Open mywoodstock opened 3 weeks ago
@sjameelTT @tarafdarTT Hey guys, any update on the ETA for this?
Mid next week
On Thu, Sep 26, 2024 at 15:20 Abhinav Sarje @.***> wrote:
@sjameelTT https://github.com/sjameelTT @tarafdarTT https://github.com/tarafdarTT Hey guys, any update on the ETA for this?
— Reply to this email directly, view it on GitHub https://github.com/tenstorrent/tt-metal/issues/12184#issuecomment-2378034610, or unsubscribe https://github.com/notifications/unsubscribe-auth/BAK3IAZITLKP5NLGTXEVL7TZYSCCFAVCNFSM6AAAAABNS4QRKSVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDGNZYGAZTINRRGA . You are receiving this because you were mentioned.Message ID: @.***>
Isolated the issue to
padded_offset_bytes = align(input_unit_size, input.buffer()->alignment());
in interleaved_to_sharded.When input width is
16
(BFLOAT16, so 32b), in ROW_MAJOR_LAYOUT, the generated sharded output has alternating rows of 0s with the data, e.g.:In this use case, the input unit size is 32b, and DRAM alignment on BH is 64b, so
padded_offset_bytes
is also 64, which is the increment value for the write ptr.Branch:
asarje/debug-halo-bh
Test:pytest "tests/ttnn/unit_tests/operations/test_maxpool2d.py::test_run_max_pool[dtype=DataType.BFLOAT16-dilation=(1, 1)-stride=(2, 2)-padding=(1, 1)-kernel_size=(3, 3)-act_shape=[1, 16, 1056, 160]-device_params={'l1_small_size': 24576}]"