Improve scheduling (#73)

- tangent_basis takes ExecutionSpace instance as argument. - Remove unecessary fences - Ensure synchronous deep_copies - Fixes - Use DDC fork
blegouix · Jan 3, 2025 · 3420667 · 3420667
1 parent be4c2a8
commit 3420667
Show file tree

Hide file tree

Showing 12 changed files with 50 additions and 59 deletions.
diff --git a/.devnote b/.devnote
@@ -18,6 +18,9 @@ SPDX-License-Identifier: GPL-3.0-or-later
 - Consider removing metric_prod_t (and every helper producing a tensor). At least, standardize API.
 - Relabelizion in different file ?
 - Streams in Laplacian
+- Clarify domain() and mem_domain() (the same)
+- There are several projection of metric on MetricIndex which is simply a metric.indices_domain()
+- swap_character_t
 
 ----- SETUP BASIC COMMAND -----
 

diff --git a/include/similie/exterior/coboundary.hpp b/include/similie/exterior/coboundary.hpp
@@ -253,16 +253,14 @@ coboundary_tensor_t<TagToAddToCochain, CochainTag, TensorType> coboundary(
             CochainTag::rank() + 1,
             typename detail::NonSpectatorDimension<
                     TagToAddToCochain,
-                    typename TensorType::non_indices_domain_t>::type,
-            typename ExecSpace::memory_space>();
+                    typename TensorType::non_indices_domain_t>::type>(exec_space);
 
     // compute the tangent K-basis for each node of the mesh. This is a local K-chain.
     auto lower_chain = tangent_basis<
             CochainTag::rank(),
             typename detail::NonSpectatorDimension<
                     TagToAddToCochain,
-                    typename TensorType::non_indices_domain_t>::type,
-            typename ExecSpace::memory_space>();
+                    typename TensorType::non_indices_domain_t>::type>(exec_space);
 
     // iterate over every node, we will work inside the tangent space associated to each of them
     ddc::parallel_for_each(

diff --git a/include/similie/exterior/codifferential.hpp b/include/similie/exterior/codifferential.hpp
@@ -197,7 +197,6 @@ codifferential_tensor_t<TagToRemoveFromCochain, CochainTag, TensorType> codiffer
             sil::tensor::upper<MetricIndex>,
             MuUpSeq,
             NuLowSeq>(exec_space, hodge_star, inv_metric);
-    exec_space.fence();
 
     // Dual tensor
     [[maybe_unused]] tensor::TensorAccessor<
@@ -219,7 +218,6 @@ codifferential_tensor_t<TagToRemoveFromCochain, CochainTag, TensorType> codiffer
             KOKKOS_LAMBDA(typename TensorType::non_indices_domain_t::discrete_element_type elem) {
                 sil::tensor::tensor_prod(dual_tensor[elem], tensor[elem], hodge_star[elem]);
             });
-    exec_space.fence();
 
     // Dual codifferential
     [[maybe_unused]] tensor::TensorAccessor<
@@ -241,7 +239,6 @@ codifferential_tensor_t<TagToRemoveFromCochain, CochainTag, TensorType> codiffer
             misc::convert_type_seq_to_t<
                     tensor::TensorAntisymmetricIndex,
                     NuLowSeq>>(exec_space, dual_codifferential, dual_tensor);
-    exec_space.fence();
 
     // Hodge star 2
     [[maybe_unused]] sil::tensor::tensor_accessor_for_domain_t<HodgeStarDomain2>
@@ -257,7 +254,6 @@ codifferential_tensor_t<TagToRemoveFromCochain, CochainTag, TensorType> codiffer
             sil::tensor::upper<MetricIndex>,
             RhoUpSeq,
             SigmaLowSeq>(exec_space, hodge_star2, inv_metric);
-    exec_space.fence();
 
