xGEMM implementation with Halide

This program is an implementation of a BLAS (Basic Linear Algebra Subprograms) routine from the third level which is the general matrix-matrix multiplication "xGEMM", of the form alpha*A*B + beta*C using Halide.

Prerequisites

Linux

You'll need the Halide to run this project, you can either:

Download the binaries directly from here https://github.com/halide/Halide/releases, get halide-linux-64-gcc53 or halide-linux-32-gcc53 depending on your architecture and then extract it.

Build Halide using LLVM :

Acquiring & building LLVM

git clone https://github.com/llvm/llvm-project.git --depth 1 -b release/10.x
mkdir llvm-build

cd llvm-build

 cmake -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=../llvm-install \
 -DLLVM_ENABLE_PROJECTS="clang;lld;clang-tools-extra" \
 -DLLVM_TARGETS_TO_BUILD="X86;ARM;NVPTX;AArch64;Mips;Hexagon" \
 -DLLVM_ENABLE_TERMINFO=OFF -DLLVM_ENABLE_ASSERTIONS=ON \
 -DLLVM_ENABLE_EH=ON -DLLVM_ENABLE_RTTI=ON -DLLVM_BUILD_32_BITS=OFF \
 ../llvm-project/llvm

cmake --build . --target install -- -j8
export LLVM_CONFIG=<path to llvm>/llvm-install/bin/llvm-config

Acquiring & building Halide
- Start by cloning the repository git clone https://github.com/halide/Halide.git
- Run make -j8 in the root directory.

More information can be found at https://github.com/halide/Halide

Usage

First clone this repository and give the script execute permission to make it runnable :

git clone https://github.com/AdelNamani/halide-xGEMM.git
cd halide-xGEMM
chmod +x script.sh

To run the program you just need to run the script.sh with the path of Halide as a parameter. For example:

With relative path : $ ./script.sh -p ../halide
With absolute path : $ ./script.sh -p /mnt/c/Users/DELL/code/halide

This script will compile both programs and run them, if everything is okay and the multiplication is giving the correct results we will have this output:

Compiling halide_xGEMM.cpp ...
Compiling test_halide_xGEMM.c ...
Executing halide_xGEMM ...
Multiplication done ...
Saving to bin/data/results.txt ...
Saved successfully !
Executing test_halide_xGEMM ...
Reading and checking results ...
The multiplication with Halide is working well.

Testing with different data

If you want to test with different matrices, you just have to change their initialization in the code, respecting these two constraints:

Dimensions must be of the form : A(n,m) B(m,l) C(n,l)
Same matrices must be in the Halide program and the C program that test the results

For example :

int  a[2][4] = {
{0, 1, 2, 3},
{4, 5, 6, 7}
};

int  b[4][5] = {
{0, 1, 2, 3, 5},
{4, 5, 6, 7, 5},
{8, 9, 10, 11, 5},
{8, 9, 10, 11, 5}
};
  
int  c[2][5] = {
{0, 1, 2, 3, 1},
{4, 5, 6, 7, 1},
};