diff options
| author | Claudius "keldu" Holeksa <mail@keldu.de> | 2026-11-11 01:16:44 +0100 |
|---|---|---|
| committer | Claudius "keldu" Holeksa <mail@keldu.de> | 2026-11-11 01:16:44 +0100 |
| commit | 0c0087ef9482634273ec1a23d3a915754f94926d (patch) | |
| tree | 972a8e48c1fae67af23b982d7a9e1a506fc8705f /examples/poiseulle_3d_gpu_minimal/sim.cpp | |
| parent | 2e63c4a31f946a8bafa92c49c1a0a116c0eb6dae (diff) | |
| download | libs-lbm-0c0087ef9482634273ec1a23d3a915754f94926d.tar.gz | |
Snapshot. Migration to Lbm Unit
Diffstat (limited to 'examples/poiseulle_3d_gpu_minimal/sim.cpp')
| -rw-r--r-- | examples/poiseulle_3d_gpu_minimal/sim.cpp | 433 |
1 files changed, 433 insertions, 0 deletions
diff --git a/examples/poiseulle_3d_gpu_minimal/sim.cpp b/examples/poiseulle_3d_gpu_minimal/sim.cpp new file mode 100644 index 0000000..0bc4642 --- /dev/null +++ b/examples/poiseulle_3d_gpu_minimal/sim.cpp @@ -0,0 +1,433 @@ +#include <kel/lbm/sycl/lbm.hpp> +#include <kel/lbm/lbm.hpp> +#include <kel/lbm/particle.hpp> + +#include <forstio/io/io.hpp> +#include <forstio/remote/filesystem/easy.hpp> +#include <forstio/codec/json/json.hpp> +#include <forstio/codec/simple.hpp> + +namespace kel { +namespace lbm { + +// Deliberately tiny: 32 x 16 x 16 = 8192 cells, so a full run finishes in +// seconds and the inlet signal reaches the outlet in ~32 steps. The full size +// case lives in poiseulle_3d_gpu. +constexpr uint64_t dim_x = 32ul; +constexpr uint64_t dim_y = 16ul; +constexpr uint64_t dim_z = 16ul; +constexpr uint64_t particle_size = 4ul; + +namespace sch { + +using LbmArgs = Args< + Struct< + Member<String, "name">, + Member<UInt64, "time_steps"> + >, + Tuple<> +>; + +using InfoChunk = Chunk<UInt8, 0u, dim_x, dim_y, dim_z>; + +template<typename T, typename Desc> +using DfChunk = Chunk<FixedArray<T,Desc::Q>, 1u, dim_x, dim_y, dim_z>; + +template<typename T, typename Desc> +using ScalarChunk = Chunk<Scalar<T>, 0u, dim_x, dim_y, dim_z>; + +template<typename T, typename Desc> +using VectorChunk = Chunk<Vector<T,Desc::D>, 0u, dim_x, dim_y, dim_z>; + +template<typename T, typename Desc> +using ChunkStruct = Struct< + Member<InfoChunk, "info">, + Member<DfChunk<T,Desc>, "dfs">, + Member<DfChunk<T,Desc>, "dfs_old">, + Member<VectorChunk<T,Desc>, "particle_N">, + Member<ScalarChunk<T,Desc>, "particle_D"> +>; + +template<typename T, typename Desc> +using MomChunk = Chunk<Vector<T,Desc::D>, 0u, dim_x, dim_y, dim_z>; + +template<typename T> +using RhoChunk = Chunk<Scalar<T>, 0u, dim_x, dim_y, dim_z>; + +template<typename T, typename Desc> +using MacroStruct = Struct< + Member<MomChunk<T,Desc>, "velocity">, + Member<RhoChunk<T>, "density">, + Member<ScalarChunk<T,Desc>, "porosity"> +>; +} + +template<typename T, typename Desc> +kel::lbm::error_or<void> setup_initial_conditions( + converter<T>& conv, + kel::lbm::data<sch::ChunkStruct<T,Desc>>& fields, + kel::lbm::data<sch::MacroStruct<T,Desc>>& macros, + kel::lbm::data<sch::FixedArray<sch::Particle<T,Desc::D>, particle_size>>& particles +){ + // Unit conversion is not exercised