     // Codifferential
     ddc::parallel_for_each(
@@ -273,7 +269,6 @@ codifferential_tensor_t<TagToRemoveFromCochain, CochainTag, TensorType> codiffer
                     codifferential_tensor[elem] *= -1;
                 }
             });
-    exec_space.fence();
 
     return codifferential_tensor;
 }

diff --git a/include/similie/exterior/laplacian.hpp b/include/similie/exterior/laplacian.hpp
@@ -45,7 +45,6 @@ TensorType codifferential_of_coboundary(
     sil::tensor::Tensor derivative_tensor(derivative_tensor_alloc);
 
     sil::exterior::deriv<LaplacianDummyIndex, CochainTag>(exec_space, derivative_tensor, tensor);
-    exec_space.fence();
 
     // Codifferential
     sil::exterior::codifferential<
@@ -54,7 +53,6 @@ TensorType codifferential_of_coboundary(
             coboundary_index_t<
                     LaplacianDummyIndex,
                     CochainTag>>(exec_space, out_tensor, derivative_tensor, inv_metric);
-    exec_space.fence();
 
     return out_tensor;
 }
@@ -90,15 +88,13 @@ TensorType coboundary_of_codifferential(
             MetricIndex,
             LaplacianDummyIndex,
             CochainTag>(exec_space, codifferential_tensor, tensor, inv_metric);
-    exec_space.fence();
 
     // Coboundary
     sil::exterior::deriv<
             LaplacianDummyIndex,
             codifferential_index_t<
                     LaplacianDummyIndex,
                     CochainTag>>(exec_space, out_tensor, codifferential_tensor);
-    exec_space.fence();
 
     return out_tensor;
 }
@@ -132,9 +128,8 @@ TensorType laplacian(
                 MetricIndex,
                 LaplacianDummyIndex2,
                 CochainTag>(exec_space, laplacian_tensor, tensor, inv_metric);
-        exec_space.fence();
     } else if constexpr (CochainTag::rank() < LaplacianDummyIndex::size()) {
-        auto tmp_alloc = ddc::create_mirror(laplacian_tensor);
+        auto tmp_alloc = ddc::create_mirror(exec_space, laplacian_tensor);
         tensor::Tensor tmp(tmp_alloc);
 
         detail::codifferential_of_coboundary<
@@ -145,7 +140,6 @@ TensorType laplacian(
                 MetricIndex,
                 LaplacianDummyIndex,
                 CochainTag>(exec_space, tmp, tensor, inv_metric);
-        exec_space.fence();
 
         ddc::parallel_for_each(
                 exec_space,
@@ -158,7 +152,6 @@ TensorType laplacian(
                 MetricIndex,
                 LaplacianDummyIndex,
                 CochainTag>(exec_space, laplacian_tensor, tensor, inv_metric);
-        exec_space.fence();
     } else {
         assert(false && "Unsupported differential form in Laplacian operator");
     }

diff --git a/include/similie/exterior/local_chain.hpp b/include/similie/exterior/local_chain.hpp
@@ -374,13 +374,14 @@ LocalChain(Head, Tail...)
 
 namespace detail {
 
-template <std::size_t K, misc::Specialization<ddc::DiscreteDomain> Dom, class MemorySpace>
+template <std::size_t K, misc::Specialization<ddc::DiscreteDomain> Dom>
 struct TangentBasis;
 
-template <std::size_t K, class... Tag, class MemorySpace>
-struct TangentBasis<K, ddc::DiscreteDomain<Tag...>, MemorySpace>
+template <std::size_t K, class... Tag>
+struct TangentBasis<K, ddc::DiscreteDomain<Tag...>>
 {
-    static auto constexpr run()
+    template <class ExecSpace>
+    static auto constexpr run(ExecSpace const& exec_space)
     {
         std::array<std::ptrdiff_t, sizeof...(Tag)> permutation
                 = {0 * ddc::type_seq_rank_v<Tag, ddc::detail::TypeSeq<Tag...>>...};
@@ -394,22 +395,22 @@ struct TangentBasis<K, ddc::DiscreteDomain<Tag...>, MemorySpace>
             basis_host(i) = ddc::DiscreteVector<Tag...>();
             ddc::detail::array(basis_host(i++)) = permutation;
         } while (std::prev_permutation(permutation.begin(), permutation.end()));
-        Kokkos::View<ddc::DiscreteVector<Tag...>*, MemorySpace> basis
-                = create_mirror_view_and_copy(MemorySpace(), basis_host);
+        Kokkos::View<ddc::DiscreteVector<Tag...>*, typename ExecSpace::memory_space> basis
+                = create_mirror_view_and_copy(exec_space, basis_host);
 
