Forum > Alternate Hardware platforms
OSX Multibeam OpenCL question
Raistmer:
last call is oclFFT one.
Joe Fox:
I will be back to working on this today and should be able to get the parameters for the oclFFT call that is causing the error. Again to reiterate though, the OPEN_CL_INTEL version works correctly with accurate test results when run on an nvidia as well as AMD based Mac machines.
Raistmer:
Intel version probably runs slower than would be ATi or NV builds. Cause code paths are different for all these 3 builds.
Also, intel uses most precise but slowest trigonometry. Other vendors have enough precision in native trig functions to allow simpler and faster math.
EDIT: in other words, if you limited in time you could build all types with same "intel" path (take care for platform selection code though!) but if you can better to use optimized path for each of vendors.
//"intel" path currently almost identical to ATi non-HD5 path. And it's not right cause iGPU has real local memory. I'm in process of changing that so soon USE_OPENCL_INTEL could give enother results for your type of testing.
Joe Fox:
Ah, thanks for the clarification! That explains what I'm observing.
Joe Fox:
I'm not sure if this information will be useful or not, but here is the debug message I entered along with the code that was causing the error:
--- Code: --- fprintf(stderr,"INFO: oclFFT done no strip. plan_ptr: %p, FtfNum=%d, fftlen=%d, NumBlockFfts=%d, chirplen=%d\n", plan, FftNum, fftlen, NumBlockFfts, chirplen);
err |= clFFT_ExecuteInterleaved_mb(cq, plan[FftNum],NumBlockFfts, clFFT_Inverse, gpu_ChirpedData,gpu_WorkData, FFTbuf, 0, NULL, NULL);
--- End code ---
And the output of that line is:
INFO: oclFFT done no strip. plan_ptr: 0x7fff4fecf060, FtfNum=0, fftlen=8, NumBlockFfts=131072, chirplen=1048576
Error in mb oclFFT_2: -49
ERROR: OpenCL kernel/call 'non-strip fft' call failed (-49) in file /Users/joe/projects/seti/sah_v7_opt/AKv8/client/analyzeFuncs.cpp near line 3823.
This is running the USE_OPENCL_HD5XXX define on an AMD Mac Pro.
I'm not sure what exactly I should be looking for, but plan looks to be non-null, and the index (FftNum) is 0, so I would expect it to be in bounds. Any insight?
Navigation
[0] Message Index
[#] Next page
[*] Previous page
Go to full version