Open Pennycook opened 4 months ago
KhronosGroup/SYCL-Docs#461 is related here. I remember that we put on hold any work on marray
support for group algorithms because of it, but perhaps our analysis was wrong and we should have added support for at least some of the overloads/algorithms. Tagging @jzc for awareness
Before we add marray
support we should decide whether we extend SPIR-V to support OpTypeArray to avoid duplicate effort. If we don't extend SPIR-V we should not forward to the SPIR-V group algorithm but instead implement the algorithm in the SYCL headers. Because it is inefficient to run the algorithm multiple times with smaller sizes instead of once for the whole array.
Note: this was partially implemented in #14364 by @steffenlarsen
Describe the bug
Section 4.17.3 of SYCL 2020 says:
...but DPC++ does not support the
marray
variants.To reproduce
Compilation gives the following error:
marray
does not satisfyis_vector_arithmetic
and there is no alternativeis_marray
check.Environment
Additional context
There doesn't appear to be a corresponding test for group algorithms with
marray
arguments in the SYCL-CTS.