Profiler OpenCL results explanations

Accelerated Computing CUDA CUDA Programming and Performance
1

Hi,
i’m performing test and collecting results with NVidia Compute Visual Profiler.
THe problem is, when executing kernel the CPU Time is comparable to GPU time.
Whereas i would expect a much much lower CPU time.
Can someone explain such strange behaviour?
Thanks in advance
V.

2

It could be anything - not a suitable parallel problem, hight register usage or high local memory usage → low occupancy on GPU, bad global memory accessing scheme,… please specify more how your kernel looks like

3

Thanks for reply,
my program try to compare two vector with size of 12 Mb
I use no local memory.
to create the vector i use
Input1P = clCreateBuffer (Contesto,CL_MEM_READ_ONLY |
CL_MEM_ALLOC_HOST_PTR,Msg_size1,NULL,&Error);
OutputP = clCreateBuffer (Contesto,CL_MEM_WRITE_ONLY |
CL_MEM_ALLOC_HOST_PTR,(Nres*sizeof(unsigned)),NULL,&Error);

    my kernel is :
    
    __kernel void Confronto(__global char* msg1,__global char* msg2,int
    size,__global unsigned *out,int ln)
    {
           int gidx,gidy,lidx,lidy,i,index,temp;
           i=0;
           gidx=get_global_id(0);
           gidy=get_global_id(1);
           index = gidy *  get_global_size(0) + gidx;
           temp=index*ln;
           while (i<ln)
           {
                   if (msg1[temp+i] != msg2[temp+i]){ out[index]=temp+i; return;}
                   i++;
           }
           out[index]=0;
    
    }
    NDRange size in order to 159 x 159;
    
    the profiler return a output like:
                                   GPU time  CPU time
    1.29072e+15     Confronto       125761    125979
    
    i don't understand why cpu time is so similar to gpu time.
    i would expect a much much lower CPU time.
    
    Thanks in advance
    V.
4

Aaa, you meant the CPU time written in OpenCL profiler! From your first post I thought, you run your OpenCL code on CPU and compare it to GPU time…

So the copy paste from the profiler help defines what is the CPU time:
CPU Time: It is sum of GPU time and CPU overhead to launch that Method. At driver generated data level, CPU Time is only CPU overhead to launch the Method for non-blocking Methods; for blocking methods it is sum of GPU time and CPU overhead. All kernel launches by default are non-blocking. But if any profiler counters are enabled kernel launches are blocking. Asynchronous memory copy requests in different streams are non-blocking.

5

I understand.
Thanks for explanation.
V.