#17477: Adopt ND coordinate system in system mesh, coordinate translation #17926
Conversation
omilyutin-tt
requested review from
cfjchu,
ayerofieiev-tt,
dmakoviichuk-tt,
tt-asaigal,
aliuTT,
abhullar-tt,
pgkeller,
tt-aho,
tt-dma and
ubcheema
as code owners
| {1, std::make_pair("device.json", SimpleMeshShape(1, 1))}, | ||
| {2, std::make_pair("N300.json", SimpleMeshShape(1, 2))}, | ||
| {8, std::make_pair("T3000.json", SimpleMeshShape(2, 4))}, | ||
| {32, std::make_pair(is_qg ? "QG.json" : "TG.json", SimpleMeshShape(8, 4))}, | ||
| {64, std::make_pair("TGG.json", SimpleMeshShape(8, 8))}, |
There was a problem hiding this comment.
Keeping these 2 together to ensure things are in sync.
| for (auto& [device_range, program] : programs_) { | ||
| if (program_idx) { | ||
| TT_ASSERT(sem_size == program.get_sem_size(device, logical_core, core_type)); | ||
| TT_ASSERT(sem_size == program.get_sem_size(mesh_device.get(), logical_core, core_type)); |
There was a problem hiding this comment.
No need to expose get_device_index directly any more.
| @@ -112,7 +115,8 @@ std::vector<chip_id_t> SystemMesh::Impl::get_mapped_physical_device_ids(const Me | |||
| auto line_coords = MeshDeviceView::get_line_coordinates( | |||
There was a problem hiding this comment.
@cfjchu is 1D case fundamentally special? Below, the 2D code attempts to "rotate" the shape and the constraint is that every dim has to fit inside SystemMesh. In ND case, do we keep the same rotating logic + keep the same special case for 1D? Would 2D (or any reduced dimensionality) be a special case too?
There was a problem hiding this comment.
Our WH mesh is 2D physically connected. So for example, when a user requests a 1x8 on a 2x4 physically connected T3000 mesh, we don't constrain the line to be bounded by the max-dim on row/col (i.e. 2, 4). Instead, we do return a line back to the user that snakes the 2D physical mesh and we follow a convention where we snake from top-left going right and then left.
There was a problem hiding this comment.
Right. So we need the special case for 1D representation as there are dependencies, and in the future we might do something similar for collapsing 3D into 2D and so on.
There is some more stuff in |
I have changes in flight for |
Yep, sg. |
…tion (#17926) ### Ticket #17477 ### Problem description TT-distributed needs to adopt ND coordinate system for mesh primitives. ### What's changed Plumbed `SimpleMeshShape` in `SystemMesh`, logical to physical coordinate translation mapping. ### Checklist - [X] [All post commit](https://github.com/tenstorrent/tt-metal/actions/runs/13395057290) - [X] New/Existing tests provide coverage for changes
…tion (#17926) ### Ticket #17477 ### Problem description TT-distributed needs to adopt ND coordinate system for mesh primitives. ### What's changed Plumbed `SimpleMeshShape` in `SystemMesh`, logical to physical coordinate translation mapping. ### Checklist - [X] [All post commit](https://github.com/tenstorrent/tt-metal/actions/runs/13395057290) - [X] New/Existing tests provide coverage for changes
…tion (#17926) ### Ticket #17477 ### Problem description TT-distributed needs to adopt ND coordinate system for mesh primitives. ### What's changed Plumbed `SimpleMeshShape` in `SystemMesh`, logical to physical coordinate translation mapping. ### Checklist - [X] [All post commit](https://github.com/tenstorrent/tt-metal/actions/runs/13395057290) - [X] New/Existing tests provide coverage for changes
Ticket
#17477
Problem description
TT-distributed needs to adopt ND coordinate system for mesh primitives.
What's changed
Plumbed
SimpleMeshShapeinSystemMesh, logical to physical coordinate translation mapping.Checklist