sorting: add to binsort support for strided views and reorg tests (ko…

…kkos#6081) * binsort: support strided views, reorg tests [skip ci] * should be working for cuda * cleanup sorting and random tests * cleanup and reorg tests for random and sort * fixes for CI, remove unused guards * update * fix integer type and makefile * fix tests to use hpp, and fix template for windows * formatting * fix typo * rename * use assert_true (sorry DLG for the CI troubles) * fix for nvhpc * address comments * remove wrong closing bracket
masterleinad · May 25, 2023 · b3f9f78 · b3f9f78
1 parent 2a5c949
commit b3f9f78
Show file tree

Hide file tree

Showing 13 changed files with 656 additions and 509 deletions.
diff --git a/algorithms/src/Kokkos_Sort.hpp b/algorithms/src/Kokkos_Sort.hpp
@@ -146,8 +146,12 @@ class BinSort {
         Kokkos::is_view<SrcViewType>::value,
         Kokkos::View<typename SrcViewType::const_data_type,
                      typename SrcViewType::array_layout,
-                     typename SrcViewType::device_type,
-                     Kokkos::MemoryTraits<Kokkos::RandomAccess> >,
+                     typename SrcViewType::device_type
+#if !defined(KOKKOS_COMPILER_NVHPC)  // FIXME_NVHPC
+                     ,
+                     Kokkos::MemoryTraits<Kokkos::RandomAccess>
+#endif
+                     >,
         typename SrcViewType::const_type>;
 
     using perm_view_type = typename PermuteViewType::const_type;
@@ -348,18 +352,16 @@ class BinSort {
         "The provided execution space must be able to access the memory space "
         "of the View argument!");
 
-    using scratch_view_type =
-        Kokkos::View<typename ValuesViewType::data_type,
-                     typename ValuesViewType::array_layout,
-                     typename ValuesViewType::device_type>;
-
     const size_t len        = range_end - range_begin;
     const size_t values_len = values_range_end - values_range_begin;
     if (len != values_len) {
       Kokkos::abort(
           "BinSort::sort: values range length != permutation vector length");
     }
 
+    using scratch_view_type =
+        Kokkos::View<typename ValuesViewType::data_type,
+                     typename ValuesViewType::device_type>;
     scratch_view_type sorted_values(
         view_alloc(exec, WithoutInitializing,
                    "Kokkos::SortImpl::BinSortFunctor::sorted_values"),

diff --git a/algorithms/unit_tests/CMakeLists.txt b/algorithms/unit_tests/CMakeLists.txt
@@ -16,35 +16,45 @@ foreach(Tag Threads;Serial;OpenMP;Cuda;HPX;HIP;SYCL;OpenMPTarget)
     set(dir ${CMAKE_CURRENT_BINARY_DIR}/${dir})
     file(MAKE_DIRECTORY ${dir})
 
-    # -------------------------
-    # Sort1d,3d, Random
-    # -------------------------
-    set(SOURCES_A)
-    if(Tag STREQUAL "OpenMP")
-      LIST(APPEND SOURCES_A
-	TestOpenMP_Sort1D.cpp
-	TestOpenMP_Sort3D.cpp
-	TestOpenMP_SortDynamicView.cpp
-	)
-    endif()
-
+    # ------------------------------------------
+    # Sort
+    # ------------------------------------------
     # Each of these inputs is an .hpp file.
     # Generate a .cpp file for each one that runs it on the current backend (Tag),
     # and add this .cpp file to the sources for UnitTest_RandomAndSort.
-    foreach(SOURCES_A_Input
-        TestRandomCommon
-        TestSortCommon
-        TestNestedSort
-    )
-      set(file ${dir}/${SOURCES_A_Input}.cpp)
+    set(ALGO_SORT_SOURCES)
+    foreach(SOURCE_Input
+	TestSort
+	TestBinSortA
+	TestBinSortB
+	TestNestedSort
+      )
+      set(file ${dir}/${SOURCE_Input}.cpp)
       # Write to a temporary intermediate file and call configure_file to avoid
       # updating timestamps triggering unnecessary rebuilds on subsequent cmake runs.
       file(WRITE ${dir}/dummy.cpp
         "#include <Test${Tag}_Category.hpp>\n"
-        "#include <${SOURCES_A_Input}.hpp>\n"
+        "#include <${SOURCE_Input}.hpp>\n"
+        )
+      configure_file(${dir}/dummy.cpp ${file})
+      list(APPEND ALGO_SORT_SOURCES ${file})
+    endforeach()
+
+    # ------------------------------------------
+    # Random
+    # ------------------------------------------
+    # do as above
+    set(ALGO_RANDOM_SOURCES)
+    foreach(SOURCE_Input
+	TestRandom
+      )
+      set(file ${dir}/${SOURCE_Input}.cpp)
+      file(WRITE ${dir}/dummy.cpp
+        "#include <Test${Tag}_Category.hpp>\n"
+        "#include <${SOURCE_Input}.hpp>\n"
         )
       configure_file(${dir}/dummy.cpp ${file})
-      list(APPEND SOURCES_A ${file})
+      list(APPEND ALGO_RANDOM_SOURCES ${file})
     endforeach()
 
