pjt222/ga104-cuda-kernels

GA104 Hand-Optimized CUDA Kernel Corpus

A measurement corpus of hand-optimized CUDA / SASS kernels targeting the RTX 3070 Ti (GA104, sm_86, Ampere). Every kernel is written without cuBLAS, cuDNN, or PyTorch in the optimized path; vendor libraries are linked only for measured comparison under kernels/reference/. This dataset is for SASS and GPU-optimization researchers — it pairs each .cu source with its compiled machine code and its disassembly, so the exact instruction stream a kernel produced on this toolchain can be studied without owning the hardware.

The four laws of GA104

The corpus is organized around four empirically derived constraints. The observations behind each are in docs/gpu_reflections.md.

1. Feed Tensor Cores continuously. Overlap loads with HMMA / IMMA. At ≥8 warps, cp.async benefit depends on the compute/load ratio — helpful when compute is short, harmful when compute is long.
2. Read each byte of DRAM at most once per kernel. im2col converts 9× re-reads to 1×; implicit GEMM eliminates the column buffer.
3. Fill the warp schedulers. 32 warps/SM is ideal, 8 sufficient; below 8 indicates a structural problem.
4. Never cross the 50 KB shared-memory cliff per block. Blocks at

   50 KB drop to 1 block/SM (4 warps), a measured 2× regression.

sm_86 only

Every artifact in this corpus is compiled for compute capability 8.6 (sm_86) only. The cubins are not portable to other architectures and will not load on a non-Ampere or non-GA10x device. This is deliberate: the corpus is a single-target measurement record, not a portable kernel library. There are no fat binaries and no PTX-only fallbacks.

Provenance

This corpus was built and published by scripts/publish_hf.R in the source repository. The stamp below is filled in at publish time.

• Source commit: cd06d0a628464a4186f702ac8ca7ee3fd629d8b9
• Build date: 2026-05-21 13:49:49 CEST
• Compiler: CUDA 13.2, nvcc V13.2.78
• Cubin build line: nvcc -arch=sm_86 -O2 --cubin
• Disassembly: cuobjdump -sass (via scripts/build.R disasm)
• Measurement hardware: RTX 3070 Ti Laptop (GA104, 46-SM bin)

What the .cubin and .sass files are

*.sm_86.cubin files are GPU machine code — ELF containers of sm_86 SASS, loaded through the CUDA Driver API (cuModuleLoad). They are not host executables; running one on a CPU does nothing. The matching *.sm_86.sass files are cuobjdump -sass disassembly of those cubins, provided so the instruction stream is greppable without a CUDA install.

The repository's local R package cuasmR reads a cubin via nvdisasm, indexes instructions by .text byte offset, and patches at the byte level — a byte-identical roundtrip (compile → disasm → reassemble) is part of its test surface. See docs/cuasm_r.md.

Headline performance

Measured on RTX 3070 Ti Laptop (GA104, sm_86, 46-SM bin), CUDA 13.2 / nvcc V13.2.78, driver 595.97. Sparse 2:4 figures are dense-equivalent (the multiply count the sparse pattern would do as dense work).

Kernel	Size	GFLOPS / TOPS	% peak
Sparse HGEMM 2:4	2048³	41,721 (eq)	24.0%
Sparse INT8 mma.sp	2048³	39,674 (eq)	11.4%
HGEMM 16-warp	4096³	31,910	18.3%
IGEMM 128×256	4096³	27,591	7.9%
Flash Attention v2	seq=1024 b=8 h=8	11,453	6.6%
Conv2d implicit GEMM	64×64, 320ch	6,687	3.8%

The headline figure is 41,721 GFLOPS dense-equivalent for sparse 2:4 HGEMM at 2048³. Full per-kernel tables and the phase progression (naive SGEMM → sparse INT8, 90.5× cumulative) are in docs/inventory.md.

File layout

The corpus distinguishes tracked sources from the generated supplement rebuilt at publish time.

Path	Contents
kernels/	Tracked .cu / .cuh kernel sources, grouped by family
kernels/**/*_handtuned.sm_86.cubin	Tracked hand-patched cubins (SASS hand-edits)
generated/	Rebuilt .sm_86.cubin + .sm_86.sass set, build output (kernels of record only — test/probe binaries and negative-result experiment variants excluded)
data/	Regenerable CSV/JSON: baselines, register audit, SASS histogram
docs/	Researcher-facing analyses and references
AGENTS.md	Hardware constants, toolchain, conventions, the four laws
SHA256SUMS	SHA-256 of every cubin in the corpus
LICENSE	MIT

generated/ is build output — it is rebuilt deterministically from the tracked kernels/ sources with the build line above. Treat kernels/ as canonical and generated/ as a convenience supplement.

Reproduction

The single source of truth and the only supported reproduction path is the GitHub repository:

https://github.com/pjt222/bare-metal

From a fresh clone on an sm_86 host: make reproduce (setup → verify → build → bench). This dataset is a published snapshot; it is not the build system.

License

MIT. See LICENSE. Vendor libraries linked under kernels/reference/ for comparison are not redistributed here — only the project's own benchmark wrappers are included.