Kokkos promotion

packages/amesos2/src/Amesos2_Tacho_decl.hpp

@@ -96,6 +96,7 @@ class TachoSolver : public SolverCore<Amesos2::TachoSolver, Matrix, Vector>
   typedef Tacho::ordinal_type                                  ordinal_type;
   typedef Tacho::size_type                                        size_type;
   typedef Kokkos::DefaultHostExecutionSpace                   HostSpaceType;
+  typedef Kokkos::TaskScheduler<HostSpaceType>                SchedulerType;
   typedef Kokkos::View<size_type*,HostSpaceType>            size_type_array;
   typedef Kokkos::View<ordinal_type*,HostSpaceType>      ordinal_type_array;
   typedef Kokkos::View<tacho_type*, HostSpaceType>         value_type_array;
@@ -203,7 +204,7 @@ class TachoSolver : public SolverCore<Amesos2::TachoSolver, Matrix, Vector>
 
   // struct holds all data necessary to make a tacho factorization or solve call
   mutable struct TACHOData {
-    typename Tacho::Solver<tacho_type,HostSpaceType> solver;
+    typename Tacho::Solver<tacho_type,SchedulerType> solver;
 
     // TODO: Implement the paramter options - confirm which we want and which have been implemented
     // int num_kokkos_threads;
@@ -224,7 +225,7 @@ class TachoSolver : public SolverCore<Amesos2::TachoSolver, Matrix, Vector>
 #else
   typedef Kokkos::Serial DeviceSpaceType;
 #endif
-  typedef typename Tacho::Solver<tacho_type,DeviceSpaceType>::value_type_matrix
+  typedef typename Tacho::Solver<tacho_type,SchedulerType>::value_type_matrix
     solve_array_t;
 
   // used as an internal workspace - possibly we can store this better in TACHOData

packages/ifpack2/src/Ifpack2_BlockTriDiContainer_impl.hpp

@@ -127,8 +127,8 @@ namespace KB = KokkosBatched::Experimental;
     using do_not_initialize_tag = Kokkos::ViewAllocateWithoutInitializing;
 
     template <typename MemoryTraitsType, Kokkos::MemoryTraitsFlags flag>
-    using MemoryTraits = Kokkos::MemoryTraits<MemoryTraitsType::Unmanaged |
-                                              MemoryTraitsType::RandomAccess |
+    using MemoryTraits = Kokkos::MemoryTraits<MemoryTraitsType::is_unmanaged |
+                                              MemoryTraitsType::is_random_access |
                                               flag>;
 
     template <typename ViewType>

packages/ifpack2/src/Ifpack2_LocalSparseTriangularSolver_def.hpp

@@ -432,7 +432,7 @@ initialize ()
     decltype(val)                           newval ("val", val.extent (0));
 
     // FIXME: The code below assumes UVM
-    crs_matrix_type::execution_space::fence();
+    typename crs_matrix_type::execution_space().fence();
     newptr(0) = 0;
     for (local_ordinal_type row = 0, rowStart = 0; row < numRows; ++row) {
       auto A_r = Alocal.row(numRows-1 - row);
@@ -445,7 +445,7 @@ initialize ()
       rowStart += numEnt;
       newptr(row+1) = rowStart;
     }
-    crs_matrix_type::execution_space::fence();
+    typename crs_matrix_type::execution_space().fence();
 
     // Reverse maps
     using map_type = typename crs_matrix_type::map_type;

packages/intrepid2/perf-test/ComputeBasis/test_hgrad.hpp

@@ -183,7 +183,7 @@ namespace Intrepid2 {
 
             flush.run();
 
-            DeviceSpaceType::fence();
+            DeviceSpaceType().fence();
             timer.reset();
 
             cts::setJacobian(jacobian, refPoints, worksetCells, cellTopo);
@@ -197,7 +197,7 @@ namespace Intrepid2 {
             fts::HGRADtransformGRAD(phyBasisGrads, jacobianInv, refBasisGrads);
             fts::multiplyMeasure(weightedBasisGrads, cellMeasure, phyBasisGrads);
 
-            DeviceSpaceType::fence();
+            DeviceSpaceType().fence();
             t_horizontal += (iwork >= 0)*timer.seconds();
           }
         }
@@ -226,12 +226,12 @@ namespace Intrepid2 {
           for (ordinal_type iwork=ibegin;iwork<nworkset;++iwork) {
             flush.run();
 
-            DeviceSpaceType::fence();
+            DeviceSpaceType().fence();
             timer.reset();
 
             Kokkos::parallel_for(policy, functor);
 
