async_work_group_copy with destination of mixed type not supported

[robinchrist]

Hello,

I tried to check a kernel which heavily depends on async_work_group_copy.

3 warnings generated. OpenCL3.opt.bc: error: async_work_group_copy with destination of mixed type not supported Please contact the developers; this is an implementation limitation

I don't know what this error means, since the destinations are always __local float4*.

[afd]

Hi Robin

Would you be able to send a small code example that issues this warning where you believe it should not?

Thanks

Ally

On 28/01/2018 23:10, Robin Christ wrote:

Hello,

I tried to check a kernel which heavily depends on async_work_group_copy.

|3 warnings generated. OpenCL3.opt.bc: error: async_work_group_copy with destination of mixed type not supported Please contact the developers; this is an implementation limitation|

I don't know what this error means, since the destinations are always __local float4*.

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/mc-imperial/gpuverify/issues/26, or mute the thread https://github.com/notifications/unsubscribe-auth/AEhSXdJDjhH3m95wL2-lc5qWaKWHu3n2ks5tPP58gaJpZM4Rv5wa.

[robinchrist]

Hi Ally,

I will need to reduce the source code, but I could provide the whole source code via personal message or email. Would that be okay for you?

[jeroenk]

I don't know what this error means

It's an internal problem in Bugle, GPUVerify's LLVM to Boogie translator. It most likely signals that Bugle's pointer analysis fails to do the right thing on its first iteration. It would be really helpful to see a reduced kernel that triggers this behaviour.

[robinchrist]

It would be really helpful to see a reduced kernel that triggers this behaviour. I will provide a reduced kernel as soon as possible.

The kernel uses a double buffering technique with two buffers, where data is asynchronously loaded into the buffer which is not used. After the calculation and the async copy finished, the pointers to the buffers are swapped. I assume Boogie is not able to "understand", that the buffers or rather the pointers to them are swapped.

[jeroenk]

Going by your description, I guess there’s a loop in which this all happens. If so, cutting down the kernel might be as simple as keeping the loop with the async call(s) and maybe one access to the array you’re actually using.

On 29 Jan 2018, at 08:48, Robin Christ notifications@github.com wrote:

It would be really helpful to see a reduced kernel that triggers this behaviour. I will provide a reduced kernel as soon as possible.

The kernel uses a double buffering technique with two buffers, where data is asynchronously loaded into the buffer which is not used. After the calculation and the async copy finished, the pointers to the buffers are swapped. I assume Boogie is not able to "understand", that the buffers or rather the pointers to them are swapped.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.

[jeroenk]

I believe I locally have a fix for this issue.

@robinchrist How did you obtain GPUVerify? Did you compile it yourself, or are you using a nightly build? I'm trying to figure out how to best get the fix to you to try.

[robinchrist]

@jeroenk I used the nightly build available from the website (packed 2015-09-25).

[jeroenk]

@robinchrist Do you use the Windows or the Linux build?

[robinchrist]

@jeroenk I am using Windows 10.

[jeroenk]

@robinchrist Thanks. In that case I'll create a new Windows nightly for you. Please give me a few days.

[jeroenk]

@robinchrist I dropped a new build here: https://github.com/jeroenk/gpuverify/releases/tag/fusion-async (grab the deploy.zip file from there).

Please let me know if this solves your problem.

[jeroenk]

@robinchrist Did the updated build work for you?

[jeroenk]

We believe we fixed this. Please open a new issue if there are still problems with this.