Merge branch 'lapack-backend' of github.com:Maison-de-la-Simulation/d…

…dc into lapack-backend
CExA-project · Jun 28, 2024 · 5b56d56 · 5b56d56
2 parents 72fc3d6 + 26225f3
commit 5b56d56
Show file tree

Hide file tree

Showing 4 changed files with 20 additions and 3 deletions.
diff --git a/include/ddc/kernels/fft.hpp b/include/ddc/kernels/fft.hpp
@@ -564,6 +564,7 @@ void core(
         }
 
         Kokkos::parallel_for(
+                "ddc_fft_normalization",
                 Kokkos::RangePolicy<ExecSpace>(
                         execSpace,
                         0,

diff --git a/include/ddc/kernels/splines/spline_builder.hpp b/include/ddc/kernels/splines/spline_builder.hpp
@@ -793,6 +793,7 @@ operator()(
         auto derivs_xmin_values = *derivs_xmin;
         auto const dx_proxy = m_dx;
         ddc::parallel_for_each(
+                "ddc_splines_hermite_compute_lower_coefficients",
                 exec_space(),
                 batch_domain(),
                 KOKKOS_LAMBDA(typename batch_domain_type::discrete_element_type j) {
@@ -810,6 +811,7 @@ operator()(
     auto const& interp_size_proxy = interpolation_domain().extents();
     auto const& nbasis_proxy = ddc::discrete_space<bsplines_type>().nbasis();
     ddc::parallel_for_each(
+            "ddc_splines_fill_rhs",
             exec_space(),
             batch_domain(),
             KOKKOS_LAMBDA(typename batch_domain_type::discrete_element_type j) {
@@ -830,6 +832,7 @@ operator()(
         auto derivs_xmax_values = *derivs_xmax;
         auto const dx_proxy = m_dx;
         ddc::parallel_for_each(
+                "ddc_splines_hermite_compute_upper_coefficients",
                 exec_space(),
                 batch_domain(),
                 KOKKOS_LAMBDA(typename batch_domain_type::discrete_element_type j) {
@@ -843,12 +846,13 @@ operator()(
     }
 
     // TODO : Consider optimizing
-    // Allocate and fill a transposed version of spline in order to get dimension of interest as last dimension (optimal for GPU, necessary for Ginkgo)
+    // Allocate and fill a transposed version of spline in order to get dimension of interest as last dimension (optimal for GPU, necessary for Ginkgo). Also select only relevant rows in case of periodic boundaries
     ddc::Chunk spline_tr_alloc(
             batched_spline_tr_domain(),
             ddc::KokkosAllocator<double, memory_space>());
     ddc::ChunkSpan spline_tr = spline_tr_alloc.span_view();
     ddc::parallel_for_each(
+            "ddc_splines_transpose_rhs",
             exec_space(),
             batch_domain(),
             KOKKOS_LAMBDA(typename batch_domain_type::discrete_element_type const j) {
@@ -864,8 +868,9 @@ operator()(
             batch_domain().size());
     // Compute spline coef
     matrix->solve(bcoef_section);
-    // Transpose back spline_tr in spline
+    // Transpose back spline_tr into spline.
     ddc::parallel_for_each(
+            "ddc_splines_transpose_back_rhs",
             exec_space(),
             batch_domain(),
             KOKKOS_LAMBDA(typename batch_domain_type::discrete_element_type const j) {
@@ -875,9 +880,10 @@ operator()(
                 }
             });
 
-    // Not sure yet of what this part do
+    // Duplicate the lower spline coefficients to the upper side in case of periodic boundaries
     if (bsplines_type::is_periodic()) {
         ddc::parallel_for_each(
+                "ddc_splines_periodic_rows_duplicate_rhs",
                 exec_space(),
                 batch_domain(),
                 KOKKOS_LAMBDA(typename batch_domain_type::discrete_element_type const j) {

diff --git a/include/ddc/kernels/splines/spline_evaluator.hpp b/include/ddc/kernels/splines/spline_evaluator.hpp
@@ -275,6 +275,7 @@ class SplineEvaluator
         batch_domain_type const batch_domain(spline_eval.domain());
 
         ddc::parallel_for_each(
+                "ddc_splines_evaluate",
                 exec_space(),
                 batch_domain,
                 KOKKOS_CLASS_LAMBDA(typename batch_domain_type::discrete_element_type const j) {
@@ -341,6 +342,7 @@ class SplineEvaluator
         batch_domain_type const batch_domain(spline_eval.domain());
 
         ddc::parallel_for_each(
+                "ddc_splines_differentiate",
                 exec_space(),
                 batch_domain,
                 KOKKOS_CLASS_LAMBDA(typename batch_domain_type::discrete_element_type const j) {
@@ -379,10 +381,12 @@ class SplineEvaluator
                 ddc::KokkosAllocator<double, memory_space>());
         ddc::ChunkSpan values = values_alloc.span_view();
         Kokkos::parallel_for(
+                "ddc_splines_integrate_bsplines",
                 Kokkos::RangePolicy<exec_space>(0, 1),
                 KOKKOS_LAMBDA(int) { ddc::discrete_space<bsplines_type>().integrals(values); });
 
         ddc::parallel_for_each(
+                "ddc_splines_integrate",
                 exec_space(),
                 batch_domain,
                 KOKKOS_LAMBDA(typename batch_domain_type::discrete_element_type const j) {

diff --git a/include/ddc/kernels/splines/spline_evaluator_2d.hpp b/include/ddc/kernels/splines/spline_evaluator_2d.hpp
@@ -380,6 +380,7 @@ class SplineEvaluator2D
         evaluation_domain_type1 const evaluation_domain1(spline_eval.domain());
         evaluation_domain_type2 const evaluation_domain2(spline_eval.domain());
         ddc::parallel_for_each(
+                "ddc_splines_evaluate_2d",
                 exec_space(),
                 batch_domain,
                 KOKKOS_CLASS_LAMBDA(typename batch_domain_type::discrete_element_type const j) {
@@ -559,6 +560,7 @@ class SplineEvaluator2D
         evaluation_domain_type1 const evaluation_domain1(spline_eval.domain());
         evaluation_domain_type2 const evaluation_domain2(spline_eval.domain());
         ddc::parallel_for_each(
+                "ddc_splines_differentiate_2d_dim_1",
                 exec_space(),
                 batch_domain,
                 KOKKOS_CLASS_LAMBDA(typename batch_domain_type::discrete_element_type const j) {
@@ -609,6 +611,7 @@ class SplineEvaluator2D
         evaluation_domain_type1 const evaluation_domain1(spline_eval.domain());
         evaluation_domain_type2 const evaluation_domain2(spline_eval.domain());
         ddc::parallel_for_each(
+                "ddc_splines_differentiate_2d_dim_2",
                 exec_space(),
                 batch_domain,
                 KOKKOS_CLASS_LAMBDA(typename batch_domain_type::discrete_element_type const j) {
@@ -659,6 +662,7 @@ class SplineEvaluator2D
         evaluation_domain_type1 const evaluation_domain1(spline_eval.domain());
         evaluation_domain_type2 const evaluation_domain2(spline_eval.domain());
         ddc::parallel_for_each(
+                "ddc_splines_cross_differentiate",
                 exec_space(),
                 batch_domain,
                 KOKKOS_CLASS_LAMBDA(typename batch_domain_type::discrete_element_type const j) {
@@ -804,13 +808,15 @@ class SplineEvaluator2D
                 ddc::KokkosAllocator<double, memory_space>());
         ddc::ChunkSpan values2 = values2_alloc.span_view();
         Kokkos::parallel_for(
+                "ddc_splines_integrate_bsplines_2d",
                 Kokkos::RangePolicy<exec_space>(0, 1),
                 KOKKOS_LAMBDA(int) {
                     ddc::discrete_space<bsplines_type1>().integrals(values1);
                     ddc::discrete_space<bsplines_type2>().integrals(values2);
                 });
 
         ddc::parallel_for_each(
+                "ddc_splines_integrate_bsplines",
                 exec_space(),
                 batch_domain,
                 KOKKOS_LAMBDA(typename batch_domain_type::discrete_element_type const j) {