Update of the perf tests based on Luc's comments

kokkos · Nov 9, 2022 · 372493d · 372493d
1 parent 5e013fe
commit 372493d
Show file tree

Hide file tree

Showing 12 changed files with 84 additions and 152 deletions.
diff --git a/perf_test/batched/sparse/CG/Functor_TestBatchedTeamVectorCG_1.hpp b/perf_test/batched/sparse/CG/Functor_TestBatchedTeamVectorCG_1.hpp
@@ -99,10 +99,10 @@ struct Functor_TestBatchedTeamVectorCG_1 {
     _handle.tmp_view = typename KrylovHandleType::TemporaryViewType(
         "", _X.extent(0), 4 * _X.extent(1));
 
-    Kokkos::TeamPolicy<DeviceType> auto_policy(
-        ceil(1. * _D.extent(0) / _N_team), Kokkos::AUTO(), Kokkos::AUTO());
-    Kokkos::TeamPolicy<DeviceType> tuned_policy(
-        ceil(1. * _D.extent(0) / _N_team), _team_size, _vector_length);
+    Kokkos::TeamPolicy<DeviceType> auto_policy(_handle.get_number_of_teams(),
+                                               Kokkos::AUTO(), Kokkos::AUTO());
+    Kokkos::TeamPolicy<DeviceType> tuned_policy(_handle.get_number_of_teams(),
+                                                _team_size, _vector_length);
     Kokkos::TeamPolicy<DeviceType> policy;
 
     if (_team_size < 1)

diff --git a/perf_test/batched/sparse/CG/Functor_TestBatchedTeamVectorCG_2.hpp b/perf_test/batched/sparse/CG/Functor_TestBatchedTeamVectorCG_2.hpp
@@ -120,10 +120,10 @@ struct Functor_TestBatchedTeamVectorCG_2 {
     _handle.tmp_view = typename KrylovHandleType::TemporaryViewType(
         "", _X.extent(0), 4 * _X.extent(1));
 
-    Kokkos::TeamPolicy<DeviceType> auto_policy(
-        ceil(1. * _D.extent(0) / _N_team), Kokkos::AUTO(), Kokkos::AUTO());
-    Kokkos::TeamPolicy<DeviceType> tuned_policy(
-        ceil(1. * _D.extent(0) / _N_team), _team_size, _vector_length);
+    Kokkos::TeamPolicy<DeviceType> auto_policy(_handle.get_number_of_teams(),
+                                               Kokkos::AUTO(), Kokkos::AUTO());
+    Kokkos::TeamPolicy<DeviceType> tuned_policy(_handle.get_number_of_teams(),
+                                                _team_size, _vector_length);
     Kokkos::TeamPolicy<DeviceType> policy;
 
     if (_team_size < 1)

diff --git a/perf_test/batched/sparse/CG/Functor_TestBatchedTeamVectorCG_3.hpp b/perf_test/batched/sparse/CG/Functor_TestBatchedTeamVectorCG_3.hpp
@@ -117,10 +117,10 @@ struct Functor_TestBatchedTeamVectorCG_3 {
 
     _handle.set_memory_strategy(0);
 
-    Kokkos::TeamPolicy<DeviceType> auto_policy(
-        ceil(1. * _D.extent(0) / _N_team), Kokkos::AUTO(), Kokkos::AUTO());
-    Kokkos::TeamPolicy<DeviceType> tuned_policy(
-        ceil(1. * _D.extent(0) / _N_team), _team_size, _vector_length);
+    Kokkos::TeamPolicy<DeviceType> auto_policy(_handle.get_number_of_teams(),
+                                               Kokkos::AUTO(), Kokkos::AUTO());
+    Kokkos::TeamPolicy<DeviceType> tuned_policy(_handle.get_number_of_teams(),
+                                                _team_size, _vector_length);
     Kokkos::TeamPolicy<DeviceType> policy;
 
     if (_team_size < 1)

diff --git a/perf_test/batched/sparse/CG/KokkosBatched_Test_CG.cpp b/perf_test/batched/sparse/CG/KokkosBatched_Test_CG.cpp
@@ -43,27 +43,11 @@
 
