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

Update Kokkos library in LAMMPS to v4.1.0

Browse Source
This commit is contained in:
Stan Gerald Moore
2023-06-29 10:42:42 -06:00
parent 170173a213
commit 330107b77b
480 changed files with 24051 additions and 23393 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
lib/kokkos/core/unit_test/incremental/Test05_ParallelReduce_RangePolicy.hpp
View File

@ -28,8 +28,11 @@ using value_type = double;
constexpr double value = 0.5;
struct ReduceFunctor {
KOKKOS_INLINE_FUNCTION
void operator()(const int i, double &UpdateSum) const {
// The functor is templated on purpose to check that the value_type deduction
// in parallel_reduce even works in this case.
template <typename IndexType, typename ValueType>
KOKKOS_INLINE_FUNCTION void operator()(const IndexType i,
ValueType &UpdateSum) const {
UpdateSum += (i + 1) * value;
}
};
@ -45,6 +48,7 @@ struct NonTrivialReduceFunctor {
NonTrivialReduceFunctor(NonTrivialReduceFunctor &&) = default;
NonTrivialReduceFunctor &operator=(NonTrivialReduceFunctor &&) = default;
NonTrivialReduceFunctor &operator=(NonTrivialReduceFunctor const &) = default;
// Also make sure that it's OK if the destructor is not device-callable.
~NonTrivialReduceFunctor() {}
};

4
lib/kokkos/core/unit_test/incremental/Test12a_ThreadScratch.hpp
View File

@ -98,6 +98,10 @@ struct ThreadScratch {
TEST(TEST_CATEGORY, IncrTest_12a_ThreadScratch) {
ThreadScratch<TEST_EXECSPACE> test;
#ifdef KOKKOS_ENABLE_OPENACC // FIXME_OPENACC
GTEST_SKIP() << "skipping since scratch memory is not yet implemented in the "
"OpenACC backend";
#endif
// FIXME_OPENMPTARGET - team_size has to be a multiple of 32 for the tests to
// pass in the Release and RelWithDebInfo builds. Does not need the team_size
// to be a multiple of 32 for the Debug builds.

4
lib/kokkos/core/unit_test/incremental/Test12b_TeamScratch.hpp
View File

@ -88,6 +88,10 @@ struct TeamScratch {
TEST(TEST_CATEGORY, IncrTest_12b_TeamScratch) {
TeamScratch<TEST_EXECSPACE> test;
#ifdef KOKKOS_ENABLE_OPENACC // FIXME_OPENACC
GTEST_SKIP() << "skipping since scratch memory is not yet implemented in the "
"OpenACC backend";
#endif
// FIXME_OPENMPTARGET - team_size has to be a multiple of 32 for the tests to
// pass in the Release and RelWithDebInfo builds. Does not need the team_size
// to be a multiple of 32 for the Debug builds.

21
lib/kokkos/core/unit_test/incremental/Test14_MDRangeReduce.hpp
View File

@ -38,6 +38,13 @@ struct MyComplex {
KOKKOS_INLINE_FUNCTION
MyComplex(const MyComplex& src) : _re(src._re), _im(src._im) {}
KOKKOS_INLINE_FUNCTION
MyComplex& operator=(const MyComplex& src) {
_re = src._re;
_im = src._im;
return *this;
}
KOKKOS_INLINE_FUNCTION
void operator+=(const MyComplex& src) {
_re += src._re;
@ -93,6 +100,8 @@ struct TestMDRangeReduce {
},
d_result);
// FIXME_OPENACC: scalar reduction variable on the device is not yet supported.
#if !defined(KOKKOS_ENABLE_OPENACC)
// Parallel reduce on a view.
Kokkos::parallel_reduce(
mdPolicy_2D,
@ -100,16 +109,23 @@ struct TestMDRangeReduce {
update_value += d_data(i, j);
},
d_resultView);
#endif
// Check correctness.
ASSERT_EQ(h_result, d_result);
// FIXME_OPENACC: scalar reduction variable on the device is not yet supported.
#if !defined(KOKKOS_ENABLE_OPENACC)
// Copy view back to host.
value_type view_result = 0.0;
Kokkos::deep_copy(view_result, d_resultView);
ASSERT_EQ(h_result, view_result);
#endif
}
// FIXME_OPENACC: custom reductions are not yet supported in the
// OpenACC backend.
#if !defined(KOKKOS_ENABLE_OPENACC)
// Custom Reduction
void reduce_custom() {
Complex_View_1D d_data("complex array", N);
@ -136,6 +152,7 @@ struct TestMDRangeReduce {
ASSERT_EQ(result._re, sum * 0.5);
ASSERT_EQ(result._im, -sum * 0.5);
}
#endif
};
// Reductions tests for MDRange policy and customized reduction.
@ -144,9 +161,13 @@ TEST(TEST_CATEGORY, incr_14_MDrangeReduce) {
test.reduce_MDRange();
// FIXME_OPENMPTARGET: custom reductions are not yet supported in the
// OpenMPTarget backend.
// FIXME_OPENACC: custom reductions are not yet supported in the
// OpenACC backend.
#if !defined(KOKKOS_ENABLE_OPENMPTARGET)
#if !defined(KOKKOS_ENABLE_OPENACC)
test.reduce_custom();
#endif
#endif
}
} // namespace Test

