Important

The tutorial is missing some compilation, linker, library settings steps. However you can download the template project, it works and has everything setup.

Introduction

• Download template project source
• Download executable

I am a developer who has been developing software using .NET and C# for several years. I have never used C or C++, and it has never been required.

I like to investigate new technologies, mainly because I am curious, but also because it could make my daily development work easier or smarter.

Recently my focus has been directed towards GPGPU on the Nvidia Cuda platform.

The programming language for Cuda is called “Cuda C”. The name implies that knowledge of C indeed is required for using GPGPU on the Cuda platform.

I discovered that there exist .NET bindings to the Cuda platform and drivers. However, I find their usage complicated and insufficient, and further more kernel development will still have to done in Cuda C.

These fact made me realise that I would have to learn a bit of C and C++ to use Cuda as it was actually intended by Nvidia. Nvidia provides many samples and suggest that a Cuda development environment on Windows could use Visual Studio and Nvidia Parallel NSight for debugging, profiling etc.

As my knowledge of C and C++ development was severely limited, so was the setting up and configuration of Visual Studio 2008 for Cuda C development.

I have read “Cuda by example…” (http://developer.nvidia.com/object/cuda-by-example.html), “Programming Massively Parallel Processors…” (http://www.nvidia.com/object/io_1264656303008.html) and “C Programming Language, 2. edition” (http://www.pearsonhighered.com/educator/product/C-Programming-Language/9780131103627.page). These books have given me the foundation to start developing using Cuda C and GPGPU.

Setting up Visual Studio 2008 and making the compiler work required some work, but here is what I did.

1. Download and install driver and toolkit

Download Cuda toolkit and the developer driver and install. A restart is probably required. (http://developer.nvidia.com/object/gpucomputing.html)

2. Start Visual Studio 2008, and create a new project of type Win32 Console Application

Give the project and solution a name (here called Cuda_Template).

3. Click Next

4. Select Console application and check the empty project, then click Finish

5. Add new item called main.cpp of type C++ File

6. Add the following code the file

#include <stdio.h> 

int main() {

    printf("Hello world...\n);
    return 0;

}

7. Build and try to run the exe file. The output should be

8. Select Project -> Custom Build Rules…

9. Select Cuda Runtime API build rule (v3.2)

10. Add a new file called kernel.cu

11. Add the following to the file kernel.cu

/* power: raise base to n-th power; n >= 0 */
__device__ int devicePower(int base, int n) {

    int p = 1;

    for (int i = 1; i <= n; ++i) {
        p = p * base;
    }

    return p;
}

__global__ void power( int *base, int *n, int *output, int threadMax ) {

    int tid = threadIdx.x + blockIdx.x * blockDim.x;

    if (tid < threadMax) {
        output[tid] = devicePower(base[tid], n[tid]);
    }

}

12. Right click the newly created file and select properties

13. Set the “Exclude From Build” and make sure that the project still builds

14. Create a new file called call_kernel.cu

15. Add the following to the file call_kernel.cu

#include <cuda_runtime_api.h>
#include "main.h"

// includes, kernels
#include <kernel.cu>

void call_kernel_power(int *base, int *n, int *output, int elementCount) {

    int *dev_base, *dev_n, *dev_output;
    int gridX = (elementCount+ThreadsPerBlock-1)/ThreadsPerBlock;

    cudaMalloc( (void**)&dev_base, elementCount * sizeof(int) );
    cudaMalloc( (void**)&dev_n, elementCount * sizeof(int) );
    cudaMalloc( (void**)&dev_output, elementCount * sizeof(int) );

    cudaMemcpy( dev_base, base, elementCount * sizeof(int), cudaMemcpyHostToDevice);
    cudaMemcpy( dev_n, n, elementCount * sizeof(int), cudaMemcpyHostToDevice);

    power<<<gridX,ThreadsPerBlock>>>(dev_base, dev_n, dev_output, elementCount);

    cudaMemcpy( output, dev_output, elementCount * sizeof(int), cudaMemcpyDeviceToHost);

    cudaFree( dev_base );
    cudaFree( dev_n );
    cudaFree( dev_output );
}

16. Create a new header file called main.h

17. Add the following content to the file

#define ThreadsPerBlock 128

#include <stdio.h>

void call_kernel_power(int *base, int *n, int *output, int elementCount);

18. Update the main.cpp file with the following:

#include "main.h"

#define N   80000

int main() {

    printf("Power Cuda kernel test from C++\n");
    printf("Testing %d elements\n", N);

    int base[N], n[N], output[N];

    for(int i = 0; i < N; i++) {
          base[i] = 2;
          n[i] = i+1;
          output[i] = 0;
      }

    call_kernel_power(base, n, output, N);

    for(int i = 0; i < N && i < 15; i++) {

          printf("%d^%d = %d\n", base[i], n[i], output[i]);

      }

      printf("Done\n");

    return 0;
}

19. That should be it…

You now have a template that you can work from. When you build the file and run it, I get this output: