perf(cco): preserve global address space in LSA pointer wrapper

[yangyuhuiling]

Summary

• document that ccoWindow_t metadata objects are allocated in device global memory
• restore LLVM addrspace(1) after the scalar FlyDSL ABI converts the window handle to uint64_t
• keep the existing cco_lsa_ptr API while allowing AMDGPU to select global_load instead of conservative generic flat_load

Performance

With the current EPv2 kernel held fixed and only the linked CCO bitcode switched:

• BF16 scatter-combine, 128 tokens/rank: 166.66 us -> 113.74 us (-31.75%)
• BF16 scatter-combine, 512 tokens/rank: 577.20 us -> 361.79 us (-37.32%)
• dispatch remained within 0.7%, confirming the change targets repeated LSA address formation in scatter-combine

Source-mapped ATT shows flat_load_dwordx3 plus vmcnt/lgkmcnt waits becoming global_load_dwordx3; the scatter kernel also drops from 24 to 16 reported VGPRs with no scratch use.

Test plan

• compile cco_device_wrapper.cpp to gfx950 device bitcode
• verify cco_lsa_ptr IR uses ptr addrspace(1) for metadata loads
• run examples/cco/python/04_flydsl_lsa_put on 2 ranks
• run EPv2 BF16 gather and scatter correctness with the PR bitcode (2 passed)
• run isolated flat/global bitcode performance A/B and source-mapped ATT

Made with Cursor