Add custom GPU kernels playbook: Vector Addition and Matrix Multiplication kernels

[sdevinenamd]

Summary

Adds a new supplemental playbook that teaches users how to write, compile, and launch custom GPU kernels on AMD hardware using PyTorch's ROCm build. The playbook covers two end-to-end flows, a high-level JIT path and a low-level native C++ extension path, applied across two examples: vector addition and matrix multiplication.

File Structure

• playbook.json — Playbook metadata
• README.md — Full playbook content covering GPU kernel fundamentals, the HIP/ROCm stack, PyTorch's CUDA→HIP remapping, and two worked examples with both compilation flows
• assets/Vector Addition/add_one_kernel.py — Flow 1: JIT-compiled kernel via torch.cuda._compile_kernel, with background GPU utilization monitoring via rocm-smi
• assets/Vector Addition/add_one_kernel.cu — Flow 2: Native HIP C++ extension (kernel + pybind11 binding)
• assets/Vector Addition/setup.py — Build script using CUDAExtension (auto-remapped to hipcc by PyTorch ROCm)
• assets/Matrix Multiplication/matmul_kernel.py — Flow 1: 2D-grid JIT matmul kernel with timing and torch.mm verification
• assets/Matrix Multiplication/matmul_kernel.cu — Flow 2: Native HIP C++ extension
• assets/Matrix Multiplication/setup.py — Build script

Playbook Structure

• Overview — Two compilation flows side by side (JIT vs. native extension).
• What is a GPU Kernel?
• PyTorch + AMD/HIP
• Installing Dependencies — Full ROCm 7.1.1 install sequence (uninstall old stack, amdgpu-install, PyTorch, and ROCm wheel)
• Example 1: Vector Addition — Flow 1 (JIT, GPU utilization monitoring) and Flow 2 (C++ extension, pybind11 binding, .so output)
• Example 2: Matrix Multiplication — Row-major memory layout math, 2D thread indexing model, Flow 1 (JIT + torch.mm verification) and Flow 2 (native extension returning output tensor)
• Next Steps

[adamlam2-amd]

pretty good! just a couple of minor things.

Address these, and then should be okay to merge for QA to check.

[danielholanda]

Looks good. Please take a look at the recommended suggestions before merging

[jammm]

Looks good overall with some minor nits here and there. This should be running on Windows too. If you face any issues, report them in this PR so we can help you fix it.

Also, I suggest renaming the folders to have underscore instead of space, e.g., `Matrix_Multiplication" instead of "Matrix Multiplication".

[sdevinenamd]

Also, I suggest renaming the folders to have underscore instead of space, e.g., `Matrix_Multiplication" instead of "Matrix Multiplication".

Addressed

[jammm]

LGTM with a couple minor nitpicks. Good to merge overall!

Also take a look at my comment in https://github.com/amd/playbooks/pull/100/changes#r3043808230