Skip to content

feat(ccu): add two-rank direct alltoall - #78

Draft
Kur0x wants to merge 2 commits into
codex/ccu-stack-06-signal-waitfrom
codex/ccu-stack-07-alltoall-2rank
Draft

feat(ccu): add two-rank direct alltoall#78
Kur0x wants to merge 2 commits into
codex/ccu-stack-06-signal-waitfrom
codex/ccu-stack-07-alltoall-2rank

Conversation

@Kur0x

@Kur0x Kur0x commented Jul 28, 2026

Copy link
Copy Markdown
Member

Summary

  • add two-rank direct CCU AllToAll program generation
  • add endpoint exchange, remote-buffer import, and route-scoped memory overrides
  • add synchronization fixes and repeated submission support

Stack

7 of 9. Depends on codex/ccu-stack-06-signal-wait.

Validation

  • git diff --check
  • focused AllToAll, backend, orchestrator, smoke, driver, and lower-layer tests
  • the suite exercised 169 tests; 168 passed before one outdated source assertion was aligned with the final route index
  • the corrected assertion was rerun and passed

Scope

Prior two-rank hardware evidence belongs to #71 and was not rerun from this host.

next.syncResourceCount = 3;
next.syncInstructionCount = std::max<uint32_t>(
next.syncInstructionCount,
TILEXR_CCU_DIRECT_ALLTOALL_INSTRUCTION_COUNT);

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This fixed capacity only covers 64 copy blocks (64 * 32 KiB = 2 MiB), while the request validation accepts any aligned nonzero bytes. For a larger all-to-all, the generated program exceeds the allocated repository/task window and fails later despite passing validation. Please derive the instruction capacity from bytes with overflow and uint16_t window checks, or reject sizes above the supported limit up front.

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants