Investigate hipSYCL stream and execution lane usage
The latter parts of the following discussion from !1833 (merged) should be addressed:
-
@mark.j.abraham started a discussion: (+6 comments) Currently hipSYCL seems to sometimes return a different
hipStream_t
, even when run on the same SYCL queue. See https://github.com/illuhad/hipSYCL/issues/614 for discussion. Hopefully we can get that fixed, otherwise we will have to set the stream every time before we callrocfft_execute
).
We should investigate whether there is performance advantage in using e.g. a common hipSYCL execution lane between stages of PME, e.g. spread and 3DFFT.