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

Update Kokkos library in LAMMPS to v3.1

Browse Source
This commit is contained in:
Stan Moore
2020-04-16 09:06:08 -06:00
parent fa6922a182
commit ba8d043c7e
554 changed files with 24628 additions and 14871 deletions
Download Patch File Download Diff File Expand all files Collapse all files

58
lib/kokkos/core/unit_test/TestRange.hpp
View File

@ -54,9 +54,10 @@ template <class ExecSpace, class ScheduleType>
struct TestRange {
typedef int value_type; ///< typedef required for the parallel_reduce
typedef Kokkos::View<int *, ExecSpace> view_type;
typedef Kokkos::View<value_type *, ExecSpace> view_type;
view_type m_flags;
view_type result_view;
struct VerifyInitTag {};
struct ResetTag {};
@ -65,9 +66,19 @@ struct TestRange {
struct VerifyOffsetTag {};
int N;
#ifndef KOKKOS_WORKAROUND_OPENMPTARGET_GCC
static const int offset = 13;
#else
int offset;
#endif
TestRange(const size_t N_)
: m_flags(Kokkos::ViewAllocateWithoutInitializing("flags"), N_), N(N_) {}
: m_flags(Kokkos::ViewAllocateWithoutInitializing("flags"), N_),
result_view(Kokkos::ViewAllocateWithoutInitializing("results"), N_),
N(N_) {
#ifdef KOKKOS_WORKAROUND_OPENMPTARGET_GCC
offset = 13;
#endif
}
void test_for() {
typename view_type::HostMirror host_flags =
@ -185,7 +196,7 @@ struct TestRange {
//----------------------------------------
void test_reduce() {
int total = 0;
value_type total = 0;
Kokkos::parallel_for(Kokkos::RangePolicy<ExecSpace, ScheduleType>(0, N),
*this);
@ -220,15 +231,31 @@ struct TestRange {
Kokkos::parallel_for(Kokkos::RangePolicy<ExecSpace, ScheduleType>(0, N),
*this);
auto check_scan_results = [&]() {
auto const host_mirror =
Kokkos::create_mirror_view_and_copy(Kokkos::HostSpace(), result_view);
for (int i = 0; i < N; ++i) {
if (((i + 1) * i) / 2 != host_mirror(i)) {
std::cout << "Error at " << i << std::endl;
EXPECT_EQ(size_t(((i + 1) * i) / 2), size_t(host_mirror(i)));
}
}
};
Kokkos::parallel_scan(
"TestKernelScan",
Kokkos::RangePolicy<ExecSpace, ScheduleType, OffsetTag>(0, N), *this);
int total = 0;
check_scan_results();
value_type total = 0;
Kokkos::parallel_scan(
"TestKernelScanWithTotal",
Kokkos::RangePolicy<ExecSpace, ScheduleType, OffsetTag>(0, N), *this,
total);
check_scan_results();
ASSERT_EQ(size_t((N - 1) * (N) / 2), size_t(total)); // sum( 0 .. N-1 )
}
@ -239,16 +266,16 @@ struct TestRange {
if (final) {
if (update != (i * (i + 1)) / 2) {
printf("TestRange::test_scan error %d : %d != %d\n", i,
(i * (i + 1)) / 2, m_flags(i));
printf("TestRange::test_scan error (%d,%d) : %d != %d\n", i, m_flags(i),
(i * (i + 1)) / 2, update);
}
result_view(i) = update;
}
}
void test_dynamic_policy() {
auto const N_no_implicit_capture = N;
#if defined(KOKKOS_ENABLE_CXX11_DISPATCH_LAMBDA)
#if !defined(KOKKOS_ENABLE_CUDA) || (8000 <= CUDA_VERSION)
auto const N_no_implicit_capture = N;
typedef Kokkos::RangePolicy<ExecSpace, Kokkos::Schedule<Kokkos::Dynamic> >
policy_t;
@ -273,7 +300,7 @@ struct TestRange {
int error = 0;
Kokkos::parallel_reduce(
Kokkos::RangePolicy<ExecSpace>(0, N),
KOKKOS_LAMBDA(const int &i, int &lsum) {
KOKKOS_LAMBDA(const int &i, value_type &lsum) {
lsum += (a(i) != (i < N_no_implicit_capture / 2 ? 1 : 10000));
},
error);
@ -300,10 +327,10 @@ struct TestRange {
count("Count", ExecSpace::concurrency());
Kokkos::View<int *, ExecSpace> a("A", N);
int sum = 0;
value_type sum = 0;
Kokkos::parallel_reduce(
policy_t(0, N),
KOKKOS_LAMBDA(const int &i, int &lsum) {
KOKKOS_LAMBDA(const int &i, value_type &lsum) {
for (int k = 0; k < (i < N_no_implicit_capture / 2 ? 1 : 10000);
k++) {
a(i)++;
@ -321,7 +348,7 @@ struct TestRange {
int error = 0;
Kokkos::parallel_reduce(
Kokkos::RangePolicy<ExecSpace>(0, N),
KOKKOS_LAMBDA(const int &i, int &lsum) {
KOKKOS_LAMBDA(const int &i, value_type &lsum) {
lsum += (a(i) != (i < N_no_implicit_capture / 2 ? 1 : 10000));
},
error);
@ -342,7 +369,6 @@ struct TestRange {
//}
}
}
#endif
#endif
}
};
@ -417,7 +443,7 @@ TEST(TEST_CATEGORY, range_scan) {
TestRange<TEST_EXECSPACE, Kokkos::Schedule<Kokkos::Dynamic> > f(0);
f.test_scan();
}
#if !defined(KOKKOS_ENABLE_CUDA) && !defined(KOKKOS_ENABLE_ROCM)
#if !defined(KOKKOS_ENABLE_CUDA) && !defined(KOKKOS_ENABLE_HIP)
{
TestRange<TEST_EXECSPACE, Kokkos::Schedule<Kokkos::Dynamic> > f(0);
f.test_dynamic_policy();
@ -432,7 +458,7 @@ TEST(TEST_CATEGORY, range_scan) {
TestRange<TEST_EXECSPACE, Kokkos::Schedule<Kokkos::Dynamic> > f(3);
f.test_scan();
}
#if !defined(KOKKOS_ENABLE_CUDA) && !defined(KOKKOS_ENABLE_ROCM)
#if !defined(KOKKOS_ENABLE_CUDA) && !defined(KOKKOS_ENABLE_HIP)
{
TestRange<TEST_EXECSPACE, Kokkos::Schedule<Kokkos::Dynamic> > f(3);
f.test_dynamic_policy();
@ -447,7 +473,7 @@ TEST(TEST_CATEGORY, range_scan) {
TestRange<TEST_EXECSPACE, Kokkos::Schedule<Kokkos::Dynamic> > f(1001);
f.test_scan();
}
#if !defined(KOKKOS_ENABLE_CUDA) && !defined(KOKKOS_ENABLE_ROCM)
#if !defined(KOKKOS_ENABLE_CUDA) && !defined(KOKKOS_ENABLE_HIP)
{
TestRange<TEST_EXECSPACE, Kokkos::Schedule<Kokkos::Dynamic> > f(1001);
f.test_dynamic_policy();