CUDA8.0+OpenCL+Win7 throws endless CL_OUT_OF_RESOURCES errors #2112
Comments
|
which gpu do you use and which cuda version?
|
|
latest 417.01 driver, CUDA10 SDK |
|
Found some Intel garbage in my environment which was not there to block proper location of nvidia stuff.. fixing and retesting... |
|
I think it is still broken I defined MEMORY in gpu.cpp as unsigned long
long but for opencl it must be unsigned long only.
|
|
I switched to building within CLion and how it does build-configurations and environments is weird. |
|
Okay, with corrected OpenCL include/lib it still complains the same on compile phase. EDIT: testing now with |
|
that no matter what I try, even ultra low intensity etc. |
|
Seems related to CL-CUDA throwing this error for SEGFAULTs so suspicion is alignments and scratchpads etc are disjointed so the kernel is hitting wrong pointers for buffers, or host is hitting wrong pointers for cross-access. CL-CUDA may have different sizeof() for some parts of structures (versus AMD)? Will test using the Khronos CL headers instead of the CUDA SDK 10.0 headers, although I expect the nvidia headers to be best for its own library... |
|
fixed in #2113 |
|
Retesting both CUDA8.0 for NVS 5400M and CUDA10.0 for 970GTX with patch applied... brb EDIT: CUDA10.0 works EDIT2: CUDA8.0 still gives clEnqueueReadBuffer error like before |
|
The fix should work for all NVIDIA GPUs. Please try to clean up your opencl cache. I tested the fix with sm_20, sm_35, sm_5X |
|
Cleared, same. This is SM_20 ... but laptop and Quadro (double edge case? haha) I believe it worked before, as I posted speeds during the CN2v2 design phase. Intensity 256 still not happy. Flat memory (mode 0) no help which I think I had to use before (no stride) I won't use this method ever, but it should work. CUDA generally faster. |
|
reduce to 128 since the fpu has nearly no compite units.
|
|
intensity 128 with strided off, same Also tried a build done with |
|
which cuda 8 version do you use? Is it the latest cuda 8?
|
|
Yes, 8.0-GA2 ( Rebased again on latest Agree it could be something with my environment or compiler, although nothing should have changed with the v14.0 toolkit and compiler. CUDA10 works great (both native and CL) even with 14.16 compiler I just updated and tried it out. Half expected it to complain without 14.15 based on older version headers having tight version slotting. |
|
ahh you are using windows. I used always linux ;-)
I have no windows system with auch an old gpu :-(
|
|
Any solution here? I have the same effect since 2.10.0 Thanks |
|
wich currency? please also provide your amd.txt and nvidia.txt
|
|
This works fine with newer versions but it's slower than CUDA directly. |
|
"currency" : "monero", It's a GTX 570 - since it doesn't work with CUDA I used OpenCL for it on 2.5.0 and it worked. Since 2.10.0 I got this error. Here are my parameters: xmr-stak.exe --noNVIDIA --openCLVendor NVIDIA --noCPU amd.txt: ], nvidia.txt is should be irrelevant since starting with -noNVIDIA I allready tried deleting the OpenCL Cache, unfortunately no luck. Any ideas and suggestions are very appreciated. Thank you. |
|
could you set the intensity to 384 and test it again.
|
|
I have tested nv-OpenCL on my CUDA8 also and it still streams the out of resources errors. Also tested all intensities even stupid small like 128/4, no difference. |
still no luck - here the complete output: |
|
Similar, I however do not get the compile errors, just the others |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Previous to this compile-time crash, it would just crash when firing up the work.
Last time I had it work at all was during CN2v2 pre-fork testing phases. Sometime after final CN2v2 and other AMD fixes, the non-AMD OpenCL broke (twice I think).
The text was updated successfully, but these errors were encountered: