bnn_on_gpu

1. Navigate to folder of respective model
2. Change desired batch size in utils.h
3. A. run using

$ bash run.bash

3. B. alternitavely: set the path to the CUDA Runtime Library of your system in Makefile

• currently commented out: the path on my personal PC on line 5, and the path on the server on line 7
• and run using

$ make
$ ./fashion_prof.o <OR ./cifar.o>

Layer configuration for FashionMNIST BNN model

L1 Conv → L2 Maxpool → L3 Step → L4 Conv → L5 Maxpool → L6 Step → L7 Flattening → L8 Gemm → L9 Step → L10 Gemm

Default starting configuration:

• Layer 1, 2, 4, 5, 8, 10 on GPU using PROFILE XYZ
• Layer 3, 6, 7, 9 on CPU.

Layer configuration for CIFAR-10 BNN model

L1 Conv → L2 Step → L3 Conv → L4 Maxpool → L5 Step → L6 Conv → L7 Step → L8 Conv → L9 Maxpool → L10 Step → L11 Conv → L12 Step → L13 Conv → L14 Maxpool → L15 Step → L16 Flattening → L17 Gemm → L18 Step → L19 Gemm

Default starting configuration:

• Layer 1, 3, 4, 6, 8, 9, 11, 13, 14, 17, 19 on GPU using PROFILE XYZ
• Layer 2, 5, 7, 10, 12, 15, 16, 18 on CPU.

Run layers on CPU only:

in net.cpp (X layer number):

• comment /* Layer X GPU */ lines
• comment out immediate /* Layer X CPU */ lines

Run layers on GPU in different parallel configurations:

• Revert previous steps in case they were performed
• Do NOT comment out the STEP layers (i.e. the layers running on CPU by default)

Available parallel configurations:

• PROFILE X – data-images
• PROFILE Y – windows
• PROFILE Z – neurons
• PROFILE XY – data-images + windows
• PROFILE XZ – data-images + neurons
• PROFILE YZ – windows + neurons
• PROFILE XYZ – data-images + windows + neurons

in cuda_kernel.cu:

• Each layer is specifically implemented for every profile iteratively
• Each profile can be found between delimiting lines: =============
• Search for desired profile using CTRL+F, comment out all lines of coude between delimiting lines
• Don't forget to comment-in previously used profiles (e.g. default profile XYZ at the bottom of the file)