tboudreaux/GridFire
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

perf(multi): Simple parallel multi zone solver

Browse Source 
Added a simple parallel multi-zone solver
This commit is contained in:
Emily Boudreaux
2025-12-18 12:47:39 -05:00
parent 4e1edfc142
commit dcfd7b60aa
27 changed files with 1018 additions and 2193 deletions
Download Patch File Download Diff File Expand all files Collapse all files

109
tests/graphnet_sandbox/main.cpp
View File

@@ -19,7 +19,7 @@
#include <clocale>
#include "gridfire/reaction/reaclib.h"
#include "gridfire/utils/gf_omp.h"
static std::terminate_handler g_previousHandler = nullptr;
@@ -31,7 +31,7 @@ gridfire::NetIn init(const double temp, const double rho, const double tMax) {
std::setlocale(LC_ALL, "");
g_previousHandler = std::set_terminate(quill_terminate_handler);
quill::Logger* logger = fourdst::logging::LogManager::getInstance().getLogger("log");
logger->set_log_level(quill::LogLevel::TraceL2);
logger->set_log_level(quill::LogLevel::Info);
using namespace gridfire;
const std::vector<double> X = {0.7081145999999999, 2.94e-5, 0.276, 0.003, 0.0011, 9.62e-3, 1.62e-3, 5.16e-4};
@@ -143,7 +143,7 @@ void log_results(const gridfire::NetOut& netOut, const gridfire::NetIn& netIn) {
}
void record_abundance_history_callback(const gridfire::solver::CVODESolverStrategy::TimestepContext& ctx) {
void record_abundance_history_callback(const gridfire::solver::PointSolverTimestepContext& ctx) {
s_wrote_abundance_history = true;
const auto& engine = ctx.engine;
// std::unordered_map<std::string, std::pair<double, double>> abundances;
@@ -224,11 +224,12 @@ void quill_terminate_handler()
std::abort();
}
void callback_main(const gridfire::solver::CVODESolverStrategy::TimestepContext& ctx) {
void callback_main(const gridfire::solver::PointSolverTimestepContext& ctx) {
record_abundance_history_callback(ctx);
}
int main() {
GF_PAR_INIT();
using namespace gridfire;
constexpr size_t breaks = 1;
@@ -239,98 +240,20 @@ int main() {
const NetIn netIn = init(temp, rho, tMax);
policy::MainSequencePolicy stellarPolicy(netIn.composition);
policy::ConstructionResults construct = stellarPolicy.construct();
auto [engine, ctx_template] = stellarPolicy.construct();
std::println("Sandbox Engine Stack: {}", stellarPolicy);
std::println("Scratch Blob State: {}", *construct.scratch_blob);
std::println("Scratch Blob State: {}", *ctx_template);
constexpr size_t runs = 1000;
auto startTime = std::chrono::high_resolution_clock::now();
// arrays to store timings
std::array<std::chrono::duration<double>, runs> setup_times;
std::array<std::chrono::duration<double>, runs> eval_times;
std::array<NetOut, runs> serial_results;
for (size_t i = 0; i < runs; ++i) {
auto start_setup_time = std::chrono::high_resolution_clock::now();
std::print("Run {}/{}\r", i + 1, runs);
solver::CVODESolverStrategy solver(construct.engine, *construct.scratch_blob);
// solver.set_callback(solver::CVODESolverStrategy::TimestepCallback(callback_main));
solver.set_stdout_logging_enabled(false);
auto end_setup_time = std::chrono::high_resolution_clock::now();
std::chrono::duration<double> setup_elapsed = end_setup_time - start_setup_time;
setup_times[i] = setup_elapsed;
auto start_eval_time = std::chrono::high_resolution_clock::now();
const NetOut netOut = solver.evaluate(netIn);
auto end_eval_time = std::chrono::high_resolution_clock::now();
serial_results[i] = netOut;
std::chrono::duration<double> eval_elapsed = end_eval_time - start_eval_time;
eval_times[i] = eval_elapsed;
// log_results(netOut, netIn);
}
auto endTime = std::chrono::high_resolution_clock::now();
std::chrono::duration<double> elapsed = endTime - startTime;
std::println("");
// Summarize serial timings
double total_setup_time = 0.0;
double total_eval_time = 0.0;
for (size_t i = 0; i < runs; ++i) {
total_setup_time += setup_times[i].count();
total_eval_time += eval_times[i].count();
}
std::println("Average Setup Time over {} runs: {:.6f} seconds", runs, total_setup_time / runs);
std::println("Average Evaluation Time over {} runs: {:.6f} seconds", runs, total_eval_time / runs);
std::println("Total Time for {} runs: {:.6f} seconds", runs, elapsed.count());
std::println("Final H-1 Abundances Serial: {}", serial_results[0].composition.getMolarAbundance(fourdst::atomic::H_1));
// OPTIONAL: Prevent CppAD from returning memory to the system
// during execution to reduce overhead (can speed up tight loops)
CppAD::thread_alloc::hold_memory(true);
std::array<NetOut, runs> parallelResults;
std::array<std::chrono::duration<double>, runs> setupTimes;
std::array<std::chrono::duration<double>, runs> evalTimes;
std::array<std::unique_ptr<gridfire::engine::scratch::StateBlob>, runs> workspaces;
for (size_t i = 0; i < runs; ++i) {
workspaces[i] = construct.scratch_blob->clone_structure();
constexpr size_t nZones = 100;
std::array<NetIn, nZones> netIns;
for (size_t zone = 0; zone < nZones; ++zone) {
netIns[zone] = netIn;
netIns[zone].temperature = 1.0e7;
}
const solver::PointSolver localSolver(engine);
solver::GridSolverContext solverCtx(*ctx_template);
const solver::GridSolver gridSolver(engine, localSolver);
// Parallel runs
startTime = std::chrono::high_resolution_clock::now();
for (size_t i = 0; i < runs; ++i) {
auto start_setup_time = std::chrono::high_resolution_clock::now();
solver::CVODESolverStrategy solver(construct.engine, *workspaces[i]);
solver.set_stdout_logging_enabled(false);
auto end_setup_time = std::chrono::high_resolution_clock::now();
std::chrono::duration<double> setup_elapsed = end_setup_time - start_setup_time;
setupTimes[i] = setup_elapsed;
auto start_eval_time = std::chrono::high_resolution_clock::now();
parallelResults[i] = solver.evaluate(netIn);
auto end_eval_time = std::chrono::high_resolution_clock::now();
std::chrono::duration<double> eval_elapsed = end_eval_time - start_eval_time;
evalTimes[i] = eval_elapsed;
}
endTime = std::chrono::high_resolution_clock::now();
elapsed = endTime - startTime;
std::println("");
// Summarize parallel timings
total_setup_time = 0.0;
total_eval_time = 0.0;
for (size_t i = 0; i < runs; ++i) {
total_setup_time += setupTimes[i].count();
total_eval_time += evalTimes[i].count();
}
std::println("Average Parallel Setup Time over {} runs: {:.6f} seconds", runs, total_setup_time / runs);
std::println("Average Parallel Evaluation Time over {} runs: {:.6f} seconds", runs, total_eval_time / runs);
std::println("Total Parallel Time for {} runs: {:.6f} seconds", runs, elapsed.count());
std::println("Final H-1 Abundances Parallel: {}", utils::iterable_to_delimited_string(parallelResults, ",", [](const auto& result) {
return result.composition.getMolarAbundance(fourdst::atomic::H_1);
}));
std::vector<NetOut> netOuts = gridSolver.evaluate(solverCtx, netIns | std::ranges::to<std::vector>());
}