 /// Kokkos headers
 #include "Kokkos_Core.hpp"
-#include "Kokkos_Timer.hpp"
-#include "Kokkos_Random.hpp"
-#include "Kokkos_UnorderedMap.hpp"
-#include "Kokkos_Sort.hpp"
 
-/// KokkosKernels headers
+#include "Kokkos_ArithTraits.hpp"
 #include "KokkosBatched_Util.hpp"
 #include "KokkosBatched_Vector.hpp"
 
-#include <Kokkos_ArithTraits.hpp>
-#include <KokkosBatched_Util.hpp>
-#include <KokkosBatched_Vector.hpp>
-#include <KokkosBatched_Copy_Decl.hpp>
-#include <KokkosBatched_AddRadial_Decl.hpp>
-#include <KokkosBatched_Gemm_Decl.hpp>
-#include <KokkosBatched_Gemv_Decl.hpp>
-#include <KokkosBatched_Trsm_Decl.hpp>
-#include <KokkosBatched_Trsv_Decl.hpp>
-#include <KokkosBatched_LU_Decl.hpp>
-#include <KokkosSparse_CrsMatrix.hpp>
-
 #include "KokkosBatched_Test_Sparse_Helper.hpp"
 
 #include "KokkosBatched_Spmv.hpp"
@@ -128,15 +112,29 @@ int main(int argc, char *argv[]) {
             << std::endl
             << "-timers           :  Filename of the output timers."
             << std::endl
-            << "-n1               :  Number of repetitions 1." << std::endl
-            << "-n2               :  Number of repetitions 2." << std::endl
+            << "-n1               :  Number of repetitions of the experience."
+            << std::endl
+            << "-n2               :  Number of the kernel calls inside one "
+               "experience."
+            << std::endl
             << "-team_size        :  Used team size." << std::endl
             << "-n_implementations:  Number of implementations to use: test "
                "all "
                "implementations [0, specified number -1]."
             << std::endl
             << "-implementation   :  Specify only one implementation at a time."
             << std::endl
+            << "                     Note: implementation 0 : use scratch pad "
+               "only for scalar temporary variable."
+            << std::endl
+            << "                     Note: implementation 1 : use scratch pad "
+               "for scalar temporary variables and for the graph of the "
+               "matrices."
+            << std::endl
+            << "                     Note: implementation 2 : use scratch pad "
+               "for scalar and vector temporary variables and for the graph of "
+               "the matrices."
+            << std::endl
             << "-l                :  Specify left layout." << std::endl
             << "-r                :  Specify right layout." << std::endl
             << "-C                :  Specify if the convergence is monitored."
@@ -273,8 +271,6 @@ int main(int argc, char *argv[]) {
           flush.run();
           exec_space().fence();
 
-          exec_space().fence();
-
           if (i_impl == 0 && layout_left) {
             t_spmv +=
                 Functor_TestBatchedTeamVectorCG_1<exec_space,

diff --git a/perf_test/batched/sparse/GMRES/Functor_TestBatchedTeamVectorGMRES_1.hpp b/perf_test/batched/sparse/GMRES/Functor_TestBatchedTeamVectorGMRES_1.hpp
@@ -147,10 +147,10 @@ struct Functor_TestBatchedTeamVectorGMRES_1 {
     _handle.tmp_view = typename KrylovHandleType::TemporaryViewType(
         "", _X.extent(0), _X.extent(1) + maximum_iteration + 3);
 
-    Kokkos::TeamPolicy<DeviceType> auto_policy(
-        ceil(1. * _D.extent(0) / _N_team), Kokkos::AUTO(), Kokkos::AUTO());
-    Kokkos::TeamPolicy<DeviceType> tuned_policy(
-        ceil(1. * _D.extent(0) / _N_team), _team_size, _vector_length);
+    Kokkos::TeamPolicy<DeviceType> auto_policy(_handle.get_number_of_teams(),
+                                               Kokkos::AUTO(), Kokkos::AUTO());
+    Kokkos::TeamPolicy<DeviceType> tuned_policy(_handle.get_number_of_teams(),
+                                                _team_size, _vector_length);
     Kokkos::TeamPolicy<DeviceType> policy;
 
     if (_team_size < 1)

diff --git a/perf_test/batched/sparse/GMRES/Functor_TestBatchedTeamVectorGMRES_2.hpp b/perf_test/batched/sparse/GMRES/Functor_TestBatchedTeamVectorGMRES_2.hpp
@@ -169,10 +169,10 @@ struct Functor_TestBatchedTeamVectorGMRES_2 {
     Kokkos::Timer timer;
     Kokkos::Profiling::pushRegion(name.c_str());
 
