Model M reconstruction: zero total gather weight differs between the reference and the confirmed behaviour

未关闭
#129 0 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

维护者通常 1 天内回复

还没有人认领这个 Issue。

评估

难度
4/5
预计耗时
3-5 天
新手友好度
30/100
Issue 类型
缺陷
描述清晰度
基本清楚
活跃度
活跃
技术栈
cpp, python

调研方向

Start with the dec function in clean/reference/post.py, where a NaN is mapped to the low clamp value, and compare it with the NaN-propagating behaviour confirmed for Model L (see test_l_recon.py for the known-answer pattern). Then find the Model M D3D12 and HIP reconstruction kernels and establish what they output for a gather with zero total weight. Done means the reference matches the confirmed NaN propagation, with a known-answer test, and the kernel output for this case is recorded in the issue.

由索引模型根据 Issue 内容生成。

描述

While settling Model L's reconstruction reference (#124) the behaviour for a gather whose total weight is exactly zero (extreme but finite coefficients, every tap underflows) was confirmed for both models: there is no special path, the NaN propagates, and the decode passes NaN through.

Model M's CPU reference (clean/reference/post.py, dec) instead maps a NaN to the low clamp value, so for this input the Model M reference and the confirmed behaviour give different history values.

To do:

  • state what Model M's D3D12 and HIP reconstruction kernels produce for this input (a synthetic case with zero total weight);
  • bring the reference, and if needed the kernels, to the confirmed behaviour, with a known-answer test as in Model L's test_l_recon.py.

This needs extreme coefficient values; no visible effect in a game has been reported. Refs #89

主要语言
C++
星标
12
派生
2
平均合并
3 小时 10 分钟
30 天内合并 PR
171

环境准备

从这里开始

  1. 先读完整个 Issue,再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库,在一个分支上完成修改。
  4. 提交 Pull Request,并在描述里引用这个 Issue 编号。

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 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。