We build parallel algorithms of forward propagation for general convolutional neural networks and implemented ZF-Net for performance testing. Several methods include tiling, diminishing control divergence and carefully setting block size are used to optimize the algorithm speed. Finally, the most optimized algorithm is 107.35% and 1287 times faster than naive parallel method and serial method respectively.

The code is the convolution part of CNN. Fully-connected layer part was implemented by my teammate and I didn't upload them.

We use the pyCUDA environment of tesseract server provided by Columbia University.

(1) report.pdf: our report that contains all details
(2) result.out: the result of our program, contains all profiling (time measuring) information
(3) convolution.py: contains all kernel code of several methods
(4) test_funcs.py: testing functions, verifying correctness and measuring time
(5) utils.py: other functions, such as automatically setting block size
(6) run.py: run this file as the main function if you know the running command
(7) cnn.sh: run this file if you don't know the running command