-    Kokkos::TeamPolicy<DeviceType> auto_policy(
-        ceil(1. * _D.extent(0) / _N_team), Kokkos::AUTO(), Kokkos::AUTO());
-    Kokkos::TeamPolicy<DeviceType> tuned_policy(
-        ceil(1. * _D.extent(0) / _N_team), _team_size, _vector_length);
+    Kokkos::TeamPolicy<DeviceType> auto_policy(_handle.get_number_of_teams(),
+                                               Kokkos::AUTO(), Kokkos::AUTO());
+    Kokkos::TeamPolicy<DeviceType> tuned_policy(_handle.get_number_of_teams(),
+                                                _team_size, _vector_length);
     Kokkos::TeamPolicy<DeviceType> policy;
 
     if (_team_size < 1)

diff --git a/perf_test/batched/sparse/GMRES/Functor_TestBatchedTeamVectorGMRES_3.hpp b/perf_test/batched/sparse/GMRES/Functor_TestBatchedTeamVectorGMRES_3.hpp
@@ -169,10 +169,10 @@ struct Functor_TestBatchedTeamVectorGMRES_3 {
     Kokkos::Timer timer;
     Kokkos::Profiling::pushRegion(name.c_str());
 
-    Kokkos::TeamPolicy<DeviceType> auto_policy(
-        ceil(1. * _D.extent(0) / _N_team), Kokkos::AUTO(), Kokkos::AUTO());
-    Kokkos::TeamPolicy<DeviceType> tuned_policy(
-        ceil(1. * _D.extent(0) / _N_team), _team_size, _vector_length);
+    Kokkos::TeamPolicy<DeviceType> auto_policy(_handle.get_number_of_teams(),
+                                               Kokkos::AUTO(), Kokkos::AUTO());
+    Kokkos::TeamPolicy<DeviceType> tuned_policy(_handle.get_number_of_teams(),
+                                                _team_size, _vector_length);
     Kokkos::TeamPolicy<DeviceType> policy;
 
     if (_team_size < 1)

diff --git a/perf_test/batched/sparse/GMRES/KokkosBatched_Test_GMRES.cpp b/perf_test/batched/sparse/GMRES/KokkosBatched_Test_GMRES.cpp
@@ -46,39 +46,18 @@
 
 /// Kokkos headers
 #include "Kokkos_Core.hpp"
-#include "Kokkos_Timer.hpp"
-#include "Kokkos_Random.hpp"
-#include "Kokkos_UnorderedMap.hpp"
-#include "Kokkos_Sort.hpp"
 
-/// KokkosKernels headers
+#include "Kokkos_ArithTraits.hpp"
 #include "KokkosBatched_Util.hpp"
 #include "KokkosBatched_Vector.hpp"
 
-#include <Kokkos_ArithTraits.hpp>
-#include <KokkosBatched_Util.hpp>
-#include <KokkosBatched_Vector.hpp>
-#include <KokkosBatched_Copy_Decl.hpp>
-#include <KokkosBatched_AddRadial_Decl.hpp>
-#include <KokkosBatched_Gemm_Decl.hpp>
-#include <KokkosBatched_Gemv_Decl.hpp>
-#include <KokkosBatched_Trsm_Decl.hpp>
-#include <KokkosBatched_Trsv_Decl.hpp>
-#include <KokkosBatched_LU_Decl.hpp>
-#include <KokkosSparse_CrsMatrix.hpp>
-
 #include "KokkosBatched_Test_Sparse_Helper.hpp"
 
