Update Kokkos library in LAMMPS to v3.6.0

This commit is contained in:
Stan Gerald Moore
2022-05-05 11:44:47 -06:00
parent bd4bbbddbe
commit b79c0bc7b4
380 changed files with 41928 additions and 8786 deletions

View File

@ -712,7 +712,8 @@ class TestDynViewAPI {
using host_view_space = typename View0::host_mirror_space;
static void run_tests() {
run_test_resize_realloc();
run_test_resize_realloc<false>();
run_test_resize_realloc<true>();
run_test_mirror();
run_test_mirror_and_copy();
run_test_scalar();
@ -722,6 +723,7 @@ class TestDynViewAPI {
run_test_subview();
run_test_subview_strided();
run_test_vector();
run_test_as_view_of_rank_n();
}
static void run_operator_test_rank12345() {
@ -738,21 +740,28 @@ class TestDynViewAPI {
TestViewOperator_LeftAndRight<int, device, 6>::testit(2, 3, 4, 2, 3, 4);
}
template <bool Initialize>
static void run_test_resize_realloc() {
dView0 drv0("drv0", 10, 20, 30);
ASSERT_EQ(drv0.rank(), 3);
ASSERT_EQ(drv0.rank(), 3u);
Kokkos::resize(drv0, 5, 10);
ASSERT_EQ(drv0.rank(), 2);
ASSERT_EQ(drv0.extent(0), 5);
ASSERT_EQ(drv0.extent(1), 10);
ASSERT_EQ(drv0.extent(2), 1);
if (Initialize)
Kokkos::resize(Kokkos::WithoutInitializing, drv0, 5, 10);
else
Kokkos::resize(drv0, 5, 10);
ASSERT_EQ(drv0.rank(), 2u);
ASSERT_EQ(drv0.extent(0), 5u);
ASSERT_EQ(drv0.extent(1), 10u);
ASSERT_EQ(drv0.extent(2), 1u);
Kokkos::realloc(drv0, 10, 20);
ASSERT_EQ(drv0.rank(), 2);
ASSERT_EQ(drv0.extent(0), 10);
ASSERT_EQ(drv0.extent(1), 20);
ASSERT_EQ(drv0.extent(2), 1);
if (Initialize)
Kokkos::realloc(Kokkos::WithoutInitializing, drv0, 10, 20);
else
Kokkos::realloc(drv0, 10, 20);
ASSERT_EQ(drv0.rank(), 2u);
ASSERT_EQ(drv0.extent(0), 10u);
ASSERT_EQ(drv0.extent(1), 20u);
ASSERT_EQ(drv0.extent(2), 1u);
}
static void run_test_mirror() {
@ -961,6 +970,199 @@ class TestDynViewAPI {
}
}
static void run_test_as_view_of_rank_n() {
Kokkos::View<int, Kokkos::HostSpace> error_flag_host("error_flag");
error_flag_host() = 0;
auto error_flag =
Kokkos::create_mirror_view_and_copy(DeviceType(), error_flag_host);
dView0 d("d");
#if defined(KOKKOS_ENABLE_CXX11_DISPATCH_LAMBDA)
// Rank 0
Kokkos::resize(d);
auto policy0 = Kokkos::RangePolicy<DeviceType>(DeviceType(), 0, 1);
View0 v0 = Kokkos::Impl::as_view_of_rank_n<0>(d);
// Assign values after calling as_view_of_rank_n() function under
// test to ensure aliasing
Kokkos::parallel_for(
policy0, KOKKOS_LAMBDA(int) { d() = 13; });
ASSERT_EQ(v0.size(), d.size());
ASSERT_EQ(v0.data(), d.data());
Kokkos::parallel_for(
