Problems with latest commit

[Imran]

Hi accULL team,

first of all, thanks a lot for the nice efforts and nice work. It was really easy for me to download, install and run some examples on GPUs by using accULL.

When i started compiling laplace example, i came to know about issues regarding 0.3 version so i committed the latest from repository (0.4).
However, with 0.4 even my simple code did not work. Attached is a complete example which compiles ok, but gives segmentation fault
on execution.

Best Regards,
imran

[Jose Lucas Grillo Lorenzo]

Dear Imran,

We are working in the last release of our compiler (accull 0.4). It will include deep changes in the management of reductions for a better support of future characteristics. 

This is the reason why the runtime reductions have changed slightly. 
In the latest commits accull reductions have a copy transfer implicitly defined in runtime.
Thus, if you comment copy(sum) in your codes you should get a better performance due to the redundancy avoided. Please confirm. Later, we will study optimizing this behavior.   
 
Notice, that right now, reductions are not supported in the OpenCL backend. 
The good news are that last week, we succeded in detect and fix the situation with your code to avoid seg faults in our run-time. You can check it with the CUDA backend.
 
Thank to your feedback we were able to detect some bugs in the new features. Now, the accull commit 7d170f372e09 is updated to support your code without modifications, although including the unnecessary extra transfer. 
 
Do not hesitate to contact us if you need further assistance. We appreciate your feedback.

Best regards,

     J. Lucas Grillo

--
You received this message because you are subscribed to the Google Groups "accULL Support" group.
To unsubscribe from this group and stop receiving emails from it, send an email to accull+un...@googlegroups.com.
To post to this group, send email to acc...@googlegroups.com.
Visit this group at http://groups.google.com/group/accull.
For more options, visit https://groups.google.com/d/optout.

[Imran Ashraf]

Dear Lucas,

Thanks a lot.

On 12/02/2014 05:36 PM, Jose Lucas Grillo Lorenzo wrote:

Dear Imran,

We are working in the last release of our compiler (accull 0.4). It will include deep changes in the management of reductions for a better support of future characteristics.

Looking forward to it.

This is the reason why the runtime reductions have changed slightly. 
In the latest commits accull reductions have a copy transfer implicitly defined in runtime.
Thus, if you comment copy(sum) in your codes you should get a better performance due to the redundancy avoided. Please confirm. Later, we will study optimizing this behavior.  

I will do a detailed testing as soon as possible and will confirm it.

Notice, that right now, reductions are not supported in the OpenCL backend. 
The good news are that last week, we succeded in detect and fix the situation with your code to avoid seg faults in our run-time. You can check it with the CUDA backend.

I dont have openCL setup, so i mostly use CUDA backend.

Thank to your feedback we were able to detect some bugs in the new features. Now, the accull commit 7d170f372e09 is updated to support your code without modifications, although including the unnecessary extra transfer.

Yes, it works. I will also do testing with some other examples.

Do not hesitate to contact us if you need further assistance. We appreciate your feedback.

Thanks a lot.
Kind Regards,
imran