diff --git a/images/img1.png b/images/img1.png new file mode 100644 index 0000000..65ded34 Binary files /dev/null and b/images/img1.png differ diff --git a/images/img2.png b/images/img2.png new file mode 100644 index 0000000..98b3012 Binary files /dev/null and b/images/img2.png differ diff --git a/images/img3.png b/images/img3.png new file mode 100644 index 0000000..83cfd42 Binary files /dev/null and b/images/img3.png differ diff --git a/project_fs_core.tex b/project_fs_core.tex index 24d32a6..37e4e93 100644 --- a/project_fs_core.tex +++ b/project_fs_core.tex @@ -1,3 +1,4 @@ +% Slides for 2025-05-13 % To create a slide, use the following: % \begin{frame}{TITLE} % BODY @@ -38,3 +39,34 @@ % \end{column} % \end{columns} % \end{frame} + +\begin{frame}{Rust Pipeline} + \begin{columns}[c] + \column{0.30\textwidth} + \centering + \includegraphics[height=0.45\textheight,keepaspectratio]{./images/img1.png} + + \column{0.44\textwidth} + \centering + \includegraphics[height=0.45\textheight,keepaspectratio]{./images/img3.png} + \column{0.30\textwidth} + \centering + \includegraphics[height=0.45\textheight,keepaspectratio]{./images/img2.png} + \end{columns} + \begin{enumerate} + \item Segment around 100 test images with truths + \item Compare normalized pixel difference from truths (with old python pipeline) + \item Look into hipify-perl converting Mitsuba + Dr.Jit CUDA code to HIP + \end{enumerate} +\end{frame} + +\begin{frame}{Tom - Underwater Renderer} + \begin{enumerate} + \item Attempted implement simple FP64 gelu with CubeCL in rust on CUDA + \item CubeCL FP64 achieved by converting to FP32, doing calculation, then scale back to FP64; loses accuracy + \item CubeCL HIP seems to be the only stable HIP binding for rust; AMD GPU only, requires rocWMMA or WMMA Intrinsics (FP16 only); also unsafe since raw bindings + \item Switching to C++ as Rust has very limited support and benifit because of the unsafe GPU calls + \item Reading PBRT book, looking into drjit-core and HIPIFY. Possible simple renderer from the ground up? + \item rocWMMA requires MI1/2/300 series for FP64 runs, unable to test running code on GPU yet + \end{enumerate} +\end{frame}