallel_vector_21.mpirun=2.output | 23 ++ tests/cuda/vector_reinit_01.cu | 119 ++++++++ tests/cuda/vector_reinit_01.output | 4 + 32 files changed, 2569 insertions(+) create mode 100644 tests/cuda/parallel_vector_01.cu create mode 100644 tests/cuda/parallel_vector_01.mpirun=2.output create mode 100644 tests/cuda/parallel_vector_02.cu create mode 100644 tests/cuda/parallel_vector_02.mpirun=2.output create mode 100644 tests/cuda/parallel_vector_03.cu create mode 100644 tests/cuda/parallel_vector_03.mpirun=2.output create mode 100644 tests/cuda/parallel_vector_03a.cu create mode 100644 tests/cuda/parallel_vector_03a.mpirun=2.output create mode 100644 tests/cuda/parallel_vector_04.cu create mode 100644 tests/cuda/parallel_vector_04.mpirun=2.output create mode 100644 tests/cuda/parallel_vector_05.cu create mode 100644 tests/cuda/parallel_vector_05.mpirun=2.output create mode 100644 tests/cuda/parallel_vector_06.cu create mode 100644 tests/cuda/parallel_vector_06.mpirun=2.output create mode 100644 tests/cuda/parallel_vector_08.cu create mode 100644 tests/cuda/parallel_vector_08.mpirun=2.output create mode 100644 tests/cuda/parallel_vector_10.cu create mode 100644 tests/cuda/parallel_vector_10.mpirun=2.output create mode 100644 tests/cuda/parallel_vector_11.cu create mode 100644 tests/cuda/parallel_vector_11.mpirun=2.output create mode 100644 tests/cuda/parallel_vector_12.cu create mode 100644 tests/cuda/parallel_vector_12.mpirun=2.output create mode 100644 tests/cuda/parallel_vector_13.cu create mode 100644 tests/cuda/parallel_vector_13.mpirun=2.output create mode 100644 tests/cuda/parallel_vector_14.cu create mode 100644 tests/cuda/parallel_vector_14.mpirun=2.output create mode 100644 tests/cuda/parallel_vector_15.cu create mode 100644 tests/cuda/parallel_vector_15.mpirun=2.output create mode 100644 tests/cuda/parallel_vector_21.cu create mode 100644 tests/cuda/parallel_vector_21.mpirun=2.output create mode 100644 tests/cuda/vector_reinit_01.cu create mode 100644 tests/cuda/vector_reinit_01.output diff --git a/tests/cuda/parallel_vector_01.cu b/tests/cuda/parallel_vector_01.cu new file mode 100644 index 0000000000..9e53fb11e9 --- /dev/null +++ b/tests/cuda/parallel_vector_01.cu @@ -0,0 +1,110 @@ +// --------------------------------------------------------------------- +// +// Copyright (C) 2018 by the deal.II authors +// +// This file is part of the deal.II library. +// +// The deal.II library is free software; you can use it, redistribute +// it, and/or modify it under the terms of the GNU Lesser General +// Public License as published by the Free Software Foundation; either +// version 2.1 of the License, or (at your option) any later version. +// The full text of the license can be found in the file LICENSE.md at +// the top level directory of deal.II. +// +// --------------------------------------------------------------------- + + +// check correct initialization of parallel vector without any ghosts + +#include +#include +#include + +#include +#include + +#include +#include + +#include "../tests.h" + + +void +test() +{ + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + unsigned int numproc = Utilities::MPI::n_mpi_processes(MPI_COMM_WORLD); + + if (myid == 0) + deallog << "numproc=" << numproc << std::endl; + + + // each processor owns 2 indices and all + // are ghosting element 1 (the second) + IndexSet local_owned(numproc * 2); + local_owned.add_range(myid * 2, myid * 2 + 2); + + LinearAlgebra::distributed::Vector v( + local_owned, local_owned, MPI_COMM_WORLD); + + // set local values + LinearAlgebra::ReadWriteVector rw_vector(local_owned); + rw_vector(myid * 2) = myid * 2.0; + rw_vector(myid * 2 + 1) = myid * 2.0 + 1.0; + v.import(rw_vector, VectorOperation::insert); + + v *= 2.0; + + rw_vector.import(v, VectorOperation::insert); + if (myid == 0) + { + deallog << myid * 2 << ":" << rw_vector(myid * 2) << std::endl; + deallog << myid * 2 + 1 << ":" << rw_vector(myid * 2 + 1) << std::endl; + } + + Assert(rw_vector(myid * 2) == myid * 4.0, ExcInternalError()); + Assert(rw_vector(myid * 2 + 1) == myid * 4.0 + 2.0, ExcInternalError()); + + // check l2 norm + const double l2_norm = v.l2_norm(); + if (myid == 0) + deallog << "L2 norm: " << l2_norm << std::endl; + + if (myid == 0) + deallog << "OK" << std::endl; +} + + + +int +main(int argc, char **argv) +{ + Utilities::MPI::MPI_InitFinalize mpi_initialization( + argc, argv, testing_max_num_threads()); + + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + deallog.push(Utilities::int_to_string(myid)); + + Utilities::CUDA::Handle cuda_handle; + // By default, all the ranks will try to access the device 0. This is fine if + // we have one rank per node _and_ one gpu per node. If we have multiple GPUs + // on one node, we need each process to access a different GPU. We assume that + // each node has the same number of GPUs. + int n_devices = 0; + cudaError_t cuda_error_code = cudaGetDeviceCount(&n_devices); + AssertCuda(cuda_error_code); + int device_id = myid % n_devices; + cuda_error_code = cudaSetDevice(device_id); + AssertCuda(cuda_error_code); + + + if (myid == 0) + { + initlog(); + deallog << std::setprecision(4); + + test(); + } + else + test(); +} diff --git a/tests/cuda/parallel_vector_01.mpirun=2.output b/tests/cuda/parallel_vector_01.mpirun=2.output new file mode 100644 index 0000000000..5f15965ac3 --- /dev/null +++ b/tests/cuda/parallel_vector_01.mpirun=2.output @@ -0,0 +1,6 @@ + +DEAL:0::numproc=2 +DEAL:0::0:0 +DEAL:0::1:2.000 +DEAL:0::L2 norm: 7.483 +DEAL:0::OK diff --git a/tests/cuda/parallel_vector_02.cu b/tests/cuda/parallel_vector_02.cu new file mode 100644 index 0000000000..7c1d0ba0bc --- /dev/null +++ b/tests/cuda/parallel_vector_02.cu @@ -0,0 +1,121 @@ +// --------------------------------------------------------------------- +// +// Copyright (C) 2018 by the deal.II authors +// +// This file is part of the deal.II library. +// +// The deal.II library is free software; you can use it, redistribute +// it, and/or modify it under the terms of the GNU Lesser General +// Public License as published by the Free Software Foundation; either +// version 2.1 of the License, or (at your option) any later version. +// The full text of the license can be found in the file LICENSE.md at +// the top level directory of deal.II. +// +// --------------------------------------------------------------------- + + +// check addition into ghosts for parallel vector + +#include +#include +#include + +#include +#include + +#include +#include + +#include "../tests.h" + + +void +test() +{ + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + unsigned int numproc = Utilities::MPI::n_mpi_processes(MPI_COMM_WORLD); + + if (myid == 0) + deallog << "numproc=" << numproc << std::endl; + + + // each processor owns 2 indices and all + // are ghosting element 1 (the second) + IndexSet local_owned(numproc * 2); + local_owned.add_range(myid * 2, myid * 2 + 2); + IndexSet local_relevant(numproc * 2); + local_relevant = local_owned; + local_relevant.add_range(1, 2); + + LinearAlgebra::distributed::Vector v( + local_owned, local_relevant, MPI_COMM_WORLD); + + // set local values and check them + LinearAlgebra::ReadWriteVector rw_vector(local_owned); + rw_vector(myid * 2) = myid * 2.0; + rw_vector(myid * 2 + 1) = myid * 2.0 + 1.0; + v.import(rw_vector, VectorOperation::add); + + v *= 2.0; + + rw_vector.import(v, VectorOperation::insert); + AssertThrow(rw_vector(myid * 2) == myid * 4.0, ExcInternalError()); + AssertThrow(rw_vector(myid * 2 + 1) == myid * 4.0 + 2.0, ExcInternalError()); + + // set ghost dof, compress + LinearAlgebra::ReadWriteVector rw_relevant_vector(numproc * 2); + rw_relevant_vector(1) = 7; + v.import(rw_relevant_vector, VectorOperation::add); + + rw_vector.import(v, VectorOperation::insert); + if (myid == 0) + { + deallog << myid * 2 << ":" << rw_vector(myid * 2) << std::endl; + deallog << myid * 2 + 1 << ":" << rw_vector(myid * 2 + 1) << std::endl; + } + + rw_relevant_vector.import(v, VectorOperation::insert); + AssertThrow(rw_relevant_vector(1) == 7. * numproc + 2, ExcInternalError()); + + // check l2 norm + const double l2_norm = v.l2_norm(); + if (myid == 0) + deallog << "L2 norm: " << l2_norm << std::endl; + + if (myid == 0) + deallog << "OK" << std::endl; +} + + + +int +main(int argc, char **argv) +{ + Utilities::MPI::MPI_InitFinalize mpi_initialization( + argc, argv, testing_max_num_threads()); + + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + deallog.push(Utilities::int_to_string(myid)); + + Utilities::CUDA::Handle cuda_handle; + // By default, all the ranks will try to access the device 0. This is fine if + // we have one rank per node _and_ one gpu per node. If we have multiple GPUs + // on one node, we need each process to access a different GPU. We assume that + // each node has the same number of GPUs. + int n_devices = 0; + cudaError_t cuda_error_code = cudaGetDeviceCount(&n_devices); + AssertCuda(cuda_error_code); + int device_id = myid % n_devices; + cuda_error_code = cudaSetDevice(device_id); + AssertCuda(cuda_error_code); + + if (myid == 0) + { + initlog(); + deallog << std::setprecision(4); + + test(); + } + else + test(); +} diff --git a/tests/cuda/parallel_vector_02.mpirun=2.output b/tests/cuda/parallel_vector_02.mpirun=2.output new file mode 100644 index 0000000000..9e9690e2a9 --- /dev/null +++ b/tests/cuda/parallel_vector_02.mpirun=2.output @@ -0,0 +1,6 @@ + +DEAL:0::numproc=2 +DEAL:0::0:0 +DEAL:0::1:16.00 +DEAL:0::L2 norm: 17.55 +DEAL:0::OK diff --git a/tests/cuda/parallel_vector_03.cu b/tests/cuda/parallel_vector_03.cu new file mode 100644 index 0000000000..1dfb25ceb2 --- /dev/null +++ b/tests/cuda/parallel_vector_03.cu @@ -0,0 +1,126 @@ +// --------------------------------------------------------------------- +// +// Copyright (C) 2018 by the deal.II authors +// +// This file is part of the deal.II library. +// +// The deal.II library is free software; you can use it, redistribute +// it, and/or modify it under the terms of the GNU Lesser General +// Public License as published by the Free Software Foundation; either +// version 2.1 of the License, or (at your option) any later version. +// The full text of the license can be found in the file LICENSE.md at +// the top level directory of deal.II. +// +// --------------------------------------------------------------------- + + +// check set of ghosts for parallel vector + +#include +#include +#include + +#include +#include + +#include +#include + +#include "../tests.h" + + +void +test() +{ + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + unsigned int numproc = Utilities::MPI::n_mpi_processes(MPI_COMM_WORLD); + + if (myid == 0) + deallog << "numproc=" << numproc << std::endl; + + + // each processor owns 2 indices and all + // are ghosting element 1 (the second) + IndexSet local_owned(numproc * 2); + local_owned.add_range(myid * 2, myid * 2 + 2); + IndexSet local_relevant(numproc * 2); + local_relevant = local_owned; + local_relevant.add_range(1, 2); + + LinearAlgebra::distributed::Vector v( + local_owned, local_relevant, MPI_COMM_WORLD); + + // set local values and check them + LinearAlgebra::ReadWriteVector rw_vector(local_owned); + rw_vector(myid * 2) = myid * 2.0; + rw_vector(myid * 2 + 1) = myid * 2.0 + 1.0; + v.import(rw_vector, VectorOperation::insert); + + v *= 2.0; + + rw_vector.import(v, VectorOperation::insert); + AssertThrow(rw_vector(myid * 2) == myid * 4.0, ExcInternalError()); + AssertThrow(rw_vector(myid * 2 + 1) == myid * 4.0 + 2.0, ExcInternalError()); + + // set ghost dof on all processors, compress + // (insert mode) + IndexSet ghost_entry(numproc * 2); + ghost_entry.add_index(1); + LinearAlgebra::ReadWriteVector rw_ghost_entry(ghost_entry); + rw_ghost_entry(1) = 7.; + v.import(rw_ghost_entry, VectorOperation::insert); + + rw_vector.import(v, VectorOperation::insert); + if (myid == 0) + { + deallog << myid * 2 << ":" << rw_vector(myid * 2) << std::endl; + deallog << myid * 2 + 1 << ":" << rw_vector(myid * 2 + 1) << std::endl; + } + // import ghosts onto all procs + v.update_ghost_values(); + rw_ghost_entry(1) = 0.; + rw_ghost_entry.import(v, VectorOperation::insert); + AssertThrow(rw_ghost_entry(1) == 7.0, ExcInternalError()); + + // check l2 norm + const double l2_norm = v.l2_norm(); + if (myid == 0) + deallog << "L2 norm: " << l2_norm << std::endl; + + if (myid == 0) + deallog << "OK" << std::endl; +} + + + +int +main(int argc, char **argv) +{ + Utilities::MPI::MPI_InitFinalize mpi_initialization( + argc, argv, testing_max_num_threads()); + + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + deallog.push(Utilities::int_to_string(myid)); + + Utilities::CUDA::Handle cuda_handle; + // By default, all the ranks will try to access the device 0. This is fine if + // we have one rank per node _and_ one gpu per node. If we have multiple GPUs + // on one node, we need each process to access a different GPU. We assume that + // each node has the same number of GPUs. + int n_devices = 0; + cudaError_t cuda_error_code = cudaGetDeviceCount(&n_devices); + AssertCuda(cuda_error_code); + int device_id = myid % n_devices; + cuda_error_code = cudaSetDevice(device_id); + AssertCuda(cuda_error_code); + + if (myid == 0) + { + initlog(); + deallog << std::setprecision(4); + + test(); + } + else + test(); +} diff --git a/tests/cuda/parallel_vector_03.mpirun=2.output b/tests/cuda/parallel_vector_03.mpirun=2.output new file mode 100644 index 0000000000..94ec6d82d7 --- /dev/null +++ b/tests/cuda/parallel_vector_03.mpirun=2.output @@ -0,0 +1,6 @@ + +DEAL:0::numproc=2 +DEAL:0::0:0 +DEAL:0::1:7.000 +DEAL:0::L2 norm: 10.05 +DEAL:0::OK diff --git a/tests/cuda/parallel_vector_03a.cu b/tests/cuda/parallel_vector_03a.cu new file mode 100644 index 0000000000..ddd3eac749 --- /dev/null +++ b/tests/cuda/parallel_vector_03a.cu @@ -0,0 +1,180 @@ +// --------------------------------------------------------------------- +// +// Copyright (C) 2018 by the deal.II authors +// +// This file is part of the deal.II library. +// +// The deal.II library is free software; you can use it, redistribute +// it, and/or modify it under the terms of the GNU Lesser General +// Public License as published by the Free Software Foundation; either +// version 2.1 of the License, or (at your option) any later version. +// The full text of the license can be found in the file LICENSE.md at +// the top level directory of deal.II. +// +// --------------------------------------------------------------------- + + +// similar to parallel_sparse_vector_03.cc, but make sure +// compress(insert) zeroes out ghosts in Release mode + +#include +#include +#include + +#include +#include + +#include +#include +#include + +#include "../tests.h" + +void +check(const unsigned int myid, const LinearAlgebra::ReadWriteVector &v) +{ + if (myid == 0) + { + AssertThrow(v(10) == 10.0, ExcInternalError()); + AssertThrow(v(11) == 0., ExcInternalError()); + AssertThrow(v(12) == 0., ExcInternalError()); + AssertThrow(v(14) == 14., ExcInternalError()); + + AssertThrow(v(5) == 55., ExcInternalError()); + } + else + { + AssertThrow(v(4) == 0., ExcInternalError()); + AssertThrow(v(5) == 55., ExcInternalError()); + AssertThrow(v(6) == 66., ExcInternalError()); + } +} + + +void +test() +{ + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + unsigned int numproc = Utilities::MPI::n_mpi_processes(MPI_COMM_WORLD); + + Assert(numproc == 2, ExcNotImplemented()); + + const unsigned int size = 20; + IndexSet local_owned(size); + IndexSet local_nonzero(size); + IndexSet local_relevant(size); + if (myid == 0) + { + local_owned.add_range(0, 10); + local_nonzero.add_range(5, 10); + local_relevant = local_owned; + local_relevant.add_range(10, 13); + local_relevant.add_range(14, 15); + } + else + { + local_owned.add_range(10, size); + local_nonzero.add_range(10, 11); + local_nonzero.add_range(13, 15); + local_relevant = local_owned; + local_relevant.add_range(4, 7); + } + + LinearAlgebra::distributed::Vector v( + local_owned, local_relevant, MPI_COMM_WORLD); + v = 0.; + + // set local values + IndexSet indexset_1(local_owned); + indexset_1.add_index(5); + indexset_1.compress(); + LinearAlgebra::ReadWriteVector rw_vector(indexset_1); + for (unsigned int i = 0; i < local_nonzero.n_elements(); i++) + rw_vector(local_nonzero.nth_index_in_set(i)) = + local_nonzero.nth_index_in_set(i); + + // set value from processor which does not own it: + rw_vector(5) = 55.; + v.import(rw_vector, VectorOperation::insert); + + // add to value from processor which has it as a ghost + // Because of limitation in import, the ReadWriteVector needs to have locally + // owned elements + IndexSet indexset_2(local_owned.size()); + LinearAlgebra::ReadWriteVector rw_add; + if (myid == 1) + { + indexset_2.add_index(6); + rw_add.reinit(indexset_2); + rw_add(6) = 60; + } + else + { + rw_add.reinit(indexset_2); + } + v.import(rw_add, VectorOperation::add); // 60 + 6 + // compress(insert) used to leave ghosts un-touched which resulted in + // the wrong 55+55 for this compress(add) operation. + + v.update_ghost_values(); + + IndexSet indexset_3(size); + if (myid == 0) + { + indexset_3.add_index(10); + indexset_3.add_index(11); + indexset_3.add_index(12); + indexset_3.add_index(14); + + indexset_3.add_index(5); + } + else + { + indexset_3.add_index(4); + indexset_3.add_index(5); + indexset_3.add_index(6); + } + + rw_vector.reinit(indexset_3); + rw_vector.import(v, VectorOperation::insert); + + check(myid, rw_vector); + + if (myid == 0) + deallog << "OK" << std::endl; +} + + + +int +main(int argc, char **argv) +{ + Utilities::MPI::MPI_InitFinalize mpi_initialization( + argc, argv, testing_max_num_threads()); + + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + deallog.push(Utilities::int_to_string(myid)); + + Utilities::CUDA::Handle cuda_handle; + // By default, all the ranks will try to access the device 0. This is fine if + // we have one rank per node _and_ one gpu per node. If we have multiple GPUs + // on one node, we need each process to access a different GPU. We assume that + // each node has the same number of GPUs. + int n_devices = 0; + cudaError_t cuda_error_code = cudaGetDeviceCount(&n_devices); + AssertCuda(cuda_error_code); + int device_id = myid % n_devices; + cuda_error_code = cudaSetDevice(device_id); + AssertCuda(cuda_error_code); + + if (myid == 0) + { + std::ofstream logfile("output"); + deallog.attach(logfile); + deallog << std::setprecision(4); + + test(); + } + else + test(); +} diff --git a/tests/cuda/parallel_vector_03a.mpirun=2.output b/tests/cuda/parallel_vector_03a.mpirun=2.output new file mode 100644 index 0000000000..be8d055f86 --- /dev/null +++ b/tests/cuda/parallel_vector_03a.mpirun=2.output @@ -0,0 +1,2 @@ + +DEAL:0::OK diff --git a/tests/cuda/parallel_vector_04.cu b/tests/cuda/parallel_vector_04.cu new file mode 100644 index 0000000000..94731f7fd1 --- /dev/null +++ b/tests/cuda/parallel_vector_04.cu @@ -0,0 +1,181 @@ +// --------------------------------------------------------------------- +// +// Copyright (C) 2018 by the deal.II authors +// +// This file is part of the deal.II library. +// +// The deal.II library is free software; you can use it, redistribute +// it, and/or modify it under the terms of the GNU Lesser General +// Public License as published by the Free Software Foundation; either +// version 2.1 of the License, or (at your option) any later version. +// The full text of the license can be found in the file LICENSE.md at +// the top level directory of deal.II. +// +// --------------------------------------------------------------------- + + +// check that operator= resets ghosts, both if they have been set and if they +// have not been set + +#include +#include +#include + +#include +#include + +#include +#include + +#include "../tests.h" + + +__global__ void +set_value(double *values_dev, unsigned int index, double val) +{ + values_dev[index] = val; +} + +void +test() +{ + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + unsigned int numproc = Utilities::MPI::n_mpi_processes(MPI_COMM_WORLD); + + if (myid == 0) + deallog << "numproc=" << numproc << std::endl; + + + // each processor owns 2 indices and all + // are ghosting element 1 (the second) + IndexSet local_owned(numproc * 2); + local_owned.add_range(myid * 2, myid * 2 + 2); + IndexSet local_relevant(numproc * 2); + local_relevant = local_owned; + local_relevant.add_range(1, 2); + + LinearAlgebra::distributed::Vector v( + local_owned, local_relevant, MPI_COMM_WORLD); + + // set