ICODE-ADC/lammps
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Updating Kokkos lib

Browse Source
This commit is contained in:
Stan Moore
2017-01-09 10:39:46 -07:00
parent 51fa33a407
commit a9f0b7d523
359 changed files with 20077 additions and 27456 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
lib/kokkos/core/perf_test/CMakeLists.txt
View File

@ -1,6 +1,6 @@
INCLUDE_DIRECTORIES(${CMAKE_CURRENT_BINRARY_DIR})
INCLUDE_DIRECTORIES(${CMAKE_CURRENT_SOURCE_DIR})
INCLUDE_DIRECTORIES(REQUIRED_DURING_INSTALLATION_TESTING ${CMAKE_CURRENT_SOURCE_DIR})
SET(SOURCES
PerfTestMain.cpp
@ -19,7 +19,7 @@ TRIBITS_ADD_EXECUTABLE(
TESTONLYLIBS kokkos_gtest
)
TRIBITS_ADD_EXECUTABLE_AND_TEST(
TRIBITS_ADD_TEST(
PerfTest
NAME PerfTestExec
COMM serial mpi

21
lib/kokkos/core/perf_test/Makefile
View File

@ -7,21 +7,18 @@ vpath %.cpp ${KOKKOS_PATH}/core/perf_test
default: build_all
echo "End Build"
ifneq (,$(findstring Cuda,$(KOKKOS_DEVICES)))
CXX = $(KOKKOS_PATH)/config/nvcc_wrapper
else
CXX = g++
endif
CXXFLAGS = -O3
LINK ?= $(CXX)
LDFLAGS ?= -lpthread
include $(KOKKOS_PATH)/Makefile.kokkos
ifeq ($(KOKKOS_INTERNAL_USE_CUDA), 1)
CXX = $(NVCC_WRAPPER)
CXXFLAGS ?= -O3
LINK = $(CXX)
LDFLAGS ?= -lpthread
else
CXX ?= g++
CXXFLAGS ?= -O3
LINK ?= $(CXX)
LDFLAGS ?= -lpthread
endif
KOKKOS_CXXFLAGS += -I$(GTEST_PATH) -I${KOKKOS_PATH}/core/perf_test
TEST_TARGETS =

17
lib/kokkos/core/perf_test/PerfTestHost.cpp
View File

@ -79,10 +79,21 @@ class host : public ::testing::Test {
protected:
static void SetUpTestCase()
{
const unsigned team_count = Kokkos::hwloc::get_available_numa_count();
const unsigned threads_per_team = 4 ;
if(Kokkos::hwloc::available()) {
const unsigned numa_count = Kokkos::hwloc::get_available_numa_count();
const unsigned cores_per_numa = Kokkos::hwloc::get_available_cores_per_numa();
const unsigned threads_per_core = Kokkos::hwloc::get_available_threads_per_core();
TestHostDevice::initialize( team_count * threads_per_team );
unsigned threads_count = 0 ;
threads_count = std::max( 1u , numa_count )
* std::max( 2u , cores_per_numa * threads_per_core );
TestHostDevice::initialize( threads_count );
} else {
const unsigned thread_count = 4 ;
TestHostDevice::initialize( thread_count );
}
}
static void TearDownTestCase()