-            DeviceSpaceType::fence();
+            DeviceSpaceType().fence();
             t_vertical += (iwork >= 0)*timer.seconds();
           }
 

packages/intrepid2/perf-test/ComputeBasis/test_hgrad_vector.hpp

@@ -189,12 +189,12 @@ namespace Intrepid2 {
           for (ordinal_type iwork=ibegin;iwork<nworkset;++iwork) {
             flush.run();
 
-            DeviceSpaceType::fence();
+            DeviceSpaceType().fence();
             timer.reset();
 
             Kokkos::parallel_for(policy, functor);
 
-            DeviceSpaceType::fence();
+            DeviceSpaceType().fence();
             t_vectorize += (iwork >= 0)*timer.seconds();
           }
 

packages/intrepid2/perf-test/DynRankView/test_01.hpp

@@ -191,13 +191,13 @@ namespace Intrepid2 {
 
           Kokkos::deep_copy(in, 1.0);                          
 
-          DeviceSpaceType::fence();
+          DeviceSpaceType().fence();
           timer.reset();
 
           for (ordinal_type i=0;i<nworkset;++i) 
             Kokkos::parallel_for( policy, FunctorType(out, in) );
 
-          DeviceSpaceType::fence();
+          DeviceSpaceType().fence();
           t_view[itest] = timer.seconds();
         }
 
@@ -214,13 +214,13 @@ namespace Intrepid2 {
           Kokkos::deep_copy(in, 1.0);                          
 
 
-          DeviceSpaceType::fence();
+          DeviceSpaceType().fence();
           timer.reset();
 
           for (ordinal_type i=0;i<nworkset;++i) 
             Kokkos::parallel_for( policy, FunctorType(out, in) );
 
-          DeviceSpaceType::fence();
+          DeviceSpaceType().fence();
           t_dynrankview[itest] = timer.seconds();
         }
       } catch (std::exception err) {

packages/intrepid2/perf-test/DynRankView/test_02.hpp

@@ -180,26 +180,26 @@ namespace Intrepid2 {
           Kokkos::deep_copy(in, 1.0);                          
           {
             *verboseStream << " -> with subview \n";
-            DeviceSpaceType::fence();
+            DeviceSpaceType().fence();
             timer.reset();
 
             typedef F_clone<ViewType,ViewType,-1> FunctorType;
             for (ordinal_type i=0;i<nworkset;++i) {
               Kokkos::parallel_for( policy, FunctorType(out, in) );
             }
-            DeviceSpaceType::fence();
+            DeviceSpaceType().fence();
             t_with_subview[itest] = timer.seconds();
           }
           {
             *verboseStream << " -> without subview \n";
-            DeviceSpaceType::fence();
+            DeviceSpaceType().fence();
             timer.reset();
 
             typedef F_clone<ViewType,ViewType,0> FunctorType;
             for (ordinal_type i=0;i<nworkset;++i) {
               Kokkos::parallel_for( policy, FunctorType(out, in) );
             }
-            DeviceSpaceType::fence();
+            DeviceSpaceType().fence();
             t_without_subview[itest] = timer.seconds();
           }
         }
@@ -210,26 +210,26 @@ namespace Intrepid2 {
           Kokkos::deep_copy(in, 1.0);                          
           {
             *verboseStream << " -> with subview \n";
-            DeviceSpaceType::fence();
+            DeviceSpaceType().fence();
             timer.reset();
 
             typedef F_clone<ViewType,ViewType,-1> FunctorType;
             for (ordinal_type i=0;i<nworkset;++i) {
               Kokkos::parallel_for( policy, FunctorType(out, in) );
             }
-            DeviceSpaceType::fence();
+            DeviceSpaceType().fence();
             t_with_subview[itest] = timer.seconds();
           }
           {
             *verboseStream << " -> without subview \n";
-            DeviceSpaceType::fence();
+            DeviceSpaceType().fence();
             timer.reset();
 
             typedef F_clone<ViewType,ViewType,1> FunctorType;
             for (ordinal_type i=0;i<nworkset;++i) {
               Kokkos::parallel_for( policy, FunctorType(out, in) );
             }
-            DeviceSpaceType::fence();
+            DeviceSpaceType().fence();
             t_without_subview[itest] = timer.seconds();
           }
         }
@@ -241,26 +241,26 @@ namespace Intrepid2 {
           Kokkos::deep_copy(in, 1.0);                          
           {
             *verboseStream << " -> with subview \n";
-            DeviceSpaceType::fence();
+            DeviceSpaceType().fence();
             timer.reset();
 
             typedef F_clone<ViewType,ViewType,-1> FunctorType;
             for (ordinal_type i=0;i<nworkset;++i) {
               Kokkos::parallel_for( policy, FunctorType(out, in) );
             }
-            DeviceSpaceType::fence();
+            DeviceSpaceType().fence();
             t_with_subview[itest] = timer.seconds();
           }
           {
             *verboseStream << " -> without subview \n";
-            DeviceSpaceType::fence();
+            DeviceSpaceType().fence();
             timer.reset();
 