         return LocalChain<
                 Simplex<K, Tag...>,
                 Kokkos::LayoutRight,
-                MemorySpace>(basis, basis.size());
+                typename ExecSpace::memory_space>(basis, basis.size());
     }
 };
 
 } // namespace detail
 
-template <std::size_t K, misc::Specialization<ddc::DiscreteDomain> Dom, class MemorySpace>
-constexpr auto tangent_basis()
+template <std::size_t K, misc::Specialization<ddc::DiscreteDomain> Dom, class ExecSpace>
+constexpr auto tangent_basis(ExecSpace const& exec_space)
 {
-    return detail::TangentBasis<K, Dom, MemorySpace>::run();
+    return detail::TangentBasis<K, Dom>::run(exec_space);
 }
 
 template <misc::Specialization<LocalChain> ChainType>

diff --git a/include/similie/tensor/determinant.hpp b/include/similie/tensor/determinant.hpp
@@ -68,6 +68,7 @@ KOKKOS_FUNCTION typename ViewType::value_type determinant(const ViewType& matrix
     return det;
 }
 
+// TODO port on GPU
 template <misc::Specialization<Tensor> TensorType>
 TensorType::element_type determinant(TensorType tensor)
 {

diff --git a/include/similie/tensor/identity_tensor.hpp b/include/similie/tensor/identity_tensor.hpp
@@ -88,8 +88,11 @@ struct TensorIdentityIndex
     KOKKOS_FUNCTION static constexpr std::array<std::size_t, rank()>
     mem_id_to_canonical_natural_ids(std::size_t mem_id)
     {
-        assert(mem_id < mem_size()); // Always false
+        assert(false);
         std::array<std::size_t, rank()> ids;
+        for (auto i = ids.begin(); i < ids.end(); ++i) {
+            *i = 0;
+        }
         return ids;
     }
 };

diff --git a/include/similie/tensor/levi_civita_tensor.hpp b/include/similie/tensor/levi_civita_tensor.hpp
@@ -99,8 +99,11 @@ struct TensorLeviCivitaIndex
     KOKKOS_FUNCTION static constexpr std::array<std::size_t, rank()>
     mem_id_to_canonical_natural_ids(std::size_t mem_id)
     {
-        assert(mem_id < mem_size()); // Always false
+        assert(false);
         std::array<std::size_t, rank()> ids;
+        for (auto i = ids.begin(); i < ids.end(); ++i) {
+            *i = 0;
+        }
         return ids;
     }
 };

diff --git a/include/similie/tensor/lorentzian_sign_tensor.hpp b/include/similie/tensor/lorentzian_sign_tensor.hpp
@@ -94,8 +94,11 @@ struct TensorLorentzianSignIndex
     KOKKOS_FUNCTION static constexpr std::array<std::size_t, rank()>
     mem_id_to_canonical_natural_ids(std::size_t mem_id)
     {
-        assert(mem_id < mem_size()); // Always false
+        assert(false);
         std::array<std::size_t, rank()> ids;
+        for (auto i = ids.begin(); i < ids.end(); ++i) {
+            *i = 0;
+        }
         return ids;
     }
 };

diff --git a/include/similie/tensor/metric.hpp b/include/similie/tensor/metric.hpp
@@ -350,6 +350,7 @@ inplace_apply_metric(ExecSpace const& exec_space, TensorType tensor, MetricType
                                 tensor)[elem]);
             });
     Kokkos::deep_copy(
+            exec_space,
             tensor.allocation_kokkos_view(),
             result.allocation_kokkos_view()); // We rely on Kokkos::deep_copy in place of ddc::parallel_deepcopy to avoid type verification of the type dimensions
 
