Batched forward convolution with Intel OneAPI

Code repository for the master thesis Implementación do algoritmo da convolución por lotes usando Intel oneAPI.

Dependencies

• Intel oneAPI Base Toolkit
  • Intel oneAPI Deep Neural Network Library (oneDNN)
  • Intel oneAPI DPC++/C++ Compiler

Local

Clone this repository:

git clone https://git.fic.udc.es/s.aguado/tfm.git && cd tfm

Set the environment:

source /opt/intel/inteloneapi/setvars.sh # replace with custom installation path 

Compile the codes:

./build debug # to print some feedback while running the codes
./build       # to run the codes in quiet mode

After running these commands, the executables should be in the bin/ folder. All of them share the same interface:

./executable (cpu|gpu) N C K H W R S

Examples:

./bin/convolution # Run convolution with default parameters in the CPU
./bin/gemm gpu    # Run convolution with default parameters in the GPU
./bin/winograd gpu 4 3 3 64 64 3 3 

Cloud

1. Sign up for Intel DevCloud for oneAPI
2. Connect via SSH from Linux/macOS
3. Submit a job to the queue
4. Advanced queue management

This repo includes some scripts ready to be launched as jobs in DevCloud.