cuMatlab

Build your cuda applications inside Matlab. You don't need the parallel computing toolbox for your customized applciations.

The following examples are provided as the boilerplates.

[cudakernels] - call a vector add cuda kernel
[cuBLAS] - run sgemm on gpu

Software

Ubuntu 14.04
Matlab 2015b
Cuda Toolkit 7.5

Examples

cudakernels

Build the mex function for the cuda files.

Please modify the Makefile to update the compilation environment.

$ cd cudakernels && make

The mex function for vectoradd_cuda kernel will be generated.

$ ls
Makefile  nvmex  nvopts.sh  vectoradd_cuda.cu  vectoradd_cuda.mexa64

Run the vectoradd_cuda() inside Matlab.

$ matlab -nodesktop -nojvm -nosplash
>> a = rand(1, 10)
a =
    0.8147    0.9058    0.1270    0.9134    0.6324    0.0975    0.2785    0.5469    0.9575    0.9649

a1=single(a)
a1 =
    0.8147    0.9058    0.1270    0.9134    0.6324    0.0975    0.2785    0.5469    0.9575    0.9649
    
>> b = vectoradd_cuda(a1)
b =
    1.8147    1.9058    1.1270    1.9134    1.6324    1.0975    1.2785    1.5469    1.9575    1.9649

cuBLAS

Test SGEMM. You can modify the MEXFILES for your kernel function in the Makefile.

Run the sgemm_cublas() inside Matlab.

$ matlab -nodesktop -nojvm -nosplash
>> A = [1 2 3 4; 5 6 7 8]
A =
     1     2     3     4
     5     6     7     8
     
>> B = [1 5; 2 6; 3 7; 4 8]
B =
     1     5
     2     6
     3     7
     4     8
     
>> C = zeros(size(A,1), size(B,2))
C =
     0     0
     0     0
>> C_out = sgemm_cublas(0,0,1,0, single(A), single(B), single(C))
C_out =
    30    70
    70   174

>> A * B
ans =
    30    70
    70   174

References

Voilà! You made it.