Files
lammps/lib/kokkos/core/unit_test/TestTeamScan.hpp
2023-11-21 15:02:12 -07:00

263 lines
9.2 KiB
C++

//@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
#include <Kokkos_Core.hpp>
#include <impl/Kokkos_Stacktrace.hpp>
#include <cstdio>
#include <cstdint>
#include <sstream>
#include <type_traits>
namespace Test {
template <class ExecutionSpace, class DataType>
struct TestTeamScan {
using execution_space = ExecutionSpace;
using value_type = DataType;
using policy_type = Kokkos::TeamPolicy<execution_space>;
using member_type = typename policy_type::member_type;
using view_type = Kokkos::View<value_type**, execution_space>;
view_type a_d;
view_type a_r;
int32_t M = 0;
int32_t N = 0;
KOKKOS_FUNCTION
void operator()(const member_type& team) const {
auto leagueRank = team.league_rank();
auto beg = 0;
auto end = N;
Kokkos::parallel_for(
Kokkos::TeamThreadRange(team, beg, end),
[&](const int i) { a_d(leagueRank, i) = leagueRank * N + i; });
Kokkos::parallel_scan(Kokkos::TeamThreadRange(team, beg, end),
[&](int i, DataType& val, const bool final) {
val += a_d(leagueRank, i);
if (final) a_r(leagueRank, i) = val;
});
}
auto operator()(int32_t _M, int32_t _N) {
std::stringstream ss;
ss << Kokkos::Impl::demangle(typeid(*this).name());
ss << "(/*M=*/" << _M << ", /*N=*/" << _N << ")";
std::string const test_id = ss.str();
M = _M;
N = _N;
a_d = view_type("a_d", M, N);
a_r = view_type("a_r", M, N);
// Set team size explicitly to check whether non-power-of-two team sizes can
// be used.
if (ExecutionSpace().concurrency() > 10000)
Kokkos::parallel_for(policy_type(M, 127), *this);
else if (ExecutionSpace().concurrency() > 2)
Kokkos::parallel_for(policy_type(M, 3), *this);
else
Kokkos::parallel_for(policy_type(M, 1), *this);
auto a_i = Kokkos::create_mirror_view(a_d);
auto a_o = Kokkos::create_mirror_view(a_r);
Kokkos::deep_copy(a_i, a_d);
Kokkos::deep_copy(a_o, a_r);
for (int32_t i = 0; i < M; ++i) {
value_type scan_ref = 0;
value_type scan_calc;
value_type abs_err = 0;
// each fp addition is subject to small loses in precision and these
// compound as loop so we set the base error to be the machine epsilon and
// then add in another epsilon each iteration. For example, with CUDA
// backend + 32-bit float + large N values (e.g. 1,000) + high
// thread-counts (e.g. 1024), this test will fail w/o epsilon
// accommodation
constexpr value_type epsilon = std::numeric_limits<value_type>::epsilon();
for (int32_t j = 0; j < N; ++j) {
scan_ref += a_i(i, j);
scan_calc = a_o(i, j);
if (std::is_integral<value_type>::value) {
ASSERT_EQ(scan_ref, scan_calc)
<< test_id
<< " calculated scan output value differs from reference at "
"indices i="
<< i << " and j=" << j;
} else {
abs_err += epsilon;
ASSERT_NEAR(scan_ref, scan_calc, abs_err)
<< test_id
<< " calculated scan output value differs from reference at "
"indices i="
<< i << " and j=" << j;
}
}
}
}
};
TEST(TEST_CATEGORY, team_scan) {
TestTeamScan<TEST_EXECSPACE, int32_t>{}(0, 0);
TestTeamScan<TEST_EXECSPACE, int32_t>{}(0, 1);
TestTeamScan<TEST_EXECSPACE, int32_t>{}(1, 0);
TestTeamScan<TEST_EXECSPACE, uint32_t>{}(99, 32);
TestTeamScan<TEST_EXECSPACE, uint32_t>{}(139, 64);
TestTeamScan<TEST_EXECSPACE, uint32_t>{}(163, 128);
TestTeamScan<TEST_EXECSPACE, int64_t>{}(433, 256);
TestTeamScan<TEST_EXECSPACE, uint64_t>{}(976, 512);
TestTeamScan<TEST_EXECSPACE, uint64_t>{}(1234, 1024);
TestTeamScan<TEST_EXECSPACE, float>{}(2596, 34);
TestTeamScan<TEST_EXECSPACE, double>{}(2596, 59);
TestTeamScan<TEST_EXECSPACE, float>{}(2596, 65);
TestTeamScan<TEST_EXECSPACE, double>{}(2596, 371);
TestTeamScan<TEST_EXECSPACE, int64_t>{}(2596, 987);
TestTeamScan<TEST_EXECSPACE, double>{}(2596, 1311);
}
// Temporary: This condition will progressively be reduced when parallel_scan
// with return value will be implemented for more backends.
#if !defined(KOKKOS_ENABLE_OPENACC)
template <class ExecutionSpace, class DataType>
struct TestTeamScanRetVal {
using execution_space = ExecutionSpace;
using value_type = DataType;
using policy_type = Kokkos::TeamPolicy<execution_space>;
using member_type = typename policy_type::member_type;
using view_1d_type = Kokkos::View<value_type*, execution_space>;
using view_2d_type = Kokkos::View<value_type**, execution_space>;
view_2d_type a_d;
view_2d_type a_r;
view_1d_type a_s;
int32_t M = 0;
int32_t N = 0;
KOKKOS_FUNCTION
void operator()(const member_type& team) const {
auto leagueRank = team.league_rank();
auto beg = 0;
auto end = N;
Kokkos::parallel_for(
Kokkos::TeamThreadRange(team, beg, end),
[&](const int i) { a_d(leagueRank, i) = leagueRank * N + i; });
DataType accum;
Kokkos::parallel_scan(
Kokkos::TeamThreadRange(team, beg, end),
[&](int i, DataType& val, const bool final) {
val += a_d(leagueRank, i);
if (final) a_r(leagueRank, i) = val;
},
accum);
// Save return value from parallel_scan
Kokkos::single(Kokkos::PerTeam(team), [&]() { a_s(leagueRank) = accum; });
}
auto operator()(int32_t _M, int32_t _N) {
std::stringstream ss;
ss << Kokkos::Impl::demangle(typeid(*this).name());
ss << "(/*M=*/" << _M << ", /*N=*/" << _N << ")";
std::string const test_id = ss.str();
M = _M;
N = _N;
a_d = view_2d_type("a_d", M, N);
a_r = view_2d_type("a_r", M, N);
a_s = view_1d_type("a_s", M);
// Set team size explicitly to check whether non-power-of-two team sizes can
// be used.
if (ExecutionSpace().concurrency() > 10000)
Kokkos::parallel_for(policy_type(M, 127), *this);
else if (ExecutionSpace().concurrency() > 2)
Kokkos::parallel_for(policy_type(M, 3), *this);
else
Kokkos::parallel_for(policy_type(M, 1), *this);
Kokkos::fence();
auto a_i = Kokkos::create_mirror_view(a_d);
auto a_o = Kokkos::create_mirror_view(a_r);
auto a_os = Kokkos::create_mirror_view(a_s);
Kokkos::deep_copy(a_i, a_d);
Kokkos::deep_copy(a_o, a_r);
Kokkos::deep_copy(a_os, a_s);
for (int32_t i = 0; i < M; ++i) {
value_type scan_ref = 0;
value_type scan_calc;
value_type abs_err = 0;
// each fp addition is subject to small loses in precision and these
// compound as loop so we set the base error to be the machine epsilon and
// then add in another epsilon each iteration. For example, with CUDA
// backend + 32-bit float + large N values (e.g. 1,000) + high
// thread-counts (e.g. 1024), this test will fail w/o epsilon
// accommodation
constexpr value_type epsilon = std::numeric_limits<value_type>::epsilon();
for (int32_t j = 0; j < N; ++j) {
scan_ref += a_i(i, j);
scan_calc = a_o(i, j);
if (std::is_integral<value_type>::value) {
ASSERT_EQ(scan_ref, scan_calc)
<< test_id
<< " calculated scan output value differs from reference at "
"indices i="
<< i << " and j=" << j;
} else {
abs_err += epsilon;
ASSERT_NEAR(scan_ref, scan_calc, abs_err)
<< test_id
<< " calculated scan output value differs from reference at "
"indices i="
<< i << " and j=" << j;
}
}
// Validate return value from parallel_scan
if (std::is_integral<value_type>::value) {
ASSERT_EQ(scan_ref, a_os(i));
} else {
ASSERT_NEAR(scan_ref, a_os(i), abs_err);
}
}
}
};
TEST(TEST_CATEGORY, team_scan_ret_val) {
TestTeamScanRetVal<TEST_EXECSPACE, int32_t>{}(0, 0);
TestTeamScanRetVal<TEST_EXECSPACE, int32_t>{}(0, 1);
TestTeamScanRetVal<TEST_EXECSPACE, int32_t>{}(1, 0);
TestTeamScanRetVal<TEST_EXECSPACE, uint32_t>{}(99, 32);
TestTeamScanRetVal<TEST_EXECSPACE, uint32_t>{}(139, 64);
TestTeamScanRetVal<TEST_EXECSPACE, uint32_t>{}(163, 128);
TestTeamScanRetVal<TEST_EXECSPACE, int64_t>{}(433, 256);
TestTeamScanRetVal<TEST_EXECSPACE, uint64_t>{}(976, 512);
TestTeamScanRetVal<TEST_EXECSPACE, uint64_t>{}(1234, 1024);
TestTeamScanRetVal<TEST_EXECSPACE, float>{}(2596, 34);
TestTeamScanRetVal<TEST_EXECSPACE, double>{}(2596, 59);
TestTeamScanRetVal<TEST_EXECSPACE, float>{}(2596, 65);
TestTeamScanRetVal<TEST_EXECSPACE, double>{}(2596, 371);
TestTeamScanRetVal<TEST_EXECSPACE, int64_t>{}(2596, 987);
TestTeamScanRetVal<TEST_EXECSPACE, double>{}(2596, 1311);
}
#endif
} // namespace Test