YaConv: Convolution with Low Cache Footprint

1. Install dependencies

sudo apt install linux-tools-common linux-tools-generic linux-tools-$(uname -r) numactl

2. Build

git clone --recurse-submodules https://github.com/ivan23kor/yaconv.git
cd yaconv && make

3. Run

1. Run either im2col or yaconv binary on a given layer:

./<BINARY> N H W C FH FW M PH PW  # outputs GFLOPS

Parameter	Explanation
N	number of images in per batch
H, W	image height and width
C	number of input channels
FH, FW	filter height and width
M	number of output channels
PH, PW	vertical and horizontal padding

2. Gather comparative data for im2col and yaconv on a pre-defined grid:

./run.sh # writes run data on disk to ./Results/

The command above will run ./im2col and ./yaconv binaries on the following parameter grid:

Parameter	Values
N	Auto-adjusted to make ~100 GFLOPS workload per layer
H, W	7 14 28 56
C	32 64 96 128 192 256 384 512 768 960 1024 1152
FH, FW	3
M	32 64 96 128 192 256 384 512 768 960 1024 1152
PH, PW	1

GFLOPS and perf stat data will be in a directory structure:

Results/
|-- im2col
|   |-- gflops
|   `-- stat
`-- yaconv
    |-- gflops
    `-- stat

For more information on the output in ./Results/, run ./run.sh -h

3. Gather comparative data for im2col or yaconv on a set of layers given in layers.txt:

./run.sh -i file  # (from layers.txt)

./layers.txt template:

7 7 1024 3 3 32 1 1
5 5 128 3 3 32 1 1
7 7 256 3 3 32 1 1

Note the empty line at the end of the file.

Publication

You can find the paper online or download a pdf from this repository.