[tenpercent]

Motivation

rocKE already provided gfx1250 K=64 FP8/BF8 WMMA and a block-scaled GEMM that applied FP16 or FP32 scales in software after the matrix instruction. This change adds the first native FP8 x FP8 SCALE and SCALE16 vertical slice, including the catalog contract, dual-engine lowering, packed E8M0 scale operands, an opt-in complete kernel path, and numerical regression coverage on gfx1250.

ISSUE ID : AICK-2281

Umbrella: AICK-2056.

Technical Details

• Add gfx1250 K=128 FP8 x FP8 wmma_scale and wmma_scale16 catalog entries with the required fragment lengths.
• Add Python and C IR builders for the five-input scaled MMA operations.
• Add byte-identical Python and C++ LLVM 23 lowering for the SCALE and SCALE16 intrinsic ABIs.
• Add matching Python and C++ HIP builtin lowering.
• Reject scaled WMMA under LLVM 20 and LLVM 22.
• Extend the gfx1250 block-scaled GEMM with opt-in SCALE and SCALE16 paths that consume packed E8M0 scales and pass them directly to the matrix instruction.
• Preserve the existing K=64 unscaled path as the default.
• Correct HIP zero extension to cast through the unsigned source type, preserving E8M0 values with the high bit set.
• Add focused Python tests, C++ lowering tests, fragment-table checks, and Python/C parity configurations.
• Extend the numerical verifier with native SCALE/SCALE16 cases and an independent float64 scaled-matmul reference rounded to BF16. Bounded inputs allow exact native-result comparisons; unwritten outputs and nonfinite values fail validation.
• Add seven GPU test configurations covering neutral, independently varying A/B, mixed, and isolated scale groups; multi-tile K=256 cases through COMGR and HIP; and a legacy WMMA smoke case. ROCKE_REQUIRE_GFX1250=1 makes a missing target fail instead of skip. Torch is not required.

This is intentionally the FP8 x FP8 prerequisite slice. Broader FP4, FP6, BF6, conversion, and native type-combination coverage remains outside this PR.

Test Plan

• Focused scaled-WMMA regression: 7 passed, 12 subtests passed.
• Broader relevant suite: 47 passed, 1 deselected, 24 subtests passed.
• C++ tests: 5/5 passed.
• Python/C++ byte identity for all 16 gfx1250 lowering configurations under LLVM 20, 22, and 23.
• SCALE and SCALE16 HIP source compiled for gfx1250.
• HIP and direct COMGR objects contain the expected SCALE and SCALE16 instructions.
• Pre-commit range checks, including the numerical-test commit.
• Independent numerical-reference tests: 16 passed.
• Numeric execution and scale-lane validation on gfx1250 hardware: 7 configurations passed, 25 comparisons with bad=0, zero skips.

Test Result

CPU, dual-engine, compiler, COMGR, and ISA-selection checks passed. GPU numerical validation passed on gfx1250 at commit 8a13045058e7e7464377bb5bffcfbaa9523813ad using the ROCm 10.0.0 release tests distribution, HIP 7.15.26333, and AMD clang/LLVM 23. Tools and HIP/COMGR libraries were explicitly pinned to that distribution.

The 25 comparisons cover SCALE and SCALE16 with every packed scale group, finite high-bit E8M0 values, independent A/B scaling, multi-tile K=256 accumulation, both COMGR and HIP compilation routes, and the legacy WMMA smoke case. GPU execution used ROCKE_BACKEND=python; C++ coverage remains the lowering tests and Python/C++ byte-identity checks above. This run does not establish C++ GPU numerical coverage or downstream provider integration.

To reproduce, activate a Python environment with NumPy, ml_dtypes, and pytest, set ROCM_PATH to the selected LLVM-23-capable installation, and run from dnn-providers/hip-kernel-provider/rocke/platform:

export ROCM_HOME="$ROCM_PATH" HIP_PATH="$ROCM_PATH"
export HIP_CLANG_PATH="$ROCM_PATH/llvm/bin"
export PATH="$ROCM_PATH/bin:$ROCM_PATH/llvm/bin:$PATH"
export LD_LIBRARY_PATH="$ROCM_PATH/lib${LD_LIBRARY_PATH:+:$LD_LIBRARY_PATH}"
export ROCKE_HIP_LIB="$ROCM_PATH/lib/libamdhip64.so"
export ROCKE_COMGR_LIB="$ROCM_PATH/lib/libamd_comgr.so"
PYTHONPATH=python HIP_VISIBLE_DEVICES=0 ROCKE_BACKEND=python \
  ROCKE_LLVM_FLAVOR=llvm23 ROCKE_REQUIRE_GFX1250=1 \
  python -m pytest -q -s tests/instances/test_gfx1250_scaled_wmma_numeric.py

Submission Checklist

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

[therock-pr-bot]

✅ All Policy Checks Passed

Check	Status	Details
📝 PR Description	✅ Pass	—
⛔ Forbidden Files	✅ Pass	—
🧪 Unit Test	✅ Pass	—
🚫 Draft PR	🔜 To Be Enabled	—
🚩 Feature Flag	🔜 To Be Enabled	—
📊 Code Coverage	🔜 To Be Enabled	—

🎉 All policy checks passed!

📖 Need help? See the Policy FAQ for details on every check and how to fix failures.

🙋 Wish to Override Policy?

[therock-pr-bot]

🎉 All checks passed! This PR is ready for review.

[tenpercent]

@copilot resolve the merge conflicts in this pull request

[Copilot]

@copilot resolve the merge conflicts in this pull request

Resolved the merge conflicts and merged develop into this branch in 19c1c6f.

[assistant-librarian]

TheRock Submodule Bump Activity

Newest first

• 2026-09-15 — Included in TheRock via ROCm/TheRock#8209.