SGEMM Optimization

English | 简体中文

Progressive CUDA SGEMM tutorial and reference implementation. The repository contains five hand-written kernel variants, cuBLAS-backed verification, a benchmark harness, and OpenSpec-governed repository rules for keeping the project compact and trustworthy.

Why this repository exists

• Show the optimization ladder clearly: naive -> tiled -> bank-conflict-free -> double-buffered -> Tensor Core WMMA
• Stay readable: each optimization lives in its own kernel file and keeps a consistent launch interface
• Stay verifiable: kernels are checked against cuBLAS, with separate tolerances for FP32 and Tensor Core paths
• Stay maintainable: the repository uses OpenSpec to keep docs, workflow, and validation rules aligned

Kernel progression

Stage	File	Main idea
Naive	src/kernels/naive_sgemm.cuh	Baseline triple-loop mapping
Tiled	src/kernels/tiled_sgemm.cuh	Shared-memory blocking
Bank-Free	src/kernels/bank_conflict_free_sgemm.cuh	[TILE_SIZE][TILE_SIZE+1] padding
Double Buffer	src/kernels/double_buffer_sgemm.cuh	Tile staging overlap and latency hiding
Tensor Core	src/kernels/tensor_core_sgemm.cuh	WMMA path with safe FP32 fallback

Quick start

git clone https://github.com/LessUp/sgemm-optimization.git
cd sgemm-optimization

# Recommended: CMake
cmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build -j$(nproc)
./build/bin/sgemm_benchmark -a
ctest --test-dir build

# Quick local alternative
make GPU_ARCH=sm_86
make benchmark
make test

Validation model

• Local GPU machine: runtime tests, correctness checks, and benchmarking
• GitHub Actions: format/style, CUDA compile validation, OpenSpec/repository checks, and Pages deployment

Standard FP32 kernels use rtol=1e-3, atol=1e-4. The Tensor Core path uses rtol=5e-2, atol=1e-2.

Read next

• Getting Started
• Learning Path
• Architecture Overview
• Benchmark Notes
• Specifications Index
• GitHub Pages site

Repository layout

src/
├── kernels/        # Five SGEMM kernel variants
├── utils/          # CUDA RAII, verification, benchmark helpers
└── main.cu         # Benchmark entry point
tests/
└── test_sgemm.cu   # Google Test suite
docs/               # Public learning-oriented documentation
openspec/           # Stable specs, changes, and workflow guidance

Development workflow

Non-trivial repository changes are expected to follow:

1. /opsx:explore
2. /opsx:propose "description"
3. /opsx:apply
4. /review
5. /opsx:archive

The stable authoritative specs live under openspec/specs/. Active implementation plans live under openspec/changes/<change>/.

License

MIT. See LICENSE.