             typedef F_clone<ViewType,ViewType,2> FunctorType;
             for (ordinal_type i=0;i<nworkset;++i) {
               Kokkos::parallel_for( policy, FunctorType(out, in) );
             }
-            DeviceSpaceType::fence();
+            DeviceSpaceType().fence();
             t_without_subview[itest] = timer.seconds();
           }
         }

packages/intrepid2/unit-test/Discretization/FunctionSpaceTools/test_01.hpp

@@ -250,7 +250,7 @@ namespace Intrepid2 {
                        transformed_value_of_basis_at_cub_points,
                        weighted_transformed_value_of_basis_at_cub_points);
 
-        DeviceSpaceType::fence();
+        DeviceSpaceType().fence();
 
         /*******************  STOP COMPUTATION ***********************/
 

packages/intrepid2/unit-test/Discretization/FunctionSpaceTools/test_02.hpp

@@ -285,7 +285,7 @@ namespace Intrepid2 {
         // apply field signs (after the fact, as a post-processing step)
         fst::applyLeftFieldSigns(mass_matrices, field_signs);
         fst::applyRightFieldSigns(mass_matrices, field_signs);
-        DeviceSpaceType::fence();
+        DeviceSpaceType().fence();
 
         /*******************  STOP COMPUTATION ***********************/
 

packages/kokkos-kernels/CHANGELOG.md

@@ -1,5 +1,22 @@
 # Change Log
 
+## [2.9.00](https://github.com/kokkos/kokkos-kernels/tree/2.9.00) (2019-06-24)
+[Full Changelog](https://github.com/kokkos/kokkos-kernels/compare/2.8.00...2.9.00)
+
+**Implemented enhancements:**
+
+- KokkosBatched: Add specialization for float2, float4 and double4 [\#427](https://github.com/kokkos/kokkos-kernels/pull/427)
+- KokkosBatched: Reduce VectorLength (16 to 8) [\#432](https://github.com/kokkos/kokkos-kernels/pull/432)
+- KokkosBatched: Remove experimental name space for batched blas [\#371](https://github.com/kokkos/kokkos-kernels/issues/371)
+- Capability: Initial sparse triangular solve capability [\#435](https://github.com/kokkos/kokkos-kernels/pull/435)
+- Capability: Add support for MAGMA GESV TPL [\#409](https://github.com/kokkos/kokkos-kernels/pull/409)
+- cuBLAS: Add CudaUVMSpace specializations for GEMM [\#397](https://github.com/kokkos/kokkos-kernels/issues/397)
+
+**Fixed bugs:**
+
+- Deprecated Code Fixes [\#411](https://github.com/kokkos/kokkos-kernels/issues/411)
+- BuildSystem: Compilation error on rzansel [\#401](https://github.com/kokkos/kokkos-kernels/issues/401)
+
 ## [2.8.00](https://github.com/kokkos/kokkos-kernels/tree/2.8.00) (2019-02-05)
 [Full Changelog](https://github.com/kokkos/kokkos-kernels/compare/2.7.24...2.8.00)
 

packages/kokkos-kernels/CMakeLists.txt

@@ -290,23 +290,64 @@ ENDIF()
 # Enable Third Party Libraries
 # ==================================================================
 
-IF (TPL_ENABLE_BLAS)
-  SET(KOKKOSKERNELS_ENABLE_TPL_BLAS ${TPL_ENABLE_BLAS})
+IF (DEFINED KOKKOSKERNELS_ENABLE_TPL_BLAS)
+  # user overriding kokkoskernels
+  IF (KOKKOSKERNELS_ENABLE_TPL_BLAS)
+    IF (NOT TPL_ENABLE_BLAS)
+      MESSAGE( WARNING "KOKKOSKERNELS_ENABLE_TPL_BLAS is ON but TPL_ENABLE_BLAS is OFF. Please set TPL_ENABLE_BLAS:BOOL=ON")
+      SET(KOKKOSKERNELS_ENABLE_TPL_BLAS OFF)  
+    ENDIF()
+  ENDIF()
+ELSE()
+  # default behavior
+  IF (TPL_ENABLE_BLAS)
+    SET(KOKKOSKERNELS_ENABLE_TPL_BLAS ${TPL_ENABLE_BLAS})
+  ENDIF()
 ENDIF()
-IF (TPL_ENABLE_MKL)
-  SET(KOKKOSKERNELS_ENABLE_TPL_MKL ${TPL_ENABLE_MKL})
+
+IF (DEFINED KOKKOSKERNELS_ENABLE_TPL_MKL)
+  # user overriding kokkoskernels
+  IF (KOKKOSKERNELS_ENABLE_TPL_MKL)
+    IF (NOT TPL_ENABLE_MKL)
+      MESSAGE( WARNING "KOKKOSKERNELS_ENABLE_TPL_MKL is ON but TPL_ENABLE_MKL is OFF. Please set TPL_ENABLE_MKL:BOOL=ON")
+      SET(KOKKOSKERNELS_ENABLE_TPL_MKL OFF)  
+    ENDIF()
+  ENDIF()
+ELSE()
+  IF (TPL_ENABLE_MKL)
+    SET(KOKKOSKERNELS_ENABLE_TPL_MKL ${TPL_ENABLE_MKL})
+  ENDIF()
 ENDIF()
 