@@ -416,14 +417,14 @@ invert_metric_t<MetricType> fill_inverse_metric(
                 exec_space,
                 inv_metric.mem_domain(),
                 KOKKOS_LAMBDA(invert_metric_t<MetricType>::discrete_element_type elem) {
-                    inv_metric(elem)
+                    inv_metric.mem(elem)
                             = 1.
-                              / metric(relabelize_indices_in<
-                                       swap_character<ddc::to_type_seq_t<
-                                               typename MetricType::accessor_t::natural_domain_t>>,
-                                       ddc::to_type_seq_t<
-                                               typename MetricType::accessor_t::natural_domain_t>>(
-                                      elem));
+                              / metric.mem(relabelize_indices_in<
+                                           swap_character<ddc::to_type_seq_t<
+                                                   typename MetricType::accessor_t::
+                                                           natural_domain_t>>,
+                                           ddc::to_type_seq_t<typename MetricType::accessor_t::
+                                                                      natural_domain_t>>(elem));
                 });
     } else if (misc::Specialization<MetricIndex, TensorSymmetricIndex>) {
         // Allocate a buffer mirroring the metric as a full matrix
@@ -502,16 +503,11 @@ invert_metric_t<MetricType> fill_inverse_metric(
                     */
 
                     ddc::for_each(
-                            tensor::swap_character<ddc::DiscreteDomain<
-                                    tensor::metric_index_1<MetricIndex>,
-                                    tensor::metric_index_2<MetricIndex>>>(
-                                    inv_metric.natural_domain()),
-                            [=](decltype(tensor::swap_character<ddc::DiscreteDomain<
-                                                 tensor::metric_index_1<MetricIndex>,
-                                                 tensor::metric_index_2<MetricIndex>>>(
-                                    inv_metric.natural_domain()))::discrete_element_type index) {
-                                // TODO do better, symmetric tensor is filled twice
-                                inv_metric(inv_metric.access_element(elem, index)) = buffer(
+                            tensor::swap_character<ddc::DiscreteDomain<MetricIndex>>(
+                                    inv_metric.domain()),
+                            [=](decltype(tensor::swap_character<ddc::DiscreteDomain<MetricIndex>>(
+                                    inv_metric.domain()))::discrete_element_type mem_index) {
+                                inv_metric.mem(elem, mem_index) = buffer(
                                         elem,
                                         tensor::relabelize_indices_in<
                                                 tensor::swap_character<ddc::detail::TypeSeq<
@@ -520,7 +516,8 @@ invert_metric_t<MetricType> fill_inverse_metric(
                                                 ddc::detail::TypeSeq<
                                                         tensor::metric_index_1<MetricIndex>,
                                                         tensor::metric_index_2<MetricIndex>>>(
-                                                index));
+                                                inv_metric.accessor().canonical_natural_element(
+                                                        mem_index)));
                             });
                 });
     }

diff --git a/include/similie/tensor/tensor_impl.hpp b/include/similie/tensor/tensor_impl.hpp
@@ -1204,7 +1204,7 @@ struct NaturalTensorProd<
     run(Tensor<ElementType,
                ddc::DiscreteDomain<HeadDDim1..., TailDDim2...>,
                LayoutStridedPolicy,
-               Kokkos::DefaultHostExecutionSpace::memory_space> prod_tensor,
+               MemorySpace> prod_tensor,
         Tensor<ElementType,
                ddc::DiscreteDomain<HeadDDim1..., ContractDDim...>,
                LayoutStridedPolicy,
@@ -1239,15 +1239,9 @@ template <
         class ElementType,
         class LayoutStridedPolicy,
         class MemorySpace>
-Tensor<ElementType,
-       ddc::DiscreteDomain<ProdDDim...>,
-       LayoutStridedPolicy,
-       Kokkos::DefaultHostExecutionSpace::memory_space>
-tensor_prod(
-        Tensor<ElementType,
-               ddc::DiscreteDomain<ProdDDim...>,
-               LayoutStridedPolicy,
-               Kokkos::DefaultHostExecutionSpace::memory_space> prod_tensor,
+Tensor<ElementType, ddc::DiscreteDomain<ProdDDim...>, LayoutStridedPolicy, MemorySpace> tensor_prod(
+        Tensor<ElementType, ddc::DiscreteDomain<ProdDDim...>, LayoutStridedPolicy, MemorySpace>
+                prod_tensor,
         Tensor<ElementType, ddc::DiscreteDomain<DDim1...>, LayoutStridedPolicy, MemorySpace>
                 tensor1,
         Tensor<ElementType, ddc::DiscreteDomain<DDim2...>, LayoutStridedPolicy, MemorySpace>

