[ROCDL] Fix wave size for gfx1250 (wave32, not wave64)

[sjfeng1999]

Motivation

Technical Details

Test Plan

Test Result

Submission Checklist

• Look over the contributing guidelines at https://github.com/ROCm/ROCm/blob/develop/CONTRIBUTING.md#pull-requests.

[Copilot]

Pull request overview

This PR corrects ROCm target wavefront sizing by introducing a centralized get_warp_size() helper and switching ROCm backend configuration to use it, ensuring gfx1250 is treated as wave32 (not wave64).

Changes:

• Add get_warp_size() to python/flydsl/runtime/device.py (wave32 for gfx10/gfx11/gfx12*; otherwise wave64).
• Update ROCm backend target detection and ROCDL wave64 option to use get_warp_size() instead of is_rdna_arch().
• Update kernel/common helper and docs to reflect the new wave-size logic (but note: the current kernels_common.py change introduces an infinite recursion bug; see PR comments).

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 1 comment.

Show a summary per file

File	Description
python/flydsl/runtime/device.py	Adds centralized wavefront-size detection used across runtime/backends.
python/flydsl/compiler/backends/rocm.py	Uses get_warp_size() for GPUTarget.warp_size and ROCDL wave64 flag selection.
kernels/common/kernels_common.py	Attempts to delegate kernel-side get_warp_size() to runtime helper (currently broken due to name shadowing/recursion).
docs/prebuilt_kernels_guide.md	Updates documentation for wave-size behavior to include gfx12* (including gfx1250).
CLAUDE.md	Updates architecture table notes to reflect corrected gfx1250 wave size behavior.

Suppressed comments (1)

kernels/common/kernels_common.py:102

• get_warp_size() currently calls itself (return get_warp_size(arch)), leading to infinite recursion. It should delegate to the runtime implementation (aliased at import time).

    """Return the wavefront/warp size for the given GPU architecture.

    CDNA (gfx9xx) uses wave64, gfx10xx/gfx11xx/gfx12xx use wave32.
    """
    return get_warp_size(arch)

---

💡 Add a code-review agent skill or configure MCP servers for context-aware, tailored reviews. Learn more in the docs.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated no new comments.

Suppressed comments (1)

python/flydsl/runtime/device.py:106

• get_warp_size() is now the source of truth for wave size, but the docstring doesn’t call out that wave size is intentionally decoupled from is_rdna_arch() (e.g., gfx1250 is wave32 even though it’s not RDNA). Adding a short note here would help prevent future callers from using is_rdna_arch() for wave-size decisions again.

    """Lanes per warp for an architecture.

    If arch is None, the current GPU arch is auto-detected.
    """