Analysis of Parallelism in Ray Tracing and Coulombic Potential Benchmarks

Last week, Fahiba and DH generated some very interesting plots which gave us a lot of insight into the different benchmarks that we were using in our project. So this week, I set out to understand the reasons behind some of the trends that they spoke about in their blog by examining in important characteristic of these benchmark programs: their degree of parallism or in other words the number of software work elements that each program issues to the GPU when executing in hardware. The following calculations have been derived from the CUDA source files and are a representation of the level of parallelism employed by 2 programs which we found interesting as per Fahiba and DH's blog.

RAY

Number of threads in each block = 16 x 8 = 128 threads

Number of thread blocks = (64/16, 64/8) = (4,8)

Thus, number of thread blocks in a queue at any given time = 4 x 8 = 32.

Coulombic Potential (CP)

Number of threads in each block = 16 x 8 = 128 threads.

VOLSIZEX = 512 VOLSIZEY = 512

BLOCKSIZEX = 16 BLOCKSIZEY = 8

UNROLLX = 1

Grid Size in the x dimension = Gsz.x = 512/16 = 32

Grid Size in the y dimension = Gsz.y = 512/8 = 64

Grid Size in the z dimension = UNROLLX = 1

Thus the number of thread blocks issued = (32, 64)

and the number of blocks in the queue at a given point in time = 32 x 64 = 2048 thread blocks.