policy0, KOKKOS_LAMBDA(int) {
if (d() != v0()) error_flag() = 1;
});
Kokkos::deep_copy(error_flag_host, error_flag);
ASSERT_EQ(error_flag_host(), 0);
// Rank 1
Kokkos::resize(d, 1);
auto policy1 =
Kokkos::RangePolicy<DeviceType>(DeviceType(), 0, d.extent(0));
View1 v1 = Kokkos::Impl::as_view_of_rank_n<1>(d);
Kokkos::parallel_for(
policy1, KOKKOS_LAMBDA(int i0) { d(i0) = i0; });
for (unsigned int rank = 0; rank < d.rank(); ++rank)
ASSERT_EQ(v1.extent(rank), d.extent(rank));
ASSERT_EQ(v1.data(), d.data());
Kokkos::parallel_for(
policy1, KOKKOS_LAMBDA(int i0) {
if (d(i0) != v1(i0)) error_flag() = 1;
});
Kokkos::deep_copy(error_flag_host, error_flag);
ASSERT_EQ(error_flag_host(), 0);
// Rank 2
Kokkos::resize(d, 1, 2);
auto policy2 = Kokkos::MDRangePolicy<DeviceType, Kokkos::Rank<2>>(
{0, 0}, {d.extent(0), d.extent(1)});
View2 v2 = Kokkos::Impl::as_view_of_rank_n<2>(d);
Kokkos::parallel_for(
policy2, KOKKOS_LAMBDA(int i0, int i1) { d(i0, i1) = i0 + 10 * i1; });
for (unsigned int rank = 0; rank < d.rank(); ++rank)
ASSERT_EQ(v2.extent(rank), d.extent(rank));
ASSERT_EQ(v2.data(), d.data());
Kokkos::parallel_for(
policy2, KOKKOS_LAMBDA(int i0, int i1) {
if (d(i0, i1) != v2(i0, i1)) error_flag() = 1;
});
Kokkos::deep_copy(error_flag_host, error_flag);
ASSERT_EQ(error_flag_host(), 0);
// Rank 3
Kokkos::resize(d, 1, 2, 3);
auto policy3 = Kokkos::MDRangePolicy<DeviceType, Kokkos::Rank<3>>(
{0, 0, 0}, {d.extent(0), d.extent(1), d.extent(2)});
View3 v3 = Kokkos::Impl::as_view_of_rank_n<3>(d);
Kokkos::parallel_for(
policy3, KOKKOS_LAMBDA(int i0, int i1, int i2) {
d(i0, i1, i2) = i0 + 10 * i1 + 100 * i2;
});
for (unsigned int rank = 0; rank < d.rank(); ++rank)
ASSERT_EQ(v3.extent(rank), d.extent(rank));
ASSERT_EQ(v3.data(), d.data());
Kokkos::parallel_for(
policy3, KOKKOS_LAMBDA(int i0, int i1, int i2) {
if (d(i0, i1, i2) != v3(i0, i1, i2)) error_flag() = 1;
});
Kokkos::deep_copy(error_flag_host, error_flag);
ASSERT_EQ(error_flag_host(), 0);
// Rank 4
Kokkos::resize(d, 1, 2, 3, 4);
auto policy4 = Kokkos::MDRangePolicy<DeviceType, Kokkos::Rank<4>>(
{0, 0, 0, 0}, {d.extent(0), d.extent(1), d.extent(2), d.extent(3)});
View4 v4 = Kokkos::Impl::as_view_of_rank_n<4>(d);
Kokkos::parallel_for(
policy4, KOKKOS_LAMBDA(int i0, int i1, int i2, int i3) {
d(i0, i1, i2, i3) = i0 + 10 * i1 + 100 * i2 + 1000 * i3;
});
for (unsigned int rank = 0; rank < d.rank(); ++rank)
ASSERT_EQ(v4.extent(rank), d.extent(rank));
ASSERT_EQ(v4.data(), d.data());
Kokkos::parallel_for(
