Merge pull request #354 from bluescarni/pr/opt

Performance optimisations

benchmark/CMakeLists.txt

@@ -69,6 +69,7 @@ ADD_HEYOKA_BENCHMARK(bandf)
 ADD_HEYOKA_BENCHMARK(bandf_odeint)
 ADD_HEYOKA_BENCHMARK(penc_comparison)
 ADD_HEYOKA_BENCHMARK(large_cfunc)
+ADD_HEYOKA_BENCHMARK(sims_flanagan_jac)
 
 if(HEYOKA_WITH_MPPP AND mp++_WITH_MPFR)
   ADD_HEYOKA_BENCHMARK(pendulum_mp)

benchmark/sims_flanagan_jac.cpp

@@ -0,0 +1,157 @@
+// Copyright 2020, 2021, 2022, 2023 Francesco Biscani ([email protected]), Dario Izzo ([email protected])
+//
+// This file is part of the heyoka library.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#include <algorithm>
+#include <array>
+#include <iostream>
+#include <iterator>
+#include <stdexcept>
+#include <tuple>
+#include <utility>
+#include <vector>
+
+#include <boost/program_options.hpp>
+
+#include <spdlog/spdlog.h>
+
+#include <heyoka/expression.hpp>
+#include <heyoka/llvm_state.hpp>
+#include <heyoka/logging.hpp>
+#include <heyoka/math/cos.hpp>
+#include <heyoka/math/kepDE.hpp>
+#include <heyoka/math/sin.hpp>
+#include <heyoka/math/sqrt.hpp>
+
+#include <fmt/ranges.h>
+
+using namespace heyoka;
+
+using v_ex_t = std::array<expression, 3>;
+
+const auto mu = make_vars("mu")[0];
+const auto tm = make_vars("t")[0];
+
+std::pair<v_ex_t, v_ex_t> make_lp(const v_ex_t &pos_0, const v_ex_t &vel_0)
+{
+    const auto &[x0, y0, z0] = pos_0;
+    const auto &[vx0, vy0, vz0] = vel_0;
+
+    auto v02 = vx0 * vx0 + vy0 * vy0 + vz0 * vz0;
+    auto r0 = sqrt(x0 * x0 + y0 * y0 + z0 * z0);
+    auto eps = v02 * 0.5 - mu / r0;
+    auto a = -mu / (2. * eps);
+
+    auto sigma0 = (x0 * vx0 + y0 * vy0 + z0 * vz0) / sqrt(mu);
+    auto s0 = sigma0 / sqrt(a);
+    auto c0 = 1. - r0 / a;
+
+    auto n = sqrt(mu / (a * a * a));
+    auto DM = n * tm;
+
+    auto DE = kepDE(s0, c0, DM);
+
+    auto cDE = cos(DE);
+    auto sDE = sin(DE);
+
+    auto r = a + (r0 - a) * cDE + sigma0 * sqrt(a) * sDE;
+
+    auto F = 1. - a / r0 * (1. - cDE);
+    auto G = a * sigma0 / sqrt(mu) * (1. - cDE) + r0 * sqrt(a / mu) * sDE;
+    auto Ft = -sqrt(mu * a) / (r * r0) * sDE;
+    auto Gt = 1. - a / r * (1. - cDE);
+
+    return {{{F * x0 + G * vx0, F * y0 + G * vy0, F * z0 + G * vz0}},
+            {{Ft * x0 + Gt * vx0, Ft * y0 + Gt * vy0, Ft * z0 + Gt * vz0}}};
+}
+
+int main(int argc, char *argv[])
+{
+    namespace po = boost::program_options;
+
+    create_logger();
+    auto logger = spdlog::get("heyoka");
+
+    set_logger_level_trace();
+
+    unsigned N{};
+
+    po::options_description desc("Options");
+
+    desc.add_options()("help", "produce help message")("N", po::value<unsigned>(&N)->default_value(20u),
+                                                       "number of segments");
+
+    po::variables_map vm;
+    po::store(po::parse_command_line(argc, argv, desc), vm);
+    po::notify(vm);
+
+    if (vm.count("help") != 0u) {
+        std::cout << desc << "\n";
+        return 0;
+    }
+
+    if (N == 0u) {
+        throw std::invalid_argument("The number of segments cannot be zero");
+    }
+
+    auto [x0, y0, z0] = make_vars("x0", "y0", "z0");
+    auto [vx0, vy0, vz0] = make_vars("vx0", "vy0", "vz0");
+
+    std::vector<v_ex_t> Dvs;
+    for (auto i = 0u; i < N; ++i) {
+        auto [Dvx, Dvy, Dvz]
+            = make_vars(fmt::format("Dvx{}", i + 1u), fmt::format("Dvy{}", i + 1u), fmt::format("Dvz{}", i + 1u));
+        Dvs.push_back(v_ex_t{Dvx, Dvy, Dvz});
+    }
+
+    auto [pos_f, vel_f] = make_lp({x0, y0, z0}, {vx0, vy0, vz0});
+
+    for (auto i = 0u; i < N; ++i) {
+        vel_f[0] += Dvs[i][0];
+        vel_f[1] += Dvs[i][1];
+        vel_f[2] += Dvs[i][2];
+
+        std::tie(pos_f, vel_f) = make_lp(pos_f, vel_f);
+    }
+
+    std::vector<expression> Dvs_list;
+    for (const auto &Dv : Dvs) {
+        Dvs_list.push_back(Dv[0]);
+        Dvs_list.push_back(Dv[1]);
+        Dvs_list.push_back(Dv[2]);
+    }
+
+    std::vector<expression> diff_vars_list = {x0, y0, z0, vx0, vy0, vz0};
+    diff_vars_list.insert(diff_vars_list.end(), Dvs_list.begin(), Dvs_list.end());
+
+    auto dt
+        = diff_tensors({pos_f[0], pos_f[1], pos_f[2], vel_f[0], vel_f[1], vel_f[2]}, kw::diff_args = diff_vars_list);
+
+    std::vector<expression> jac;
+    auto jac_sr = dt.get_derivatives(1);
+    std::transform(jac_sr.begin(), jac_sr.end(), std::back_inserter(jac), [](const auto &p) { return p.second; });
+
+    llvm_state s;
+
+    std::vector<expression> vars_list = {x0, y0, z0, vx0, vy0, vz0, mu, tm};
+    vars_list.insert(vars_list.end(), Dvs_list.begin(), Dvs_list.end());
+
+    logger->trace("Adding cfunc");
+
+    add_cfunc<double>(s, "func", jac, kw::vars = vars_list);
+
+    logger->trace("Compiling cfunc");
+
+    s.compile();
+
+    logger->trace("Looking up cfunc");
+
+    auto *fptr
+        = reinterpret_cast<void (*)(double *, const double *, const double *, const double *)>(s.jit_lookup("func"));
+
+    logger->trace("All done");
+}

