Open rtawfik01 opened 1 month ago
Added the standalone test here: #10057
Does this manifest as
"compute_kernel_api/pack_untilize.h:28:48: error: too many arguments to function 'void llk_pack_untilize_init() [with long unsigned int block_ct_dim = 1]'
28 | PACK(( llk_pack_untilize_init<block_ct_dim>(ocb) ));"
Does this manifest as
"compute_kernel_api/pack_untilize.h:28:48: error: too many arguments to function 'void llk_pack_untilize_init() [with long unsigned int block_ct_dim = 1]' 28 | PACK(( llk_pack_untilize_init<block_ct_dim>(ocb) ));"
That's the first part, the second part is after the compilation error it mismatches, but that's because pack_untilize algorithm was not ported to the new blackhole packer functionality
Pushed here: https://github.com/tenstorrent/tt-metal/pull/10422 and now all of these tests pass:
pytest -svv tests/ttnn/unit_tests/operations/test_activation.py
pytest -svv tests/ttnn/unit_tests/test_to_layout.py
pytest -svv tests/ttnn/unit_tests/operations/test_elt_binary.py
TT_METAL_WATCHER=5 TT_METAL_DPRINT_CORES=0,0 TT_METAL_SLOW_DISPATCH_MODE=1 ./build/test/tt_metal/test_untilize
Will remove "P1_critical" label. But @ttmtrajkovic there is still some testing missing:
MATH(( llk_math_hw_configure_disaggregated<true>() ));
should be called, because it needs to be on the math thread, for max pool tests the above config needs to be done when initializing the reduce
Update on this issue.
Have some fixes on this branch: rtawfik/reduce_untilize_a_b
With the testing that is missing:
MATH(( llk_math_hw_configure_disaggregated<true>() ));
will be added to the api init calls for the pack_untilize & fused reduce
Pack untilize kernel for blackhole is incomplete, due to packer implementation changing for blackhole. This issue is 2 parts:
[x] Have a standalone C++ test for pack untilize, similar to unpack_untilize:
[ ] Bringup pack untilize kernel for blackhole -> check if this will give better performance similar to wormhole_b0
fyi @abhullar-tt @ttmtrajkovic