policy4, KOKKOS_LAMBDA(int i0, int i1, int i2, int i3) {
if (d(i0, i1, i2, i3) != v4(i0, i1, i2, i3)) error_flag() = 1;
});
Kokkos::deep_copy(error_flag_host, error_flag);
ASSERT_EQ(error_flag_host(), 0);
// Rank 5
Kokkos::resize(d, 1, 2, 3, 4, 5);
auto policy5 = Kokkos::MDRangePolicy<DeviceType, Kokkos::Rank<5>>(
{0, 0, 0, 0, 0},
{d.extent(0), d.extent(1), d.extent(2), d.extent(3), d.extent(4)});
View5 v5 = Kokkos::Impl::as_view_of_rank_n<5>(d);
Kokkos::parallel_for(
policy5, KOKKOS_LAMBDA(int i0, int i1, int i2, int i3, int i4) {
d(i0, i1, i2, i3, i4) =
i0 + 10 * i1 + 100 * i2 + 1000 * i3 + 10000 * i4;
});
for (unsigned int rank = 0; rank < d.rank(); ++rank)
ASSERT_EQ(v5.extent(rank), d.extent(rank));
ASSERT_EQ(v5.data(), d.data());
Kokkos::parallel_for(
policy5, KOKKOS_LAMBDA(int i0, int i1, int i2, int i3, int i4) {
if (d(i0, i1, i2, i3, i4) != v5(i0, i1, i2, i3, i4)) error_flag() = 1;
});
Kokkos::deep_copy(error_flag_host, error_flag);
ASSERT_EQ(error_flag_host(), 0);
// Rank 6
Kokkos::resize(d, 1, 2, 3, 4, 5, 6);
auto policy6 = Kokkos::MDRangePolicy<DeviceType, Kokkos::Rank<6>>(
{0, 0, 0, 0, 0, 0}, {d.extent(0), d.extent(1), d.extent(2), d.extent(3),
d.extent(4), d.extent(5)});
View6 v6 = Kokkos::Impl::as_view_of_rank_n<6>(d);
Kokkos::parallel_for(
policy6, KOKKOS_LAMBDA(int i0, int i1, int i2, int i3, int i4, int i5) {
d(i0, i1, i2, i3, i4, i5) =
i0 + 10 * i1 + 100 * i2 + 1000 * i3 + 10000 * i4 + 100000 * i5;
});
for (unsigned int rank = 0; rank < d.rank(); ++rank)
ASSERT_EQ(v6.extent(rank), d.extent(rank));
ASSERT_EQ(v6.data(), d.data());
Kokkos::parallel_for(
policy6, KOKKOS_LAMBDA(int i0, int i1, int i2, int i3, int i4, int i5) {
if (d(i0, i1, i2, i3, i4, i5) != v6(i0, i1, i2, i3, i4, i5))
error_flag() = 1;
});
Kokkos::deep_copy(error_flag_host, error_flag);
ASSERT_EQ(error_flag_host(), 0);
// Rank 7
Kokkos::resize(d, 1, 2, 3, 4, 5, 6, 7);
// MDRangePolicy only accepts Rank < 7
#if 0
auto policy7 = Kokkos::MDRangePolicy<DeviceType, Kokkos::Rank<7>>(
{0, 0, 0, 0, 0, 0, 0},
{d.extent(0), d.extent(1), d.extent(2), d.extent(3), d.extent(4),
d.extent(5), d.extent(6)});
View7 v7 = Kokkos::Impl::as_view_of_rank_n<7>(d);
Kokkos::parallel_for(
policy7,
KOKKOS_LAMBDA(int i0, int i1, int i2, int i3, int i4, int i5, int i6) {
d(i0, i1, i2, i3, i4, i5, i6) = i0 + 10 * i1 + 100 * i2 + 1000 * i3 +
10000 * i4 + 100000 * i5 +
1000000 * i6;
});
for (unsigned int rank = 0; rank < d.rank(); ++rank)
ASSERT_EQ(v7.extent(rank), d.extent(rank));
