#include "config/config.hpp"

Classes
struct	BHData

Macros
#define	WARPSIZE 32
	Barnes-Hut warp size.

#define	MAXDEPTH 32
	Maximal depth of the Barnes-Hut tree branching.

Barnes-Hut thread count for different kernels.
#define	THREADS1 512

#define	THREADS2 1024

#define	THREADS3 1024

#define	THREADS4 1024

#define	THREADS5 256

Barnes-Hut block factor for different kernels.
block count = factor * number of blocks
#define	FACTOR1 2

#define	FACTOR2 1

#define	FACTOR3 1 /* must all be resident at the same time */

#define	FACTOR4 1 /* must all be resident at the same time */

#define	FACTOR5 4

Functions
void	setBHPrecision (float epssq, float itolsq)
	Barnes-Hut parameters setter.

void	allocBHmemCopy (int nbodies, BHData *bh_data)
	An allocation of the GPU device memory and an initialization where it is needed.

void	deallocBH (BHData *bh_data)
	A deallocation of the GPU device memory.

void	fill_bh_data (float const r, float const dip, BHData const *bh_data)
	Copy Barnes-Hut data to bhpara and copy particle data.

void	initBHgpu (int blocks)
	Barnes-Hut CUDA initialization.

void	buildBoxBH (int blocks)
	Building Barnes-Hut spatial min/max position box.

void	buildTreeBH (int blocks)
	Building Barnes-Hut tree in a linear child array representation of octant cells and particles inside.

void	summarizeBH (int blocks)
	Calculate octant cells masses and cell index counts. Determine cells centers of mass and total dipole moments on all possible levels of the Barnes-Hut tree.

void	sortBH (int blocks)
	Sort particle indexes according to the Barnes-Hut tree representation. Crucial for the per-warp performance tuning of `forceCalculationKernel` and `energyCalculationKernel`.

void	forceBH (BHData bh_data, float k, float f, float *torque)
	Barnes-Hut force calculation.

void	energyBH (BHData bh_data, float k, float E)
	Barnes-Hut energy calculation.

Macro Definition Documentation

◆ FACTOR1

#define FACTOR1 2

Definition at line 73 of file barnes_hut_gpu_cuda.cuh.

◆ FACTOR2

#define FACTOR2 1

Definition at line 74 of file barnes_hut_gpu_cuda.cuh.

◆ FACTOR3

#define FACTOR3 1 /* must all be resident at the same time */

Definition at line 75 of file barnes_hut_gpu_cuda.cuh.

◆ FACTOR4

#define FACTOR4 1 /* must all be resident at the same time */

Definition at line 76 of file barnes_hut_gpu_cuda.cuh.

◆ FACTOR5

#define FACTOR5 4

Definition at line 77 of file barnes_hut_gpu_cuda.cuh.

◆ MAXDEPTH

#define MAXDEPTH 32

Maximal depth of the Barnes-Hut tree branching.

Definition at line 83 of file barnes_hut_gpu_cuda.cuh.

◆ THREADS1

#define THREADS1 512

Definition at line 63 of file barnes_hut_gpu_cuda.cuh.

◆ THREADS2

#define THREADS2 1024

Definition at line 64 of file barnes_hut_gpu_cuda.cuh.

◆ THREADS3

#define THREADS3 1024

Definition at line 65 of file barnes_hut_gpu_cuda.cuh.

◆ THREADS4

#define THREADS4 1024

Definition at line 66 of file barnes_hut_gpu_cuda.cuh.

◆ THREADS5

#define THREADS5 256

Definition at line 67 of file barnes_hut_gpu_cuda.cuh.

◆ WARPSIZE

#define WARPSIZE 32

Barnes-Hut warp size.

Definition at line 81 of file barnes_hut_gpu_cuda.cuh.

Function Documentation

◆ allocBHmemCopy()

void allocBHmemCopy	(	int	nbodies,
		BHData *	bh_data
	)

An allocation of the GPU device memory and an initialization where it is needed.

Definition at line 1177 of file barnes_hut_gpu_cuda.cu.

