libcumatrix

C++ Interface of CUDA-based GPU matrix library.

Prerequisite

You need

• An NVIDIA GPU that supports CUDA (Ex: GTX-660)
• Linux / Ubuntu
• Install NVIDIA CUDA toolkit (at least CUDA 5.0)

Quick Start

Just type "make" and run the following programs:

1. ./benchmark
2. ./example1
3. ./example2

How to Use It in My C++ Codes

Include device_matrix.h and add -I flags to your g++ command

Here's how you use it in your program.

#include <device_matrix.h>

int main() {
  // create a 10 by 20 matrix lies in GPU memory
  // (all filled with garbage value)
  device_matrix<float> A(10, 20);
  A.print();
  
  // Fill A with 0
  A.fillwith(0);

  A += 10;
  A.print();
}

You can use typedef to save yourself some typing

#include <device_matrix.h>

typedef device_matrix<float> mat;

int main() {
  // create a 10 by 20 matrix lies in GPU memory
  // (all filled with value 1.234)
  mat A(10, 20, 1.234);
  A.print();
  
  A += 10;
  A.print();
}

Here's how you add -I flags and libraries in your g++ command

On linking, libcumatrix needs:

1. -lcuda (libcuda.so )
2. -lcublas (libcublas.so)
3. -lcudart (libcudart.so)

So like this

  g++ -o my_program main.cpp -I libcumatrix/include -lcuda -lcublas -lcudart -lcumatrix

By default, the root directories containing CUDA is located at "/usr/local/cuda". If you wish to change it, open the "Makefile" and change CUDA_ROOT to where it lies.

Others

For an older version of CUDA sdk, Thrust Library did not lie directly in "/include/thrust" (something like /include/thrust-1.5.1). Create a symbolic link for it using the following command:

ln -s /usr/local/cuda/include/thrust-1.5.1 /usr/local/cuda/include/thrust