doc/changelog.rst

@@ -13,6 +13,14 @@ New
   system (`#352 <https://github.com/bluescarni/heyoka/pull/352>`__).
   Taylor derivatives are not implemented yet.
 
+Changes
+~~~~~~~
+
+- Substantial performance improvements in the computation of
+  derivative tensors of large expressions with a high degree
+  of internal redundancy
+  (`#354 <https://github.com/bluescarni/heyoka/pull/354>`__).
+
 Fix
 ~~~
 

doc/conf.py.in

@@ -63,6 +63,8 @@ html_theme_options = {
     "repository_url": "https://github.com/bluescarni/heyoka",
     "use_repository_button": True,
     "use_issues_button": True,
+    # See: https://github.com/pydata/pydata-sphinx-theme/issues/1492
+    "navigation_with_keys": False,
 }
 
 

include/heyoka/detail/fast_unordered.hpp

@@ -0,0 +1,67 @@
+// Copyright 2020, 2021, 2022, 2023 Francesco Biscani ([email protected]), Dario Izzo ([email protected])
+//
+// This file is part of the heyoka library.
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef HEYOKA_DETAIL_FAST_UNORDERED_HPP
+#define HEYOKA_DETAIL_FAST_UNORDERED_HPP
+
+#include <heyoka/config.hpp>
+
+#include <boost/version.hpp>
+
+#if (BOOST_VERSION / 100000 > 1) || (BOOST_VERSION / 100000 == 1 && BOOST_VERSION / 100 % 1000 >= 81)
+
+#define HEYOKA_HAVE_BOOST_UNORDERED_FLAT
+
+#endif
+
+#if defined(HEYOKA_HAVE_BOOST_UNORDERED_FLAT)
+
+#include <boost/unordered/unordered_flat_map.hpp>
+#include <boost/unordered/unordered_flat_set.hpp>
+
+#else
+
+#include <unordered_map>
+#include <unordered_set>
+
+#endif
+
+HEYOKA_BEGIN_NAMESPACE
+
+namespace detail
+{
+
+template <typename... Args>
+using fast_uset =
+#if defined(HEYOKA_HAVE_BOOST_UNORDERED_FLAT)
+    boost::unordered_flat_set<Args...>
+#else
+    std::unordered_set<Args...>
+#endif
+    ;
+
+template <typename... Args>
+using fast_umap =
+#if defined(HEYOKA_HAVE_BOOST_UNORDERED_FLAT)
+    boost::unordered_flat_map<Args...>
+#else
+    std::unordered_map<Args...>
+#endif
+    ;
+
+} // namespace detail
+
+HEYOKA_END_NAMESPACE
+
+#if defined(HEYOKA_HAVE_BOOST_UNORDERED_FLAT)
+
+#undef HEYOKA_HAVE_BOOST_UNORDERED_FLAT
+
+#endif
+
+#endif

include/heyoka/detail/func_cache.hpp

@@ -16,47 +16,20 @@
 #define HEYOKA_DETAIL_FUNC_CACHE_HPP
 
 #include <heyoka/config.hpp>
-
-#include <boost/version.hpp>
-
-#if (BOOST_VERSION / 100000 > 1) || (BOOST_VERSION / 100000 == 1 && BOOST_VERSION / 100 % 1000 >= 81)
-
-#include <boost/unordered/unordered_flat_map.hpp>
-#include <boost/unordered/unordered_flat_set.hpp>
+#include <heyoka/detail/fast_unordered.hpp>
 
 HEYOKA_BEGIN_NAMESPACE
 
 namespace detail
 {
 
-using funcptr_set = boost::unordered_flat_set<const void *>;
+using funcptr_set = fast_uset<const void *>;
 
 template <typename T>
-using funcptr_map = boost::unordered_flat_map<const void *, T>;
+using funcptr_map = fast_umap<const void *, T>;
 
 } // namespace detail
 
 HEYOKA_END_NAMESPACE
 
-#else
-
-#include <unordered_map>
-#include <unordered_set>
-
-HEYOKA_BEGIN_NAMESPACE
-
-namespace detail
-{
-
-using funcptr_set = std::unordered_set<const void *>;
-
-template <typename T>
-using funcptr_map = std::unordered_map<const void *, T>;
-
-} // namespace detail
-
-HEYOKA_END_NAMESPACE
-
-#endif
-
 #endif

include/heyoka/expression.hpp

@@ -501,6 +501,8 @@ class HEYOKA_DLL_PUBLIC dtens
 
     [[nodiscard]] subrange get_derivatives(std::uint32_t, std::uint32_t) const;
     [[nodiscard]] subrange get_derivatives(std::uint32_t) const;
+    [[nodiscard]] std::vector<expression> get_gradient() const;
+    [[nodiscard]] std::vector<expression> get_jacobian() const;
