From: Peter Munch Date: Mon, 23 Mar 2020 09:09:53 +0000 (+0100) Subject: Inherit LinearAlgebra::SharedMPI::MemorySpaceData from ::dealii::MemorySpace::MemoryS... X-Git-Url: https://gitweb.dealii.org/cgi-bin/gitweb.cgi?a=commitdiff_plain;h=15c9ec4e7abc233afc768f5a3c4193a29e3a8c50;p=dealii.git Inherit LinearAlgebra::SharedMPI::MemorySpaceData from ::dealii::MemorySpace::MemorySpaceData and use TBBPartitioner --- diff --git a/include/deal.II/base/memory_space.h b/include/deal.II/base/memory_space.h index 9042ac105e..425182e059 100644 --- a/include/deal.II/base/memory_space.h +++ b/include/deal.II/base/memory_space.h @@ -64,14 +64,14 @@ namespace MemorySpace * Copy the active data (values for Host and values_dev for CUDA) to @p begin. * If the data is on the device it is moved to the host. */ - void + virtual void copy_to(Number *begin, std::size_t n_elements) { (void)begin; (void)n_elements; } - void + virtual void /** * Copy the data in @p begin to the active data of the structure (values for * Host and values_dev for CUDA). The pointer @p begin must be on the host. @@ -85,7 +85,7 @@ namespace MemorySpace /** * Pointer to data on the host. */ - std::unique_ptr values; + std::unique_ptr> values; /** * Pointer to data on the device. @@ -117,19 +117,19 @@ namespace MemorySpace : values(nullptr, &std::free) {} - void + virtual void copy_to(Number *begin, std::size_t n_elements) { std::copy(values.get(), values.get() + n_elements, begin); } - void + virtual void copy_from(Number *begin, std::size_t n_elements) { std::copy(begin, begin + n_elements, values.get()); } - std::unique_ptr values; + std::unique_ptr> values; // This is not used but it allows to simplify the code until we start using // CUDA-aware MPI. @@ -157,7 +157,7 @@ namespace MemorySpace , values_dev(nullptr, Utilities::CUDA::delete_device_data) {} - void + virtual void copy_to(Number *begin, std::size_t n_elements) { const cudaError_t cuda_error_code = @@ -168,7 +168,7 @@ namespace MemorySpace AssertCuda(cuda_error_code); } - void + virtual void copy_from(Number *begin, std::size_t n_elements) { const cudaError_t cuda_error_code = @@ -179,8 +179,8 @@ namespace MemorySpace AssertCuda(cuda_error_code); } - std::unique_ptr values; - std::unique_ptr values_dev; + std::unique_ptr> values; + std::unique_ptr values_dev; }; diff --git a/include/deal.II/lac/la_parallel_vector.templates.h b/include/deal.II/lac/la_parallel_vector.templates.h index e5ad13c720..148dbef146 100644 --- a/include/deal.II/lac/la_parallel_vector.templates.h +++ b/include/deal.II/lac/la_parallel_vector.templates.h @@ -140,7 +140,7 @@ namespace LinearAlgebra reinterpret_cast(&new_val), 64, sizeof(Number) * new_alloc_size); - data.values.reset(new_val); + data.values = {new_val, [](Number *&data) { std::free(data); }}; allocated_size = new_alloc_size; } diff --git a/include/deal.II/lac/la_sm_vector.h b/include/deal.II/lac/la_sm_vector.h index 87ba0660fa..4e89827b01 100644 --- a/include/deal.II/lac/la_sm_vector.h +++ b/include/deal.II/lac/la_sm_vector.h @@ -78,9 +78,10 @@ namespace LinearAlgebra { template struct MemorySpaceData + : public ::dealii::MemorySpace::MemorySpaceData { void - copy_to(Number *begin, std::size_t n_elements) + copy_to(Number *begin, std::size_t n_elements) override { Assert(false, ExcNotImplemented()); (void)begin; @@ -88,19 +89,16 @@ namespace LinearAlgebra } void - copy_from(Number *begin, std::size_t n_elements) + copy_from(Number *begin, std::size_t n_elements) override { Assert(false, ExcNotImplemented()); (void)begin; (void)n_elements; } - std::unique_ptr> values; MPI_Win *values_win = nullptr; std::vector others; - - std::unique_ptr values_dev; }; @@ -485,6 +483,13 @@ namespace LinearAlgebra mutable MemorySpaceData data; + /** + * For parallel loops with TBB, this member variable stores the affinity + * information of loops. + */ + mutable std::shared_ptr<::dealii::parallel::internal::TBBPartitioner> + thread_loop_partitioner; + // needed? mutable MemorySpaceData import_data; diff --git a/include/deal.II/lac/la_sm_vector.templates.h b/include/deal.II/lac/la_sm_vector.templates.h index 873c6a6488..41841b283a 100644 --- a/include/deal.II/lac/la_sm_vector.templates.h +++ b/include/deal.II/lac/la_sm_vector.templates.h @@ -116,6 +116,19 @@ namespace LinearAlgebra [&data](Number *&) { MPI_Win_free(data.values_win); }}; data.values_win = win; } + + template + static void + linfty_norm_local(const ::dealii::MemorySpace::MemorySpaceData< + Number, + ::dealii::MemorySpace::Host> &data, + const unsigned int size, + RealType & max) + { + for (size_type i = 0; i < size; ++i) + max = + std::max(numbers::NumberTraits::abs(data.values[i]), max); + } }; } // namespace internal @@ -141,6 +154,9 @@ namespace LinearAlgebra allocated_size, data, comm_sm); + + thread_loop_partitioner = + std::make_shared<::dealii::parallel::internal::TBBPartitioner>(); } @@ -192,6 +208,8 @@ namespace LinearAlgebra // call these methods and hence do not need to have the storage. import_data.values.reset(); import_data.values_dev.reset(); + + thread_loop_partitioner = v.thread_loop_partitioner; } @@ -272,9 +290,15 @@ namespace LinearAlgebra { reinit(v, true); + thread_loop_partitioner = v.thread_loop_partitioner; + const size_type this_size = local_size(); if (this_size > 0) - std::memcpy(this->begin(), v.begin(), this_size * sizeof(Number)); + { + dealii::internal::VectorOperations:: + functions::copy( + thread_loop_partitioner, partitioner->local_size(), v.data, data); + } } @@ -407,9 +431,15 @@ namespace LinearAlgebra else must_update_ghost_values |= vector_is_ghosted; + thread_loop_partitioner = c.thread_loop_partitioner; + const size_type this_size = partitioner->local_size(); if (this_size > 0) - std::memcpy(this->begin(), c.begin(), this_size * sizeof(Number)); + { + dealii::internal::VectorOperations:: + functions::copy( + thread_loop_partitioner, this_size, c.data, data); + } if (must_update_ghost_values) update_ghost_values(); @@ -563,8 +593,14 @@ namespace LinearAlgebra { const size_type this_size = local_size(); if (this_size > 0) - std::fill_n(data.values.get(), this_size, s); + { + dealii::internal::VectorOperations:: + functions::set( + thread_loop_partitioner, this_size, s, data); + } + // if we call Vector::operator=0, we want to zero out all the entries + // plus ghosts. if (s == Number()) zero_out_ghosts(); @@ -611,11 +647,9 @@ namespace LinearAlgebra AssertDimension(local_size(), v.local_size()); - auto values = this->begin(); - const auto values_other = v.begin(); - - for (unsigned int i = 0; i < partitioner->local_size(); i++) - values[i] -= values_other[i]; + dealii::internal::VectorOperations:: + functions::subtract_vector( + thread_loop_partitioner, partitioner->local_size(), v.data, data); if (vector_is_ghosted) update_ghost_values(); @@ -654,11 +688,9 @@ namespace LinearAlgebra if (a == Number(0.)) return; - auto values = this->begin(); - const auto values_other = v.begin(); - - for (unsigned int i = 0; i < partitioner->local_size(); i++) - values[i] += a * values_other[i]; + dealii::internal::VectorOperations:: + functions::add_av( + thread_loop_partitioner, partitioner->local_size(), a, v.data, data); } @@ -726,22 +758,22 @@ namespace LinearAlgebra { // Downcast. Throws an exception if invalid. using VectorType = Vector; - Assert(dynamic_cast(&vv) != nullptr, + Assert((dynamic_cast(&vv) != nullptr), ExcVectorTypeNotCompatible()); const VectorType &v = dynamic_cast(vv); + AssertIsFinite(x); AssertIsFinite(a); AssertDimension(local_size(), v.local_size()); - // nothing to do if a is zero - if (a == Number(0.)) - return; - - auto values = this->begin(); - const auto values_other = v.begin(); - - for (unsigned int i = 0; i < partitioner->local_size(); i++) - values[i] = x * values[i] + a * values_other[i]; + dealii::internal::VectorOperations:: + functions::sadd_xav( + thread_loop_partitioner, + partitioner->local_size(), + x, + a, + v.data, + data); } @@ -815,11 +847,9 @@ namespace LinearAlgebra AssertDimension(local_size(), v.local_size()); - auto values = this->begin(); - const auto values_other = v.begin(); - - for (unsigned int i = 0; i < partitioner->local_size(); i++) - values[i] *= values_other[i]; + dealii::internal::VectorOperations:: + functions::scale( + thread_loop_partitioner, local_size(), v.data, data); if (vector_is_ghosted) update_ghost_values(); @@ -841,11 +871,10 @@ namespace LinearAlgebra AssertIsFinite(a); AssertDimension(local_size(), v.local_size()); - auto values = this->begin(); - const auto values_other = v.begin(); + dealii::internal::VectorOperations:: + functions::equ_au( + thread_loop_partitioner, partitioner->local_size(), a, v.data, data); - for (unsigned int i = 0; i < partitioner->local_size(); i++) - values[i] = a * values_other[i]; if (vector_is_ghosted) update_ghost_values(); @@ -890,15 +919,9 @@ namespace LinearAlgebra AssertDimension(partitioner->local_size(), v.partitioner->local_size()); - real_type sum = Number(); - - auto values = this->begin(); - auto values_other = v.begin(); - - for (unsigned int i = 0; i < partitioner->local_size(); i++) - sum += values[i] * values_other[i]; - - return sum; + return dealii::internal::VectorOperations:: + functions::dot( + thread_loop_partitioner, partitioner->local_size(), v.data, data); } @@ -927,12 +950,12 @@ namespace LinearAlgebra typename Vector::real_type Vector::norm_sqr_local() const { - real_type sum = Number(); + real_type sum; - auto values = this->begin(); - for (unsigned int i = 0; i < partitioner->local_size(); i++) - sum += values[i] * values[i]; + dealii::internal::VectorOperations:: + functions::norm_2( + thread_loop_partitioner, partitioner->local_size(), sum, data); AssertIsFinite(sum); @@ -1030,12 +1053,12 @@ namespace LinearAlgebra typename Vector::real_type Vector::linfty_norm_local() const { - real_type max = Number(); - - auto values = this->begin(); + real_type max = 0.; - for (unsigned int i = 0; i < partitioner->local_size(); i++) - max = std::max(std::abs(values[i]), max); + const size_type local_size = partitioner->local_size(); + internal::la_parallel_vector_templates_functions< + Number, + MemorySpaceType>::linfty_norm_local(data, local_size, max); return max; } @@ -1067,18 +1090,9 @@ namespace LinearAlgebra AssertDimension(vec_size, v.local_size()); AssertDimension(vec_size, w.local_size()); - real_type sum = Number(); - - auto values = this->begin(); - auto values_v = v.begin(); - auto values_w = w.begin(); - - for (unsigned int i = 0; i < partitioner->local_size(); i++) - { - const auto temp = values[i] + a * values_v[i]; - values[i] = temp; - sum += temp * values_w[i]; - } + Number sum = dealii::internal::VectorOperations:: + functions::add_and_dot( + thread_loop_partitioner, vec_size, a, v.data, w.data, data); AssertIsFinite(sum);