Hi everyone,

The OpenCL version of this app has been launched as a beta app.

AMD/Intel users can contribute with their GPUs for this app now.

Please let us know if you encounter any issues. We'll be monitoring and pull it out of beta once it seems stable.

We've released v1.01 of this app to address computation errors.

Things appear to be working on my end (nvidia, windows)

On 1080Ti it works fine, but is ~ 40% slower than cuda app.

on AMD RX590 GME I get the following error on all Wu's, Einstein@home does work correctly with opencl on this card

Wu #14354898

<core_client_version>8.2.4</core_client_version>
<![CDATA[
<message>
Incorrect function.
(0x1) - exit code 1 (0x1)</message>
<stderr_txt>
boinc gpu 0 gpuindex: 0
No checkpoint to load
Error building program: CL_BUILD_PROGRAM_FAILURE

</stderr_txt>
]]>

---

on AMD RX6600 is work fine...

---

On 1080Ti it works fine, but is ~ 40% slower than cuda app.

Same, don't send the OCL tasks to Nvidia cards. Just remove the app. All my cards are slower with the 5080 being over twice as long, 6:35 to 14:26

Cuda:
Average processing rate 0.67 GFLOPS

OCL:
Average processing rate 0.31 GFLOPS

On 1080Ti it works fine, but is ~ 40% slower than cuda app.

Same, don't send the OCL tasks to Nvidia cards. Just remove the app. All my cards are slower with the 5080 being over twice as long, 6:35 to 14:26

Cuda:
Average processing rate 0.67 GFLOPS

OCL:
Average processing rate 0.31 GFLOPS

Nice for you but the Cuda version does not work on my GTX750Ti card and the OpenCL works OK.

So the version should stay but the Admin could perhaps create it as a different app giving the user a choice.

Maybe an extra check box could be added under settings "allow OpenCL apps" with an explanation box below stating that this box needs to be checked to receive OpenCL work for AMD, Intel or Nvidia GPUs older than RTX 2000.

---

On 1080Ti it works fine, but is ~ 40% slower than cuda app.

Same, don't send the OCL tasks to Nvidia cards. Just remove the app. All my cards are slower with the 5080 being over twice as long, 6:35 to 14:26

Cuda:
Average processing rate 0.67 GFLOPS

OCL:
Average processing rate 0.31 GFLOPS

You can use something like this in de app_config.yml to run 2 (or more) open_cl tasks at the same time:

  <app_version>
    <app_name>repeatdecor</app_name>
    <plan_class>opencl_nvidia</plan_class>
    <ngpus>0.5</ngpus>
    <avg_ncpus>0.9</avg_ncpus>
  </app_version>
  <app_version>
    <app_name>repeatdecor</app_name>
    <plan_class>cuda</plan_class>
    <ngpus>1.0</ngpus>
    <avg_ncpus>0.9</avg_ncpus>
  </app_version>

I can confirm that the OpenCL app is MUCH slower than the CUDA app.

Getting issues with my AMD RX580. Computational Error straight up
Updated to the latest drivers and still the same issues.

I guess I'll lurk the forums and wait for a fix

Hi everyone,

I want to stress that the OpenCL app is in active development and is in beta.

Nvidia users, if you want to run only CUDA for the time being, you can opt out of beta.

That said, the next version should resolve most/all of the performance regression. Thank you for your patience.

What about OpenCL on MacOS?
I'm running Catalina (10.15) on an old (pre-AVX) Mac Pro with an RX570.
It's doing fine on Amicable Numbers, Einstein and PrimeGrid.
Wouldn't call it fast, though. ;-)

---

- - - - - - - - - -
Greetings, Jens

I'm interested :)

I don't know right now how easily I can get it running, but maybe I can cross-compile for darwin. I'm concerned about whether there's a limitation with Metal on newer macs - but it'd be cool to get this running on apple silicon's GPU for example

On 1080Ti it works fine, but is ~ 40% slower than cuda app.

Same, don't send the OCL tasks to Nvidia cards. Just remove the app. All my cards are slower with the 5080 being over twice as long, 6:35 to 14:26

Cuda:
Average processing rate 0.67 GFLOPS

OCL:
Average processing rate 0.31 GFLOPS

You can use something like this in de app_config.yml to run 2 (or more) open_cl tasks at the same time:

  <app_version>
    <app_name>repeatdecor</app_name>
    <plan_class>opencl_nvidia</plan_class>
    <ngpus>0.5</ngpus>
    <avg_ncpus>0.9</avg_ncpus>
  </app_version>
  <app_version>
    <app_name>repeatdecor</app_name>
    <plan_class>cuda</plan_class>
    <ngpus>1.0</ngpus>
    <avg_ncpus>0.9</avg_ncpus>
  </app_version>

I already am running 2x on every card for WUProp hours.

An app_info will be needed to force CUDA.

@boysanic there's a small issue, choosing the gpu doesn't work right if multiple platforms (nvidia+amd) are present

For CUDA tasks, the platform is always CUDA, so using aid.gpu_device_num for indexing is fine. For OpenCL, however, there can be multiple platforms, so you should check those as well. The solution is straightforward:

#include "boinc_opencl.h"

cl_platform_id platformId = nullptr;
cl_device_id deviceId = nullptr;
int result = boinc_get_opencl_ids(&deviceId, &platformId);
if(CL_SUCCESS != result) {
  // error
}
...

One of my team buddies is running on a not that old (AVX capable) Mac Pro with a CAL AMD Radeon Pro 580X Compute Engine (basically the same as an RX 580). IIRC, his OS is Ventura.

If you go down to High Sierra (10.13) you'd have CUDA as well.

---

- - - - - - - - - -
Greetings, Jens

I think there is a problem with the validation after a restart from a checkpoint.

If a Wu runs in one go from start to finish is validates just fine.
But if it has been suspendend and restarted from a checkpoint it is marked as invalid.

Thanks for the report! I'll look into it.

on AMD RX590 GME I get the following error on all Wu's, Einstein@home does work correctly with opencl on this card

Wu #14354898

<core_client_version>8.2.4</core_client_version>
<![CDATA[
<message>
Incorrect function.
(0x1) - exit code 1 (0x1)</message>
<stderr_txt>
boinc gpu 0 gpuindex: 0
No checkpoint to load
Error building program: CL_BUILD_PROGRAM_FAILURE

</stderr_txt>
]]>

I have the same on AMD 5800h integrated GPU, don't have any stand-alone GPU's to try.