MCMCProcessorGPU Class Reference

Class responsible for performing MCMC Processing with GPU. More...

Public Member Functions
	MCMCProcessorGPU ()
	constructor More...
virtual	~MCMCProcessorGPU ()
	destructor More...
__host__ void	InitGPU_AutoCorr (int n_Entries, int n_Pars, const int n_Lags)
	KS: Initialiser, here we allocate memory for variables and copy constants. More...
__host__ void	CopyToGPU_AutoCorr (float *ParStep_cpu, float *NumeratorSum_cpu, float *ParamSums_cpu, float *DenomSum_cpu)
	KS: Copy necessary variables from CPU to GPU. More...
__host__ void	RunGPU_AutoCorr (float *NumeratorSum_cpu, float *DenomSum_cpu)
	KS: This call the main kernel responsible for calculating LagL and later copy results back to CPU. More...
__host__ void	CleanupGPU_AutoCorr ()
	KS: free memory on gpu. More...

Private Attributes
float *	ParStep_gpu
	Value of each param at GPU. More...
float *	NumeratorSum_gpu
float *	ParamSums_gpu
float *	DenomSum_gpu
int	h_nLag
int	h_nDraws
int	h_nEntries

Detailed Description

Class responsible for performing MCMC Processing with GPU.

Author

Kamil Skwarczynski

Definition at line 29 of file gpuMCMCProcessorUtils.cuh.

Constructor & Destructor Documentation

MCMCProcessorGPU()

MCMCProcessorGPU::MCMCProcessorGPU

(

)

constructor

Definition at line 16 of file gpuMCMCProcessorUtils.cu.

                                   {
  ParStep_gpu = nullptr;
  NumeratorSum_gpu = nullptr;
  ParamSums_gpu = nullptr;
  DenomSum_gpu = nullptr;
 
  h_nLag     = -1;
  h_nDraws   = -1;
  h_nEntries = -1;
}

~MCMCProcessorGPU()

MCMCProcessorGPU::~MCMCProcessorGPU ( )

virtual

destructor

Definition at line 27 of file gpuMCMCProcessorUtils.cu.

                                    {
}

Member Function Documentation

CleanupGPU_AutoCorr()

__host__ void MCMCProcessorGPU::CleanupGPU_AutoCorr

(

)

KS: free memory on gpu.

Definition at line 191 of file gpuMCMCProcessorUtils.cu.

                                                    {
// *********************************
  if(ParStep_gpu)      cudaFree(ParStep_gpu);
  CudaCheckError();
  if(NumeratorSum_gpu) cudaFree(NumeratorSum_gpu);
  CudaCheckError();
  if(ParamSums_gpu)    cudaFree(ParamSums_gpu);
  CudaCheckError();
  if(DenomSum_gpu)     cudaFree(DenomSum_gpu);
  CudaCheckError();
 
  printf(" Cleared memory at GPU, I am free \n");
}

CopyToGPU_AutoCorr()

__host__ void MCMCProcessorGPU::CopyToGPU_AutoCorr	(	float *	ParStep_cpu,
		float *	NumeratorSum_cpu,
		float *	ParamSums_cpu,
		float *	DenomSum_cpu
	)

KS: Copy necessary variables from CPU to GPU.

Parameters

ParStep_cpu	Parameter value at each step
NumeratorSum_cpu	Sum used for nominator of autocorrelation calculations
ParamSums_cpu	Overall sum for each parameter over all steps
DenomSum_cpu	Sum used for denominator of autocorrelation calculations

Definition at line 79 of file gpuMCMCProcessorUtils.cu.

                                                 {
// ******************************************************
  //store value of parameter for each step
  cudaMemcpy(ParStep_gpu, ParStep_cpu, h_nDraws*h_nEntries*sizeof(float), cudaMemcpyHostToDevice);
  CudaCheckError();
 
  //Mean value for a given parameter
  cudaMemcpy(ParamSums_gpu, ParamSums_cpu, h_nDraws*sizeof(float), cudaMemcpyHostToDevice);
  CudaCheckError();
 
  //Numerator which is directly used for calculating LagL
  cudaMemcpy(NumeratorSum_gpu, NumeratorSum_cpu, h_nLag*h_nDraws*sizeof(float), cudaMemcpyHostToDevice);
  CudaCheckError();
 
  //Denominator which is directly used for calculating LagL
  cudaMemcpy(DenomSum_gpu, DenomSum_cpu, h_nLag*h_nDraws*sizeof(float), cudaMemcpyHostToDevice);
  CudaCheckError();
}

InitGPU_AutoCorr()

__host__ void MCMCProcessorGPU::InitGPU_AutoCorr	(	int	n_Entries,
		int	n_Pars,
		const int	n_Lags
	)

KS: Initialiser, here we allocate memory for variables and copy constants.

