Skip to content

Hierarchos-Native v0.2.5 — RDNA3 P3 Portable Release

Latest

Choose a tag to compare

@necat101 necat101 released this 22 Nov 01:57
d0f48c9

Hierarchos-Native v0.2.5 — RDNA3 P3 Portable Release

Windows x64, extract-and-run. The executables are statically linked against the MSVC C runtime, so no Microsoft Visual C++ Redistributable is required, and no Rust, Cargo, Python, PyTorch, Visual Studio or shader compiler is needed to run them.

This release ships the production P3 Full FP16 WMMA epilogue qualified on the measured AMD RDNA3 adapter. It supersedes the RC2 documentation, which described an older source commit and did not ship P3.

What changed

The Full FP16 WMMA arm (--wmma-mode full-fp16) now defaults to the P3 epilogue. Bank assignment, the K-slice stream, the two-wave workgroup structure, the cooperative multiply, the A/B staging and the barrier placement are unchanged. The epilogue merges the eight binary16 bank partials in float32 and rounds once to binary16, instead of seven binary16 roundings. split8, two-bank and two-wave remain selectable through HIERARCHOS_WMMA_FULL_FP16_VARIANT as rollback and comparison paths.

Embedded module: wmma_gemm_f16_f16acc_twowave_p3.spv, SHA-256 2aff4123544757b0d3aaf8eb1fe967a0c00c692eb73a74e49af802d7b5c517d2, 23,280 B, runtime identity spirv-fnv1a64:0x076216ad97d7605f — byte-identical to the module qualified in Phase LXXIX.

WMMA is opt-in

Global WMMA remains off by default (WmmaMode::Off). --wmma-mode full-fp16 explicitly enables the Full FP16 path; P3 is the default variant inside that mode. The FP32 Vulkan path and the Mixed arm are unchanged.

Admission requires AMD RDNA3-or-newer plus the complete cooperative feature contract (16×16×16 subgroup-scope f16/f16→f16 tile, subgroup forced to 32, 16-bit storage, shaderFloat16), and at least WMMA_MIN_PROFITABLE_MACS = 16_000_000 multiply-accumulates. The floor is unchanged and is applied both at admission and to each recorded batch. The hard numerical gate remains ULP_GATE_LIMITS = 0.03125; the historical 25 % tile16 headroom target remains informational and is not met.

Accuracy

  • 37/37 numerical gates at all eight qualification seeds on the production route.
  • 32/32 preserved K=96 outputs bit-identical to the Phase LXXIX GPU P3 tensors; inputs identical.
  • Maximum tile16 distance 0.03125 — exactly at the unchanged hard gate, passing (<=).
  • Maximum REF64F16 error 0.03035259 → 0.02593803; 31/32 cases improved, 0/32 worsened.

Measured performance (per-dispatch kernel timings)

P3 vs the preserved Main baseline. 24 paired blocks per shape, position-balanced ABBA/BAAB ordering, GPU timestamp queries around the compute dispatch, strictly serial execution, corrected 95 % paired-t CI. Ratio = Main ms / P3 ms; > 1.0 means P3 is faster.

Shape Main ms P3 ms Speedup Corrected 95 % CI
2048×768×2048 5.4193 5.3109 1.0243 [1.0159, 1.0326]
512×2048×768 1.4945 1.3332 1.1246 [1.1167, 1.1325]
512×768×768 0.6088 0.5486 1.1163 [1.0879, 1.1446]
128×8192×768 1.8102 1.4260 1.2722 [1.2584, 1.2860]

Every point estimate and every corrected lower bound is above 1.0, independently per shape. These are per-dispatch GPU GEMM timings, not end-to-end training-step speedups; no end-to-end speedup is claimed.

The 16×128×32 direct-kernel control is a disclosed regression (0.7566, [0.7504, 0.7627]). At 65,536 MACs it is roughly 244× below the 16M floor and fails the tile-fill requirement, so the production route never dispatches WMMA there. It is excluded by the unchanged predicate, not hidden.

Packaged dispatch and acceptance

The packaged P3 dispatch was exercised from the extracted, relocated package with no variant override. The extracted trainer reported cooperative_dispatches=8 and shader wmma_gemm_f16_f16acc_twowave_p3; the packaged logits executable reported 3 cooperative dispatches under full-fp16 and 0 under off, and matched FP32 within max abs 0.000738 / RMS 0.000122. The default training run reports requested:"off" with zero cooperative dispatches.

Package inventory is exactly the 16 allowlisted runtime files; no source, Cargo metadata, PDB, fixtures or model weights are included.

Requirements

  • Windows x64. Windows 11 x64 is the measured acceptance target; Windows 10 x64 should work but was not measured.
  • A Vulkan 1.x loader and a Vulkan compute-capable GPU driver. The loader (vulkan-1.dll) is supplied by the graphics environment and is not bundled.
  • Enough RAM and disk for the selected model. No model weights are bundled.
  • No administrator rights are required for normal operation.

Hardware qualification boundary

All measurements were taken on one adapter: AMD Radeon Graphics, PCI 1002:15bf (RDNA3, Phoenix), Vulkan 1.4.344, driver 2.0.388. This release makes no claim for other RDNA3 parts, RDNA3.5, RDNA4 or non-AMD adapters. The route is gated by the existing capability contract; no new hardware admission was added.

Assets

File Bytes SHA-256
Hierarchos-Native-Windows-x64-Portable-v0.2.5.zip 34,044,138 7e9d33d21616d3cbe0fca2d38c9745014a3e40d1c608dddcb46fa8586f25dc2d
BUILD_MANIFEST.json 2,736 37b3646073f378bdd6a5fdcaf396b758637457b46e1f6755dc59e752c6ada763
SHA256SUMS 202 53ac080eaec7de95ce2d9364af8b37f9a9ee4288ced2750346704b63c50f4bf4

SHA256SUMS covers the ZIP and BUILD_MANIFEST.json. BUILD_MANIFEST.json records the packaging source commit c2775e0, the qualified source commit b165c299, the P3 artifact identity, and the byte count and SHA-256 of all seven executables.

Known limitations

  • No independent second-device or clean-machine certification. Acceptance ran on the same Windows 11 / AMD RDNA3 host that produced the package. No second physical CPU/GPU and no physical network isolation were available; CPU compatibility is analytic only (default x86_64-pc-windows-msvc target, no target-cpu=native).
  • No code signing. The executables are unsigned and carry no Authenticode signature.
  • GUI interaction unverified. The GUI executable ships but was not driven; reported as BLOCKED, not passed.
  • Different-drive relocation and Windows 10 x64 were not tested.
  • No RGP capture or occupancy measurement; register/LDS figures in the historical record are static code-object facts.
  • The 16×128×32 direct-kernel regression is measured and excluded by policy, not fixed.
  • cargo fmt --check is not clean on the untouched baseline and is not an enforced repository gate.