#0: added api for generating corerangeset from given subcoregrid #15841
Conversation
kpaigwar
requested review from
ayerofieiev-tt,
dmakoviichuk-tt,
cfjchu,
TT-BrianLiu,
abhullar-tt,
pgkeller,
aliuTT,
tt-aho,
tt-dma,
tt-asaigal and
ubcheema
as code owners
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
tt_metal/common/work_split.cpp
Outdated
| CoreRangeSet num_cores_to_corerangeset_in_subcoregrids( | ||
| const CoreCoord start_core, | ||
| const uint32_t target_num_cores, | ||
| const CoreRangeSet subcoregrids, |
There was a problem hiding this comment.
the const qualified parameter subcoregrids is copied for each invocation; consider making it a reference
| const CoreRangeSet subcoregrids, | |
| const CoreRangeSet& subcoregrids, |
tt_metal/common/work_split.cpp
Outdated
| uint32_t subcoregrid_height = subcoregrid.grid_size().y; | ||
|
|
||
| if (row_wise) { | ||
| for (uint32_t y = current_start_core.y; y <= subcoregrid.end_coord.y; ++y) { |
There was a problem hiding this comment.
could you split this function into two subfunctions? row_wise and col_wise. this for loop is very large.
YOu simple can create two lambdas at least and call on of them in the cycle.
There was a problem hiding this comment.
@dmakoviichuk-tt converted them to lambdas now.
kpaigwar
force-pushed
the
kpaigwar/numcores_in_subcoregrids
branch
from
48716a0 to
c4ac298
Compare
tt_metal/common/work_split.cpp
Outdated
| CoreRangeSet num_cores_to_corerangeset_in_subcoregrids( | ||
| const CoreCoord start_core, | ||
| const uint32_t target_num_cores, | ||
| const CoreRangeSet subcoregrids, |
There was a problem hiding this comment.
Nit. Can we name as sub_core_grids to make it similar to naming of other args? Should also be const ref
| # SPDX-FileCopyrightText: © 2024 Tenstorrent Inc. | ||
|
|
There was a problem hiding this comment.
Nit. Can you rename the file to match the function name. ex num_cores instead of numcores
kpaigwar
force-pushed
the
kpaigwar/numcores_in_subcoregrids
branch
from
e395f08 to
ea9f9cb
Compare
Problem description
To address the need for generating a CoreRangeSet from non-rectangular subcoregrids (e.g., grids of sizes [3x10, 2x10]), I added a new API
ttnn.numcores_to_corerangeset_in_subcoregridsthat can handle multiple disjoint or irregular core regions, unlike the existingttnn.numcores_to_corerangesetfunction, which assumes a single rectangular core grid.Input Parameters
start_core: The starting coordinate (CoreCoord) for selecting cores.
target_num_cores: The desired number of cores to include in the resulting CoreRangeSet.
subcoregrids: A CoreRangeSet representing disjoint or irregular sub-core grids within which cores can be selected.
row_wise: A boolean indicating whether to prioritize selecting cores row by row (True) or column by column (False).
Output
A CoreRangeSet that contains a compact selection of cores starting from start_core and spanning up to target_num_cores cores within the given subcoregrids.
Checklist