ASSERT_EQ(v7.data(), d.data());
Kokkos::parallel_for(
policy7,
KOKKOS_LAMBDA(int i0, int i1, int i2, int i3, int i4, int i5, int i6) {
if (d(i0, i1, i2, i3, i4, i5, i6) != v7(i0, i1, i2, i3, i4, i5, i6))
error_flag() = 1;
});
Kokkos::deep_copy(error_flag_host, error_flag);
ASSERT_EQ(error_flag_host(), 0);
#endif // MDRangePolict Rank < 7
#endif // defined(KOKKOS_ENABLE_CXX11_DISPATCH_LAMBDA)
// Error checking test
EXPECT_ANY_THROW({ auto v_copy = Kokkos::Impl::as_view_of_rank_n<2>(d); });
}
static void run_test_scalar() {
using hView0 = typename dView0::HostMirror; // HostMirror of DynRankView is
// a DynRankView
@ -1071,10 +1273,10 @@ class TestDynViewAPI {
dView0 d_uninitialized(
Kokkos::view_alloc(Kokkos::WithoutInitializing, "uninit"), 10, 20);
ASSERT_NE(d_uninitialized.data(), nullptr);
ASSERT_EQ(d_uninitialized.rank(), 2);
ASSERT_EQ(d_uninitialized.extent(0), 10);
ASSERT_EQ(d_uninitialized.extent(1), 20);
ASSERT_EQ(d_uninitialized.extent(2), 1);
ASSERT_EQ(d_uninitialized.rank(), 2u);
ASSERT_EQ(d_uninitialized.extent(0), 10u);
ASSERT_EQ(d_uninitialized.extent(1), 20u);
ASSERT_EQ(d_uninitialized.extent(2), 1u);
dView0 dx, dy, dz;
hView0 hx, hy, hz;
@ -1107,8 +1309,8 @@ class TestDynViewAPI {
ASSERT_EQ(dy.extent(0), unsigned(N1)); // Okay with UVM
ASSERT_EQ(hx.extent(0), unsigned(N1));
ASSERT_EQ(hy.extent(0), unsigned(N1));
ASSERT_EQ(dx.rank(), 3); // Okay with UVM
ASSERT_EQ(hx.rank(), 3);
ASSERT_EQ(dx.rank(), 3u); // Okay with UVM
ASSERT_EQ(hx.rank(), 3u);
dx = dView0("dx", N0, N1, N2, N3);
dy = dView0("dy", N0, N1, N2, N3);
@ -1119,15 +1321,15 @@ class TestDynViewAPI {
ASSERT_EQ(dy.extent(0), unsigned(N0));
ASSERT_EQ(hx.extent(0), unsigned(N0));
ASSERT_EQ(hy.extent(0), unsigned(N0));
ASSERT_EQ(dx.rank(), 4);
ASSERT_EQ(dy.rank(), 4);
ASSERT_EQ(hx.rank(), 4);
ASSERT_EQ(hy.rank(), 4);
ASSERT_EQ(dx.rank(), 4u);
ASSERT_EQ(dy.rank(), 4u);
ASSERT_EQ(hx.rank(), 4u);
ASSERT_EQ(hy.rank(), 4u);
ASSERT_EQ(dx.use_count(), size_t(1));
ASSERT_EQ(dx.use_count(), 1);
dView0_unmanaged unmanaged_dx = dx;
ASSERT_EQ(dx.use_count(), size_t(1));
ASSERT_EQ(dx.use_count(), 1);
dView0_unmanaged unmanaged_from_ptr_dx = dView0_unmanaged(
dx.data(), dx.extent(0), dx.extent(1), dx.extent(2), dx.extent(3));
@ -1139,24 +1341,24 @@ class TestDynViewAPI {
}
const_dView0 const_dx = dx;
ASSERT_EQ(dx.use_count(), size_t(2));
ASSERT_EQ(dx.use_count(), 2);
{
const_dView0 const_dx2;
const_dx2 = const_dx;
ASSERT_EQ(dx.use_count(), size_t(3));
ASSERT_EQ(dx.use_count(), 3);