 };
 
 HEYOKA_DLL_PUBLIC std::ostream &operator<<(std::ostream &, const dtens &);

src/expression_basic.cpp

@@ -27,6 +27,7 @@
 #include <variant>
 #include <vector>
 
+#include <boost/container_hash/hash.hpp>
 #include <boost/safe_numerics/safe_integer.hpp>
 
 #include <llvm/IR/Function.h>
@@ -494,12 +495,60 @@ std::size_t hash(funcptr_map<std::size_t> &func_map, const expression &ex) noexc
         ex.value());
 }
 
-// NOLINTNEXTLINE(bugprone-exception-escape)
+// NOLINTNEXTLINE(bugprone-exception-escape,misc-no-recursion)
 std::size_t hash(const expression &ex) noexcept
 {
-    detail::funcptr_map<std::size_t> func_map;
+    // NOTE: we implement an optimisation here: if either the expression is **not** a function,
+    // or if it is a function and none of its arguments are functions (i.e., it is a non-recursive
+    // expression), then we do not recurse into detail::hash() and we do not create/update the
+    // funcptr_map cache. This allows us to avoid memory allocations, in an effort to optimise the
+    // computation of hashes in decompositions, where the elementary subexpressions are
+    // non-recursive by definition.
+    // NOTE: we must be very careful with this optimisation: we need to make sure the hash computation
+    // result is the same whether the optimisation is enabled or not. That is, the hash
+    // **must** be the same regardless of whether a non-recursive expression is standalone
+    // or it is part of a larger expression.
+    return std::visit(
+        // NOLINTNEXTLINE(misc-no-recursion)
+        [&ex](const auto &v) {
+            using type = detail::uncvref_t<decltype(v)>;
+
+            if constexpr (std::is_same_v<type, func>) {
+                const auto no_func_args = std::none_of(v.args().begin(), v.args().end(), [](const auto &x) {
+                    return std::holds_alternative<func>(x.value());
+                });
+
+                if (no_func_args) {
+                    // NOTE: it is very important that here we replicate *exactly* the logic
+                    // of func::hash(), so that we produce the same hash value that would
+                    // be produced if ex were part of a larger expression.
+                    std::size_t seed = std::hash<std::string>{}(v.get_name());
+
+                    for (const auto &arg : v.args()) {
+                        // NOTE: for non-function arguments, calling hash(arg)
+                        // is identical to calling detail::hash(func_map, arg)
+                        // (as we do in func::hash()):
+                        // both ultimately end up using directly the std::hash
+                        // specialisation on the arg.
+                        boost::hash_combine(seed, hash(arg));
+                    }
 
-    return hash(func_map, ex);
+                    return seed;
+                } else {
+                    // The regular, non-optimised path.
+                    detail::funcptr_map<std::size_t> func_map;
+
+                    return hash(func_map, ex);
+                }
+            } else {
+                // ex is not a function, compute its hash directly
+                // via a std::hash specialisation.
+                // NOTE: this is the same logic implemented in detail::hash(),
+                // except we avoid the unnecessary creation of a funcptr_map.
+                return std::hash<type>{}(v);
+            }
+        },
+        ex.value());
 }
 
 } // namespace detail