Parameters

n_Entries	Total number of entries in mcmc chain
n_Pars	Number of relevant parameters
n_Lags	Value of Lag in autocreation calculation

Definition at line 32 of file gpuMCMCProcessorUtils.cu.

                                            {
// *******************************************
  // Write to the global statics (h_* denotes host stored variable)
  h_nDraws = n_Pars;
  h_nLag = n_Lags;
  h_nEntries = n_Entries;
 
  // Copy the constants
  cudaMemcpyToSymbol(d_nLag,   &h_nLag,   sizeof(h_nLag));
  CudaCheckError();
 
  cudaMemcpyToSymbol(d_nDraws, &h_nDraws, sizeof(h_nDraws));
  CudaCheckError();
 
  cudaMemcpyToSymbol(d_nEntries, &h_nEntries, sizeof(h_nEntries));
  CudaCheckError();
 
  // Allocate chunks of memory to GPU
  //Numerator which is directly used for calculating LagL
  cudaMalloc((void **) &NumeratorSum_gpu, h_nLag*h_nDraws*sizeof(float));
  CudaCheckError();
 
  //Denominator which is directly used for calculating LagL
  cudaMalloc((void **) &DenomSum_gpu, h_nLag*h_nDraws*sizeof(float));
  CudaCheckError();
 
  //Mean value for a given parameter
  cudaMalloc((void **) &ParamSums_gpu, h_nDraws*sizeof(float));
  CudaCheckError();
 
  //store value of parameter for each step
  cudaMalloc((void **) &ParStep_gpu, h_nDraws*h_nEntries*sizeof(float*));
  CudaCheckError();
 
  printf(" Allocated in total %f MB for autocorrelations calculations on GPU\n",
         static_cast<double>(sizeof(float) * (h_nLag * h_nDraws + h_nLag * h_nDraws + h_nDraws + h_nDraws * h_nEntries)) / 1.0e6);
}

RunGPU_AutoCorr()

__host__ void MCMCProcessorGPU::RunGPU_AutoCorr	(	float *	NumeratorSum_cpu,
		float *	DenomSum_cpu
	)

KS: This call the main kernel responsible for calculating LagL and later copy results back to CPU.

Parameters

NumeratorSum_cpu	Sum used for nominator of autocorrelation calculations
DenomSum_cpu	Sum used for denominator of autocorrelation calculations

Definition at line 159 of file gpuMCMCProcessorUtils.cu.

                                                                      {
// *****************************************
  dim3 block_size;
  dim3 grid_size;
 
  block_size.x = _BlockSize_;
  grid_size.x = (h_nLag*h_nDraws / block_size.x) + 1;
 
  EvalOnGPU_AutoCorr<<<grid_size, block_size>>>(
      ParStep_gpu,
      ParamSums_gpu,
      NumeratorSum_gpu,
      DenomSum_gpu);
  CudaCheckError();
 
  printf(" Finished calculating now copying results back to CPU \n");
 
  //KS: Finally copy paste memory from GPU to CPU
  cudaMemcpy(NumeratorSum_cpu, NumeratorSum_gpu, h_nLag*h_nDraws*sizeof(float), cudaMemcpyDeviceToHost);
  CudaCheckError();
 
  cudaMemcpy(DenomSum_cpu, DenomSum_gpu, h_nLag*h_nDraws*sizeof(float), cudaMemcpyDeviceToHost);
  CudaCheckError();
}

Member Data Documentation

DenomSum_gpu

float* MCMCProcessorGPU::DenomSum_gpu

private

Definition at line 72 of file gpuMCMCProcessorUtils.cuh.

h_nDraws

int MCMCProcessorGPU::h_nDraws

private

Definition at line 76 of file gpuMCMCProcessorUtils.cuh.

h_nEntries

int MCMCProcessorGPU::h_nEntries

private

Definition at line 77 of file gpuMCMCProcessorUtils.cuh.

h_nLag

int MCMCProcessorGPU::h_nLag

private

Definition at line 75 of file gpuMCMCProcessorUtils.cuh.

NumeratorSum_gpu

float* MCMCProcessorGPU::NumeratorSum_gpu

private

Definition at line 70 of file gpuMCMCProcessorUtils.cuh.

ParamSums_gpu

float* MCMCProcessorGPU::ParamSums_gpu

private

Definition at line 71 of file gpuMCMCProcessorUtils.cuh.

ParStep_gpu

float* MCMCProcessorGPU::ParStep_gpu

private

Value of each param at GPU.

Definition at line 69 of file gpuMCMCProcessorUtils.cuh.

---

The documentation for this class was generated from the following files:

• Fitters/gpuMCMCProcessorUtils.cuh
• Fitters/gpuMCMCProcessorUtils.cu