[gfx1250][tdm] Add K-loop hoist descriptor helpers and modify build script

[XingerZhu]

…t tweaks

TDM (python/flydsl/expr/rocdl/tdm_ops.py)

• Add K-loop hoist helpers that let MoE-style K-reduction loops patch only dgroup0 lane 2 (the per-K-tile global addr_lo) of an otherwise hoisted base descriptor, instead of rebuilding the full 4xi32 descriptor every iteration:
  • update_tensor_descriptor_2d_addr_lo
  • update_tensor_gather_descriptor_addr_lo (carry-unsafe fast path; lo-32 add only)
  • update_tensor_descriptor_2d_addr_lo_hi
  • update_tensor_gather_descriptor_addr_lo_hi
  • update_tensor_descriptor_2d_addr64
  • update_tensor_gather_descriptor_addr64 (carry-safe variants via the new add_addr_with_carry helper, needed when per-CTA base + cumulative K-tile delta can cross a 4 GiB boundary in lo-32-bit arithmetic; otherwise the descriptor silently aliases into the wrong 4 GiB page and the GPU deadlocks in amdgpu_mes_reg_write_reg_wait with no recoverable signal).

Build scripts

• scripts/build.sh: forward HIP_PLATFORM (default amd) to cmake as a cache var, because the shipped /opt/rocm/lib/cmake/hip/hip-config.cmake has if("OFF") ... set(hip_HIPCONFIG_EXECUTABLE) ... and cannot auto-detect the platform.
• scripts/build_llvm.sh: enable lld in LLVM_ENABLE_PROJECTS, turn on MLIR_ENABLE_ROCM_RUNNER, and pass HIP_PLATFORM=amd for the same reason.

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

Adds new Python-side TDM (gfx1250) descriptor “K-loop hoist” helpers to cheaply patch the per-iteration global address portion of an otherwise-hoisted descriptor (including a carry-safe 64-bit variant), and updates build scripts to better interoperate with ROCm HIP CMake configuration.

Changes:

• Introduce descriptor patching helpers in tdm_ops.py to update dgroup0 lane 2 (addr_lo) and optionally lane 3 (addr_hi with type bits preserved), plus an add_addr_with_carry utility for carry-safe address updates.
• Update scripts/build.sh to forward HIP_PLATFORM into the FlyDSL CMake configure.
• Update scripts/build_llvm.sh to enable lld, turn on MLIR_ENABLE_ROCM_RUNNER, and pass HIP_PLATFORM.

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated 1 comment.

File	Description
scripts/build.sh	Prints/defaults HIP_PLATFORM and forwards it to FlyDSL’s CMake configure.
scripts/build_llvm.sh	Enables additional LLVM/MLIR components and sets HIP_PLATFORM for the LLVM/MLIR build.
python/flydsl/expr/rocdl/tdm_ops.py	Adds K-loop descriptor hoist/patch helpers, including a carry-safe addr_hi update path.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.