Skip to content

Linux/RADV: Model K's default head stage differs from HIP on 0.6.5 and matches after #288; Model L check results #289

Open

Description

@j0nsh1n

About #288 (and #284). We checked main 11d3bbc against 0.6.5 (2b31cbc) on an RX 9070 XT under Linux: Wine 11.21, vkd3d-proton 5b1089c8 with dxil-spirv PR #362 (for the AGS ops), Mesa RADV 26.99.99. Synthetic models only. One caveat for every "vs HIP" number below: our HIP side was built with ROCm 10.1 on Linux and run on the same card, not with the HIP SDK the project builds with, so a difference could come from either side's compiler.

Model K, whole frame. Our own replay of the runtime, output 200x120, 6 frames, input 130x78 except two frames at 100x60. #283, #285, #286 and #287 each give the same bytes as 0.6.5. #288 is the first commit that changes them: the output images of the two reset frames are equal, but the hidden history after frame 0 already differs by 1-5 FP16 steps in about 7 % of elements, and later frames follow. 0.6.5 repeated its own bytes across two runs, main across three.

Model K, single kernels. k_port_check.py, narrow family, 2 seeds x 6 sizes, D3D12 side replayed under Wine, HIP side on the same card. "Identical" is the checker's own rule: every defined byte equal with valid slack sentinels, or a difference only in the NaN sign bit with the same payload.

Kernel 0.6.5 vs HIP 11d3bbc vs HIP
k_stage2_d0_hp (default) 8 of 12 identical 12 of 12
k_stage2_d0_hidden (full head, not default) 6 of 12 8 of 12
the other narrow-stage kernels 12 of 12 12 of 12

The d0 differences are in the hidden-history output only, in the low bits of FP32 words; the head output is identical in every case. We did not isolate the cause; the FP32 order or contraction of the 0.2 * (c0 + ... + c4) sum under ACO is one candidate. So the whole-frame change from #288 is, for the default head stage, a move onto our HIP build's bytes; k_stage2_d0_hidden still differs in 4 of 12 cases on main.

Model L: a question. For #284's kernels that ran, the D3D12 outputs match 0.6.5 byte for byte. The stage-dump replays of l_hyb_stage_check did not run: l_swin_block_c32_*_hyb_dump are not in the kernel manifest. The checks' direct D3D12-vs-HIP comparisons differ in both versions on this stack:

  • l_hyb_block_check: 14 of 28 identical.
  • l_d3d12_ops_check: l_project_gather 10 of 36, l_dec_project 2 of 36, l_coef 3 of 12, l_coef_pack12 3 of 12 (its other kernels all identical).
  • Fused vs HIP sequence: l_enclin_check 0 of 28, l_fold_check 21 of 84, l_linfuse_check 10 of 54. The sequential D3D12-vs-HIP legs of the same three checks are 10 of 28, 62 of 84 and 54 of 54.

We can't tell yet whether that is RADV, our HIP build, or our harness. Do these checks pass byte for byte on Windows with the official HIP SDK?

Also seen:

  • Model M matches 0.6.5 on every dumped buffer that is stable from run to run, at three sizes (93 of 93, 87 of 87, 66 of 66 files), although ten of its hybrid kernels' DXIL changed with wg_hyb.hlsli in Model L: exact speedup in the hybrid chain ends (-72.75 us at 1080p Q, -112.5 us at 1440p Q) #284.
  • Stock vkd3d-proton in GE-Proton 11-11 and 11-12 creates the same 42 of 157 kernels.
  • l_q7_block_c32_{enc,dec} crash the pipeline-creation process on every pair we tried, including the patched one (the opt-in q7 reference route; the default D3D12 route is hybrid). We did not look into where it crashes.

Happy to run anything else on this card.

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Projects

    No projects

      Milestone

      No milestone

      Relationships

      None yet

      Development

      No branches or pull requests

      Issue actions