Forum > GPU crunching

Modified SETI MB CUDA + opt AP package for full GPU utilization

<< < (10/58) > >>

Raistmer:
Well, 6.06 sources differ form my own.
For example:

Mine:
          // TODO: Deallocate these at the end of the function
#ifdef USE_CUDA
         if(gSetiUseCudaDevice)
            cudaAcc_fftwf_plan_dft_1d(FftNum, FftLen, NumDataPoints);
   
From Berkeley's repository:

#ifdef USE_CUDA
            // Create FFT plan configuration
         if(gSetiUseCudaDevice)
            {
            if(cudaAcc_fftwf_plan_dft_1d(FftNum, FftLen, NumDataPoints))
                {
                    // If were here, something went wrong in the plan.
                    // Possibly ran out of video memory.  Destroy what we
                    // have and fallback to host CPU processing.
                    fprintf(stderr, "setiathome_CUDA: CUDA runtime ERROR in plan FFT. Falling back to HOST CPU processing...\n");
                    cudaAcc_free();
                    gSetiUseCudaDevice = 0;

                }
            }

#endif

Will rebuild mod with new sources and re-enable access to it for testing.

Raistmer:
Ok, here is a new package, based on current (380) revision of Berkeley's SETI CUDA branch.


[attachment deleted by admin]

Jason G:
Does 6.06 fix the memory leak at least ?

Raistmer:

--- Quote from: Jason G on 25 Dec 2008, 06:27:03 pm ---Does 6.06 fix the memory leak at least ?

--- End quote ---
I think it fixes nothing. But wanna check this for sure.
Joe's proposal about leak fixing still pending. I didn't check this in sources still but pretty sure that modification not included in sources by Berkeley.
rev 380 from 17 December ....

ADDON: now I will use this "up to date"mod to check broken results standalone too.
I didn't see any problems with PG* VHARs before, only with VLARs....

And first E-mailed sources differ from Berkeley's rep ones... So it's worth to try...

Raistmer:
Here is new build with logging of AR of overflowed tasks.
It will create and append later file r_debug.txt on C:\ and will write AR of overflowed task there.
Sure there will be "legal" overflows too, but any overflows not validated against CPU result worth to report.

Just replace CUDA MB executable from my last package with this one. Name remains same.

Logging mod is:

  if (e == RESULT_OVERFLOW) {
        fprintf(stderr, "SETI@Home Informational message -9 result_overflow\n");
        fprintf(stderr, "NOTE: The number of results detected exceeds the storage space allocated.\n");
      {//R: Logging of AR of overflowed task
         FILE* f=fopen("C:\\r_debug.txt","a");
         if(f){
            fprintf(f,"AR of overflowed task is\t%f\n",swi.angle_range);
            fclose(f);
         }
      }


[attachment deleted by admin]

Navigation

[0] Message Index

[#] Next page

[*] Previous page