104
lib/kokkos/core/unit_test/incremental/Test16_ParallelScan.hpp
View File

@ -26,11 +26,60 @@ namespace Test {
using value_type = double;
const int N = 10;
template <typename ExecSpace>
struct TrivialScanFunctor {
Kokkos::View<value_type *, ExecSpace> d_data;
KOKKOS_FUNCTION
void operator()(const int i, value_type &update_value,
const bool final) const {
const value_type val_i = d_data(i);
if (final) d_data(i) = update_value;
update_value += val_i;
}
};
template <typename ExecSpace>
struct NonTrivialScanFunctor {
Kokkos::View<value_type *, ExecSpace> d_data;
KOKKOS_FUNCTION
void operator()(const int i, value_type &update_value,
const bool final) const {
const value_type val_i = d_data(i);
if (final) d_data(i) = update_value;
update_value += val_i;
}
NonTrivialScanFunctor(const Kokkos::View<value_type *, ExecSpace> &data)
: d_data(data) {}
NonTrivialScanFunctor(NonTrivialScanFunctor const &) = default;
NonTrivialScanFunctor(NonTrivialScanFunctor &&) = default;
NonTrivialScanFunctor &operator=(NonTrivialScanFunctor &&) = default;
NonTrivialScanFunctor &operator=(NonTrivialScanFunctor const &) = default;
// Also make sure that it's OK if the destructor is not device-callable.
~NonTrivialScanFunctor() {}
};
template <typename ExecSpace>
struct GenericExclusiveScanFunctor {
Kokkos::View<value_type *, ExecSpace> d_data;
template <typename IndexType, typename ValueType>
KOKKOS_FUNCTION void operator()(const IndexType i, ValueType &update_value,
const bool final) const {
const ValueType val_i = d_data(i);
if (final) d_data(i) = update_value;
update_value += val_i;
}
};
template <class ExecSpace>
struct TestScan {
// 1D View of double
using View_1D = typename Kokkos::View<value_type *, ExecSpace>;
template <typename FunctorType>
void parallel_scan() {
View_1D d_data("data", N);
@ -39,15 +88,9 @@ struct TestScan {
Kokkos::RangePolicy<ExecSpace>(0, N),
KOKKOS_LAMBDA(const int i) { d_data(i) = i * 0.5; });
// Exclusive parallel_scan call.
Kokkos::parallel_scan(
Kokkos::RangePolicy<ExecSpace>(0, N),
KOKKOS_LAMBDA(const int i, value_type &update_value, const bool final) {
const value_type val_i = d_data(i);
if (final) d_data(i) = update_value;
update_value += val_i;
});
// Exclusive parallel_scan call
Kokkos::parallel_scan(Kokkos::RangePolicy<ExecSpace>(0, N),
FunctorType{d_data});
// Copy back the data.
auto h_data =
@ -63,9 +106,50 @@ struct TestScan {
}
};
template <class ExecSpace>
struct TestScanWithTotal {
// 1D View of double
using View_1D = typename Kokkos::View<value_type *, ExecSpace>;
View_1D d_data = View_1D("data", N);
template <typename IndexType>
KOKKOS_FUNCTION void operator()(IndexType i) const {
d_data(i) = i * 0.5;
}
template <typename FunctorType>
void parallel_scan() {
// Initialize data.
Kokkos::parallel_for(Kokkos::RangePolicy<ExecSpace>(0, N), *this);
value_type total;
// Exclusive parallel_scan call
Kokkos::parallel_scan(Kokkos::RangePolicy<ExecSpace>(0, N),
FunctorType{d_data}, total);
// Copy back the data.
auto h_data =
Kokkos::create_mirror_view_and_copy(Kokkos::HostSpace(), d_data);
// Check Correctness
ASSERT_EQ(h_data(0), 0.0);
value_type upd = h_data(0);
for (int i = 1; i < N; ++i) {
upd += (i - 1) * 0.5;
ASSERT_EQ(h_data(i), upd);
}
ASSERT_EQ(total, N * (N - 1) * 0.25);
}
};
TEST(TEST_CATEGORY, IncrTest_16_parallelscan) {
TestScan<TEST_EXECSPACE> test;
test.parallel_scan();
test.parallel_scan<TrivialScanFunctor<TEST_EXECSPACE>>();
test.parallel_scan<NonTrivialScanFunctor<TEST_EXECSPACE>>();
TestScanWithTotal<TEST_EXECSPACE> test_total;
test_total.parallel_scan<TrivialScanFunctor<TEST_EXECSPACE>>();
test_total.parallel_scan<NonTrivialScanFunctor<TEST_EXECSPACE>>();
test_total.parallel_scan<GenericExclusiveScanFunctor<TEST_EXECSPACE>>();
}
} // namespace Test