const_dx2 = dy;
ASSERT_EQ(dx.use_count(), size_t(2));
ASSERT_EQ(dx.use_count(), 2);
const_dView0 const_dx3(dx);
ASSERT_EQ(dx.use_count(), size_t(3));
ASSERT_EQ(dx.use_count(), 3);
dView0_unmanaged dx4_unmanaged(dx);
ASSERT_EQ(dx.use_count(), size_t(3));
ASSERT_EQ(dx.use_count(), 3);
}
ASSERT_EQ(dx.use_count(), size_t(2));
ASSERT_EQ(dx.use_count(), 2);
ASSERT_NE(dx.data(), nullptr);
ASSERT_NE(const_dx.data(), nullptr);
@ -1336,18 +1538,18 @@ class TestDynViewAPI {
// View - DynRankView Interoperability tests
// deep_copy from view to dynrankview
const int testdim = 4;
constexpr size_t testdim = 4;
dView0 dxx("dxx", testdim);
View1 vxx("vxx", testdim);
auto hvxx = Kokkos::create_mirror_view(vxx);
for (int i = 0; i < testdim; ++i) {
for (size_t i = 0; i < testdim; ++i) {
hvxx(i) = i;
}
Kokkos::deep_copy(vxx, hvxx);
Kokkos::deep_copy(dxx, vxx);
auto hdxx = Kokkos::create_mirror_view(dxx);
Kokkos::deep_copy(hdxx, dxx);
for (int i = 0; i < testdim; ++i) {
for (size_t i = 0; i < testdim; ++i) {
ASSERT_EQ(hvxx(i), hdxx(i));
}
@ -1362,7 +1564,7 @@ class TestDynViewAPI {
ASSERT_EQ(rank(hdxx), rank(hvdxx));
ASSERT_EQ(hvdxx.extent(0), testdim);
ASSERT_EQ(hdxx.extent(0), hvdxx.extent(0));
for (int i = 0; i < testdim; ++i) {
for (size_t i = 0; i < testdim; ++i) {
ASSERT_EQ(hvxx(i), hvdxx(i));
}
}
@ -1432,51 +1634,51 @@ class TestDynViewAPI {
unsigned order[] = {6, 5, 4, 3, 2, 1, 0},
dimen[] = {N0, N1, N2, 2, 2, 2, 2}; // LayoutRight equivalent
sdView d7("d7", Kokkos::LayoutStride::order_dimensions(7, order, dimen));
ASSERT_EQ(d7.rank(), 7);
ASSERT_EQ(d7.rank(), 7u);
sdView ds0 = Kokkos::subdynrankview(d7, 1, 1, 1, 1, 1, 1, 1);
ASSERT_EQ(ds0.rank(), 0);
ASSERT_EQ(ds0.rank(), 0u);
// Basic test - ALL
sdView dsALL = Kokkos::subdynrankview(
d7, Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL(),
Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL());
ASSERT_EQ(dsALL.rank(), 7);
ASSERT_EQ(dsALL.rank(), 7u);
// Send a value to final rank returning rank 6 subview
sdView dsm1 =
Kokkos::subdynrankview(d7, Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL(),
Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL(), 1);
ASSERT_EQ(dsm1.rank(), 6);
ASSERT_EQ(dsm1.rank(), 6u);
// Send a std::pair as argument to a rank
sdView dssp = Kokkos::subdynrankview(
d7, Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL(),
Kokkos::ALL(), Kokkos::ALL(), std::pair<unsigned, unsigned>(1, 2));
ASSERT_EQ(dssp.rank(), 7);
ASSERT_EQ(dssp.rank(), 7u);
