eliahburns / cuda_example

CUDA Ray Trace Example

Testing was performed on the Midway supercomputer at UChicago.

For a refresher (or intro) on what ray tracing is, check Wikipedia's page on ray tracing

Usage

There's two versions in this project, which are largely similar. A standard version written using only C. And a version which uses some extra CUDA code to offload the most computationally intensive sections of the code to a GPU--in the case of my tests a Tesla K80 GPU was used.

note: This is a simply a demonstration of some low level programming using C and CUDA, and then analysing differences in performance--if interested in using ray tracing in another program, there's better off the shelf options.

All generated data is saved to sphere.bin.

In the image directory, (and to the left) there's a plot illustrating a scaling study performed for the occupancy of a Telsa K80 GPU. For a fixed problem size, the code was run using powers of 2, from 2 to 1024 threads per block. Precautions had to be taken at certain points to insure that the max thread or max block size was not exceeded--some of this adjusting may be responsible for the curve's early dip in execution time and then slight increase before it begins to taper off again.

There's also a plot (again to the left) that shows the execution times for serial and CUDA versions side by side. We see the serial version increase in time rapidly, while the CUDA version stays largely steady and its linear slope is quite negligible in comparison to the serial version. Because this program's primary computational task is to simply randomly generate vectors (rays of light)--a problem considered embarrassingly parallel--it makes sense that we would see immense reductions in execution time in comparision to our serial version.

Generated Image

We can take the binary data file and use python and matplotlib to display the final product, which is a single sphere and a single light source observed through our first person perspective.

Compilation

To compile the CUDA version (ray_cude.cu):

$ make cuda

To compile the standard version (raytrace.c):

$ make

Execution:

After compiling, run either version of the executable file from the command line, including the maximum number of rays that you would like to be generated, along with the number of grid points:

$ ./ray_cuda <num_rays> <num_grid_pts>

$ ./raytrace <num_rays> <num_grid_pts>

note: the number of grid points will be equal to the product of the two dimensions of the grid, e.g. for a 4x4 grid we should enter 16 as the number of grid points.

Requirements

To compile run the CUDA version of this software we need to have the CUDA Developer Toolkit installed.