diff --git a/vendor/ddc b/vendor/ddc
+2 −4		.clang-tidy
+9 −9		benchmarks/deepcopy.cpp
+14 −9		benchmarks/splines.cpp
+0 −34		docs/first_steps.md
+0 −269		examples/.clang-format
+2 −6		examples/CMakeLists.txt
+37 −72		examples/characteristics_advection.cpp
+16 −30		examples/game_of_life.cpp
+59 −107		examples/heat_equation.cpp
+33 −64		examples/heat_equation_spectral.cpp
+100 −151		examples/non_uniform_heat_equation.cpp
+77 −0		examples/pdi_event.cpp
+49 −108		examples/uniform_heat_equation.cpp
+8 −4		include/ddc/chunk.hpp
+1 −1		include/ddc/chunk_common.hpp
+4 −2		include/ddc/chunk_span.hpp
+16 −2		include/ddc/detail/kokkos.hpp
+2 −2		include/ddc/detail/type_seq.hpp
+7 −7		include/ddc/discrete_domain.hpp
+2 −2		include/ddc/discrete_element.hpp
+2 −0		include/ddc/discrete_vector.hpp
+35 −0		include/ddc/for_each.hpp
+19 −47		include/ddc/kernels/fft.hpp
+8 −7		include/ddc/kernels/splines/constant_extrapolation_rule.hpp
+22 −22		include/ddc/kernels/splines/spline_builder.hpp
+19 −19		include/ddc/kernels/splines/spline_builder_2d.hpp
+5 −7		include/ddc/kernels/splines/spline_evaluator.hpp
+6 −8		include/ddc/kernels/splines/spline_evaluator_2d.hpp
+1 −1		include/ddc/kernels/splines/splines_linear_problem.hpp
+2 −2		include/ddc/kernels/splines/splines_linear_problem_sparse.hpp
+4 −4		include/ddc/non_uniform_point_sampling.hpp
+6 −2		include/ddc/parallel_deepcopy.hpp
+12 −17		include/ddc/periodic_sampling.hpp
+67 −1		include/ddc/transform_reduce.hpp
+11 −16		include/ddc/uniform_point_sampling.hpp
+1 −1		tests/chunk.cpp
+41 −0		tests/chunk_span.cpp
+6 −6		tests/discrete_domain.cpp
+44 −11		tests/fft/fft.cpp
+70 −0		tests/for_each.cpp
+25 −9		tests/parallel_transform_reduce.cpp
+116 −124		tests/splines/batched_2d_spline_builder.cpp
+66 −62		tests/splines/batched_spline_builder.cpp
+6 −6		tests/splines/evaluator_2d.hpp
+91 −87		tests/splines/extrapolation_rule.cpp
+2 −2		tests/splines/knots_as_interpolation_points.cpp
+19 −19		tests/splines/non_periodic_spline_builder.cpp
+11 −11		tests/splines/periodic_spline_builder.cpp
+6 −6		tests/splines/periodic_spline_builder_ordered_points.cpp
+38 −34		tests/splines/periodicity_spline_builder.cpp
+17 −17		tests/splines/spline_builder.cpp
+3 −3		tests/tagged_vector.cpp
+38 −0		tests/transform_reduce.cpp