Creating a CuTe TV Layout in NvFuser

[rdspring1]

This PR creates a GitHub markdown / CPP hybrid tutorial that shows how to create CuTe Thread-Value Layouts via NvFuser.

TODO:

• Extract stride of loop iterDomains from TensorIndexer.
• Separate Thread and Value dimensions based on thread parallelization.
• Diagrams have different colors for threads. Apply colors based on values and replication patterns.

[github-actions]

Review updated until commit 67d9c1e

Description

• Add tutorial on CuTe Thread-Value layouts in NvFuser

• Implement example TV layouts: simple, vectorized, WGMMA

• Include visual diagrams for TV layout examples

• Integrate tutorial into CMake test suite

---

Changes walkthrough 📝

	Relevant files
Documentation	tutorial_cute_tv_layout.cpp Add hybrid tutorial file link                                                        tests/cpp/tutorial_cute_tv_layout.cpp Added symlink to cute_tv_layout.md as a hybrid CPP/Markdown tutorial Enables the file to be both compiled and rendered as documentation +1/-0      cute_tv_layout.md Create CuTe TV layout tutorial                                                      doc/dev/cute_tv_layout.md Created comprehensive tutorial on CuTe TV layouts Added three example layouts: simple, vectorized, and Hopper WGMMA Included SVG diagrams for visual explanation Provided detailed code comments and layout algebra background +461/-0
Configuration changes	CMakeLists.txt Register new tutorial in build                                                      CMakeLists.txt Added tutorial_cute_tv_layout.cpp to the list of tutorial sources Ensures the new tutorial is included in the test suite +1/-0

---

PR Reviewer Guide 🔍

Here are some key observations to aid the review process:

🧪 PR contains tests

⚡ Recommended focus areas for review Possible Issue The comment in the code indicates that the input and output tensors are column-major with shape (16, 64) and stride (1, 16), but the actual tensor dimensions defined in the test are (64, 24), which may lead to confusion or incorrect assumptions about the layout. // The input and output tensors are column-major with shape (16, 64) and stride (1, 16). __global__ void CUDAGeneratedKernel(Tensor<__bfloat, 2, 2> T0, Tensor<__bfloat, 2, 2> T1) { Possible Issue The PTX WGMMA comment references a (16, 64) shape, but the defined tensor dimensions are (64, 24), indicating a mismatch between the documented example and the actual implementation. // The input and output tensors are column-major with shape (16, 64) and stride (1, 16). __global__ void CUDAGeneratedKernel(Tensor<__bfloat, 2, 2> T0, Tensor<__bfloat, 2, 2> T1) { #pragma unroll for(nvfuser_index_t i0 = 0LL; i0 < 64LL; ++i0) { nvfuser_index_t i1; i1 = 16LL * i0; #pragma unroll for(nvfuser_index_t i2 = 0LL; i2 < 16LL; ++i2) { nvfuser_index_t i3; i3 = i1 + i2; T1[i3] = T0[i3]; } } } */

[zasdfgbnm]

Let's merge this. It's a good reference material for learning and understanding CuTe.

[rdspring1]

!build