Merge branch 'develop' into develop

crtrott · Oct 1, 2024 · 7e8140c · 7e8140c
2 parents 9572296 + f90a1e1
commit 7e8140c
Show file tree

Hide file tree

Showing 27 changed files with 739 additions and 87 deletions.
diff --git a/.jenkins b/.jenkins
@@ -233,7 +233,6 @@ pipeline {
                                 -DKokkos_ENABLE_DEPRECATED_CODE_4=OFF \
                                 -DKokkos_ENABLE_TESTS=ON \
                                 -DKokkos_ENABLE_CUDA=ON \
-                                -DKokkos_ENABLE_CUDA_LAMBDA=ON \
                                 -DKokkos_ENABLE_OPENMP=ON \
                               .. && \
                               make -j8 && ctest --verbose'''
@@ -430,7 +429,6 @@ pipeline {
                                 -DKokkos_ENABLE_TESTS=ON \
                                 -DKokkos_ENABLE_BENCHMARKS=ON \
                                 -DKokkos_ENABLE_CUDA=ON \
-                                -DKokkos_ENABLE_CUDA_LAMBDA=ON \
                                 -DKokkos_ENABLE_TUNING=ON \
                                 -DKokkos_ARCH_VOLTA70=ON \
                               .. && \
@@ -470,7 +468,6 @@ pipeline {
                                 -DKokkos_ENABLE_TESTS=ON \
                                 -DKokkos_ENABLE_BENCHMARKS=ON \
                                 -DKokkos_ENABLE_CUDA=ON \
-                                -DKokkos_ENABLE_CUDA_LAMBDA=ON \
                                 -DKokkos_ENABLE_LIBDL=OFF \
                                 -DKokkos_ENABLE_OPENMP=ON \
                                 -DKokkos_ENABLE_IMPL_MDSPAN=OFF \

diff --git a/Makefile.targets b/Makefile.targets
@@ -73,6 +73,8 @@ Kokkos_HIP_Space.o: $(KOKKOS_CPP_DEPENDS) $(KOKKOS_PATH)/core/src/HIP/Kokkos_HIP
 	$(CXX) $(KOKKOS_CPPFLAGS) $(KOKKOS_CXXFLAGS) $(CXXFLAGS) -c $(KOKKOS_PATH)/core/src/HIP/Kokkos_HIP_Space.cpp
 Kokkos_HIP_Instance.o: $(KOKKOS_CPP_DEPENDS) $(KOKKOS_PATH)/core/src/HIP/Kokkos_HIP_Instance.cpp
 	$(CXX) $(KOKKOS_CPPFLAGS) $(KOKKOS_CXXFLAGS) $(CXXFLAGS) -c $(KOKKOS_PATH)/core/src/HIP/Kokkos_HIP_Instance.cpp
+Kokkos_HIP_ZeroMemset.o: $(KOKKOS_CPP_DEPENDS) $(KOKKOS_PATH)/core/src/HIP/Kokkos_HIP_ZeroMemset.cpp
+	$(CXX) $(KOKKOS_CPPFLAGS) $(KOKKOS_CXXFLAGS) $(CXXFLAGS) -c $(KOKKOS_PATH)/core/src/HIP/Kokkos_HIP_ZeroMemset.cpp
 Lock_Array_HIP.o: $(KOKKOS_CPP_DEPENDS) $(KOKKOS_PATH)/tpls/desul/src/Lock_Array_HIP.cpp
 	$(CXX) $(KOKKOS_CPPFLAGS) $(KOKKOS_CXXFLAGS) $(CXXFLAGS) -c $(KOKKOS_PATH)/tpls/desul/src/Lock_Array_HIP.cpp
 endif

diff --git a/containers/unit_tests/TestDynRankViewTypedefs.cpp b/containers/unit_tests/TestDynRankViewTypedefs.cpp
@@ -167,7 +167,7 @@ constexpr bool test_view_typedefs_impl() {
   static_assert(std::is_same_v<typename ViewType::data_handle_type, typename ViewType::pointer_type>);
   static_assert(std::is_same_v<typename ViewType::reference, typename ViewType::reference_type>);
   return true;
-};
+}
 
 // Helper function to unpack data type and other args from the View, and pass them on
 template<class T, class ... ViewArgs>

diff --git a/containers/unit_tests/TestErrorReporter.hpp b/containers/unit_tests/TestErrorReporter.hpp
@@ -149,7 +149,6 @@ struct ErrorReporterDriver : public ErrorReporterDriverBase<DeviceType> {
   }
 };
 