References BHData::blocks, BHData::child, BHData::count, cuda_get_device(), cuda_get_device_props(), cuda_safe_mem, BHData::err, FACTOR1, BHData::mass, BHData::max_lps, BHData::maxp, BHData::minp, BHData::nbodies, BHData::nnodes, BHData::r, BHData::sort, BHData::start, and BHData::u.

◆ buildBoxBH()

void buildBoxBH ( int blocks )

Building Barnes-Hut spatial min/max position box.

Definition at line 1040 of file barnes_hut_gpu_cuda.cu.

References block(), cuda_safe_mem, FACTOR1, KERNELCALL, and THREADS1.

◆ buildTreeBH()

void buildTreeBH ( int blocks )

Building Barnes-Hut tree in a linear child array representation of octant cells and particles inside.

Definition at line 1054 of file barnes_hut_gpu_cuda.cu.

References block(), cuda_safe_mem, FACTOR2, KERNELCALL, and THREADS2.

◆ deallocBH()

void deallocBH ( BHData * bh_data )

A deallocation of the GPU device memory.

Definition at line 1152 of file barnes_hut_gpu_cuda.cu.

References BHData::child, BHData::count, cuda_safe_mem, BHData::err, BHData::mass, BHData::max_lps, BHData::maxp, BHData::minp, BHData::r, BHData::sort, BHData::start, and BHData::u.

Referenced by DipolarBarnesHutGpu::~DipolarBarnesHutGpu().

◆ energyBH()

void energyBH	(	BHData *	bh_data,
		float	k,
		float *	E
	)

Barnes-Hut energy calculation.

Definition at line 1113 of file barnes_hut_gpu_cuda.cu.

References block(), BHData::blocks, cuda_safe_mem, energySum, BHData::err, FACTOR5, KERNELCALL_shared, and THREADS5.

Referenced by DipolarBarnesHutGpu::long_range_energy().

◆ fill_bh_data()

void fill_bh_data	(	float const *	r,
		float const *	dip,
		BHData const *	bh_data
	)

Copy Barnes-Hut data to bhpara and copy particle data.

Parameters

r	device particle positions to copy
dip	device particle dipoles to copy
bh_data	Barnes-Hut container

Definition at line 1258 of file barnes_hut_gpu_cuda.cu.

References bhpara, cuda_safe_mem, BHData::nbodies, BHData::r, and BHData::u.

◆ forceBH()

void forceBH	(	BHData *	bh_data,
		float	k,
		float *	f,
		float *	torque
	)

Barnes-Hut force calculation.

Definition at line 1094 of file barnes_hut_gpu_cuda.cu.

References block(), BHData::blocks, cuda_safe_mem, BHData::err, f, FACTOR5, KERNELCALL, THREADS5, and torque.

Referenced by DipolarBarnesHutGpu::add_long_range_forces().

◆ initBHgpu()

void initBHgpu ( int blocks )

Barnes-Hut CUDA initialization.

Definition at line 1019 of file barnes_hut_gpu_cuda.cu.

References block(), FACTOR5, initializationKernel(), KERNELCALL, and THREADS5.

◆ setBHPrecision()

void setBHPrecision	(	float	epssq,
		float	itolsq
	)

Barnes-Hut parameters setter.

Definition at line 1145 of file barnes_hut_gpu_cuda.cu.

References cuda_safe_mem, epssqd, and itolsqd.

Referenced by DipolarBarnesHutGpu::on_activation().

◆ sortBH()

void sortBH ( int blocks )

Sort particle indexes according to the Barnes-Hut tree representation. Crucial for the per-warp performance tuning of forceCalculationKernel and energyCalculationKernel.

Definition at line 1082 of file barnes_hut_gpu_cuda.cu.

References block(), cuda_safe_mem, FACTOR4, KERNELCALL, and THREADS4.

◆ summarizeBH()

void summarizeBH ( int blocks )

Calculate octant cells masses and cell index counts. Determine cells centers of mass and total dipole moments on all possible levels of the Barnes-Hut tree.

Definition at line 1068 of file barnes_hut_gpu_cuda.cu.

References block(), cuda_safe_mem, FACTOR3, KERNELCALL, and THREADS3.

Classes

Macros

Functions