DPCPP Atomic/Reduce on float/double

[jrouzaud]

Dear all,

We are currently working on porting our OpenMP/C++ code to DPCPP for evaluation purpose. We were very interested by collective reduce+atomic operations.

But, it seems that reduce and atomic (buffer access:mode:atomic) are not available for floating point format. Is there any plan to support it ? An experimental branch available ?

Thanks in advance,

Jonathan Rouzaud

[GouthamK_Intel]

Hi Jonathan,

As per the SYCL - 1.2.1 standards, for atomics the data types which are permitted to use are int, unsigned int, long, unsigned long, long long, unsigned long long and float. Though float is only available for the store, load, and exchange member functions.

However, we are forwarding your query to the concerned team for more detailed information.

Thanks & Regards

Goutham

[jrouzaud]

Hi,

Thanks a lot. Indeed, I could reduce the precision of the floating point operation nevertheless, I will wait for the accumulate/add to be available.

Regards,

Jonathan