Forum Discussion
New ContributorLoad/store cannot be vectorised - local memory
That compiler warning in particular is a very misleading warning and it does not always point to an actual problem in your code. Looking at the report, both the load from and the store to global memory are coalesced into 128-bit accesses which points to correct vectorization. The local buffer "tile" is also replicated by 28 times to provide fully-parallel non-stallable accesses. 4 times of it is because your code has 4 non-coalescable reads on line 28, and one coalescable write on line 19 (each Block RAM has two ports, writes are connected to all replicas while reads are connected to one, resulting in a replication factor of 4 for 4 reads and one write). The buffer is also replicated by 7 extra times to support 7 work-groups running concurrently in the same compute unit; this latter replication factor is a compiler decision that cannot be overridden by the user. All in all there is nothing wrong with your code and I would say you can safely ignore the warning.
New ContributorHello HRZ.
Thank you for your answer, it was really helpful. Just one more question: I'm not sure I understand the 7 work-groups running concurrently in the same compute unit part. AFAIK a work-group maps to a single CU. I wondered in the past about what happened when the pipeline had more stages than the size of the work-group, but no one gave me the answer. Will the first stages remain idle whilst the current work-group hasn't finished yet or will a new work-group enter the pipeline? I assume what happens is the latter for performance and that's the reason why the buffer is replicated 7 times, so that up to 7 work-groups can be running concurrently and the pipeline is full at every moment. Am I right?
Edit: the numbers you are providing don't match my report's. For tile, mine says there is a single bank replicated twice and each replica has 5 private copies. I'm using AOC 19.3. Are you using the same compiler version?
Thank you
Frequent ContributorYes, essentially, on top of the work-item pipelining within the same work-group, the compiler also pipelines multiple work-groups within the same CU to keep the CU as busy as possible. I think the number "7" is calculated by dividing the CU pipeline depth by the work-group size.
I generated the report with both v16.1.2 and 19.4, both gave me the same replication factors. Can you archive and attach your report folder so that I would take a look at it?