in this reduced case. + (void) conv; + + auto& info_f = fields.template get<"info">(); + // Set everything as walls + iterator<Desc::D>::apply( + [&](auto& index){ + info_f.at(index).set(1u); + }, + {}, + info_f.get_dims(), + {} + ); + // Fluid + iterator<Desc::D>::apply( + [&](auto& index){ + info_f.at(index).set(2u); + }, + {}, + info_f.get_dims(), + {{1u,1u,1u}} + ); + + // Inflow + iterator<Desc::D>::apply( + [&](auto& index){ + info_f.at(index).set(3u); + }, + {{0u,0u,0u}}, + {{1u,dim_y,dim_z}}, + {{0u,1u,1u}} + ); + + // Outflow + iterator<Desc::D>::apply( + [&](auto& index){ + info_f.at(index).set(4u); + }, + {{dim_x-1u,0u,0u}}, + {{dim_x,dim_y,dim_z}}, + {{0u,1u,1u}} + ); + // + auto& df_f = fields.template get<"dfs_old">(); + auto& rho_f = macros.template get<"density">(); + auto& mom_f = macros.template get<"velocity">(); + auto& por_f = macros.template get<"porosity">(); + + iterator<Desc::D>::apply( + [&](auto& index){ + auto& df = df_f.at(index); + auto& rho = rho_f.at(index); + por_f.at(index).at({}) = {1}; + rho.at({}) = {1}; + auto& mom = mom_f.at(index); + auto eq = equilibrium<T,Desc>(rho,mom); + + df = eq; + }, + {},// 0-index + df_f.get_dims() + ); + + iterator<Desc::D>::apply( + [&](auto& index){ + auto& df = df_f.at(index); + auto& rho = rho_f.at(index); + rho.at({}) = {1}; + auto& mom = mom_f.at(index); + mom.at({{0u}}) = 0.01; + auto eq = equilibrium<T,Desc>(rho,mom); + + df = eq; + }, + {},// 0-index + df_f.get_dims(), + {{1u,1u,1u}} + ); + + // No particles in this case -- the array is carried only so the schema + // matches poiseulle_3d_gpu. + (void) particles; + + return kel::lbm::make_void(); +} + +template<typename T, typename Desc> +kel::lbm::error_or<void> step( + converter<T>& conv, + kel::lbm::data<sch::Ptr<sch::ChunkStruct<T,Desc>>,encode::Sycl<kel::lbm::enc::Native>>& fields, + kel::lbm::data<sch::Ptr<sch::MacroStruct<T,Desc>>,encode::Sycl<kel::lbm::enc::Native>>& macros, + kel::lbm::data<sch::UInt64> t_i, + device& dev +){ + auto& q = dev.get_handle(); + auto& info_f = fields.template get<"info">(); + + // auto coll_ev = + q.submit([&](acpp::sycl::handler& h){ + // Need nicer things to handle the flow. I see improvement here + component<T,Desc,cmpt::BGK, encode::Sycl<kel::lbm::enc::Native>> collision{conv.template kinematic_viscosity_si_to_tau<Desc>({0.1})}; + // component<T,Desc,cmpt::HLBM,encode::Sycl<kel::lbm::enc::Native>> collision{0.6}; + component<T,Desc,cmpt::BounceBack,encode::Sycl<kel::lbm::enc::Native>> bb; + + component<T,Desc,cmpt::ZouHeHorizontal<true>,encode::Sycl<kel::lbm::enc::Native>> flow_in{ + [&](){ + uint64_t target_t_i = 256u; + if(t_i.get() < target_t_i){ + return 1.0 + (0.001 / target_t_i) * t_i.get(); + } + return 1.001; + }() + }; + component<T,Desc,cmpt::ZouHeHorizontal<false>,encode::Sycl<kel::lbm::enc::Native>> flow_out{1.0}; + + h.parallel_for(acpp::sycl::range<Desc::D>{dim_x,dim_y,dim_z}, [=](acpp::sycl::id<Desc::D> idx){ + kel::lbm::data<sch::FixedArray<sch::UInt64,Desc::D>> index; + for(uint64_t i = 0u; i < Desc::D; ++i){ + index.at({{i}}).set(idx[i]); + } + + auto info = info_f.at(index); + + switch(info.get()){ + case 0u: + break; + case 1u: + bb.apply(fields,index,t_i); + break; + case 2u: + collision.apply(fields,macros,index,t_i); + break; + case 3u: + flow_in.apply(fields,index,t_i); + collision.apply(fields,macros,index,t_i); + break; + case 4u: + flow_out.apply(fields,index,t_i); + collision.apply(fields,macros,index,t_i); + break; + default: + break; + } + }); + }).wait(); + + q.submit([&](acpp::sycl::handler& h){ + component<T,Desc,cmpt::Stream,encode::Sycl<kel::lbm::enc::Native>> stream; + + h.parallel_for(acpp::sycl::range<Desc::D>{dim_x,dim_y,dim_z}, [=](acpp::sycl::id<Desc::D> idx){ + kel::lbm::data<sch::FixedArray<sch::UInt64,Desc::D>> index; + for(uint64_t i = 0u; i < Desc::D; ++i){ + index.at({{i}}).set(idx[i]); + } + + auto info = info_f.at(index); + + if(info.get() > 0u){ + stream.apply(fields,index,t_i); + } + }); + }).wait(); + + // Step + /* + q.submit([&](acpp::sycl::handler& h){ + // h.depends_on(collision_ev); + }).wait(); + */ + + return kel::lbm::make_void(); +} +} +} + +template<typename T, typename Desc> +kel::lbm::error_or<void> lbm_main(int argc, char** argv){ + using namespace kel::lbm; + + kel::lbm::data<sch::LbmArgs> args; + { + auto& args_n = args.template get<"args">(); + auto& name = args_n.template get<"name">(); + name.set("poiseulle_3d_gpu_minimal"); + auto& steps_arg = args_n.template get<"time_steps">(); + steps_arg.set(64u); + } + { + auto eov = kel::lbm::parse_args(args,argc,argv); + if(eov.is_error()){ + return eov; + } + } + auto eo_lbm_env = init_lbm_env(args); + if(eo_lbm_env.is_error()){ + return std::move(eo_lbm_env.get_error()); + } + const auto& lbm_env = eo_lbm_env.get_value(); + + auto& out_dir = lbm_env.name_dir; + + { + std::error_code ec; + std::filesystem::create_directories(out_dir,ec); + if(ec != std::errc{}){ + return kel::lbm::make_error<kel::lbm::err::critical>("Could not create output directory"); + } + } + + converter<T> conv { + // delta_x + {1.0}, + // delta_t + {1.0} + }; + + print_lbm_meta<T,Desc>(conv,{0.1},{0.01},{1024.0}); + + auto lbm_data_ptr = kel::lbm::heap<kel::lbm::data<sch::ChunkStruct<T,Desc>>>(); + auto lbm_macro_data_ptr = kel::lbm::heap<kel::lbm::data<sch::MacroStruct<T,Desc>>>(); + auto lbm_particle_data_ptr = kel::lbm::heap<kel::lbm::data<sch::FixedArray<sch::Particle<T,Desc::D>, particle_size>>>(); + + std::cout<<"Estimated Bytes: "<<memory_estimate<sch::ChunkStruct<T,Desc>,sch::MacroStruct<T,Desc>>().get()<<std::endl; + + auto eo_aio = saw::setup_async_io(); + if(eo_aio.is_error()){ + return kel::lbm::make_error<kel::lbm::err::critical>( + std::string{eo_aio.get_error().get_message()}); + } + auto& aio = eo_aio.get_value(); + saw::wait_scope wait{aio.event_loop}; + + bool krun = true; + bool print_status = false; + aio.event_port.on_signal(saw::Signal::Terminate).then([&](){ + krun = false; + }).detach(); + aio.event_port.on_signal(saw::Signal::User1).then([&](){ + print_status = true; + }).detach(); + + device dev; + + auto& sycl_q = dev.get_handle(); + + sycl_q.wait(); + { + auto eov = setup_initial_conditions<T,Desc>(conv,*lbm_data_ptr,*lbm_macro_data_ptr,*lbm_particle_data_ptr); + if(eov.is_error()){ + return eov; + } + } + + kel::lbm::data<sch::ChunkStruct<T,Desc>, encode::Sycl<kel::lbm::enc::Native>> lbm_sycl_data{sycl_q}; + kel::lbm::data<sch::MacroStruct<T,Desc>, encode::Sycl<kel::lbm::enc::Native>> lbm_sycl_macro_data{sycl_q}; + kel::lbm::data<sch::FixedArray<sch::Particle<T,Desc::D>, particle_size>, encode::Sycl<kel::lbm::enc::Native>> lbm_sycl_particle_data{sycl_q}; + sycl_q.wait(); + + auto lsd_view = make_view(lbm_sycl_data); + auto lsdm_view = make_view(lbm_sycl_macro_data); + { + auto eov = dev.copy_to_device(*lbm_data_ptr,lbm_sycl_data); + if(eov.is_error()){ + return eov; + } + } + { + auto eov = dev.copy_to_device(*lbm_macro_data_ptr,lbm_sycl_macro_data); + if(eov.is_error()){ + return eov; + } + } + { + auto eov = dev.copy_to_device(*lbm_particle_data_ptr,lbm_sycl_particle_data); + if(eov.is_error()){ + return eov; + } + } + sycl_q.wait(); + kel::lbm::data<sch::UInt64> time_steps = args.template get<"args">().template get<"time_steps">(); + + for(kel::lbm::data<sch::UInt64> i{0u}; i < time_steps and krun; ++i){ + { + { + auto eov = write_vtk_file(out_dir,"t",i.get(),*lbm_macro_data_ptr); + if(eov.is_error()){ + return eov; + } + } + /*{ + std::string file_name = "p_"; + file_name += std::to_string(i.get()); + file_name += ".json"; + auto eov = saw::easy::encode_and_write_file<sch::FixedArray<sch::Particle<T,Desc::D>,particle_size>,saw::encode::Json>(out_dir, *lbm_particle_data_ptr); + if(eov.is_error()){ + return eov; + } + }*/ + } + { + auto eov = step<T,Desc>(conv,lsd_view,lsdm_view,i,dev); + if(eov.is_error()){ + return eov; + } + } + { + auto eov = dev.copy_to_host(lbm_sycl_macro_data,*lbm_macro_data_ptr); + if(eov.is_error()){ + return eov; + } + } + { + auto eov = dev.copy_to_host(lbm_sycl_data,*lbm_data_ptr); + if(eov.is_error()){ + return eov; + } + } + + wait.poll(); + if(print_status){ + std::cout<<"Status: "<<i.get()<<" of "<<time_steps.get()<<" - "<<(i.template cast_to<sch::Float64>().get() * 100 / time_steps.get())<<"%"<<std::endl; + print_status = false; + } + } + sycl_q.wait(); + { + auto eov = write_vtk_file(out_dir,"t",time_steps.get(),*lbm_macro_data_ptr); + if(eov.is_error()){ + return eov; + } + } + + /* + iterator<Desc::D>::apply( + [&](auto& index){ + std::cout<<index.at({0u}).get()<<" "<<index.at({1u}).get()<<" "<<lbm_data.get<"info">().at(index).template cast_to<sch::UInt16>().get()<<"\n"; + }, + {{0u,0u}}, + {{dim_x, dim_y}} + ); + */ + + sycl_q.wait(); + return kel::lbm::make_void(); +} + +using FloatT = kel::lbm::sch::Float32; + +int main(int argc, char** argv){ + auto eov = lbm_main<FloatT,kel::lbm::sch::D3Q27>(argc, argv); + if(eov.is_error()){ + auto& err = eov.get_error(); + std::cerr<<"[Error] "<<err.get_category(); + auto err_msg = err.get_message(); + if(err_msg.size() > 0u){ + std::cerr<<" - "<<err_msg; + } + std::cerr<<std::endl; + return err.get_id(); + } + return 0; +} |