 #include "KokkosBatched_Spmv.hpp"
 #include "KokkosBatched_CrsMatrix.hpp"
 #include "KokkosBatched_Krylov_Handle.hpp"
 #include "KokkosBatched_GMRES.hpp"
 #include "KokkosBatched_JacobiPrec.hpp"
-#include "KokkosBatched_Dot.hpp"
-#include "KokkosBatched_Util.hpp"
-#include "KokkosBatched_Dot_Internal.hpp"
-#include "KokkosBatched_Spmv_Serial_Impl.hpp"
-#include "KokkosBatched_Copy_Decl.hpp"
 
 typedef Kokkos::DefaultExecutionSpace exec_space;
 typedef typename exec_space::memory_space memory_space;
@@ -150,15 +129,28 @@ int main(int argc, char *argv[]) {
             << "-other_level      :  Select the scratch pad level (if used) "
                "where everything except the Arnoldi vectors are stored."
             << std::endl
-            << "-n1               :  Number of repetitions 1." << std::endl
-            << "-n2               :  Number of repetitions 2." << std::endl
+            << "-n1               :  Number of repetitions of the experience."
+            << std::endl
+            << "-n2               :  Number of the kernel calls inside one "
+               "experience."
+            << std::endl
             << "-team_size        :  Used team size." << std::endl
             << "-n_implementations:  Number of implementations to use: test "
                "all "
                "implementations [0, specified number -1]."
             << std::endl
             << "-implementation   :  Specify only one implementation at a time."
             << std::endl
+            << "                     Note: implementation 0 : does not use "
+               "scratch pad."
+            << std::endl
+            << "                     Note: implementation 1 : use scratch pad "
+               "for the graph and for the diagonal entries of the matrices."
+            << std::endl
+            << "                     Note: implementation 2 : use scratch pad "
+               "for the graph and for the diagonal entries of the matrices and "
+               "for the temporary variable but not for the Arnoldi vectors."
+            << std::endl
             << "-l                :  Specify left layout." << std::endl
             << "-r                :  Specify right layout." << std::endl
             << "-P                :  Specify if a Jacobi preconditioner is "
@@ -318,8 +310,6 @@ int main(int argc, char *argv[]) {
           flush.run();
           exec_space().fence();
 
-          exec_space().fence();
-
           if (i_impl == 0 && layout_left) {
             if (use_preconditioner)
               t_spmv += Functor_TestBatchedTeamVectorGMRES_1<

diff --git a/perf_test/batched/sparse/KokkosBatched_Test_Sparse_Helper.hpp b/perf_test/batched/sparse/KokkosBatched_Test_Sparse_Helper.hpp
@@ -105,21 +105,6 @@ void readArrayFromMM(std::string name, const XType &x) {
   input.close();
 
   Kokkos::deep_copy(x, x_h);
-
-  /*
-    std::ofstream myfile;
-    myfile.open("x-data.txt");
-
-
-    for (size_t i = 0; i < x_h.extent(0); ++i) {
-      for (size_t j = 0; j < x_h.extent(1); ++j) {
-        myfile << std::setprecision (15) << x_h(i, j) << " ";
-      }
-      myfile << std::endl;
-    }
-
-    myfile.close();
-    */
 }
 
 template <class AType>
@@ -195,21 +180,6 @@ void readCRSFromMM(std::string name, const VType &V, const IntType &r,
   Kokkos::deep_copy(V, V_h);
   Kokkos::deep_copy(r, r_h);
   Kokkos::deep_copy(c, c_h);
-
-  /*
-    std::ofstream myfile;
-    myfile.open("a-data.txt");
-
-
-    for (size_t i = 0; i < nrows; ++i) {
-      for (size_t j = r_h(i); j < r_h(i+1); ++j) {
-        myfile << std::setprecision (15) << i+1 << " " << c_h(j)+1 << " " <<
-    V_h(0, j) << std::endl;
-      }
-    }
-
-    myfile.close();
-    */
 }
 
 template <class VType, class IntType>
@@ -238,17 +208,4 @@ void getInvDiagFromCRS(const VType &V, const IntType &r, const IntType &c,
   }
 
   Kokkos::deep_copy(diag, diag_values_host);
-
-  /*
-    std::ofstream myfile;
-    myfile.open("a-diag.txt");
-
-
-    for (size_t i = 0; i < BlkSize; ++i) {
-      myfile << std::setprecision (15) << i+1 << " " << diag_values_host(0, i)
-    << std::endl;
-    }
-
-    myfile.close();
-    */
 }
diff --git a/perf_test/batched/sparse/SPMV/KokkosBatched_Test_SPMV.cpp b/perf_test/batched/sparse/SPMV/KokkosBatched_Test_SPMV.cpp
@@ -44,30 +44,12 @@
 