local values and check them + LinearAlgebra::ReadWriteVector rw_vector(local_owned); + rw_vector(myid * 2) = myid * 2.0; + rw_vector(myid * 2 + 1) = myid * 2.0 + 1.0; + + v.import(rw_vector, VectorOperation::insert); + v *= 2.0; + + rw_vector.import(v, VectorOperation::insert); + AssertThrow(rw_vector(myid * 2) == myid * 4.0, ExcInternalError()); + AssertThrow(rw_vector(myid * 2 + 1) == myid * 4.0 + 2.0, ExcInternalError()); + + // set ghost dof on remote process, no compress called. Since we don't want to + // call compress we cannot use import + auto partitioner = v.get_partitioner(); + if (myid > 0) + { + unsigned int local_index = partitioner->global_to_local(1); + double * values_dev = v.get_values(); + set_value<<<1, 1>>>(values_dev, local_index, 7); + } + + unsigned int allocated_size = local_relevant.n_elements(); + std::vector v_host(allocated_size); + Utilities::CUDA::copy_to_host(v.get_values(), v_host); + + AssertThrow(v_host[partitioner->global_to_local(myid * 2)] == myid * 4.0, + ExcInternalError()); + AssertThrow(v_host[partitioner->global_to_local(myid * 2 + 1)] == + myid * 4.0 + 2.0, + ExcInternalError()); + + if (myid > 0) + AssertThrow(v_host[partitioner->global_to_local(1)] == 7.0, + ExcInternalError()); + + // reset to zero + v = 0; + + Utilities::CUDA::copy_to_host(v.get_values(), v_host); + AssertThrow(v_host[partitioner->global_to_local(myid * 2)] == 0., + ExcInternalError()); + AssertThrow(v_host[partitioner->global_to_local(myid * 2 + 1)] == 0., + ExcInternalError()); + + // check that everything remains zero also + // after compress + v.compress(VectorOperation::add); + + Utilities::CUDA::copy_to_host(v.get_values(), v_host); + AssertThrow(v_host[partitioner->global_to_local(myid * 2)] == 0., + ExcInternalError()); + AssertThrow(v_host[partitioner->global_to_local(myid * 2 + 1)] == 0., + ExcInternalError()); + + // set element 1 on owning process to + // something nonzero + if (myid == 0) + { + unsigned int local_index = partitioner->global_to_local(1); + double * values_dev = v.get_values(); + set_value<<<1, 1>>>(values_dev, local_index, 2); + } + if (myid > 0) + { + Utilities::CUDA::copy_to_host(v.get_values(), v_host); + AssertThrow(v_host[partitioner->global_to_local(1)] == 0., + ExcInternalError()); + } + + // check that all processors get the correct + // value again, and that it is erased by + // operator= + v.update_ghost_values(); + + Utilities::CUDA::copy_to_host(v.get_values(), v_host); + AssertThrow(v_host[partitioner->global_to_local(1)] == 2., + ExcInternalError()); + + v = 0; + Utilities::CUDA::copy_to_host(v.get_values(), v_host); + AssertThrow(v_host[partitioner->global_to_local(1)] == 0., + ExcInternalError()); + + if (myid == 0) + deallog << "OK" << std::endl; +} + + + +int +main(int argc, char **argv) +{ + Utilities::MPI::MPI_InitFinalize mpi_initialization( + argc, argv, testing_max_num_threads()); + + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + deallog.push(Utilities::int_to_string(myid)); + + Utilities::CUDA::Handle cuda_handle; + // By default, all the ranks will try to access the device 0. This is fine if + // we have one rank per node _and_ one gpu per node. If we have multiple GPUs + // on one node, we need each process to access a different GPU. We assume that + // each node has the same number of GPUs. + int n_devices = 0; + cudaError_t cuda_error_code = cudaGetDeviceCount(&n_devices); + AssertCuda(cuda_error_code); + int device_id = myid % n_devices; + cuda_error_code = cudaSetDevice(device_id); + AssertCuda(cuda_error_code); + + + if (myid == 0) + { + initlog(); + deallog << std::setprecision(4); + + test(); + } + else + test(); +} diff --git a/tests/cuda/parallel_vector_04.mpirun=2.output b/tests/cuda/parallel_vector_04.mpirun=2.output new file mode 100644 index 0000000000..1cf39ab251 --- /dev/null +++ b/tests/cuda/parallel_vector_04.mpirun=2.output @@ -0,0 +1,3 @@ + +DEAL:0::numproc=2 +DEAL:0::OK diff --git a/tests/cuda/parallel_vector_05.cu b/tests/cuda/parallel_vector_05.cu new file mode 100644 index 0000000000..206c6c468e --- /dev/null +++ b/tests/cuda/parallel_vector_05.cu @@ -0,0 +1,122 @@ +// --------------------------------------------------------------------- +// +// Copyright (C) 2018 by the deal.II authors +// +// This file is part of the deal.II library. +// +// The deal.II library is free software; you can use it, redistribute +// it, and/or modify it under the terms of the GNU Lesser General +// Public License as published by the Free Software Foundation; either +// version 2.1 of the License, or (at your option) any later version. +// The full text of the license can be found in the file LICENSE.md at +// the top level directory of deal.II. +// +// --------------------------------------------------------------------- + + +// check that compress(add) with zero add does not change the vector entry + +#include +#include +#include + +#include +#include + +#include +#include + +#include "../tests.h" + + +void +test() +{ + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + unsigned int numproc = Utilities::MPI::n_mpi_processes(MPI_COMM_WORLD); + + if (myid == 0) + deallog << "numproc=" << numproc << std::endl; + + + // each processor owns 2 indices and all + // are ghosting element 1 (the second) + IndexSet local_owned(numproc * 2); + local_owned.add_range(myid * 2, myid * 2 + 2); + IndexSet local_relevant(numproc * 2); + local_relevant = local_owned; + local_relevant.add_range(1, 2); + + LinearAlgebra::distributed::Vector v( + local_owned, local_relevant, MPI_COMM_WORLD); + + // set local values and check them + LinearAlgebra::ReadWriteVector rw_vector(local_owned); + rw_vector(myid * 2) = myid * 2.0; + rw_vector(myid * 2 + 1) = myid * 2.0 + 1.0; + + v.import(rw_vector, VectorOperation::insert); + v *= 2.0; + + rw_vector.import(v, VectorOperation::insert); + Assert(rw_vector(myid * 2) == myid * 4.0, ExcInternalError()); + Assert(rw_vector(myid * 2 + 1) == myid * 4.0 + 2.0, ExcInternalError()); + + // set ghost dof to zero on remote processors, + // compress + IndexSet ghost_set(numproc * 2); + LinearAlgebra::ReadWriteVector ghost_vector; + if (myid > 0) + { + ghost_set.add_index(1); + ghost_vector.reinit(ghost_set); + ghost_vector(1) = 0; + } + else + ghost_vector.reinit(ghost_set); + + + v.import(ghost_vector, VectorOperation::add); + + // check that nothing has changed + rw_vector.import(v, VectorOperation::insert); + Assert(rw_vector(myid * 2) == myid * 4.0, ExcInternalError()); + Assert(rw_vector(myid * 2 + 1) == myid * 4.0 + 2.0, ExcInternalError()); + + if (myid == 0) + deallog << "OK" << std::endl; +} + + + +int +main(int argc, char **argv) +{ + Utilities::MPI::MPI_InitFinalize mpi_initialization( + argc, argv, testing_max_num_threads()); + + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + deallog.push(Utilities::int_to_string(myid)); + + Utilities::CUDA::Handle cuda_handle; + // By default, all the ranks will try to access the device 0. This is fine if + // we have one rank per node _and_ one gpu per node. If we have multiple GPUs + // on one node, we need each process to access a different GPU. We assume that + // each node has the same number of GPUs. + int n_devices = 0; + cudaError_t cuda_error_code = cudaGetDeviceCount(&n_devices); + AssertCuda(cuda_error_code); + int device_id = myid % n_devices; + cuda_error_code = cudaSetDevice(device_id); + AssertCuda(cuda_error_code); + + if (myid == 0) + { + initlog(); + deallog << std::setprecision(4); + + test(); + } + else + test(); +} diff --git a/tests/cuda/parallel_vector_05.mpirun=2.output b/tests/cuda/parallel_vector_05.mpirun=2.output new file mode 100644 index 0000000000..1cf39ab251 --- /dev/null +++ b/tests/cuda/parallel_vector_05.mpirun=2.output @@ -0,0 +1,3 @@ + +DEAL:0::numproc=2 +DEAL:0::OK diff --git a/tests/cuda/parallel_vector_06.cu b/tests/cuda/parallel_vector_06.cu new file mode 100644 index 0000000000..a1ec4beb9b --- /dev/null +++ b/tests/cuda/parallel_vector_06.cu @@ -0,0 +1,168 @@ +// --------------------------------------------------------------------- +// +// Copyright (C) 2018 by the deal.II authors +// +// This file is part of the deal.II library. +// +// The deal.II library is free software; you can use it, redistribute +// it, and/or modify it under the terms of the GNU Lesser General +// Public License as published by the Free Software Foundation; either +// version 2.1 of the License, or (at your option) any later version. +// The full text of the license can be found in the file LICENSE.md at +// the top level directory of deal.II. +// +// --------------------------------------------------------------------- + + +// check global reduction operation (norms, operator==, operator!