#!/usr/bin/env python # WARNING WARNING WARNING # This is a very fragile script. # Don't change read_test2.cu and expect this to continue to work! import sys, re # --- Constants from read_test2 --- # Memory access patterns in the order they are performed access_patterns = ['random', 'broadcast', 'linear'] calls = 10 blocks = 500 threads = 256 # per block iterations = 1000 # per thread reads = 10 # per iteration memory_read = calls * blocks * threads * iterations * reads * 4 # bytes # --- Parse the profile log --- access_times = [] counts = { } inside_chunk = False time_regex = re.compile(r'(?P[a-zA-Z]+)_speedtest.+gputime=\[ (?P