-#if !defined(KOKKOS_ENABLE_CUDA) || defined(KOKKOS_ENABLE_CUDA_LAMBDA)
 template <typename DeviceType>
 struct ErrorReporterDriverUseLambda
     : public ErrorReporterDriverBase<DeviceType> {
@@ -178,7 +177,6 @@ struct ErrorReporterDriverUseLambda
     driver_base::check_expectations(reporter_capacity, test_size);
   }
 };
-#endif
 
 #ifdef KOKKOS_ENABLE_OPENMP
 struct ErrorReporterDriverNativeOpenMP
@@ -205,8 +203,7 @@ struct ErrorReporterDriverNativeOpenMP
 
 // FIXME_MSVC MSVC just gets confused when using the base class in the
 // KOKKOS_CLASS_LAMBDA
-#if !defined(KOKKOS_COMPILER_MSVC) && \
-    (!defined(KOKKOS_ENABLE_CUDA) || defined(KOKKOS_ENABLE_CUDA_LAMBDA))
+#ifndef KOKKOS_COMPILER_MSVC
 TEST(TEST_CATEGORY, ErrorReporterViaLambda) {
   TestErrorReporter<ErrorReporterDriverUseLambda<TEST_EXECSPACE>>();
 }

diff --git a/containers/unit_tests/TestOffsetView.hpp b/containers/unit_tests/TestOffsetView.hpp
@@ -74,7 +74,6 @@ void test_offsetview_construction() {
   ASSERT_EQ(ov.extent(0), 5u);
   ASSERT_EQ(ov.extent(1), 5u);
 
-#if defined(KOKKOS_ENABLE_CUDA_LAMBDA) || !defined(KOKKOS_ENABLE_CUDA)
   {
     Kokkos::Experimental::OffsetView<Scalar*, Device> offsetV1("OneDOffsetView",
                                                                range0);
@@ -156,7 +155,6 @@ void test_offsetview_construction() {
   }
 
   ASSERT_EQ(OVResult, answer) << "Bad data found in OffsetView";
-#endif
 
   {
     offset_view_type ovCopy(ov);
@@ -191,7 +189,6 @@ void test_offsetview_construction() {
     range3_type rangePolicy3DZero(point3_type{{0, 0, 0}},
                                   point3_type{{extent0, extent1, extent2}});
 
-#if defined(KOKKOS_ENABLE_CUDA_LAMBDA) || !defined(KOKKOS_ENABLE_CUDA)
     int view3DSum = 0;
     Kokkos::parallel_reduce(
         rangePolicy3DZero,
@@ -214,7 +211,6 @@ void test_offsetview_construction() {
 
     ASSERT_EQ(view3DSum, offsetView3DSum)
         << "construction of OffsetView from View and begins array broken.";
-#endif
   }
   view_type viewFromOV = ov.view();
 
@@ -239,7 +235,6 @@ void test_offsetview_construction() {
     view_type aView("aView", ov.extent(0), ov.extent(1));
     Kokkos::deep_copy(aView, ov);
 
-#if defined(KOKKOS_ENABLE_CUDA_LAMBDA) || !defined(KOKKOS_ENABLE_CUDA)
     int sum = 0;
     Kokkos::parallel_reduce(
         rangePolicy2D,
@@ -249,7 +244,6 @@ void test_offsetview_construction() {
         sum);
 
     ASSERT_EQ(sum, 0) << "deep_copy(view, offsetView) broken.";
-#endif
   }
 
   {  // test view to  offsetview deep copy
@@ -258,7 +252,6 @@ void test_offsetview_construction() {
     Kokkos::deep_copy(aView, 99);
     Kokkos::deep_copy(ov, aView);
 
-#if defined(KOKKOS_ENABLE_CUDA_LAMBDA) || !defined(KOKKOS_ENABLE_CUDA)
     int sum = 0;
     Kokkos::parallel_reduce(
         rangePolicy2D,
@@ -268,7 +261,6 @@ void test_offsetview_construction() {
         sum);
 
     ASSERT_EQ(sum, 0) << "deep_copy(offsetView, view) broken.";
-#endif
   }
 }
 
@@ -521,7 +513,6 @@ void test_offsetview_subview() {
       ASSERT_EQ(offsetSubview.begin(1), 0);
       ASSERT_EQ(offsetSubview.end(1), 9);
 
-#if defined(KOKKOS_ENABLE_CUDA_LAMBDA) || !defined(KOKKOS_ENABLE_CUDA)
       using range_type = Kokkos::MDRangePolicy<Device, Kokkos::Rank<2>,
                                                Kokkos::IndexType<int> >;
       using point_type = typename range_type::point_type;
@@ -547,7 +538,6 @@ void test_offsetview_subview() {
           sum);
 
