Haim Avron

To do the same on a Linux the one running the script should have root privileges. I currently do not have MATLAB on a Linux for which I have root privileges...

Ok, hopefully I'm not overstaying my welcome on this forum, but I've found out more answers to my cutCreateTimer issue and I wanted to put them out there incase someone else falls into this issue.
It appears that somehow the cuda_mex compilation couldn't find my cutil libraries. This is not necessarily it's fault, I don't know much about make files, linking, etc. so it's fairly probable that I screwed something up somewhere. There is likely a short, easy, and correct way to fix this, but since I don't know it, here's my work around. If you find your way into the CUDA SDK folder you will find a Visual Studio solution for compiling everything cutil. If you change the output of these libraries to the cuda standard library ( likely something along the lines of CUDA/bin) and then rebuild the solution the cutil32 libraries will then be found by the compiler/linker. Will keep posting if I find out a more elegant solution.

Still having trouble with the cutCreateTimer. Here is an example. I have modified the example.cu that comes with this package to create this timer. In addition, the cutil32.lib and cutil32D.lib are linked in the setup_cuda_mex.m file. If anyone has a solution, I would be very excited to hear it.

#include <stdio.h>
#include <cuda.h>
#include <cutil.h>

#include "mex.h"

__global__ void square_array(float *a, int N)
{
  int idx = blockIdx.x * blockDim.x + threadIdx.x;
  if (idx<N) a[idx] = a[idx] * a[idx];
}

void mexFunction(int nargout, mxArray *argout[], int nargin, const mxArray *argin[])
{
  float *input, *cuda_mem;
  int N;
  size_t size;
unsigned int hTimer;

cutCreateTimer(&hTimer);

  argout[0] = mxDuplicateArray(argin[0]);
  input = (float *)mxGetPr(argout[0]);
  N = mxGetN(argout[0]) * mxGetM(argout[0]);
  size = N * sizeof(float);
  cudaMalloc((void **) &cuda_mem, size);
  cudaMemcpy(cuda_mem, input, size, cudaMemcpyHostToDevice);
  int block_size = 4;
  int n_blocks = N/block_size + (N%block_size == 0 ? 0:1);
  square_array <<< n_blocks, block_size >>> (cuda_mem, N);
  cudaMemcpy(input, cuda_mem, size, cudaMemcpyDeviceToHost);
  cudaFree(cuda_mem);
}

If you add cutil32.lib to

        if (strcmp(name, 'LINKFLAGS'))
            value = [value ' cudart.lib cutil32.lib'];
        end

in setup_cuda_mex.m, the linker will properly access cutil based functions. I spent a lot of time trying to figure out why I would get errors in the linking involving various cuda based functions even though the compiler knew the address to the header files...