Thread Subject: Compiling CUDA C/C++ mex code under linux

Hi all

There a several methods available on the file exchange for compiling CUDA C/C++ code into mex files under Windows, but none that I've come across work for linux. However, I've found a nice, easy way to do it, which I'll share with you, though I must confess I haven't tested it extensively.

The idea is to use Nvidia's nvcc compiler to convert CUDA C/C++ code into standard C++ code, then use mex after that. The first stage looks something like:

system(sprintf('nvcc -I"%s/extern/include" --cuda "mexfun.cu" --output-file "mexfun.cpp"', matlabroot));

Then the second stage is roughly:

mex -I/opt/cuda/include -L/opt/cuda/lib -lcudart mexfun.cpp

Obviously you need to set the various paths and file/function names to suit your needs.

HTH,
Oliver

PS Does anyone think this approach could reduce the efficiency of the resulting machine code? I do wonder if it doesn't limit the level of optimization that can be applied.

Oliver!

Five stars, thanks very much. I was SO close to this solution, but with things still going wrong - your commands below just sorted me right out :)

Re. Performance, I don't think it'll have a detrimental effect. The NVCC compiler will use it's standard options for the device code; and gcc (or whatever c/c++ compiler you use) will use the options that mex passes it to compile the part of the code which runs on the host.

If you were concerned about the host code performance, you can always adjust the mexopts.sh file (in matlabroot/bin/) or the call to mex to pass various performance flags (O3 etc) in to the compiler.

Cheers again!

Tom Clark





"Oliver Woodford" <o.j.woodford.98@cantab.net> wrote in message <h9shtd$q9g$1@fred.mathworks.com>...
> Hi all
>
> There a several methods available on the file exchange for compiling CUDA C/C++ code into mex files under Windows, but none that I've come across work for linux. However, I've found a nice, easy way to do it, which I'll share with you, though I must confess I haven't tested it extensively.
>
> The idea is to use Nvidia's nvcc compiler to convert CUDA C/C++ code into standard C++ code, then use mex after that. The first stage looks something like:
>
> system(sprintf('nvcc -I"%s/extern/include" --cuda "mexfun.cu" --output-file "mexfun.cpp"', matlabroot));
>
> Then the second stage is roughly:
>
> mex -I/opt/cuda/include -L/opt/cuda/lib -lcudart mexfun.cpp
>
> Obviously you need to set the various paths and file/function names to suit your needs.
>
> HTH,
> Oliver
>
> PS Does anyone think this approach could reduce the efficiency of the resulting machine code? I do wonder if it doesn't limit the level of optimization that can be applied.

hi all,

     I'm trying to use your method to call CUDA code as function in MATLAB. I succeeded to obtain the mex file, but I have some problems:


Mex file entry point is missing. Please check the (case-sensitive)
spelling of mexFunction (for C MEX-files), or the (case-insensitive)
spelling of MEXFUNCTION (for FORTRAN MEX-files).
Invalid MEX-file



can you please help me?

Thank you

"Andrei" wrote:
> hi all,
>
> I'm trying to use your method to call CUDA code as function in MATLAB. I succeeded to obtain the mex file, but I have some problems:
>
>
> Mex file entry point is missing. Please check the (case-sensitive)
> spelling of mexFunction (for C MEX-files), or the (case-insensitive)
> spelling of MEXFUNCTION (for FORTRAN MEX-files).
> Invalid MEX-file
>
>
>
> can you please help me?
>
> Thank you

The error is quite helpful. Do you have a mexFunction function in your mex file?

"Oliver Woodford" wrote in message <ja5ll3$s5c$1@newscl01ah.mathworks.com>...
> "Andrei" wrote:
> > hi all,
> >
> > I'm trying to use your method to call CUDA code as function in MATLAB. I succeeded to obtain the mex file, but I have some problems:
> >
> >
> > Mex file entry point is missing. Please check the (case-sensitive)
> > spelling of mexFunction (for C MEX-files), or the (case-insensitive)
> > spelling of MEXFUNCTION (for FORTRAN MEX-files).
> > Invalid MEX-file
> >
> >
> >
> > can you please help me?
> >
> > Thank you
>
> The error is quite helpful. Do you have a mexFunction function in your mex file?


I have a CUDA file, using your commands I generated the mex file.

 I have to write a mex file? I thought that the commands above generates the mex file.

Maybe you can post some example files.

Thanks

"Andrei" wrote:
>
> I have a CUDA file, using your commands I generated the mex file.
>
> I have to write a mex file? I thought that the commands above generates the mex file.

Andrei,

The commands I give compile a .cu file into a .cpp file. They do not create a mexFunction host function (which is the gateway function between MATLAB and any C/C++/CUDA code) in the resulting .cpp file. You either need to have a mexFunction function in the .cu file or in another .c/.cpp file which you compile along with the converted .cu file in the last step.

I have included a simple example of a .cu file containing such a function below, which generates an array of 256 random numbers and squares them on the gpu, then checks these against the result computed on the gpu.

This topic is a little outside the scope of this thread, so if you need further help then I suggest you look at MATLABs documentation on mex files, and/or post a new question to this newsgroup or on MATLAB Answers.

HTH,
Oliver

===================================================

#include "mex.h"

// CUDA kernel which squares a float
__global__ void sq(float *d_buffer)
{
d_buffer[threadIdx.x] *= d_buffer[threadIdx.x];
}

// Function which interfaces with MATLAB
void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
{
    if (nrhs != 0)
        mexErrMsgTxt("No inputs expected.");
    if (nlhs != 0)
        mexErrMsgTxt("No outputs expected.");

// Generate the data
float h_buffer[256];
for (int a = 0; a < 256; ++a)
h_buffer[a] = float(rand()) / RAND_MAX;

// Copy to the gpu
float *d_buffer;
cudaMalloc((void **)&d_buffer, 256 * sizeof(*d_buffer));
cudaMemcpy(d_buffer, h_buffer, 256 * sizeof(*d_buffer), cudaMemcpyHostToDevice);

// Call the CUDA kernel
sq<<<1, 256>>>(d_buffer);

// Copy from gpu
float h_buffer2[256];
cudaMemcpy(h_buffer2, d_buffer, 256 * sizeof(*d_buffer), cudaMemcpyDeviceToHost);

// Check result
for (int a = 0; a < 256; ++a) {
if (h_buffer2[a] != h_buffer[a] * h_buffer[a])
mexErrMsgTxt("Error in calculation!");
}
mexPrintf("Test passed!\n");
}

"Oliver Woodford" wrote in message <jagi1s$1j8$1@newscl01ah.mathworks.com>...
> "Andrei" wrote:
> >
> > I have a CUDA file, using your commands I generated the mex file.
> >
> > I have to write a mex file? I thought that the commands above generates the mex file.
>
> Andrei,
>
> The commands I give compile a .cu file into a .cpp file. They do not create a mexFunction host function (which is the gateway function between MATLAB and any C/C++/CUDA code) in the resulting .cpp file. You either need to have a mexFunction function in the .cu file or in another .c/.cpp file which you compile along with the converted .cu file in the last step.
>
> I have included a simple example of a .cu file containing such a function below, which generates an array of 256 random numbers and squares them on the gpu, then checks these against the result computed on the gpu.
>
> This topic is a little outside the scope of this thread, so if you need further help then I suggest you look at MATLABs documentation on mex files, and/or post a new question to this newsgroup or on MATLAB Answers.
>
> HTH,
> Oliver
>
> ===================================================
>
> #include "mex.h"
>
> // CUDA kernel which squares a float
> __global__ void sq(float *d_buffer)
> {
> d_buffer[threadIdx.x] *= d_buffer[threadIdx.x];
> }
>
> // Function which interfaces with MATLAB
> void mexFunction(int nlhs, mxArray *plhs[], int nrhs, const mxArray *prhs[])
> {
> if (nrhs != 0)
> mexErrMsgTxt("No inputs expected.");
> if (nlhs != 0)
> mexErrMsgTxt("No outputs expected.");
>
> // Generate the data
> float h_buffer[256];
> for (int a = 0; a < 256; ++a)
> h_buffer[a] = float(rand()) / RAND_MAX;
>
> // Copy to the gpu
> float *d_buffer;
> cudaMalloc((void **)&d_buffer, 256 * sizeof(*d_buffer));
> cudaMemcpy(d_buffer, h_buffer, 256 * sizeof(*d_buffer), cudaMemcpyHostToDevice);
>
> // Call the CUDA kernel
> sq<<<1, 256>>>(d_buffer);
>
> // Copy from gpu
> float h_buffer2[256];
> cudaMemcpy(h_buffer2, d_buffer, 256 * sizeof(*d_buffer), cudaMemcpyDeviceToHost);
>
> // Check result
> for (int a = 0; a < 256; ++a) {
> if (h_buffer2[a] != h_buffer[a] * h_buffer[a])
> mexErrMsgTxt("Error in calculation!");
> }
> mexPrintf("Test passed!\n");
> }


Thank you very much.