=) on +// parallel vector + +#include +#include +#include + +#include +#include + +#include +#include + +#include "../tests.h" + + +void +test() +{ + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + unsigned int numproc = Utilities::MPI::n_mpi_processes(MPI_COMM_WORLD); + + if (myid == 0) + deallog << "numproc=" << numproc << std::endl; + + + // each processor owns 2 indices and all processors are + // ghosting element 1 (the second) + IndexSet local_owned(std::min(16U, numproc * 2)); + local_owned.add_range(myid * 2, myid * 2 + 2); + IndexSet local_relevant(numproc * 2); + local_relevant = local_owned; + local_relevant.add_range(1, 2); + + LinearAlgebra::distributed::Vector v( + local_owned, local_owned, MPI_COMM_WORLD); + + // set local values + LinearAlgebra::ReadWriteVector rw_vector(local_owned); + { + rw_vector(myid * 2) = myid * 2.0; + rw_vector(myid * 2 + 1) = myid * 2.0 + 1.0; + } + v.import(rw_vector, VectorOperation::insert); + v *= 2.0; + { + rw_vector.import(v, VectorOperation::insert); + AssertThrow(rw_vector(myid * 2) == myid * 4.0, ExcInternalError()); + AssertThrow(rw_vector(myid * 2 + 1) == myid * 4.0 + 2.0, + ExcInternalError()); + } + + // check l2 norm + { + const double l2_norm = v.l2_norm(); + if (myid == 0) + deallog << "l2 norm: " << l2_norm << std::endl; + } + + // check l1 norm + { + const double l1_norm = v.l1_norm(); + if (myid == 0) + deallog << "l1 norm: " << l1_norm << std::endl; + } + + // check linfty norm + { + const double linfty_norm = v.linfty_norm(); + if (myid == 0) + deallog << "linfty norm: " << linfty_norm << std::endl; + } + + // check mean value (should be equal to l1 + // norm divided by vector size here since we + // have no negative entries) + { + const double mean = v.mean_value(); + if (myid == 0) + deallog << "Mean value: " << mean << std::endl; + + Assert(std::fabs(mean * v.size() - v.l1_norm()) < 1e-15, + ExcInternalError()); + } + // check inner product + { + const double norm_sqr = v.l2_norm() * v.l2_norm(); + AssertThrow(std::fabs(v * v - norm_sqr) < 1e-15, ExcInternalError()); + LinearAlgebra::distributed::Vector v2; + v2 = v; + AssertThrow(std::fabs(v2 * v - norm_sqr) < 1e-15, ExcInternalError()); + + const double inner_prod = v * v2; + if (myid == 0) + deallog << "Inner product: " << inner_prod << std::endl; + } + + // check all_zero + { + bool allzero = v.all_zero(); + if (myid == 0) + deallog << " v==0 ? " << allzero << std::endl; + LinearAlgebra::distributed::Vector v2; + v2.reinit(v); + allzero = v2.all_zero(); + if (myid == 0) + deallog << " v2==0 ? " << allzero << std::endl; + + v2.import(rw_vector, VectorOperation::insert); + allzero = v2.all_zero(); + if (myid == 0) + deallog << " v2==0 ? " << allzero << std::endl; + } + + if (myid == 0) + deallog << "OK" << std::endl; +} + + + +int +main(int argc, char **argv) +{ + Utilities::MPI::MPI_InitFinalize mpi_initialization( + argc, argv, testing_max_num_threads()); + + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + deallog.push(Utilities::int_to_string(myid)); + + Utilities::CUDA::Handle cuda_handle; + // By default, all the ranks will try to access the device 0. This is fine if + // we have one rank per node _and_ one gpu per node. If we have multiple GPUs + // on one node, we need each process to access a different GPU. We assume that + // each node has the same number of GPUs. + int n_devices = 0; + cudaError_t cuda_error_code = cudaGetDeviceCount(&n_devices); + AssertCuda(cuda_error_code); + int device_id = myid % n_devices; + cuda_error_code = cudaSetDevice(device_id); + AssertCuda(cuda_error_code); + + + if (myid == 0) + { + initlog(); + deallog << std::setprecision(4); + + test(); + } + else + test(); +} diff --git a/tests/cuda/parallel_vector_06.mpirun=2.output b/tests/cuda/parallel_vector_06.mpirun=2.output new file mode 100644 index 0000000000..61a766349d --- /dev/null +++ b/tests/cuda/parallel_vector_06.mpirun=2.output @@ -0,0 +1,11 @@ + +DEAL:0::numproc=2 +DEAL:0::l2 norm: 7.483 +DEAL:0::l1 norm: 12.00 +DEAL:0::linfty norm: 6.00 +DEAL:0::Mean value:3.000 +DEAL:0::Inner product: 56.00 +DEAL:0:: v==0 ? 0 +DEAL:0:: v2==0 ? 1 +DEAL:0:: v2==0 ? 0 +DEAL:0::OK diff --git a/tests/cuda/parallel_vector_08.cu b/tests/cuda/parallel_vector_08.cu new file mode 100644 index 0000000000..993a32e55c --- /dev/null +++ b/tests/cuda/parallel_vector_08.cu @@ -0,0 +1,150 @@ +// --------------------------------------------------------------------- +// +// Copyright (C) 2018 by the deal.II authors +// +// This file is part of the deal.II library. +// +// The deal.II library is free software; you can use it, redistribute +// it, and/or modify it under the terms of the GNU Lesser General +// Public License as published by the Free Software Foundation; either +// version 2.1 of the License, or (at your option) any later version. +// The full text of the license can be found in the file LICENSE.md at +// the top level directory of deal.II. +// +// --------------------------------------------------------------------- + + +// check parallel_vector::copy_from to update ghost values. Same vector layout +// as in parallel_vector_07.cc + +#include +#include +#include + +#include +#include + +#include +#include + +#include "../tests.h" + + +void +test() +{ + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + unsigned int numproc = Utilities::MPI::n_mpi_processes(MPI_COMM_WORLD); + + if (myid == 0) + deallog << "numproc=" << numproc << std::endl; + + const unsigned int set = 200; + AssertIndexRange(numproc, set - 2); + const unsigned int local_size = set - myid; + unsigned int global_size = 0; + unsigned int my_start = 0; + for (unsigned int i = 0; i < numproc; ++i) + { + global_size += set - i; + if (i < myid) + my_start += set - i; + } + // each processor owns some indices and all + // are ghosting elements from three + // processors (the second). some entries + // are right around the border between two + // processors + IndexSet local_owned(global_size); + local_owned.add_range(my_start, my_start + local_size); + IndexSet local_relevant(global_size); + local_relevant = local_owned; + unsigned int ghost_indices[10] = { + 1, 2, 13, set - 3, set - 2, set - 1, set, set + 1, set + 2, set + 3}; + local_relevant.add_indices(&ghost_indices[0], &ghost_indices[0] + 10); + + // v has ghosts, w has none. set some entries + // on w, copy into v and check if they are + // there + LinearAlgebra::distributed::Vector v( + local_owned, local_relevant, MPI_COMM_WORLD); + LinearAlgebra::distributed::Vector w( + local_owned, local_owned, MPI_COMM_WORLD); + + // set a few of the local elements + LinearAlgebra::ReadWriteVector rw_vector(local_owned); + for (unsigned i = 0; i < local_size; ++i) + rw_vector.local_element(i) = 2.0 * (i + my_start); + w.import(rw_vector, VectorOperation::insert); + + v = w; + v.update_ghost_values(); + + // check local values for correctness + rw_vector.import(v, VectorOperation::insert); + for (unsigned int i = 0; i < local_size; ++i) + AssertThrow(rw_vector.local_element(i) == 2.0 * (i + my_start), + ExcInternalError()); + + // check non-local entries on all processors + LinearAlgebra::ReadWriteVector ghost_vector(local_relevant); + ghost_vector.import(v, VectorOperation::insert); + for (unsigned int i = 0; i < 10; ++i) + AssertThrow(ghost_vector(ghost_indices[i]) == 2. * ghost_indices[i], + ExcInternalError()); + + // now the same again, but import ghosts automatically because v had ghosts + // set before calling operator = + v.reinit(local_owned, local_relevant, MPI_COMM_WORLD); + v.update_ghost_values(); + v = w; + + // check local values for correctness + rw_vector.import(v, VectorOperation::insert); + for (unsigned int i = 0; i < local_size; ++i) + AssertThrow(rw_vector.local_element(i) == 2.0 * (i + my_start), + ExcInternalError()); + + // check non-local entries on all processors + ghost_vector.import(v, VectorOperation::insert); + for (unsigned int i = 0; i < 10; ++i) + AssertThrow(ghost_vector(ghost_indices[i]) == 2. * ghost_indices[i], + ExcInternalError()); + + if (myid == 0) + deallog << "OK" << std::endl; +} + + + +int +main(int argc, char **argv) +{ + Utilities::MPI::MPI_InitFinalize mpi_initialization( + argc, argv, testing_max_num_threads()); + + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + deallog.push(Utilities::int_to_string(myid)); + + Utilities::CUDA::Handle cuda_handle; + // By default, all the ranks will try to access the device 0. This is fine if + // we have one rank per node _and_ one gpu per node. If we have multiple GPUs + // on one node, we need each process to access a different GPU. We assume that + // each node has the same number of GPUs. + int n_devices = 0; + cudaError_t cuda_error_code = cudaGetDeviceCount(&n_devices); + AssertCuda(cuda_error_code); + int device_id = myid % n_devices; + cuda_error_code = cudaSetDevice(device_id); + AssertCuda(cuda_error_code); + + if (myid == 0) + { + initlog(); + deallog << std::setprecision(4); + + test(); + } + else + test(); +} diff --git a/tests/cuda/parallel_vector_08.mpirun=2.output b/tests/cuda/parallel_vector_08.mpirun=2.output new file mode 100644 index 0000000000..1cf39ab251 --- /dev/null +++ b/tests/cuda/parallel_vector_08.mpirun=2.output @@ -0,0 +1,3 @@ + +DEAL:0::numproc=2 +DEAL:0::OK diff --git a/tests/cuda/parallel_vector_10.cu b/tests/cuda/parallel_vector_10.cu new file mode 100644 index 0000000000..4f08b45640 --- /dev/null +++ b/tests/cuda/parallel_vector_10.cu @@ -0,0 +1,116 @@ +// --------------------------------------------------------------------- +// +// Copyright (C) 2018 by the deal.II authors +// +// This file is part of the deal.II library. +// +// The deal.II library is free software; you can use it, redistribute +// it, and/or modify it under the terms of the GNU Lesser General +// Public License as published by the Free Software Foundation; either +// version 2.1 of the License, or (at your option) any later version. +// The full text of the license can be found in the file LICENSE.md at +// the top level directory of deal.II. +// +// --------------------------------------------------------------------- + + +// check operator= when we do some operations with ghosts + +#include +#include +#include + +#include +#include + +#include +#include + +#include "../tests.h" + + +void +test() +{ + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + unsigned int numproc = Utilities::MPI::n_mpi_processes(MPI_COMM_WORLD); + + if (myid == 0) + deallog << "numproc=" << numproc << std::endl; + + + // each processor owns 2 indices and all + // are ghosting element 1 (the second) + IndexSet local_owned(numproc * 2); + local_owned.add_range(myid * 2, myid * 2 + 2); + IndexSet local_relevant(numproc * 2); + local_relevant = local_owned; + local_relevant.add_range(1, 2); + + LinearAlgebra::distributed::Vector v( + local_owned, local_relevant, MPI_COMM_WORLD); + LinearAlgebra::distributed::Vector w(v); + + // set local values and check them + LinearAlgebra::ReadWriteVector rw_vector(local_owned); + rw_vector(myid * 2) = myid * 2.0; + rw_vector(myid * 2 + 1) = myid * 2.0 + 1.0; + v.import(rw_vector, VectorOperation::insert); + + v.update_ghost_values(); + + // check that the value of the ghost is 1.0 + IndexSet ghost_set(numproc * 2); + ghost_set.add_index(1); + LinearAlgebra::ReadWriteVector ghost_vector(ghost_set); + ghost_vector.import(v, VectorOperation::insert); + AssertThrow(ghost_vector(1) == 1., ExcInternalError()); + + // copy vector + w = v; + v *= 2.0; + + v.update_ghost_values(); + w.update_ghost_values(); + ghost_vector.import(v, VectorOperation::insert); + AssertThrow(ghost_vector(1) == 2., ExcInternalError()); + ghost_vector.import(w, VectorOperation::insert); + AssertThrow(ghost_vector(1) == 1., ExcInternalError()); + + if (myid == 0) + deallog << "OK" << std::endl; +} + + + +int +main(int argc, char **argv) +{ + Utilities::MPI::MPI_InitFinalize mpi_initialization( + argc, argv, testing_max_num_threads()); + + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + deallog.push(Utilities::int_to_string(myid)); + + Utilities::CUDA::Handle cuda_handle; + // By default, all the ranks will try to access the device 0. This is fine if + // we have one rank per node _and_ one gpu per node. If we have multiple GPUs + // on one node, we need each process to access a different GPU. We assume that + // each node has the same number of GPUs. + int n_devices = 0; + cudaError_t cuda_error_code = cudaGetDeviceCount(&n_devices); + AssertCuda(cuda_error_code); + int device_id = myid % n_devices; + cuda_error_code = cudaSetDevice(device_id); + AssertCuda(cuda_error_code); + + if (myid == 0) + { + initlog(); + deallog << std::setprecision(4); + + test(); + } + else + test(); +} diff --git a/tests/cuda/parallel_vector_10.mpirun=2.output b/tests/cuda/parallel_vector_10.mpirun=2.output new file mode 100644 index 0000000000..1cf39ab251 --- /dev/null +++ b/tests/cuda/parallel_vector_10.mpirun=2.output @@ -0,0 +1,3 @@ + +DEAL:0::numproc=2 +DEAL:0::OK diff --git a/tests/cuda/parallel_vector_11.cu b/tests/cuda/parallel_vector_11.cu new file mode 100644 index 0000000000..64840468e6 --- /dev/null +++ b/tests/cuda/parallel_vector_11.cu @@ -0,0 +1,258 @@ +// --------------------------------------------------------------------- +// +// Copyright (C) 2018 by the deal.II authors +// +// This file is part of the deal.II library. +// +// The deal.II library is free software; you can use it, redistribute +// it, and/or modify it under the terms of the GNU Lesser General +// Public License as published by the Free Software Foundation; either +// version 2.1 of the License, or (at your option) any later version. +// The full text of the license can be found in the file LICENSE.md at +// the top level directory of deal.II. +// +// --------------------------------------------------------------------- + + +// check that add, sadd, equ, scale work correctly on a vector where some +// processor do not own any degrees of freedom + +#include +#include +#include + +#include +#include + +#include +#include + +#include "../tests.h" + + +void +test() +{ + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + unsigned int numproc = Utilities::MPI::n_mpi_processes(MPI_COMM_WORLD); + + if (myid == 0) + deallog << "numproc=" << numproc << std::endl; + + // global size: 20, local_size: 3 as long as + // less than 20 + const unsigned int local_size = 3; + const unsigned int global_size = std::min(20U, local_size * numproc); + const int my_start = std::min(local_size * myid, global_size); + const int my_end = std::min(local_size * (myid + 1), global_size); + const int actual_local_size = my_end - my_start; + + IndexSet local_owned(global_size); + if (my_end > my_start) + local_owned.add_range(static_cast(my_start), + static_cast(my_end)); + IndexSet local_relevant(global_size); + local_relevant = local_owned; + local_relevant.add_index(2); + + LinearAlgebra::distributed::Vector v( + local_owned, local_relevant, MPI_COMM_WORLD); + AssertDimension(static_cast(actual_local_size), v.local_size()); + LinearAlgebra::distributed::Vector w(v), x(v), + y(v); + + // set local elements + LinearAlgebra::ReadWriteVector v_rw(local_owned); + LinearAlgebra::ReadWriteVector w_rw(local_owned); + LinearAlgebra::ReadWriteVector x_rw(local_owned); + for (int i = 0; i < actual_local_size; ++i) + { + v_rw.local_element(i) = i + my_start; + w_rw.local_element(i) = 1000 + 2 * (my_start + i); + x_rw.local_element(i) = 10000; + } + v.import(v_rw, VectorOperation::insert); + w.import(w_rw, VectorOperation::insert); + x.import(x_rw, VectorOperation::insert); + + y = v; + LinearAlgebra::ReadWriteVector y_rw(local_owned); + y_rw.import(y, VectorOperation::insert); + for (int i = 0; i < actual_local_size; ++i) + AssertThrow(y_rw.local_element(i) == i + my_start, ExcInternalError()); + + if (myid == 0) + deallog << "Check add (scalar): "; + y.add(42); + y_rw.import(y, VectorOperation::insert); + for (int i = 0; i < actual_local_size; ++i) + AssertThrow(y_rw.local_element(i) == i + my_start + 42, ExcInternalError()); + if (myid == 0) + deallog << "OK" << std::endl; + + if (myid == 0) + deallog << "Check add (vector): "; + y.add(1., w); + y_rw.import(y, VectorOperation::insert); + for (int i = 0; i < actual_local_size; ++i) + AssertThrow(y_rw.local_element(i) == 3 * (i + my_start) + 1042, + ExcInternalError()); + if (myid == 0) + deallog << "OK" << std::endl; + + if (myid == 0) + deallog << "Check add (factor, vector): "; + y.add(-1., w); + y_rw.import(y, VectorOperation::insert); + for (int i = 0; i < actual_local_size; ++i) + AssertThrow(y_rw.local_element(i) == i + my_start + 42, ExcInternalError()); + if (myid == 0) + deallog << "OK" << std::endl; + + if (myid == 0) + deallog << "Check add (factor, vector, factor, vector): "; + y.add(2., w, -0.5, x); + y_rw.import(y, VectorOperation::insert); + for (int i = 0; i < actual_local_size; ++i) + AssertThrow(y_rw.local_element(i) == 5 * (i + my_start) + 2042 - 5000, + ExcInternalError()); + if (myid == 0) + deallog << "OK" << std::endl; + + if (myid == 0) + deallog << "Check sadd (factor, factor, vector): "; + y = v; + y.sadd(-3., 2., v); + y_rw.import(y, VectorOperation::insert); + for (int i = 0; i < actual_local_size; ++i) + AssertThrow(y_rw.local_element(i) == (-i - my_start), ExcInternalError()); + if (myid == 0) + deallog << "OK" << std::endl; + + if (myid == 0) + deallog << "Check sadd (factor, factor, vector, factor, vector): "; + y.sadd(2., 3., v, 2., w); + y_rw.import(y, VectorOperation::insert); + for (int i = 0; i < actual_local_size; ++i) + { + AssertThrow(y_rw.local_element(i) == 5 * (i + my_start) + 2000, + ExcInternalError()); + } + if (myid == 0) + deallog << "OK" << std::endl; + + if (myid == 0) + deallog + << "Check sadd (factor, factor, vector, factor, vector, factor, vector): "; + y.sadd(-1., 1., v, 2., w); + y.add(2., x); + y_rw.import(y, VectorOperation::insert); + for (int i = 0; i < actual_local_size; ++i) + AssertThrow(y_rw.local_element(i) == 20000, ExcInternalError()); + if (myid == 0) + deallog << "OK" << std::endl; + + if (myid == 0) + deallog << "Check add (factor, vector_1, factor, vector_1): "; + y = 0; + y.add(1., v, 3., v); + y_rw.import(y, VectorOperation::insert); + for (int i = 0; i < actual_local_size; ++i) + AssertThrow(y_rw.local_element(i) == 4 * (i + my_start), + ExcInternalError()); + if (myid == 0) + deallog << "OK" << std::endl; + + if (myid == 0) + deallog << "Check operator * (scalar): "; + x *= 2.; + x_rw.import(x, VectorOperation::insert); + for (int i = 0; i < actual_local_size; ++i) + AssertThrow(x_rw.local_element(i) == 20000., ExcInternalError()); + if (myid == 0) + deallog << "OK" << std::endl; + + if (myid == 0) + deallog << "Check operator / (scalar): "; + x /= 2.; + x_rw.import(x, VectorOperation::insert); + for (int i = 0; i < actual_local_size; ++i) + AssertThrow(x_rw.local_element(i) == 10000., ExcInternalError()); + if (myid == 0) + deallog << "OK" << std::endl; + + if (myid == 0) + deallog << "Check scale (vector): "; + y.scale(x); + y_rw.import(y, VectorOperation::insert); + for (int i = 0; i < actual_local_size; ++i) + AssertThrow(y_rw.local_element(i) == 40000. * (i + my_start), + ExcInternalError()); + if (myid == 0) + deallog << "OK" << std::endl; + + if (myid == 0) + deallog << "Check equ (factor, vector): "; + y.equ(10., x); + y_rw.import(y, VectorOperation::insert); + for (int i = 0; i < actual_local_size; ++i) + AssertThrow(y_rw.local_element(i) == 100000., ExcInternalError()); + if (myid == 0) + deallog << "OK" << std::endl; + + if (myid == 0) + deallog << "Check equ (factor, vector, factor, vector): "; + y.equ(10., v, -2., w); + y_rw.import(y, VectorOperation::insert); + for (int i = 0; i < actual_local_size; ++i) + AssertThrow(y_rw.local_element(i) == 6. * (i + my_start) - 2000, + ExcInternalError()); + if (myid == 0) + deallog << "OK" << std::endl; + + if (myid == 0) + deallog << "Check equ (factor, vector, factor, vector, factor, vector): "; + y.equ(10., v, -2., w); + y.add(3., x); + y_rw.import(y, VectorOperation::insert); + for (int i = 0; i < actual_local_size; ++i) + AssertThrow(y_rw.local_element(i) == 6. * (i + my_start) + 28000, + ExcInternalError()); + if (myid == 0) + deallog << "OK" << std::endl; +} + + + +int +main(int argc, char **argv) +{ + Utilities::MPI::MPI_InitFinalize mpi_initialization( + argc, argv, testing_max_num_threads()); + + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + deallog.push(Utilities::int_to_string(myid)); + + Utilities::CUDA::Handle cuda_handle; + // By default, all the ranks will try to access the device 0. This is fine if + // we have one rank per node _and_ one gpu per node. If we have multiple GPUs + // on one node, we need each process to access a different GPU. We assume that + // each node has the same number of GPUs. + int n_devices = 0; + cudaError_t cuda_error_code = cudaGetDeviceCount(&n_devices); + AssertCuda(cuda_error_code); + int device_id = myid % n_devices; + cuda_error_code = cudaSetDevice(device_id); + AssertCuda(cuda_error_code); + + + if (myid == 0) + { + initlog(); + deallog << std::setprecision(4); + + test(); + } + else + test(); +} diff --git a/tests/cuda/parallel_vector_11.mpirun=2.output b/tests/cuda/parallel_vector_11.mpirun=2.output new file mode 100644 index 0000000000..631cc4daf5 --- /dev/null +++ b/tests/cuda/parallel_vector_11.mpirun=2.output @@ -0,0 +1,16 @@ + +DEAL:0::numproc=2 +DEAL:0::Check add (scalar): OK +DEAL:0::Check add (vector): OK +DEAL:0::Check add (factor, vector): OK +DEAL:0::Check add (factor, vector, factor, vector): OK +DEAL:0::Check sadd (factor, factor, vector): OK +DEAL:0::Check sadd (factor, factor, vector, factor, vector): OK +DEAL:0::Check sadd (factor, factor, vector, factor, vector, factor, vector): OK +DEAL:0::Check add (factor, vector_1, factor, vector_1): OK +DEAL:0::Check operator * (scalar): OK +DEAL:0::Check operator / (scalar): OK +DEAL:0::Check scale (vector): OK +DEAL:0::Check equ (factor, vector): OK +DEAL:0::Check equ (factor, vector, factor, vector): OK +DEAL:0::Check equ (factor, vector, factor, vector, factor, vector): OK diff --git a/tests/cuda/parallel_vector_12.cu b/tests/cuda/parallel_vector_12.cu new file mode 100644 index 0000000000..29e5c4da36 --- /dev/null +++ b/tests/cuda/parallel_vector_12.cu @@ -0,0 +1,241 @@ +// --------------------------------------------------------------------- +// +// Copyright (C) 2018 by the deal.II authors +// +// This file is part of the deal.II library. +// +// The deal.II library is free software; you can use it, redistribute +// it, and/or modify it under the terms of the GNU Lesser General +// Public License as published by the Free Software Foundation; either +// version 2.1 of the License, or (at your option) any later version. +// The full text of the license can be found in the file LICENSE.md at +// the top level directory of deal.II. +// +// --------------------------------------------------------------------- + + +// check LinearAlgebra::distributed::Vector::swap + +#include +#include +#include + +#include +#include + +#include +#include + +#include "../tests.h" + +DeclException2(ExcNonEqual, + double, + double, + << "Left compare: " << arg1 << ", right compare: " << arg2); + +void +test() +{ + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + unsigned int numproc = Utilities::MPI::n_mpi_processes(MPI_COMM_WORLD); + + if (myid == 0) + deallog << "numproc=" << numproc << std::endl; + + // vector 0: + // global size: 20, local_size: 3 as long as + // less than 20 + const unsigned int local_size0 = 3; + const unsigned int global_size0 = std::min(20U, local_size0 * numproc); + const unsigned int my_start0 = std::min(local_size0 * myid, global_size0); + const unsigned int my_end0 = std::min(local_size0 * (myid + 1), global_size0); + const unsigned int actual_local_size0 = my_end0 - my_start0; + + IndexSet local_owned0(global_size0); + if (my_end0 > my_start0) + local_owned0.add_range(static_cast(my_start0), + static_cast(my_end0)); + IndexSet local_relevant0(global_size0); + local_relevant0 = local_owned0; + local_relevant0.add_index(2); + if (numproc > 2) + local_relevant0.add_index(8); + + LinearAlgebra::distributed::Vector v0( + local_owned0, local_relevant0, MPI_COMM_WORLD); + + // vector1: local size 4 + const unsigned int local_size1 = 4; + const unsigned int global_size1 = local_size1 * numproc; + const int my_start1 = local_size1 * myid; + const int my_end1 = local_size1 * (myid + 1); + + IndexSet local_owned1(global_size1); + local_owned1.add_range(static_cast(my_start1), + static_cast(my_end1)); + IndexSet local_relevant1(global_size1); + local_relevant1 = local_owned1; + local_relevant1.add_index(0); + local_relevant1.add_index(2); + if (numproc > 2) + { + local_relevant1.add_index(8); + local_relevant1.add_index(10); + } + + LinearAlgebra::distributed::Vector v1( + local_owned1, local_relevant1, MPI_COMM_WORLD); + + v0 = 1; + v1 = 2; + // check assignment in initial state + LinearAlgebra::ReadWriteVector v0_rw(local_owned0); + v0_rw.import(v0, VectorOperation::insert); + LinearAlgebra::ReadWriteVector v1_rw(local_owned1); + v1_rw.import(v1, VectorOperation::insert); + for (unsigned int i = 0; i < v0.local_size(); ++i) + AssertThrow(v0_rw.local_element(i) == 1., + ExcNonEqual(v0_rw.local_element(i), 1.)); + for (unsigned int i = 0; i < v1.local_size(); ++i) + AssertThrow(v1_rw.local_element(i) == 2., + ExcNonEqual(v1_rw.local_element(i), 2.)); + + // check ghost elements in initial state + v0.update_ghost_values(); + v1.update_ghost_values(); + LinearAlgebra::ReadWriteVector v0_ghost_rw(local_relevant0); + v0_ghost_rw.import(v0, VectorOperation::insert); + LinearAlgebra::ReadWriteVector v1_ghost_rw(local_relevant1); + v1_ghost_rw.import(v1, VectorOperation::insert); + AssertThrow(v0_ghost_rw(2) == 1., ExcNonEqual(v0_ghost_rw(2), 1.)); + if (numproc > 2) + AssertThrow(v0_ghost_rw(8) == 1., ExcNonEqual(v0_ghost_rw(8), 2.)); + AssertThrow(v1_ghost_rw(0) == 2., ExcNonEqual(v1_ghost_rw(0), 2.)); + AssertThrow(v1_ghost_rw(2) == 2., ExcNonEqual(v1_ghost_rw(2), 2.)); + if (numproc > 2) + { + AssertThrow(v1_ghost_rw(8) == 2., ExcNonEqual(v1_ghost_rw(8), 2.)); + AssertThrow(v1_ghost_rw(10) == 2., ExcNonEqual(v1_ghost_rw(10), 2.)); + } + MPI_Barrier(MPI_COMM_WORLD); + if (myid == 0) + deallog << "Initial set and ghost update OK" << std::endl; + + // now swap v1 and v0 + v0.swap(v1); + AssertDimension(v0.local_size(), local_size1); + AssertDimension(v1.local_size(), actual_local_size0); + AssertDimension(v0.size(), global_size1); + AssertDimension(v1.size(), global_size0); + v1_rw.import(v0, VectorOperation::insert); + for (unsigned int i = 0; i < local_size1; ++i) + AssertThrow(v1_rw.local_element(i) == 2., + ExcNonEqual(v1_rw.local_element(i), 2.)); + v0_rw.import(v1, VectorOperation::insert); + for (unsigned int i = 0; i < actual_local_size0; ++i) + AssertThrow(v0_rw.local_element(i) == 1., + ExcNonEqual(v0_rw.local_element(i), 1.)); + MPI_Barrier(MPI_COMM_WORLD); + if (myid == 0) + deallog << "First swap OK" << std::endl; + v0.update_ghost_values(); + v1.update_ghost_values(); + v0_ghost_rw.import(v1, VectorOperation::insert); + AssertThrow(v0_ghost_rw(2) == 1., ExcNonEqual(v0_ghost_rw(2), 1.)); + if (numproc > 2) + AssertThrow(v0_ghost_rw(8) == 1., ExcNonEqual(v0_ghost_rw(8), 1.)); + v1_ghost_rw.import(v0, VectorOperation::insert); + AssertThrow(v1_ghost_rw(0) == 2., ExcNonEqual(v1_ghost_rw(0), 2.)); + AssertThrow(v1_ghost_rw(2) == 2., ExcNonEqual(v1_ghost_rw(2), 2.)); + if (numproc > 2) + { + AssertThrow(v1_ghost_rw(8) == 2., ExcNonEqual(v1_ghost_rw(8), 2.)); + AssertThrow(v1_ghost_rw(10) == 2., ExcNonEqual(v1_ghost_rw(10), 2.)); + } + MPI_Barrier(MPI_COMM_WORLD); + if (myid == 0) + deallog << "Ghost values after first swap OK" << std::endl; + + // now set the vectors to some different + // values and check the ghost values again + v0 = 7.; + v1 = 42.; + v0.update_ghost_values(); + v1.update_ghost_values(); + v0_ghost_rw.import(v1, VectorOperation::insert); + AssertThrow(v0_ghost_rw(2) == 42., ExcNonEqual(v0_ghost_rw(2), 42.)); + if (numproc > 2) + AssertThrow(v0_ghost_rw(8) == 42., ExcNonEqual(v0_ghost_rw(8), 42.)); + v1_ghost_rw.import(v0, VectorOperation::insert); + AssertThrow(v1_ghost_rw(0) == 7., ExcNonEqual(v1_ghost_rw(0), 7.)); + AssertThrow(v1_ghost_rw(2) == 7., ExcNonEqual(v1_ghost_rw(2), 7.)); + if (numproc > 2) + { + AssertThrow(v1_ghost_rw(8) == 7., ExcNonEqual(v1_ghost_rw(8), 7.)); + AssertThrow(v1_ghost_rw(10) == 7., ExcNonEqual(v1_ghost_rw(10), 7.)); + } + MPI_Barrier(MPI_COMM_WORLD); + if (myid == 0) + deallog << "Ghost values after re-set OK" << std::endl; + + // swap with an empty vector + LinearAlgebra::distributed::Vector v2; + v2.swap(v0); + AssertDimension(v0.size(), 0); + AssertDimension(v2.size(), global_size1); + AssertDimension(v2.local_size(), local_size1); + v1_rw.import(v2, VectorOperation::insert); + for (int i = my_start1; i < my_end1; ++i) + AssertThrow(v1_rw(i) == 7., ExcNonEqual(v1_rw(i), 7.)); + MPI_Barrier(MPI_COMM_WORLD); + if (myid == 0) + deallog << "Second swap OK" << std::endl; + v2 = -1.; + v2.update_ghost_values(); + v1_ghost_rw.import(v2, VectorOperation::insert); + AssertThrow(v1_ghost_rw(0) == -1., ExcNonEqual(v1_ghost_rw(0), -1.)); + AssertThrow(v1_ghost_rw(2) == -1., ExcNonEqual(v1_ghost_rw(2), -1.)); + if (numproc > 2) + { + AssertThrow(v1_ghost_rw(8) == -1., ExcNonEqual(v1_ghost_rw(8), -1.)); + AssertThrow(v1_ghost_rw(10) == -1., ExcNonEqual(v1_ghost_rw(10), -1.)); + } + MPI_Barrier(MPI_COMM_WORLD); + if (myid == 0) + deallog << "Ghost values after second swap OK" << std::endl; +} + + + +int +main(int argc, char **argv) +{ + Utilities::MPI::MPI_InitFinalize mpi_initialization( + argc, argv, testing_max_num_threads()); + + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + deallog.push(Utilities::int_to_string(myid)); + + Utilities::CUDA::Handle cuda_handle; + // By default, all the ranks will try to access the device 0. This is fine if + // we have one rank per node _and_ one gpu per node. If we have multiple GPUs + // on one node, we need each process to access a different GPU. We assume that + // each node has the same number of GPUs. + int n_devices = 0; + cudaError_t cuda_error_code = cudaGetDeviceCount(&n_devices); + AssertCuda(cuda_error_code); + int device_id = myid % n_devices; + cuda_error_code = cudaSetDevice(device_id); + AssertCuda(cuda_error_code); + + + if (myid == 0) + { + initlog(); + deallog << std::setprecision(4); + + test(); + } + else + test(); +} diff --git a/tests/cuda/parallel_vector_12.mpirun=2.output b/tests/cuda/parallel_vector_12.mpirun=2.output new file mode 100644 index 0000000000..ad4bdea2aa --- /dev/null +++ b/tests/cuda/parallel_vector_12.mpirun=2.output @@ -0,0 +1,8 @@ + +DEAL:0::numproc=2 +DEAL:0::Initial set and ghost update OK +DEAL:0::First swap OK +DEAL:0::Ghost values after first swap OK +DEAL:0::Ghost values after re-set OK +DEAL:0::Second swap OK +DEAL:0::Ghost values after second swap OK diff --git a/tests/cuda/parallel_vector_13.cu b/tests/cuda/parallel_vector_13.cu new file mode 100644 index 0000000000..097ef2ed9d --- /dev/null +++ b/tests/cuda/parallel_vector_13.cu @@ -0,0 +1,137 @@ +// --------------------------------------------------------------------- +// +// Copyright (C) 2018 by the deal.II authors +// +// This file is part of the deal.II library. +// +// The deal.II library is free software; you can use it, redistribute +// it, and/or modify it under the terms of the GNU Lesser General +// Public License as published by the Free Software Foundation; either +// version 2.1 of the License, or (at your option) any later version. +// The full text of the license can be found in the file LICENSE.md at +// the top level directory of deal.II. +// +// --------------------------------------------------------------------- + + +// check for ghosts on parallel vector: similar to parallel_vector_03, but +// setting where one ghost is zero and should not have an effect on vector +// entries + +#include +#include +#include + +#include +#include + +#include +#include + +#include "../tests.h" + + +void +test() +{ + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + unsigned int numproc = Utilities::MPI::n_mpi_processes(MPI_COMM_WORLD); + + if (myid == 0) + deallog << "numproc=" << numproc << std::endl; + + + // each processor owns 2 indices and all + // are ghosting element 1 (the second) + IndexSet local_owned(numproc * 2); + local_owned.add_range(myid * 2, myid * 2 + 2); + IndexSet local_relevant(numproc * 2); + local_relevant = local_owned; + local_relevant.add_range(1, 2); + + LinearAlgebra::distributed::Vector v( + local_owned, local_relevant, MPI_COMM_WORLD); + + // set local values and check them + LinearAlgebra::ReadWriteVector rw_vector(local_owned); + rw_vector(myid * 2) = myid * 2.0; + rw_vector(myid * 2 + 1) = myid * 2.0 + 1.0; + + v.import(rw_vector, VectorOperation::insert); + v *= 2.0; + v.add(1.0); + + rw_vector.import(v, VectorOperation::insert); + AssertThrow(rw_vector(myid * 2) == myid * 4.0 + 1, ExcInternalError()); + AssertThrow(rw_vector(myid * 2 + 1) == myid * 4.0 + 3.0, ExcInternalError()); + + // set ghost dof on all processors, compress + // (insert mode) + IndexSet index(numproc * 2); + index.add_index(1); + LinearAlgebra::ReadWriteVector local_rw_vector(index); + local_rw_vector(1) = 7; + v.import(local_rw_vector, VectorOperation::insert); + + { + rw_vector.import(v, VectorOperation::insert); + deallog << myid * 2 << ":" << rw_vector(myid * 2) << std::endl; + deallog << myid * 2 + 1 << ":" << rw_vector(myid * 2 + 1) << std::endl; + } + + local_rw_vector(1) = -7; + v.import(local_rw_vector, VectorOperation::insert); + + { + rw_vector.import(v, VectorOperation::insert); + deallog << myid * 2 << ":" << rw_vector(myid * 2) << std::endl; + deallog << myid * 2 + 1 << ":" << rw_vector(myid * 2 + 1) << std::endl; + } + + // import ghosts onto all procs + v.update_ghost_values(); + local_rw_vector.import(v, VectorOperation::insert); + AssertThrow(local_rw_vector(1) == -7.0, ExcInternalError()); + + // check l2 norm + const double l2_norm = v.l2_norm(); + if (myid == 0) + deallog << "L2 norm: " << l2_norm << std::endl; + + if (myid == 0) + deallog << "OK" << std::endl; +} + + + +int +main(int argc, char **argv) +{ + Utilities::MPI::MPI_InitFinalize mpi_initialization( + argc, argv, testing_max_num_threads()); + + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + deallog.push(Utilities::int_to_string(myid)); + + Utilities::CUDA::Handle cuda_handle; + // By default, all the ranks will try to access the device 0. This is fine if + // we have one rank per node _and_ one gpu per node. If we have multiple GPUs + // on one node, we need each process to access a different GPU. We assume that + // each node has the same number of GPUs. + int n_devices = 0; + cudaError_t cuda_error_code = cudaGetDeviceCount(&n_devices); + AssertCuda(cuda_error_code); + int device_id = myid % n_devices; + cuda_error_code = cudaSetDevice(device_id); + AssertCuda(cuda_error_code); + + if (myid == 0) + { + initlog(); + deallog << std::setprecision(4); + + test(); + } + else + test(); +} diff --git a/tests/cuda/parallel_vector_13.mpirun=2.output b/tests/cuda/parallel_vector_13.mpirun=2.output new file mode 100644 index 0000000000..848c89cf4e --- /dev/null +++ b/tests/cuda/parallel_vector_13.mpirun=2.output @@ -0,0 +1,8 @@ + +DEAL:0::numproc=2 +DEAL:0::0:1.000 +DEAL:0::1:7.000 +DEAL:0::0:1.000 +DEAL:0::1:-7.000 +DEAL:0::L2 norm: 11.14 +DEAL:0::OK diff --git a/tests/cuda/parallel_vector_14.cu b/tests/cuda/parallel_vector_14.cu new file mode 100644 index 0000000000..955b562131 --- /dev/null +++ b/tests/cuda/parallel_vector_14.cu @@ -0,0 +1,185 @@ +// --------------------------------------------------------------------- +// +// Copyright (C) 2018 by the deal.II authors +// +// This file is part of the deal.II library. +// +// The deal.II library is free software; you can use it, redistribute +// it, and/or modify it under the terms of the GNU Lesser General +// Public License as published by the Free Software Foundation; either +// version 2.1 of the License, or (at your option) any later version. +// The full text of the license can be found in the file LICENSE.md at +// the top level directory of deal.II. +// +// --------------------------------------------------------------------- + + +// check that handling of ghost elements in parallel distributed vectors works +// appropriately when assigning from ghosted to non-ghosted vectors + +#include +#include +#include + +#include +#include + +#include +#include + +#include "../tests.h" + + +void +test() +{ + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + unsigned int numproc = Utilities::MPI::n_mpi_processes(MPI_COMM_WORLD); + + if (myid == 0) + deallog << "numproc=" << numproc << std::endl; + + + // processor 0 and 1 own 2 indices each, higher processors nothing, all are + // ghosting global elements 1 and 3 + IndexSet local_owned(std::min(numproc * 2, 4U)); + if (myid < 2) + local_owned.add_range(myid * 2, myid * 2 + 2); + IndexSet local_relevant(local_owned.size()); + local_relevant = local_owned; + local_relevant.add_range(1, 2); + if (numproc > 1) + local_relevant.add_range(3, 4); + + // run this twice, once where the vectors have called update_ghost_values + // and once where they have not + for (unsigned int run = 0; run < 2; ++run) + { + LinearAlgebra::distributed::Vector v( + local_owned, local_relevant, MPI_COMM_WORLD); + + // set local values + LinearAlgebra::ReadWriteVector rw_vector(local_owned); + if (myid < 2) + { + rw_vector(myid * 2) = myid * 2.0; + rw_vector(myid * 2 + 1) = myid * 2.0 + 1.0; + } + + v.import(rw_vector, VectorOperation::insert); + + LinearAlgebra::distributed::Vector w(v), u(v); + u = 0; + + v *= 2.0; + v.add(1.0); + + if (run == 1) + { + v.update_ghost_values(); + w.update_ghost_values(); + u.update_ghost_values(); + } + + rw_vector.import(v, VectorOperation::insert); + if (myid < 2) + { + Assert(rw_vector(myid * 2) == myid * 4.0 + 1, ExcInternalError()); + Assert(rw_vector(myid * 2 + 1) == myid * 4.0 + 3.0, + ExcInternalError()); + } + + // copy vector content to non-ghosted vectors, manually created. + LinearAlgebra::distributed::Vector v_dist( + local_owned, MPI_COMM_WORLD), + w_dist(v_dist), u_dist(v_dist); + + v_dist = v; + w_dist = w; + u_dist = u; + + u_dist.add(1.0, v_dist, -1.0, w); + + // copy back to a ghosted vector and update ghost values there + u = u_dist; + u.update_ghost_values(); + + rw_vector.import(u_dist, VectorOperation::insert); + if (myid < 2) + { + Assert(rw_vector(myid * 2) == myid * 2.0 + 1, ExcInternalError()); + Assert(rw_vector(myid * 2 + 1) == myid * 2.0 + 2.0, + ExcInternalError()); + } + + rw_vector.import(u, VectorOperation::insert); + if (myid < 2) + { + Assert(rw_vector(myid * 2) == myid * 2.0 + 1, ExcInternalError()); + Assert(rw_vector(myid * 2 + 1) == myid * 2.0 + 2.0, + ExcInternalError()); + } + + IndexSet u_ghost_set(local_owned.size()); + u_ghost_set.add_index(1); + u_ghost_set.add_index(3); + LinearAlgebra::ReadWriteVector u_ghost_vector(u_ghost_set); + u_ghost_vector.import(u, VectorOperation::insert); + Assert(u_ghost_vector(1) == 2., ExcInternalError()); + if (numproc > 1) + { + if (run == 1) + { + IndexSet v_ghost_set(local_owned.size()); + v_ghost_set.add_index(3); + LinearAlgebra::ReadWriteVector v_ghost_vector( + v_ghost_set); + v_ghost_vector.import(v, VectorOperation::insert); + Assert(v_ghost_vector(3) == 7., ExcInternalError()); + } + Assert(u_ghost_vector(3) == 4., ExcInternalError()); + } + + // check l2 norm + const double l2_norm = u.l2_norm(); + if (myid == 0 && run == 1) + deallog << "L2 norm: " << l2_norm << std::endl; + } + + if (myid == 0) + deallog << "OK" << std::endl; +} + + + +int +main(int argc, char **argv) +{ + Utilities::MPI::MPI_InitFinalize mpi_initialization( + argc, argv, testing_max_num_threads()); + + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + deallog.push(Utilities::int_to_string(myid)); + + Utilities::CUDA::Handle cuda_handle; + // By default, all the ranks will try to access the device 0. This is fine if + // we have one rank per node _and_ one gpu per node. If we have multiple GPUs + // on one node, we need each process to access a different GPU. We assume that + // each node has the same number of GPUs. + int n_devices = 0; + cudaError_t cuda_error_code = cudaGetDeviceCount(&n_devices); + AssertCuda(cuda_error_code); + int device_id = myid % n_devices; + cuda_error_code = cudaSetDevice(device_id); + AssertCuda(cuda_error_code); + + if (myid == 0) + { + initlog(); + deallog << std::setprecision(4); + + test(); + } + else + test(); +} diff --git a/tests/cuda/parallel_vector_14.mpirun=2.output b/tests/cuda/parallel_vector_14.mpirun=2.output new file mode 100644 index 0000000000..12e7a746ab --- /dev/null +++ b/tests/cuda/parallel_vector_14.mpirun=2.output @@ -0,0 +1,4 @@ + +DEAL:0::numproc=2 +DEAL:0::L2 norm: 5.477 +DEAL:0::OK diff --git a/tests/cuda/parallel_vector_15.cu b/tests/cuda/parallel_vector_15.cu new file mode 100644 index 0000000000..973b0313ed --- /dev/null +++ b/tests/cuda/parallel_vector_15.cu @@ -0,0 +1,135 @@ +// --------------------------------------------------------------------- +// +// Copyright (C) 2018 by the deal.II authors +// +// This file is part of the deal.II library. +// +// The deal.II library is free software; you can use it, redistribute +// it, and/or modify it under the terms of the GNU Lesser General +// Public License as published by the Free Software Foundation; either +// version 2.1 of the License, or (at your option) any later version. +// The full text of the license can be found in the file LICENSE.md at +// the top level directory of deal.II. +// +// --------------------------------------------------------------------- + + +// check that handling of ghost elements in parallel distributed vectors works +// appropriately when creating a vector from a non-ghosted source vector using +// the assignment operator + +#include +#include +#include + +#include +#include + +#include +#include + +#include "../tests.h" + + +void +test() +{ + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + unsigned int numproc = Utilities::MPI::n_mpi_processes(MPI_COMM_WORLD); + + if (myid == 0) + deallog << "numproc=" << numproc << std::endl; + + // processor 0 and 1 own 2 indices each, higher processors nothing, all are + // ghosting global elements 1 and 3 + IndexSet local_owned(std::min(numproc * 2, 4U)); + if (myid < 2) + local_owned.add_range(myid * 2, myid * 2 + 2); + IndexSet local_relevant(local_owned.size()); + local_relevant = local_owned; + local_relevant.add_range(1, 2); + if (numproc > 1) + local_relevant.add_range(3, 4); + + LinearAlgebra::distributed::Vector v( + local_owned, local_relevant, MPI_COMM_WORLD); + + // set local values + LinearAlgebra::ReadWriteVector rw_vector(local_owned); + if (myid < 2) + { + rw_vector(myid * 2) = myid * 2.0; + rw_vector(myid * 2 + 1) = myid * 2.0 + 1.0; + } + + v.import(rw_vector, VectorOperation::insert); + + if (myid == 0) + deallog << "v has ghost elements: " << v.has_ghost_elements() << std::endl; + + LinearAlgebra::distributed::Vector w, x; + w = v; + if (myid == 0) + deallog << "w has ghost elements: " << w.has_ghost_elements() << std::endl; + + v.update_ghost_values(); + w = v; + if (myid == 0) + deallog << "w has ghost elements: " << w.has_ghost_elements() << std::endl; + + v.zero_out_ghosts(); + w = v; + if (myid == 0) + deallog << "w has ghost elements: " << w.has_ghost_elements() << std::endl; + + w.zero_out_ghosts(); + w = v; + if (myid == 0) + deallog << "w has ghost elements: " << w.has_ghost_elements() << std::endl; + + v.update_ghost_values(); + x = v; + if (myid == 0) + deallog << "x has ghost elements: " << x.has_ghost_elements() << std::endl; + + x.zero_out_ghosts(); + if (myid == 0) + deallog << "x has ghost elements: " << x.has_ghost_elements() << std::endl; + + if (myid == 0) + deallog << "OK" << std::endl; +} + + + +int +main(int argc, char **argv) +{ + Utilities::MPI::MPI_InitFinalize mpi_initialization( + argc, argv, testing_max_num_threads()); + + unsigned int myid = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + deallog.push(Utilities::int_to_string(myid)); + + Utilities::CUDA::Handle cuda_handle; + // By default, all the ranks will try to access the device 0. This is fine if + // we have one rank per node _and_ one gpu per node. If we have multiple GPUs + // on one node, we need each process to access a different GPU. We assume that + // each node has the same number of GPUs. + int n_devices = 0; + cudaError_t cuda_error_code = cudaGetDeviceCount(&n_devices); + AssertCuda(cuda_error_code); + int device_id = myid % n_devices; + cuda_error_code = cudaSetDevice(device_id); + AssertCuda(cuda_error_code); + + if (myid == 0) + { + initlog(); + deallog << std::setprecision(4); + + test(); + } + else + test(); +} diff --git a/tests/cuda/parallel_vector_15.mpirun=2.output b/tests/cuda/parallel_vector_15.mpirun=2.output new file mode 100644 index 0000000000..39f1d154ae --- /dev/null +++ b/tests/cuda/parallel_vector_15.mpirun=2.output @@ -0,0 +1,10 @@ + +DEAL:0::numproc=2 +DEAL:0::v has ghost elements: 0 +DEAL:0::w has ghost elements: 0 +DEAL:0::w has ghost elements: 1 +DEAL:0::w has ghost elements: 1 +DEAL:0::w has ghost elements: 0 +DEAL:0::x has ghost elements: 1 +DEAL:0::x has ghost elements: 0 +DEAL:0::OK diff --git a/tests/cuda/parallel_vector_21.cu b/tests/cuda/parallel_vector_21.cu new file mode 100644 index 0000000000..f8e75de788 --- /dev/null +++ b/tests/cuda/parallel_vector_21.cu @@ -0,0 +1,104 @@ +// --------------------------------------------------------------------- +// +// Copyright (C) 2018 by the deal.II authors +// +// This file is part of the deal.II library. +// +// The deal.II library is free software; you can use it, redistribute +// it, and/or modify it under the terms of the GNU Lesser General +// Public License as published by the Free Software Foundation; either +// version 2.1 of the License, or (at your option) any later version. +// The full text of the license can be found in the file LICENSE.md at +// the top level directory of deal.II. +// +// --------------------------------------------------------------------- + +// Check that the data range representing ghosts is really initialized to zero +// when doing reinit() from another vector and manually setting the local +// range + +#include +#include +#include + +#include +#include + +#include +#include + +#include "../tests.h" + +__global__ void +set(double *v) +{ + v[threadIdx.x] = 1.; +} + +void +test() +{ + unsigned int my_id = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + unsigned int n_procs = Utilities::MPI::n_mpi_processes(MPI_COMM_WORLD); + + IndexSet locally_owned(n_procs * 2); + locally_owned.add_range(my_id * 2, my_id * 2 + 2); + IndexSet ghost_set(n_procs * 2); + ghost_set.add_index(0); + ghost_set.add_index(2); + + LinearAlgebra::distributed::Vector v( + locally_owned, ghost_set, MPI_COMM_WORLD); + + // create vector without actually setting the entries since they will be + // overwritten soon anyway + LinearAlgebra::distributed::Vector v2; + v2.reinit(v, true); + + // set locally owned range of v2 manually + set<<<1, v2.local_size()>>>(v2.get_values()); + + // add entries to ghost values + // Because of limitation in import, the IndexSet of the ReadWriteVector needs + // to have the local elements. + IndexSet workaround_set(locally_owned); + workaround_set.add_index(0); + workaround_set.add_index(2); + workaround_set.compress(); + LinearAlgebra::ReadWriteVector rw_vector(workaround_set); + rw_vector(0) += 1.; + rw_vector(2) += 1.; + v2.import(rw_vector, VectorOperation::add); + + // now we should have the correct data, not some uninitialized trash that + // resided in the ghost range + v2.print(deallog.get_file_stream()); + + v2.update_ghost_values(); + v2.print(deallog.get_file_stream()); +} + + + +int +main(int argc, char **argv) +{ + Utilities::MPI::MPI_InitFinalize mpi_initialization( + argc, argv, testing_max_num_threads()); + + unsigned int my_id = Utilities::MPI::this_mpi_process(MPI_COMM_WORLD); + Utilities::CUDA::Handle cuda_handle; + // By default, all the ranks will try to access the device 0. This is fine if + // we have one rank per node _and_ one gpu per node. If we have multiple GPUs + // on one node, we need each process to access a different GPU. We assume that + // each node has the same number of GPUs. + int n_devices = 0; + cudaError_t cuda_error_code = cudaGetDeviceCount(&n_devices); + AssertCuda(cuda_error_code); + int device_id = my_id % n_devices; + cuda_error_code = cudaSetDevice(device_id); + AssertCuda(cuda_error_code); + + MPILogInitAll log; + test(); +} diff --git a/tests/cuda/parallel_vector_21.mpirun=2.output b/tests/cuda/parallel_vector_21.mpirun=2.output new file mode 100644 index 0000000000..71d2ddbee6 --- /dev/null +++ b/tests/cuda/parallel_vector_21.mpirun=2.output @@ -0,0 +1,23 @@ + +Process #0 +Local range: [0, 2), global size: 4 +Vector data: +3.000e+00 1.000e+00 +Process #0 +Local range: [0, 2), global size: 4 +Vector data: +3.000e+00 1.000e+00 +Ghost entries (global index / value): +(2/3.000e+00) + +Process #1 +Local range: [2, 4), global size: 4 +Vector data: +3.000e+00 1.000e+00 +Process #1 +Local range: [2, 4), global size: 4 +Vector data: +3.000e+00 1.000e+00 +Ghost entries (global index / value): +(0/3.000e+00) + diff --git a/tests/cuda/vector_reinit_01.cu b/tests/cuda/vector_reinit_01.cu new file mode 100644 index 0000000000..df834427b8 --- /dev/null +++ b/tests/cuda/vector_reinit_01.cu @@ -0,0 +1,119 @@ +// --------------------------------------------------------------------- +// +// Copyright (C) 2018 by the deal.II authors +// +// This file is part of the deal.II library. +// +// The deal.II library is free software; you can use it, redistribute +// it, and/or modify it under the terms of the GNU Lesser General +// Public License as published by the Free Software Foundation; either +// version 2.1 of the License, or (at your option) any later version. +// The full text of the license can be found in the file LICENSE.md at +// the top level directory of deal.II. +// +// --------------------------------------------------------------------- + + +// check that LinearAlgebra::distributed::Vector::reinit does not carry over any +// state that can lead to invalid memory access. In this test, the MPI +// communicator is deleted. + + +#include + +#include +#include +#include +#include + +#include "../tests.h" + +template +void +do_test() +{ + IndexSet set(5); + set.add_range(0, 5); + + VectorType v1, v2; + + { + MPI_Comm communicator = + Utilities::MPI::duplicate_communicator(MPI_COMM_WORLD); + v1.reinit(set, communicator); + deallog << "reinit: " << v1.size() << " "; + +#ifdef DEAL_II_WITH_MPI + MPI_Comm_free(&communicator); +#endif + } + + { + MPI_Comm communicator = + Utilities::MPI::duplicate_communicator(MPI_COMM_WORLD); + v2.reinit(set, communicator); + v1.reinit(v2); + deallog << v1.size() << " "; + v1.reinit(v2); + deallog << v1.size() << std::endl; + +#ifdef DEAL_II_WITH_MPI + MPI_Comm_free(&communicator); +#endif + } + + { + MPI_Comm communicator = + Utilities::MPI::duplicate_communicator(MPI_COMM_WORLD); + v2.reinit(set, communicator); + v1 = v2; + deallog << "assign " << v1.size() << " "; + v1 = v2; + deallog << v1.size() << std::endl; + +#ifdef DEAL_II_WITH_MPI + MPI_Comm_free(&communicator); +#endif + } + + { + MPI_Comm communicator = + Utilities::MPI::duplicate_communicator(MPI_COMM_WORLD); + GrowingVectorMemory memory; + typename VectorMemory::Pointer v3(memory); + v1.reinit(set, communicator); + v3->reinit(v1); + deallog << "reinit pool " << v1.size() << " " << v3->size() << " "; + +#ifdef DEAL_II_WITH_MPI + MPI_Comm_free(&communicator); +#endif + } + + { + MPI_Comm communicator = + Utilities::MPI::duplicate_communicator(MPI_COMM_WORLD); + GrowingVectorMemory memory; + typename VectorMemory::Pointer v3(memory); + v1.reinit(set, communicator); + v3->reinit(v1); + deallog << "reinit pool " << v3->size() << std::endl; + +#ifdef DEAL_II_WITH_MPI + MPI_Comm_free(&communicator); +#endif + } +} + +int +main(int argc, char **argv) +{ + Utilities::MPI::MPI_InitFinalize mpi_initialization( + argc, argv, testing_max_num_threads()); + Utilities::CUDA::Handle cuda_handle; + + initlog(); + + do_test< + LinearAlgebra::distributed::Vector>(); +} diff --git a/tests/cuda/vector_reinit_01.output b/tests/cuda/vector_reinit_01.output new file mode 100644 index 0000000000..747b8affb7 --- /dev/null +++ b/tests/cuda/vector_reinit_01.output @@ -0,0 +1,4 @@ + +DEAL::reinit: 5 5 5 +DEAL::assign 5 5 +DEAL::reinit pool 5 5 reinit pool 5 -- 2.39.5