Under Proton (RADV/ACO), three Model M kernels go over the 192-VGPR bound

未關閉
#121 2 則留言 0 個 reaction 已指派 0 人 在 GitHub 檢視

維護者通常 1 天內回覆

還沒有人認領這個 Issue。

評估

難度
5/5
預估耗時
一週以上
新手友好度
15/100
Issue 類型
缺陷
描述清晰度
需要釐清
活躍度
活躍
技術堆疊
cpp

研究方向

The issue asks for a maintainer decision, not a code change: restructure the kernels, report the SPIR-V to Mesa ACO, or leave it. Start with clean/dx12/vgpr_limits.json and clean/dx12/README.md, then the kernels under clean/dx12/kernels built by build_dx12.py. Reproducing the VGPR counts needs RADV_DEBUG=shaderstats under Wine with native_direct_test. Done means a chosen path, agreed with the maintainer, and no code change to the kernels has been made yet.

由索引模型根據 Issue 內容生成。

描述

area: linux area: runtime performance

What

Under Proton, Mesa's ACO compiler needs more VGPRs than clean/dx12/vgpr_limits.json allows for three of Model M's heaviest kernels. Per vgpr_limits.json's own note, they then fit one 256-lane workgroup per CU instead of two:

Kernel VGPRs under ACO Bound Waves/SIMD GPU time per frame
swin_block_c64_nc 240 192 6 0.30 ms
swin_block_c96_pm_rev_nc 256 192 5 0.26 ms
bneck/swin_bneck6_mbo_nc 256 192 5 0.43 ms
swin_block_c96_add_xe_xo_nc 192 192 8 0.21 ms
d1p12/swin_block_c64_add_dec0_nc 192 192 8 0.43 ms

That's about 1.0 ms of Model M's roughly 2.3 ms on the direct route, at 1706×960 → 2560×1440. The repository's notes put a lost workgroup per CU at +38 µs (swin_block_c96_add_xe_* at 215 VGPRs) to +103 µs (dec0 at 201/215 VGPRs) per frame at 1440p Quality. So this is probably a few tenths of a millisecond on Linux. I haven't measured that directly: I can't force ACO to 192 VGPRs to compare.

It isn't loop unrolling. That was my first guess, because of the "unroll loops despite [loop]" note in clean/dx12/README.md, so I checked:

  • swin_block_c64_nc's DXIL has no loops at all: DXC already unrolled it, so AMD's Windows compiler gets the same straight-line code and fits it in ≤192 VGPRs (the build's RGA check).
  • swin_block_c96_pm_rev_nc and the bottleneck keep their loops under ACO (2 and 8 back-edges).

So it's ACO's register allocation on these large kernels. Before scheduling it's already about 198 VGPRs for c64, and 194 and 230 for the other two.

How it was measured (synthetic data only)

  • RX 9070 XT, Mesa RADV 27.0.0-devel, GE-Proton 11-11's Wine, vkd3d-proton 5b1089c8 with dxil-spirv #362 (dxil-spirv 114e664e plus the PR).
  • Kernels built from main's clean/dx12/kernels (unchanged since a8500cb) with build_dx12.py.
  • native_direct_test (direct route, check mode, 3 frames) run under Wine with RADV_DEBUG=shaderstats (VGPRs, waves, LDS) and RADV_DEBUG=shaders (ACO's ISA).
  • Kernels matched by their WMMA count: each SPIR-V cooperative-matrix multiply-add becomes two v_wmma in ACO's ISA (736, 560 and 1312 for the three), and by LDS size where that could be read.
  • GPU times are from WG_D3D12_PROFILE in the same harness. Profiling adds overhead, so treat them as rough.

What could be done (your call):

  • Restructure these kernels so they need fewer registers on any compiler. That would change the Windows DXIL too.
  • Report it to Mesa (ACO) with these kernels' SPIR-V as a reproducer. They're WrongGPU's own MIT-licensed code, so the SPIR-V can be shared. I can put that together if you'd rather it went that way.
  • Leave it. On Linux this is a few tenths of a millisecond, not a correctness problem.

Related: #29, #10.

主要語言
C++
星號
12
分支
2
平均合併
3 小時 10 分鐘
30 天內合併 PR
171

環境準備

從這裡開始

  1. 先讀完整個 Issue,再讀專案的貢獻指南。
  2. 在 Issue 下留言說明你要接手 —— 這能避免兩個人做同樣的事。
  3. Fork 儲存庫,在一個分支上完成修改。
  4. 送出 Pull Request,並在描述裡引用這個 Issue 編號。

phantoslayer21/wronggpu 的其他 Issue

查看 phantoslayer21/wronggpu 的全部 Issue

相似的 Issue

更多 C++ Issue

把新 issue 寄到你的電子郵件信箱

精選適合新手參與的 GitHub issue 摘要。