Skip to content

Submit/07_nlm_denoise 2026InfiniTensor_CUDA/龚舟飞 - #71

Open
blackbook537 wants to merge 4 commits into
InfiniTensor:2026-summer-projectfrom
blackbook537:submit/nlm-denoise-blackbook537
Open

blackbook537 wants to merge 4 commits into
InfiniTensor:2026-summer-projectfrom
blackbook537:submit/nlm-denoise-blackbook537

Conversation

@blackbook537

Copy link
Copy Markdown

项目

07_nlm_denoise/blackbook537:实时图像非局部均值降噪(CUDA)

主要内容

  • 实现 NLM V0 naive、V1 shared memory、V2 模板展开三个 kernel
  • 支持灰度/RGB、参数文件、1080p/4K、CPU 参考及 MAE/PSNR 验证
  • 支持 NVIDIA CUDA、Iluvatar CoreX、MetaX MACA、Moore MUSA
  • 增加统一 benchmark、质量扫描、环境采集和一键复现实验脚本
  • 整理多平台原始 CSV、日志、校验和及压缩报告图表

验证结果

  • Host 单元测试:12/12 PASS
  • V0/V1/V2 对 CPU:MAE=0、PSNR=∞
  • Iluvatar MR-V100:48/48 配置通过,1080p V2 e2e 721.346 ms
  • MetaX C500:完整矩阵通过,1080p 最快版本 V1 e2e 420.733 ms
  • MTT S4000:完整矩阵通过,1080p V2 e2e 471.324 ms
  • RTX 4090 D:历史完整流程及 Nsight Systems 证据已整理
  • small/base/large 和 σ=10/25/50 质量实验已完成

Consolidate host and GPU sources, flatten headers, and simplify the
four-platform build and reproduction entry points. Retain required
validation and curate historical evidence with byte-preserving hashes.

Add an editable report, a six-page PDF, image provenance, and submission
notes describing statistical conventions and remaining algorithm caveats.

Validated on RTX 3060 Laptop: CUDA build, 12 host checks, CPU/GPU equality,
1080p output equivalence, small end-to-end reproduction, and C++11 syntax.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant