Thread 'OpenCL GPU applications in beta testing'
Message boards : News : OpenCL GPU applications in beta testing
Message board moderation
| Author | Message |
|---|---|
|
Send message Joined: 21 Oct 26 Posts: 67 Credit: 217,462,082 RAC: 14,913,736 |
Hello all, I have put up OpenCL GPU versions of our application that should help support some more devices not currently supported by the CUDA and HIP applications. These are intended for: Intel GPUs Nvidia GPUs with drivers older than CUDA 11 AMD GPUs that are not supported by our HIP app All systems and devices are still subject to the same basic requirements, so check that you meet them: you need:
|
TamagochNew member Send message Joined: 10 Nov 26 Posts: 4 Credit: 2,110,933 RAC: 162,694 |
I've got all tasks failed on Vega 56 (Host ID 1939): <message> Incorrect function. (0x1) - exit code 1 (0x1)</message> <stderr_txt> Kessler@home 1.06, x86_64-w64-mingw32, BOINC build; math: CORE-MATH with hardware FMA3 fused multiply-add scenario kh1_12111431010 seed 370: 35865 objects, 14610 steps of 5.0 days error: Explicit load/store type does not match pointee type of pointer operand </stderr_txt> Should I check anything on my side first?
|
|
Send message Joined: 21 Oct 26 Posts: 67 Credit: 217,462,082 RAC: 14,913,736 |
can you 1. make sure you are on the latest driver for your GPU 2. tell me what driver you are running? |
|
New member Send message Joined: 10 Nov 26 Posts: 1 Credit: 45,094 RAC: 4,075 |
Thank you for your efforts to make more gpus supported. I've just seen many wu's were downloaded and errored out and boinc stopped getting new ones thanks to daily limit, all with the exactly the same error above. AMD Radeon(TM) Graphics; 4GB VRAM; OpenCL 2.0 LEO Debris Environment Monte Carlo Sim 1.06 windows_x86_64 (opencl_amd) (ID 56) amd driver version : 31.0.21925.1001 from boinc: 16/11/2026 18:15:52 | | OpenCL: AMD/ATI GPU 0: AMD Radeon(TM) Graphics (driver version 3584.0 (PAL,HSAIL), device version OpenCL 2.0 AMD-APP (3584.0), 3.06GB, 3.06GB available, 1.15 TFLOPS peak) <core_client_version>8.2.11</core_client_version> <![CDATA[ <message> Yanlış işlev. (0x1) - exit code 1 (0x1)</message> <stderr_txt> Kessler@home 1.06, x86_64-w64-mingw32, BOINC build; math: CORE-MATH with hardware FMA3 fused multiply-add scenario kh1_12011023111 seed 352: 35865 objects, 14610 steps of 5.0 days error: Explicit load/store type does not match pointee type of pointer operand </stderr_txt> ]]> |
|
New member Send message Joined: 15 Nov 26 Posts: 1 Credit: 5,520,122 RAC: 507,505 |
I am getting errors on all of my Intel GPUs. I am using the latest drivers. <core_client_version>8.2.15</core_client_version> <![CDATA[ <message> Incorrect function. (0x1) - exit code 1 (0x1)</message> <stderr_txt> Kessler@home 1.06, x86_64-w64-mingw32, BOINC build; math: CORE-MATH with hardware FMA3 fused multiply-add scenario kh1_21021334211 seed 151: 35865 objects, 14610 steps of 5.0 days error: GPU: Intel(R) Arc(TM) B580 Graphics lacks cl_khr_fp64 (OpenCL) 2026-11-16 14:07:55 (13136): called boinc_finish(1) </stderr_txt> ]]> |
|
Send message Joined: 21 Oct 26 Posts: 67 Credit: 217,462,082 RAC: 14,913,736 |
In reply to SR97's message of 16 Nov 2026: I am getting errors on all of my Intel GPUs. I am using the latest drivers. thank you, i saw this error in the results also. it's a bug. And we should have isolated the issue with AMD GPUs also. new builds will be up soon and I'll reset everyone's daily quota who got locked out from too many bad tasks. thanks for the responses everyone! |
|
Send message Joined: 21 Oct 26 Posts: 67 Credit: 217,462,082 RAC: 14,913,736 |
Both AMD and Intel OpenCL apps will be back shortly, with a caveat for AMD. The problem with Intel was a small bug, the buffer reading the GPU capabilities was too short for the full list, leading it to wrongly conclude it didnt have FP64 support when it really did. increased the space available for that and it should work now on GPUs with FP64. the problem with the AMD systems is that the legacy driver/compiler on these old AMD GPUs in most cases cannot handle the 128-bit integer arithmetic that exists in many functions of our app. the few systems that are working fine have newer drivers (including my own test system) so it slipped through. the short term fix is that a new scheduler rule will be in place to exclude AMD systems on the old incompatible drivers. the longer term fix will be that these 128-bit function need to be rewritten as pairs of 64-bit instead. that should open up these old devices finally, but I cannot comment yet what kind of impact that will have on performance. will update soon. thanks again. |
