Commit 82acde64 authored by Yury Hayeu's avatar Yury Hayeu
Browse files

Optimization fixes for tnl-grid and parallel for

parent 66648e73

src/Benchmarks/HeatEquationGrid/HeatmapNDimGrid/implementation.h

+11 −7
Original line number Diff line number Diff line
@@ -359,11 +359,14 @@ bool HeatmapSolver<Real>::solve(const HeatmapSolver<Real>::Parameters &params) c 


   const Real hx = params.xDomainSize / (Real)grid.getDimension(0);

   const Real hy = params.yDomainSize / (Real)grid.getDimension(1);

   const Real hx_inv = 1 / (hx * hx);

   const Real hy_inv = 1 / (hy * hy);

   const Real hx_inv = 1. / (hx * hx);

   const Real hy_inv = 1. / (hy * hy);



   auto entitiesCount = grid.getEntitiesCount(0);

   auto timestep = params.timeStep ? params.timeStep : std::min(hx * hx, hy * hy);

   auto xDomainSize = params.xDomainSize;

   auto yDomainSize = params.yDomainSize;

   auto sigma = params.sigma;



   TNL::Containers::Array<Real, Device> ux(entitiesCount), // data at step u

                                        aux(entitiesCount);// data at step u + 1
@@ -378,10 +381,10 @@ bool HeatmapSolver<Real>::solve(const HeatmapSolver<Real>::Parameters &params) c 
      auto position = entity.getCoordinates();

      auto index = entity.getIndex();



      auto x = position[0] * hx - params.xDomainSize / 2;

      auto y = position[1] * hx - params.yDomainSize / 2;

      auto x = position[0] * hx - xDomainSize / 2;

      auto y = position[1] * hx - yDomainSize / 2;



      uxView[index] = exp(params.sigma * (x * x + y * y));

      uxView[index] = exp(sigma * (x * x + y * y));

   };



   const Container<2, bool> direction{ false, false };
@@ -398,9 +401,10 @@ bool HeatmapSolver<Real>::solve(const HeatmapSolver<Real>::Parameters &params) c 


   auto next = [=] __cuda_callable__(const GridEntity<2, int>& entity) mutable {

      auto index = entity.getIndex();

      auto center = 2 * uxView[index];



      auxView[index] = (uxView[index - 1] - 2 * uxView[index] + uxView[index + 1]) * hx_inv +

                       (uxView[index - xDimension] - 2 * uxView[index] + uxView[index + xDimension]) * hy_inv;

      auxView[index] = (uxView[index - 1] - center + uxView[index + 1]) * hx_inv +

                       (uxView[index - xDimension] - center + uxView[index + xDimension]) * hy_inv;

   };



   auto update = [=] __cuda_callable__(int i) mutable {

src/Benchmarks/HeatEquationGrid/HeatmapParallelFor/implementation.h

+15 −9
Original line number Diff line number Diff line
@@ -39,21 +39,26 @@ bool HeatmapSolver<Real>::solve(const HeatmapSolver<Real>::Parameters &params) c 


   const Real hx = params.xDomainSize / (Real)params.xSize;

   const Real hy = params.yDomainSize / (Real)params.ySize;

   const Real hx_inv = 1 / (hx * hx);

   const Real hy_inv = 1 / (hy * hy);

   const Real hx_inv = 1. / (hx * hx);

   const Real hy_inv = 1. / (hy * hy);



   auto timestep = params.timeStep ? params.timeStep : std::min(hx * hx, hy * hy);



   auto uxView = ux.getView(), auxView = aux.getView();



   auto xSize = params.xSize;

   auto xDomainSize = params.xDomainSize;

   auto yDomainSize = params.yDomainSize;

   auto sigma = params.sigma;



   auto init = [=] __cuda_callable__(int i, int j) mutable

   {

      auto index = j * params.xSize + i;

      auto index = j * xSize + i;



      auto x = i * hx - params.xDomainSize / 2;

      auto y = j * hy - params.yDomainSize / 2;

      auto x = i * hx - xDomainSize / 2.;

      auto y = j * hy - yDomainSize / 2.;



      uxView[index] = exp(params.sigma * (x * x + y * y));

      uxView[index] = exp(sigma * (x * x + y * y));

   };



   TNL::Algorithms::ParallelFor2D<Device>::exec(1, 1, params.xSize - 1, params.ySize - 1, init);
@@ -63,10 +68,11 @@ bool HeatmapSolver<Real>::solve(const HeatmapSolver<Real>::Parameters &params) c 


   auto next = [=] __cuda_callable__(int i, int j) mutable

   {

      auto index = j * params.xSize + i;

      auto index = j * xSize + i;

      auto center = 2 * uxView[index];



      auxView[index] = (uxView[index - 1] - 2 * uxView[index] + uxView[index + 1]) * hx_inv +

                       (uxView[index - params.xSize] - 2 * uxView[index] + uxView[index + params.xSize]) * hy_inv;

      auxView[index] = (uxView[index - 1] - center + uxView[index + 1]) * hx_inv +

                       (uxView[index - xSize] - center + uxView[index + xSize]) * hy_inv;

   };



   auto update = [=] __cuda_callable__(int i) mutable

src/Benchmarks/HeatEquationGrid/HeatmapTNLGrid/implementation.h

+13 −8
Original line number Diff line number Diff line
@@ -26,21 +26,25 @@ bool HeatmapSolver<Real>::solve(const HeatmapSolver<Real>::Parameters& params) c 


   const Real hx = params.xDomainSize / (Real)params.xSize;

   const Real hy = params.yDomainSize / (Real)params.ySize;

   const Real hx_inv = 1 / (hx * hx);

   const Real hy_inv = 1 / (hy * hy);

   const Real hx_inv = 1. / (hx * hx);

   const Real hy_inv = 1. / (hy * hy);



   auto timestep = params.timeStep ? params.timeStep : std::min(hx * hx, hy * hy);



   auto uxView = ux.getView(),

        auxView = aux.getView();



   auto xDomainSize = params.xDomainSize;

   auto yDomainSize = params.yDomainSize;

   auto sigma = params.sigma;



   auto init = [=] __cuda_callable__(const typename Grid2D::EntityType<0> &entity) mutable {

      auto index = entity.getIndex();



      auto x = entity.getCoordinates().x() * hx - params.xDomainSize / 2;

      auto y = entity.getCoordinates().y() * hy - params.yDomainSize / 2;

      auto x = entity.getCoordinates().x() * hx - xDomainSize / 2.;

      auto y = entity.getCoordinates().y() * hy - yDomainSize / 2.;



      uxView[index] = exp(params.sigma * (x * x + y * y));

      uxView[index] = exp(sigma * (x * x + y * y));

   };



   grid.template forInterior<0>(init);
@@ -48,12 +52,13 @@ bool HeatmapSolver<Real>::solve(const HeatmapSolver<Real>::Parameters& params) c 
   if (!writeGNUPlot("data.txt", params, ux))

      return false;



   auto width = grid.getDimensions().x() + 1;

   auto next = [=] __cuda_callable__(const typename Grid2D::EntityType<0>&entity) mutable {

      auto index = entity.getIndex();

      auto width = grid.getDimensions().x() + 1;

      auto center = 2 * uxView[index];



      auxView[index] = (uxView[index - 1] - 2 * uxView[index] + uxView[index + 1]) * hx_inv +

                        (uxView[index - width] - 2 * uxView[index] + uxView[index + width]) * hy_inv;

      auxView[index] = (uxView[index - 1] - center + uxView[index + 1]) * hx_inv +

                        (uxView[index - width] - center + uxView[index + width]) * hy_inv;

   };



   auto update = [=] __cuda_callable__(const typename Grid2D::EntityType<0>&entity) mutable {