This is a proof-of-concept project that compares the differences in runtime between a CPU and GPU executing similar implementations of the Boyer Moore pattern search algorithm to count the number of substrings in a very large input string. I test to see how the number of threads impacts runtime, and compare the CPU and GPU against each other.
Both implementations first read the entire input .txt file into a buffer, so that time spent reading the file into memory does not affect our runtime results. The input string is then divided into chunks for each thread to process. (4 threads = 4 chunks, 8 threads = 8 chunks, etc) A shared count variable is incremented each time a thread finds an instance of the search pattern. Execute with:
CPUTextSearch.exe <path to file> <search pattern> <number of threads>CUDATextSearch.exe <path to file> <search pattern> <number of thread blocks> <number of threads>
Ran tests on 3 large .txt files: the entirity of Shakespeare's works, Moby Dick, and a file containing a compilation of data from Wikipedia. The graphs below show an average runtime for searching the files for the substring "Thou" timed using chrono::high_resolution_clock. Honestly I should've tested this more but I wanted to start another project.
The CPU and GPU are comparable in performance. The CPU performs better under normal circumstances, but when looking for optimal performance the GPU is superior - at least on my computer and with these test cases. All tests were run on my XPS-15 laptop running an 8 core i7 and nvidia 1050ti.