GPU610/DPS915 | Student List | Group and Project Index | Student Resources | Glossary

CUDA PI Calcuation (Monte Carlo)

Team Pi CUDA

Welcome to GPU610AA Fall 2013 Team Pi CUDA Page.

My name is Peter Huang and I'm a student in the GPU610 class for the Fall Semester of 2013. Having no background whatsoever in parallel programming, I've decided to choose something that is out of my scope of understanding and interest (video game programming) to challenge myself. Thus, I've decided to investigate the benefits of parallel programming applied to the Monte Carlo statistical method to approximating the value of pi.

Announcements

N/A

Team Members

Peter Huang

eMail All

Progress

Assignment 1

Introduction

For the initial profiling, I've decided to investigate the Monte Carlo Statistics Methodology of approximating the value of Pi. A brief explanation of Monte Carlo Pi calculation can be found here: https://www.youtube.com/watch?v=VJTFfIqO4TU

Source File(s)

Link: https://drive.google.com/file/d/0B8GUuIUqdEJES3VEOGRnYmRNaEk

Code Snippet

   Serial Pi Calculation Algorithm 
    
   // loops through user amount of rounds of sets of points
   for(i = 0; i < points; i++)
   {
      x = randNum();
      y = randNum();
      
      // check if point resides within the circle
      if (((x*x) + (y*y)) <= 1.0)
      {
         score++;
      }     
   }
   // calculate pi
   pi = 4.0 * (float)score/(float)points;

Compilation and Running

Serial Results

Assignment 2

Introduction

In Phase 2, I've parallelized the serial program to run on a custom kernel on a CUDA-enabled device.

Source File(s)

Link: https://drive.google.com/file/d/0B8GUuIUqdEJEbDBRNkhWYnpGSnM

Code Snippet

    Working Kernel Parallel CUDA Pi Calculation
    __global__ void findPi(float *estimatedPi, curandState *states, unsigned int taskElements, float seed)
    {
        unsigned int task_id = blockDim.x * blockIdx.x + threadIdx.x; // linear sequence of threads x-axis
        int score = 0;
        float xCoord;
        float yCoord;
           
        // 'random' generated value using curand, initialize curand using task_id, and seed parameter
        curand_init(seed, task_id, 0, &states[task_id]);
         
        // tally number of task elements
        for(int i = 0; i < taskElements; i++)
        {
          
           // assigned each point coordinate values
           xCoord = curand_uniform (&states[task_id]);
           yCoord = curand_uniform (&states[task_id]);
            
           // determine if coordinate is within the circle
           if((xCoord*xCoord + yCoord*yCoord) <= 1.0f)
           {
              score++;
           }
         }
          
         // estimated value of pi for this particular task
         estimatedPi[task_id] = (4.0f * score) / (float)taskElements;
         }
    }

Compilation and Running

Parallel (CUDA) Results

Serial VS CUDA

Conclusion

Using CUDA technology and parallelizing the serial code in the original code, there is an enormous increase in performance (lower execution time) to calculate , as high as 1372%. In the next (final) phase, an attempt to investigate if shared memory, optimal memory allocation, minimizing said memory access time, and other optimization factors would provide a further increase (lower execution time) in performance for pi_cuda.

Assignment 3

Agenda

N/A

Progress

N/A

Meetings

N/A

Discussion

N/A

CDOT Wiki ^β

GPU610/DPS915 CUDA PI

CUDA PI Calcuation (Monte Carlo)

Team Pi CUDA

Announcements

Team Members