// Send a kokkos::pair as argument to a rank; take default layout as input
dView0 dd0("dd0", N0, N1, N2, 2, 2, 2, 2); // default layout
ASSERT_EQ(dd0.rank(), 7);
ASSERT_EQ(dd0.rank(), 7u);
sdView dtkp = Kokkos::subdynrankview(
dd0, Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL(),
Kokkos::ALL(), Kokkos::ALL(), Kokkos::pair<unsigned, unsigned>(0, 1));
ASSERT_EQ(dtkp.rank(), 7);
ASSERT_EQ(dtkp.rank(), 7u);
// Return rank 7 subview, taking a pair as one argument, layout stride input
sdView ds7 = Kokkos::subdynrankview(
d7, Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL(), Kokkos::ALL(),
Kokkos::ALL(), Kokkos::ALL(), Kokkos::pair<unsigned, unsigned>(0, 1));
ASSERT_EQ(ds7.rank(), 7);
ASSERT_EQ(ds7.rank(), 7u);
// Default Layout DynRankView
dView dv6("dv6", N0, N1, N2, N3, 2, 2);
ASSERT_EQ(dv6.rank(), 6);
ASSERT_EQ(dv6.rank(), 6u);
// DynRankView with LayoutRight
using drView = Kokkos::DynRankView<T, Kokkos::LayoutRight, device>;
drView dr5("dr5", N0, N1, N2, 2, 2);
ASSERT_EQ(dr5.rank(), 5);
ASSERT_EQ(dr5.rank(), 5u);
// LayoutStride but arranged as LayoutRight
// NOTE: unused arg_layout dimensions must be set toKOKKOS_INVALID_INDEX so
@ -1489,7 +1691,7 @@ class TestDynViewAPI {
ls.dimension[6] = KOKKOS_INVALID_INDEX;
ls.dimension[7] = KOKKOS_INVALID_INDEX;
sdView d5("d5", ls);
ASSERT_EQ(d5.rank(), 5);
ASSERT_EQ(d5.rank(), 5u);
// LayoutStride arranged as LayoutRight - commented out as example that
// fails unit test
@ -1522,7 +1724,7 @@ class TestDynViewAPI {
sdView ds5 = Kokkos::subdynrankview(d5, Kokkos::ALL(), Kokkos::ALL(),
Kokkos::ALL(), Kokkos::ALL(),
Kokkos::pair<unsigned, unsigned>(0, 1));
ASSERT_EQ(ds5.rank(), 5);
ASSERT_EQ(ds5.rank(), 5u);
// Pass in extra ALL arguments beyond the rank of the DynRank View.
// This behavior is allowed - ignore the extra ALL arguments when
@ -1554,7 +1756,7 @@ class TestDynViewAPI {
Kokkos::ALL(), 0, Kokkos::ALL());
ASSERT_EQ(ds4.rank(), ds4plus.rank());
ASSERT_EQ(ds4.rank(), 4);
ASSERT_EQ(ds4.rank(), 4u);
ASSERT_EQ(ds4.extent(0), ds4plus.extent(0));
ASSERT_EQ(ds4.extent(4), ds4plus.extent(4));
ASSERT_EQ(ds4.extent(5), ds4plus.extent(5));
@ -1601,8 +1803,8 @@ class TestDynViewAPI {
ASSERT_EQ(yl4.extent(1), xl4.extent(3));
ASSERT_EQ(yr4.extent(0), xr4.extent(1));
ASSERT_EQ(yr4.extent(1), xr4.extent(3));
ASSERT_EQ(yl4.rank(), 2);
ASSERT_EQ(yr4.rank(), 2);
ASSERT_EQ(yl4.rank(), 2u);
ASSERT_EQ(yr4.rank(), 2u);
ASSERT_EQ(&yl4(4, 4) - &xl4(1, 4, 2, 4), 0);
ASSERT_EQ(&yr4(4, 4) - &xr4(1, 4, 2, 4), 0);