 IF(${Kokkos_ENABLE_Cuda})
-  IF (NOT KOKKOSKERNELS_ENABLE_TPL_BLAS)
-    SET(KOKKOSKERNELS_ENABLE_TPL_BLAS ON)
-    LIST(APPEND TPL_LIST "BLAS")
-  ENDIF()
   # CUBLAS is ON by default when CUDA is enabled
-  SET(KOKKOSKERNELS_ENABLE_TPL_CUBLAS ON)
+  IF (NOT DEFINED KOKKOSKERNELS_ENABLE_TPL_CUBLAS)  
+    SET(KOKKOSKERNELS_ENABLE_TPL_CUBLAS ON)
+  ENDIF()
+
   # Tribit provides TPL mechanism for CUSPARSE; thus, use it
-  IF (TPL_ENABLE_CUSPARSE)
-    SET(KOKKOSKERNELS_ENABLE_TPL_CUSPARSE ${TPL_ENABLE_CUSPARSE})
+  IF (DEFINED KOKKOSKERNELS_ENABLE_TPL_CUSPARSE)  
+    IF (NOT TPL_ENABLE_CUSPARSE)
+      MESSAGE( WARNING "KOKKOSKERNELS_ENABLE_TPL_CUSPARSE is ON but TPL_ENABLE_CUSPARSE is OFF. Please set TPL_ENABLE_CUSPARSE:BOOL=ON")
+      SET(KOKKOSKERNELS_ENABLE_TPL_CUSPARSE OFF)  
+    ENDIF()
+  ELSE()
+    IF (TPL_ENABLE_CUSPARSE)
+      SET(KOKKOSKERNELS_ENABLE_TPL_CUSPARSE ${TPL_ENABLE_CUSPARSE})
+    ENDIF()
+  ENDIF()
+
+  IF (DEFINED KOKKOSKERNELS_ENABLE_TPL_MAGMA)
+    IF (KOKKOSKERNELS_ENABLE_TPL_MAGMA)  
+      IF (NOT TPL_ENABLE_MAGMA)
+        MESSAGE( WARNING "KOKKOSKERNELS_ENABLE_TPL_MAGMA is ON but TPL_ENABLE_MAGMA is OFF. Please set TPL_ENABLE_MAGMA:BOOL=ON")
+        SET(KOKKOSKERNELS_ENABLE_TPL_MAGMA OFF)  
+      ENDIF()
+    ENDIF()
+  ELSE()
+    IF (TPL_ENABLE_MAGMA)
+      SET(KOKKOSKERNELS_ENABLE_TPL_MAGMA ${TPL_ENABLE_MAGMA})
+    ENDIF()
   ENDIF()
 ENDIF()
 
@@ -322,6 +363,18 @@ ENDIF()
 IF (KOKKOSKERNELS_ENABLE_TPL_CUBLAS)
   LIST(APPEND TPL_LIST "CUBLAS")
 ENDIF()
+IF (KOKKOSKERNELS_ENABLE_TPL_MAGMA)
+  LIST(APPEND TPL_LIST "MAGMA")
+  IF (F77_BLAS_MANGLE STREQUAL "(name,NAME) name ## _")
+    SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DADD_ -fopenmp -lgfortran")
+  ELSEIF (F77_BLAS_MANGLE STREQUAL "(name,NAME) NAME")
+    SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DUPCASE -fopenmp -lgfortran")
+  ELSEIF (F77_BLAS_MANGLE STREQUAL "(name,NAME) name")
+    SET(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DNOCHANGE -fopenmp -lgfortran")
+  ELSE ()
+    MESSAGE(FATAL_ERROR "F77_BLAS_MANGLE ${F77_BLAS_MANGLE} detected while MAGMA only accepts Fortran mangling that is one of single underscore (-DADD_), uppercase (-DUPCASE), and no change (-DNOCHANGE)")
+  ENDIF()
+ENDIF()
 
 # ==================================================================
 # Fortran Complex BLAS