Closed pgkeller closed 1 month ago
This is now POR. The large tmp buffers in prefetch/dispatch need to go on the stack, stack size should be increased as needed (based on watcher). Then stack size needs to bump up to match the data @tt-dma collected
@tt-asaigal is this done?
@tt-asaigal is this done?
Yes, changes for this were merged to main here: https://github.com/tenstorrent/tt-metal/commit/df4d0131f9f4b455c92b63d40c113eff2d658e02.
I can close the issue.
This is pending more data on kernel stack usage
Prefetcher/dispatcher put a table in global memory (copied to local) that chews up a lot of global space. For this, we shrunk the stack size. However, there are kernels that put large tables on the stack. Having some kernels to A and others B leads to out-of-local-memory errors, we need consistent guidelines.
Plan is to make the guidelines to be to put these items on the stack. That is more clear and, for trisc* necessary.