Cannons Algorithm in DARTS using Dataflow Software Pipelining

Cannon's Algorithm is a distributed algorithm for matrix multiplication of two-dimensional meshes, first described in 1969 by Lynn Elliot Cannon. This repository contains its implementation using DARTS runtime using and without using Dataflow Software Pipelining techniques.

Prerequisites

Please follow the instructions on the respective dependency website to install them.

cmake : https://cmake.org/
HWLOC : http://www.open-mpi.org/projects/hwloc/
Intel Thread Building Blocks (Optional) : https://github.com/oneapi-src/oneTBB
Pthreads.

Steps to Install

Clone cannon-dfswp repository. $ git clone https://github.com/sraskar/cannon-dfswp.git

Enter the directory and Create a build directory.

$ cd cannon-dfswp
$ mkdir build 
$ cd build

Use cmake to configure the repository. Please make sure to add relevant paths for MKL and HWLOC in the camke command using -DMKL_ROOT_DIR=/path/to/mkl_root and -DHWLOC_ROOT=/path/to/hwloc_root flags.
```
$ CC=gcc CXX=g++ cmake -DCBLAS=OFF -DMKL=ON -DMKL_ROOT_DIR=/path/to/mkl_root -DHWLOC_ROOT=/path/to/hwloc_root ../
```
Use make to compile code.
```
$ make 
```

`cmake` Options

Debugging Option Debug ie. compile with -g : -DDEBUG=ON By default cmake compiles in release mode ie. compiles with -O3
Install directory: -DCMAKE_INSTALL_PREFIX=/path/to/directory/ By default, Darts will install in bash variable HOME/darts
DMKL=ON -DMKL_ROOT_DIR=/path/to/mkl_root This option will enable execution with MKL version on the specified path.
cmake cannot find a hwloc/TBB First, make sure you have hwloc/TBB installed on your system. Second cmake will look for hwloc/TBB in specific directories. To find if you installation is located in on of these directories, navigate in your source to the cmake/Modules directory and view the FindHwloc.cmake/FindTBB.cmake file. You may add your search path to these files rather than moving your installation.

Running Cannons Algorithm

The source code for Cannons Algorithm is Present in the directory apps/Cannon
After make command, code is built in the directory build/apps/Cannon
The code needs to be executed with the required command line arguments. $ ./CannonTP_test numSU numCU Matsize sqrtNumTiles SqrtNumElementPerTile where
- numSU: Number of Synchronization Units
- numCU: Number of Compute Units
- Matsize: Size of Square Matrix
- sqrtNumTiles: Square root of the total number of tiles for the matrix
- SqrtNumElementPerTile: Square root of Number of elements per Tile for matrix.
For example, ./cannonTP_test 1 1 64 2 32
mmCannonTime.sh script contains some configurations.

Running Cannons Algorithm with Dataflow Software Pipelining

The source code for Cannons Algorithm with dataflow Software Pipelining is Present in the directory apps/Cannon-dfswp
After make command, code is built in the directory build/apps/Cannon-dfswp
- The code needs to be executed with the required command line arguments. $ ./cannon_dfswp numSU numCU Matsize sqrtNumTiles SqrtNumElementPerTile. The meaning of these command line arguments is the same as the one mentioned above.
mmCannonTime.sh script contains some configurations.