       ASSERT_EQ(sum, 6 * (e0 - b0) * (e1 - b1));
-#endif
     }
 
     // slice 2
@@ -644,7 +634,6 @@ void test_offsetview_subview() {
   }
 }
 
-#if defined(KOKKOS_ENABLE_CUDA_LAMBDA) || !defined(KOKKOS_ENABLE_CUDA)
 template <class InputIt, class T, class BinaryOperation>
 KOKKOS_INLINE_FUNCTION T std_accumulate(InputIt first, InputIt last, T init,
                                         BinaryOperation op) {
@@ -748,7 +737,6 @@ void test_offsetview_offsets_rank3() {
 
   ASSERT_EQ(0, errors);
 }
-#endif
 
 TEST(TEST_CATEGORY, offsetview_construction) {
   test_offsetview_construction<int, TEST_EXECSPACE>();
@@ -767,7 +755,6 @@ TEST(TEST_CATEGORY, offsetview_subview) {
   test_offsetview_subview<int, TEST_EXECSPACE>();
 }
 
-#if defined(KOKKOS_ENABLE_CUDA_LAMBDA) || !defined(KOKKOS_ENABLE_CUDA)
 TEST(TEST_CATEGORY, offsetview_offsets_rank1) {
   test_offsetview_offsets_rank1<TEST_EXECSPACE>();
 }
@@ -779,7 +766,6 @@ TEST(TEST_CATEGORY, offsetview_offsets_rank2) {
 TEST(TEST_CATEGORY, offsetview_offsets_rank3) {
   test_offsetview_offsets_rank3<TEST_EXECSPACE>();
 }
-#endif
 
 }  // namespace Test
 

diff --git a/containers/unit_tests/TestUnorderedMap.hpp b/containers/unit_tests/TestUnorderedMap.hpp
@@ -534,8 +534,6 @@ TEST(TEST_CATEGORY, UnorderedMap_shallow_copyable_on_device) {
   ASSERT_EQ(1u, test_map_copy.m_map.size());
 }
 
-#if !defined(KOKKOS_ENABLE_CUDA) || \
-    (defined(KOKKOS_ENABLE_CUDA) && defined(KOKKOS_ENABLE_CUDA_LAMBDA))
 void test_unordered_map_device_capture() {
   TestMapCopy::map_type map;
 
@@ -549,7 +547,6 @@ void test_unordered_map_device_capture() {
 TEST(TEST_CATEGORY, UnorderedMap_lambda_capturable) {
   test_unordered_map_device_capture();
 }
-#endif
 
 /**
  * @test This test ensures that an @ref UnorderedMap can be built

diff --git a/core/perf_test/CMakeLists.txt b/core/perf_test/CMakeLists.txt
@@ -128,9 +128,7 @@ endif()
 
 kokkos_add_benchmark(PerformanceTest_Benchmark SOURCES ${BENCHMARK_SOURCES})
 
-if(NOT KOKKOS_ENABLE_CUDA OR KOKKOS_ENABLE_CUDA_LAMBDA)
-  kokkos_add_benchmark(Benchmark_Atomic_MinMax SOURCES test_atomic_minmax_simple.cpp)
-endif()
+kokkos_add_benchmark(Benchmark_Atomic_MinMax SOURCES test_atomic_minmax_simple.cpp)
 
 # FIXME_NVHPC
 if(NOT KOKKOS_CXX_COMPILER_ID STREQUAL NVHPC)

diff --git a/core/perf_test/PerfTest_ViewCopy_Raw.cpp b/core/perf_test/PerfTest_ViewCopy_Raw.cpp
@@ -18,7 +18,6 @@
 
 namespace Test {
 
-#if defined(KOKKOS_ENABLE_CUDA_LAMBDA) || !defined(KOKKOS_ENABLE_CUDA)
 BENCHMARK(ViewDeepCopy_Raw<Kokkos::LayoutLeft, Kokkos::LayoutLeft>)
     ->ArgName("N")
     ->Arg(10)
@@ -38,6 +37,5 @@ BENCHMARK(ViewDeepCopy_Raw<Kokkos::LayoutRight, Kokkos::LayoutLeft>)
     ->ArgName("N")
     ->Arg(10)
     ->UseManualTime();
-#endif
 
 }  // namespace Test
diff --git a/core/perf_test/PerfTest_ViewFill_Raw.cpp b/core/perf_test/PerfTest_ViewFill_Raw.cpp
@@ -18,7 +18,6 @@
 
 namespace Test {
 
-#if defined(KOKKOS_ENABLE_CUDA_LAMBDA) || !defined(KOKKOS_ENABLE_CUDA)
 BENCHMARK(ViewFill_Raw<Kokkos::LayoutLeft>)
     ->ArgName("N")
     ->Arg(N)
@@ -28,6 +27,5 @@ BENCHMARK(ViewFill_Raw<Kokkos::LayoutRight>)
     ->ArgName("N")
     ->Arg(N)
     ->UseManualTime();
-#endif
 
 }  // namespace Test
diff --git a/core/perf_test/PerfTest_ViewResize_Raw.cpp b/core/perf_test/PerfTest_ViewResize_Raw.cpp
@@ -18,7 +18,6 @@
 
 namespace Test {
 
-#if defined(KOKKOS_ENABLE_CUDA_LAMBDA) || !defined(KOKKOS_ENABLE_CUDA)
 BENCHMARK(ViewResize_NoInit_Raw<Kokkos::LayoutLeft>)
     ->ArgName("N")
     ->Arg(N)
@@ -30,6 +29,5 @@ BENCHMARK(ViewResize_NoInit_Raw<Kokkos::LayoutRight>)
     ->Arg(N)
     ->UseManualTime()
     ->Iterations(R);
-#endif
 
 }  // namespace Test
diff --git a/core/src/Cuda/Kokkos_Cuda_Instance.cpp b/core/src/Cuda/Kokkos_Cuda_Instance.cpp
@@ -687,16 +687,6 @@ void Cuda::print_configuration(std::ostream &os, bool /*verbose*/) const {
   os << "  KOKKOS_ENABLE_CUDA: yes\n";
 
   os << "Cuda Options:\n";
-  os << "  KOKKOS_ENABLE_CUDA_LAMBDA: ";
-#ifdef KOKKOS_ENABLE_CUDA_LAMBDA
-  os << "yes\n";
-#else
-  os << "no\n";
-#endif
-#ifdef KOKKOS_ENABLE_DEPRECATED_CODE_4
-  os << "  KOKKOS_ENABLE_CUDA_LDG_INTRINSIC: ";
-  os << "yes\n";
-#endif
   os << "  KOKKOS_ENABLE_CUDA_RELOCATABLE_DEVICE_CODE: ";
 #ifdef KOKKOS_ENABLE_CUDA_RELOCATABLE_DEVICE_CODE
   os << "yes\n";

diff --git a/core/src/HIP/Kokkos_HIP_ZeroMemset.cpp b/core/src/HIP/Kokkos_HIP_ZeroMemset.cpp
@@ -0,0 +1,36 @@
+//@HEADER
+// ************************************************************************
+//
+//                        Kokkos v. 4.0
+//       Copyright (2022) National Technology & Engineering
+//               Solutions of Sandia, LLC (NTESS).
+//
+// Under the terms of Contract DE-NA0003525 with NTESS,
+// the U.S. Government retains certain rights in this software.
+//
+// Part of Kokkos, under the Apache License v2.0 with LLVM Exceptions.
+// See https://kokkos.org/LICENSE for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//@HEADER
+
+#ifndef KOKKOS_IMPL_PUBLIC_INCLUDE
+#define KOKKOS_IMPL_PUBLIC_INCLUDE
+#endif
+
+#include <HIP/Kokkos_HIP_ZeroMemset.hpp>
+#include <HIP/Kokkos_HIP_ParallelFor_Range.hpp>
+
+namespace Kokkos {
+namespace Impl {
+
+// alternative to hipMemsetAsync, which sets the first `cnt` bytes of `dst` to 0
+void zero_with_hip_kernel(const HIP& exec_space, void* dst, size_t cnt) {
+  Kokkos::parallel_for(
+      "Kokkos::ZeroMemset via parallel_for",
+      Kokkos::RangePolicy<Kokkos::HIP>(exec_space, 0, cnt),
+      KOKKOS_LAMBDA(size_t i) { static_cast<char*>(dst)[i] = 0; });
+}
+
+}  // namespace Impl
+}  // namespace Kokkos
diff --git a/core/src/HIP/Kokkos_HIP_ZeroMemset.hpp b/core/src/HIP/Kokkos_HIP_ZeroMemset.hpp
@@ -23,11 +23,22 @@
 namespace Kokkos {
 namespace Impl {
 
+// hipMemsetAsync sets the first `cnt` bytes of `dst` to the provided value
+void zero_with_hip_kernel(const HIP& exec_space, void* dst, size_t cnt);
+
 template <>
 struct ZeroMemset<HIP> {
   ZeroMemset(const HIP& exec_space, void* dst, size_t cnt) {
+    // in ROCm <= 6.2.0, hipMemsetAsync on a host-allocated pointer
+    // returns an invalid value error, but accessing the data via a
+    // GPU kernel works.
+#if defined(KOKKOS_ENABLE_IMPL_HIP_UNIFIED_MEMORY) && \
+    defined(KOKKOS_ARCH_AMD_GFX942)
+    zero_with_hip_kernel(exec_space, dst, cnt);
+#else
     KOKKOS_IMPL_HIP_SAFE_CALL(
         hipMemsetAsync(dst, 0, cnt, exec_space.hip_stream()));
+#endif
   }
 };