#18332: Update BN Kernel

[VirdhatchaniKN]

Ticket

#18332

Problem description

The current BN implementation splits tensor reading between NCRISC (Input tensor) and BRISC (input stat tensor) , while BRISC also handles writing. This stalls BRISC from writing until compute is done, and also writes are more expensive. It also creates performance inconsistencies.

What's changed

Moved all the input reads to reader file

Checklist

• All post-commit tests
• Blackhole Post commit
• Full new models - Link to test
• (Single-card) Demo tests
• (Single-card) Device perf regressions - Same as in main
• Single-card Model perf tests

[sjameelTT]

How is the functionality atm, did it solve your problem with bigger dimensions?

[sjameelTT]

this works, but you could also consider issuing the reads for both the mean tile and the var tile, and adding the barrier for both of them:

cb_reserve_back(cb_id_batch_mean, onetile);
cb_reserve_back(cb_id_batch_var, onetile);
noc_async_read_tile(tile_offset_stat, batch_mean, l1_write_addr);
noc_async_read_tile(tile_offset_stat, batch_var, l1_batch_var_write_addr);
uint32_t l1_write_addr = get_write_ptr(cb_id_batch_mean);
uint32_t l1_batch_var_write_addr = get_write_ptr(cb_id_batch_var);
noc_async_read_barrier();
FILL_TILE_WITH_FIRST_ELEMENT(cb_id_batch_mean);
FILL_TILE_WITH_FIRST_ELEMENT(cb_id_batch_var);
cb_push_back(cb_id_batch_mean, onetile);
cb_push_back(cb_id_batch_var, onetile);

You will be able to amortize both of the reads with each other. The way you have it currently should work too though, so it's fine, and let's get this working completely before you make micro-optimizations.

[VirdhatchaniKN]

Sure. I will update this change. But this is not solving the hang issue that we are working on. Need to work more on that.

[VirdhatchaniKN]

Hi @sjameelTT , I've updated the files

[tt-aho]

Is there a reason this is a loop of one? Might as well remove the loop unless you expect this to actually change?

Overall I think it makes things clearer for tile_regs_wait() to come after tile_regs_commit() just for readability, even if your setup works for single loop. This loop and how it is set up with the tile_regs call also doesn't work if the loop is > 1.