 /// Kokkos headers
 #include "Kokkos_Core.hpp"
-#include "Kokkos_Timer.hpp"
-#include "Kokkos_Random.hpp"
-#include "Kokkos_UnorderedMap.hpp"
-#include "Kokkos_Sort.hpp"
-
-/// KokkosKernels headers
+#include "Kokkos_ArithTraits.hpp"
 #include "KokkosBatched_Util.hpp"
-#include "KokkosBatched_Vector.hpp"
-
-#include <Kokkos_ArithTraits.hpp>
-#include <KokkosBatched_Util.hpp>
-#include <KokkosBatched_Vector.hpp>
-#include <KokkosBatched_Copy_Decl.hpp>
-#include <KokkosBatched_AddRadial_Decl.hpp>
-#include <KokkosBatched_Gemm_Decl.hpp>
-#include <KokkosBatched_Gemv_Decl.hpp>
-#include <KokkosBatched_Trsm_Decl.hpp>
-#include <KokkosBatched_Trsv_Decl.hpp>
-#include <KokkosBatched_LU_Decl.hpp>
-#include <KokkosSparse_CrsMatrix.hpp>
-#include <KokkosBatched_SPMV_View.hpp>
 
 #include "KokkosBatched_Test_Sparse_Helper.hpp"
 
+#include "KokkosBatched_SPMV_View.hpp"
 #include "KokkosBatched_Spmv.hpp"
 
 typedef Kokkos::DefaultExecutionSpace exec_space;
@@ -106,9 +88,6 @@ struct Functor_TestBatchedTeamVectorSpmv {
 
   template <typename MemberType>
   KOKKOS_INLINE_FUNCTION void operator()(const MemberType &member) const {
-    // int team_size = member.team_size();
-    // printf("team_size %d\n", team_size);
-    // std::cout << "member.team_size() = " << member.team_size() << std::endl;
     const int first_matrix =
         static_cast<int>(member.league_rank()) * _matrices_per_team;
     const int N = _D.extent(0);
@@ -209,15 +188,36 @@ int main(int argc, char *argv[]) {
             << std::endl
             << "-timers           :  Filename of the output timers."
             << std::endl
-            << "-n1               :  Number of repetitions 1." << std::endl
-            << "-n2               :  Number of repetitions 2." << std::endl
+            << "-n1               :  Number of repetitions of the experience."
+            << std::endl
+            << "-n2               :  Number of the kernel calls inside one "
+               "experience."
+            << std::endl
             << "-team_size        :  Used team size." << std::endl
             << "-n_implementations:  Number of implementations to use: test "
                "all "
                "implementations [0, specified number -1]."
             << std::endl
             << "-implementation   :  Specify only one implementation at a time."
             << std::endl
+            << "                     Note: implementation 0 : use a Team "
+               "approach where a Team have to apply N_team SPMV. A given team "
+               "applies N_team SPMV sequentially and uses a ThreadRange over "
+               "the row and a VectorRange over the non zero entries of a given "
+               "row."
+            << std::endl
+            << "                     Note: implementation 1 : use a Team "
+               "approach where a Team have to apply N_team SPMV. A given team "
+               "uses a fused thread vector range policy to loop over the "
+               "independent fibers."
+            << std::endl
+            << "                     Note: implementation 2 : same as "
+               "implementation 1 but using scratch pad for the graph."
+            << std::endl
+            << "                     Note: implementation 3 : same as "
+               "implementation 1 but using the kernels from "
+               "batched/sparse/impl/*."
+            << std::endl
             << "-l                :  Specify left layout." << std::endl
             << "-r                :  Specify right layout." << std::endl
             << "-N_team           :  Specify the number of systems per team."
@@ -348,9 +348,8 @@ int main(int argc, char *argv[]) {
           timer.reset();
           exec_space().fence();
 
-          int N_team          = i_impl == 0 ? 1 : N_team_potential;
-          N_team              = N_team_potential;
-          int number_of_teams = i_impl == 0 ? N : ceil(1. * N / N_team);
+          int N_team          = N_team_potential;
+          int number_of_teams = ceil(static_cast<double>(N) / N_team);
 
           if (layout_left) {
             using policy_type = Kokkos::TeamPolicy<exec_space>;