Curated References

This is the detailed catalog behind the Resources Hub. The goal is not to dump links; it is to show which source helps with which kind of SGEMM question.

How to use this page

• Start with the Resources Hub if you need help choosing a route.
• Use this page when you already know the category of source you need.
• Continue to Further Reading Routes when the right next topic matters more than the right citation.

Official CUDA and NVIDIA docs

These are the sources to open when you need precise constraints, terminology, or API behavior.

• CUDA C++ Programming Guide
  Best for execution model, synchronization, memory hierarchy, and launch semantics.
• CUDA C++ Best Practices Guide
  Best for optimization heuristics, memory access advice, and profiling-oriented sanity checks.
• CUDA Runtime API
  Best for stream, event, launch, and runtime behavior details when the implementation questions get concrete.
• CUDA Programming Guide: WMMA section
  Best for fragment types, shape constraints, and the mechanics behind Tensor Core usage.
• NVIDIA Developer Blog: Programming Tensor Cores in CUDA 9
  Best for a high-level explanation of why WMMA programming looks different from scalar CUDA code.
• NVIDIA Mixed-Precision Training Guide
  Best for understanding where mixed-precision wins come from and what extra constraints or conversion costs appear.

Why this shelf matters:

• It anchors whitepaper claims in vendor-defined behavior instead of community folklore.
• It helps you explain why unsupported shapes, alignment limits, and fallback rules are engineering constraints, not arbitrary policy.

Papers and performance mental models

Open this shelf when you want the design logic behind SGEMM optimization rather than raw API detail.

• Anatomy of High-Performance Matrix Multiplication (GotoBLAS paper)
  The best first paper here for understanding why blocking and hierarchy dominate matrix multiplication performance.
• BLIS papers and project entry point
  Useful when you want to compare teaching-oriented kernels with a production CPU framework built around explicit packing and control trees.
• Nsight Compute roofline charts guide
  Useful when you need a stronger language for arithmetic intensity and the boundary between memory-bound and compute-bound behavior.

Why this shelf matters:

• It gives readers a mental model for why the kernel ladder is ordered the way it is.
• It helps you discuss performance ceilings without reducing everything to one benchmark number.

Exemplary codebases and production-grade samples

Open this shelf when you want to compare this repository's explanatory style with industrial-strength implementations.

• CUTLASS: Fast Linear Algebra in CUDA C++
  Best for seeing how a production CUDA GEMM library structures tiling, pipelines, and architecture-specific specialization.
• BLIS Framework
  Best for understanding how GEMM decomposition and packing ideas generalize beyond CUDA.
• CUDA Samples: matrixMul example
  Best for a smaller official example that is easier to compare with this project's early kernel stages.

Why this shelf matters:

• It shows where this repository is deliberately simplified for teaching.
• It gives interview follow-up material when someone asks what the "next production step" would look like.

Profiler, tooling, and diagnosis references

Open this shelf when a benchmark number stops being self-explanatory and you need evidence.

• Nsight Compute Documentation
  Best for kernel-level counters, memory behavior, roofline views, and occupancy analysis.
• Nsight Systems Documentation
  Best for timeline-level reasoning, launch gaps, overlap, and host/device interaction.
• CUDA Occupancy Calculator (archived official workbook)
  Best for turning block size, shared memory, and register usage into an explicit occupancy trade-off discussion.
• Compute Sanitizer Documentation
  Best for catching correctness and memory issues before performance conclusions become trustworthy.

Why this shelf matters:

• It supports metric-driven diagnosis instead of guesswork.
• It connects directly to the repository's Diagnosis Loop, Benchmark Scope, and Reproducibility pages.

Engineering workflow and validation discipline

Open this shelf when the question is about proving correctness, structuring builds, or keeping claims reproducible.

• GoogleTest Documentation
  Best for understanding the local correctness harness and tolerance-oriented testing vocabulary.
• CMake Documentation
  Best for build-system expectations, generator behavior, and reproducible local setup.
• OpenSpec documentation
  Best for understanding the spec-governed documentation and change workflow used in this repository.

Why this shelf matters:

• It reinforces the repository's local-GPU versus hosted-CI boundary model.
• It explains why "performance proof" and "repository integrity" live on different evidence surfaces.

Next-step study routes

Use these when you know you want to keep learning, but not which topic should come first.

• Further Reading Routes for curated paths on tiling, occupancy, roofline thinking, Tensor Core constraints, and profiling.
• CUDA Memory Cheat Sheet for a fast memory refresher before reopening kernel code.
• Resources Hub for scenario-based entry points back into the rest of the site.