     # ------------------------------------------
@@ -145,6 +155,26 @@ foreach(Tag Threads;Serial;OpenMP;Cuda;HPX;HIP;SYCL;OpenMPTarget)
   endif()
 endforeach()
 
+# FIXME_OPENMPTARGET This test causes internal compiler errors as of 09/01/22
+# when compiling for Intel's Xe-HP GPUs.
+# FRIZZI: 04/26/2023: not sure if the compilation error is still applicable
+# but we conservatively leave this guard on
+if(NOT (KOKKOS_ENABLE_OPENMPTARGET AND KOKKOS_CXX_COMPILER_ID STREQUAL IntelLLVM))
+  KOKKOS_ADD_EXECUTABLE_AND_TEST(
+    UnitTest_Sort
+    SOURCES
+    UnitTestMain.cpp
+    ${ALGO_SORT_SOURCES}
+  )
+
+  KOKKOS_ADD_EXECUTABLE_AND_TEST(
+    UnitTest_Random
+    SOURCES
+    UnitTestMain.cpp
+    ${ALGO_RANDOM_SOURCES}
+  )
+endif()
+
 # FIXME_OPENMPTARGET These tests cause internal compiler errors as of 09/01/22
 # when compiling for Intel's Xe-HP GPUs.
 if(KOKKOS_ENABLE_OPENMPTARGET AND KOKKOS_CXX_COMPILER_ID STREQUAL IntelLLVM)
@@ -160,17 +190,6 @@ if(KOKKOS_ENABLE_OPENMPTARGET AND KOKKOS_CXX_COMPILER_ID STREQUAL IntelLLVM)
   )
 endif()
 
-# FIXME_OPENMPTARGET This test causes internal compiler errors as of 09/01/22
-# when compiling for Intel's Xe-HP GPUs.
-if(NOT (KOKKOS_ENABLE_OPENMPTARGET AND KOKKOS_CXX_COMPILER_ID STREQUAL IntelLLVM))
-  KOKKOS_ADD_EXECUTABLE_AND_TEST(
-    AlgorithmsUnitTest_RandomAndSort
-    SOURCES
-      UnitTestMain.cpp
-      ${SOURCES_A}
-  )
-endif()
-
 foreach(ID A;B;C;D;E)
   KOKKOS_ADD_EXECUTABLE_AND_TEST(
     AlgorithmsUnitTest_StdSet_${ID}

diff --git a/algorithms/unit_tests/Makefile b/algorithms/unit_tests/Makefile
@@ -27,10 +27,8 @@ TARGETS =
 
 tmp := $(foreach device, $(KOKKOS_DEVICELIST), \
   $(if $(filter Test$(device).cpp, $(shell ls Test$(device).cpp 2>/dev/null)),,\
-    $(shell echo "\#include <Test"${device}"_Category.hpp>" > Test$(device).cpp); \
-    $(shell echo "\#include <TestRandomCommon.hpp>" >> Test$(device).cpp); \
-    $(shell echo "\#include <TestSortCommon.hpp>" >> Test$(device).cpp); \
-  ) \
+     $(shell echo "\#include <Test"${device}"_Category.hpp>" > Test$(device).cpp); \
+   ) \
 )
 
 ifeq ($(KOKKOS_INTERNAL_USE_CUDA), 1)
@@ -52,7 +50,7 @@ ifeq ($(KOKKOS_INTERNAL_USE_THREADS), 1)
 endif
 
 ifeq ($(KOKKOS_INTERNAL_USE_OPENMP), 1)
-	OBJ_OPENMP = TestOpenMP.o TestOpenMP_Sort1D.o TestOpenMP_Sort3D.o TestOpenMP_SortDynamicView.o UnitTestMain.o gtest-all.o
+	OBJ_OPENMP = TestOpenMP.o UnitTestMain.o gtest-all.o
 	TARGETS += KokkosAlgorithms_UnitTest_OpenMP
 	TEST_TARGETS += test-openmp
 endif