Hi everyone,

1.03 of this app is now live, and I've confirmed it at least launches this time before putting out this news post.

Please give it a try on your machines. On my 3090 is seems to be a bit faster than the CUDA version in my limited testing. But YMMV.

The openCL Wus are a bit, i think around 3 to 5 sec, slower than the Cuda Wus. It`s on my 4090 and on my 2080ti the same thing.

This version crashed on my 750Ti card. with the error

Error waiting for kernel completion: CL_INVALID_COMMAND_QUEUE
Error reading final checksum: CL_OUT_OF_RESOURCES

1080Ti ~ 4% slower than CUDA app.

On my RTX 4060, OpenCL is on average about 7 seconds slower than CUDA (759 seconds versus 752 seconds, which amounts to roughly a 1% difference).

I hope you split the repeatdecor applications choice selection out in the Project Preferences list before you move the new OpenCL applications out of beta status. I don't want to mix OpenCL applications with default CUDA applications on my gpus. OpenCL apps don't run, only error.

Before I got a chance to untoggle test applications, I errored over 160 OpenCL tasks when I got sent that app and tasks.

You cannot run OpenCL apps along with CUDA apps when you are running the Nvidia mps-server application and environment.

I'm not sure that we have an option to do that since the project preferences are per-app, not per-app version, to my understanding.

The best I could do is discontinue the nvidia variant of the opencl app. It just complicates things on my end since I'll have to maintain both apps. And there are some folks with the opposite problem as you where OpenCL runs and CUDA does not.

A dirty hack but make one of them a separate app name, not just a plan class.

A dirty hack but make one of them a separate app name, not just a plan class.

that would essentially make it a new subproject, and probably make cross validation impossible on the server side.

I'm not sure that we have an option to do that since the project preferences are per-app, not per-app version, to my understanding.

The best I could do is discontinue the nvidia variant of the opencl app. It just complicates things on my end since I'll have to maintain both apps. And there are some folks with the opposite problem as you where OpenCL runs and CUDA does not.

you could just keep the nvidia_opencl app in beta status. that way folks who dont want the nvidia_opencl app can just leave beta turned off, and folks who do want the opencl app can enable beta to get that app.

That's a good point Ian (& Steve? Not sure what you prefer for a shortened form of your name.)

The best path forward is likely to keep opencl_nvidia in beta perpetually and allow those with issues on CUDA to opt into the opencl like that.

I'm not aware of a boinc-native way of adding the checkbox for opencl in project preferences otherwise

I thought you could do that with the server platform choices. If you can't achieve that, I will be forced to move all apps onto the anonymous platform which makes things a little bit more complicated.

Or just remove the OpenCL on Nvidia app.

OpenCL apps never approach the speed and efficiency of standard CUDA apps if built from the same codebase.

They can't since OpenCL code on Nvidia requires extra translation to perform the direct computations that CUDA code natively uses.

I'm not sure that we have an option to do that since the project preferences are per-app, not per-app version, to my understanding.

The best I could do is discontinue the nvidia variant of the opencl app. It just complicates things on my end since I'll have to maintain both apps. And there are some folks with the opposite problem as you where OpenCL runs and CUDA does not.

you could just keep the nvidia_opencl app in beta status. that way folks who dont want the nvidia_opencl app can just leave beta turned off, and folks who do want the opencl app can enable beta to get that app.

But what happens when he releases a new beta CUDA app for some new project then which requires you to toggle beta apps back on? Catch-22 situation then.

Seems like this project doesn’t leave stuff in beta for long. Or you turn off the offending subproject with the checkbox selection until it’s cuda app is out of beta.

Anonymous platform is disabled here. So that’s a no-go. You won’t get any work if you try anonymous platform. I’ve tried before.

That's a good point Ian (& Steve? Not sure what you prefer for a shortened form of your name.)

The best path forward is likely to keep opencl_nvidia in beta perpetually and allow those with issues on CUDA to opt into the opencl like that.

I'm not aware of a boinc-native way of adding the checkbox for opencl in project preferences otherwise

Good plan but can you then please go back to the 1.01 version without the performance mprovements the Cuda users wanted because it was to slow.
Version 1.01 worked on my older card, The 1.03 version not.

How do you get the opencl app to run on NVIDIA GPUs? I want to time it myself since the CUDA app keeps erroring out with "PTX JIT compilation is disabled" and if the opencl app has similar timings then it should be a fine band-aid fix.

You've got your environent incorrectly set then.

I somehow disabled PTX JIT compilation when I was doing something. I deleted the variable that disabled it and now it runs fine.

With the sip lumpsum calculator, compare SIP and lumpsum investments. A useful guide for better decision-making.