Sharing an implementation of SGEMM using GPU and OpenCL

function test(runs)
    variable runs
   
    variable m
    variable dim = 8192
   
    make/o/n=(runs) CPUTakes, GPUTakes
    make/o/n=(dim,dim) AA, BB, CC, DD
   
    variable timer, GPUTook, CPUTook
   
    for (m=0;m<runs;m=m+1)
       
        timer=startMSTimer
        GPU_SGEMM(AA,BB,CC)
        GPUTook=StopMSTimer(timer)/1e6
       
        GPUTakes[m]=GPUTook
       
        MultiThreadingControl setMode=8
        timer=startMSTimer
        MatrixOp/O DD = AA x BB
        CPUTook=StopMSTimer(timer)/1e6 
        MultiThreadingControl setMode=0
       
        CPUTakes[m]=CPUTook
        endfor
       
    variable/g CPUTake=mean(CPUTakes)
    variable/g GPUTake=mean(GPUTakes)
   
    variable/g CPUGFLOPS = 2*dim*dim*dim/CPUTake*1e-9
    variable/g GPUGFLOPS = 2*dim*dim*dim/GPUTake*1e-9
   
    print "CPU SGEMM GFLOPS = "+num2str(CPUGFLOPS)
    print "GPU SGEMM GFLOPS = "+num2str(GPUGFLOPS)
   
    killwaves/Z CPUTakes, GPUTakes, AA, BB, CC, DD
   
end