tuning

CExA-project · Nov 12, 2023 · 9619c34 · 9619c34
1 parent 33cd935
commit 9619c34
Show file tree

Hide file tree

Showing 3 changed files with 8 additions and 8 deletions.
diff --git a/benchmarks/splines.cpp b/benchmarks/splines.cpp
@@ -179,13 +179,13 @@ static void characteristics_advection(benchmark::State& state)
 
 #ifdef KOKKOS_ENABLE_CUDA
 std::string chip = "gpu";
-int cols_per_par_chunk_ref = 1024;
-int par_chunks_per_seq_chunk_ref = 160;
+int cols_per_par_chunk_ref = 65535;
+int par_chunks_per_seq_chunk_ref = 1;
 unsigned int preconditionner_max_block_size_ref = 1u;
 #elif defined(KOKKOS_ENABLE_OPENMP)
 std::string chip = "cpu";
 int cols_per_par_chunk_ref = 512;
-int par_chunks_per_seq_chunk_ref = 160;
+int par_chunks_per_seq_chunk_ref = Kokkos::OpenMP().concurrency();
 unsigned int preconditionner_max_block_size_ref = 8u;
 #elif defined(KOKKOS_ENABLE_SERIAL)
 std::string chip = "cpu";

diff --git a/benchmarks/splines_plot.py b/benchmarks/splines_plot.py
@@ -39,7 +39,7 @@
     bandwidth = [group_data["bytes_per_second"][i] for i in range(len(ny))]
     plt.plot(ny, bandwidth, marker='o', markersize=5, label=f'nx={nx}')
 
-x = np.linspace(min(ny), 10*min(ny))
+x = np.linspace(min(ny), 20*min(ny))
 plt.plot(x, np.mean([data_groups[nx]["bytes_per_second"][0] for nx in nx_values])/min(ny)*x, linestyle='--', color='black', label='perfect scaling')
 
 # Plotting the data

diff --git a/include/ddc/kernels/splines/matrix_sparse.hpp b/include/ddc/kernels/splines/matrix_sparse.hpp
@@ -76,12 +76,12 @@ class Matrix_Sparse : public Matrix
 #endif
 #ifdef KOKKOS_ENABLE_CUDA
             if (std::is_same_v<ExecSpace, Kokkos::Cuda>) {
-                m_cols_per_par_chunk = 1024;
+                m_cols_per_par_chunk = 65535;
             }
 #endif
 #ifdef KOKKOS_ENABLE_HIP
             if (std::is_same_v<ExecSpace, Kokkos::HIP>) {
-                m_cols_per_par_chunk = 1024;
+                m_cols_per_par_chunk = 65535;
             }
 #endif
         }
@@ -101,12 +101,12 @@ class Matrix_Sparse : public Matrix
 #endif
 #ifdef KOKKOS_ENABLE_CUDA
             if (std::is_same_v<ExecSpace, Kokkos::Cuda>) {
-                m_par_chunks_per_seq_chunk = Kokkos::DefaultHostExecutionSpace().concurrency();
+                m_par_chunks_per_seq_chunk = 1;
             }
 #endif
 #ifdef KOKKOS_ENABLE_HIP
             if (std::is_same_v<ExecSpace, Kokkos::HIP>) {
-                m_par_chunks_per_seq_chunk = Kokkos::DefaultHostExecutionSpace().concurrency();
+                m_par_chunks_per_seq_chunk = 1;
             }
 #endif
         }