Cuda-Matrix-Multiplication

A basic example of how to do matrix multiplcation with cuda and cuBLAS library. In order for the example to work you must have an Nvidia GPU supporting CUDA.
Multiplcations are done by using cublasGemmEx method.

Overview

This simple app does the basic matrix multplication A X B = C

A = rowsA * rank
B = rank * colsB
C = rowsA * colsB

Arrays are 2d and are declared using single raw pointers

float* A = new float[sizeA];

In order to access them we follow this trick

for (size_t i = 0; i < rows; ++i)
{
    for (size_t j = 0; j < cols; ++j)
    {
        cout << A[j * rows + i] << " ";
    }
    cout << endl;
}

The could be either 16 or 32 bit floats. The result array is always 32 bit.

The time difference beetween the GPU(device) and CPU(host and single threaded) are very big.

Also Nvidia states the after cuBLAS version 11 tensor cores will be used automatically, link.

Build

Linux

Make sure you have all the required dependencies for cuda to compile and work.

sudo apt install nvidia-cuda-toolkit

Build

make all

Run the executable

./main.out