Qingyang-Lucas-Fang/OpenAssembler

选用指南

AI 根据 README 整理 · 2026/10/31 20:42:16

使用方式

通过命令行使用，如 openassembler compile 命令可编译 PTX 到 cubin，openassembler gpu run 可在真实 GPU 上运行，并支持 --help 查看完整命令列表。

查看 README 依据

multiply-high lowering: **13 → 2–3 instructions**. - **Strassen / AlphaTensor fast matrix multiplication**: 97 published low-rank factorizations of matrix multiply (AlphaTensor, DeepMind/Nature 2022) imported, verified exactly over the rationals, compiled by our backend, and measured against cuBLAS on an A100 at n = 8192. - **`openassembler`, a CLI** that ties all of this together: compile a PTX file straight to a cubin with our backend, verify it against a reference interpreter, and benchmark it on any NVIDIA GPU you can reach — your own box, an SSH host, or a cloud GPU.

Run it on a real GPU — any Linux box with an NVIDIA driver, no CUDA toolkit needed htere: ```sh openassembler gpu run --backend user@your-gpu-box --which ours ``` ...or on this machine, if it has an NVIDIA GPU: ```sh openassembler gpu run --backend local --which ours ``` ...or on a cloud A100 via Modal (what produced the numbers below): ```sh openassembler gpu run --backend modal --dry-run # cost estimate, $0 spent openassembler gpu run --backend modal ``` Run `openassembler --help` for the full command list (`compile`, `verify`,

部署要求

需要 Linux 环境（或 macOS 虚拟机）以及 NVIDIA CUDA 工具包作为黑盒参照，运行 GPU 需要 NVIDIA 驱动及 GPU（如 A100），编译时需安装 Python 包。

查看 README 依据

--- ## Quick start ```sh git clone https://github.com/Qingyang-Lucas-Fang/OpenAssembler && cd OpenAssembler pip install -e . ``` One-time: fetch NVIDIA's own CUDA toolkit as a black-box oracle (never vendored in this repo — see NOTICE). Needs a Linux box or a VM (e.g. colima on macOS); see REPRODUCING.md if you're not on Linux. ```sh env/fetch_cuda.sh ``` Compile a PTX kernel straight to a cubin with our backend, then verify it against a reference PTX interpreter: ```sh openassembler compile examples/div_by_const.ptx -o out.cubin --verify ``` --- ## Commands for different devices

已知限制

OpenAssembler 并未完全取代 ptxas，仍依赖 ptxas 生成的 ELF 容器；编译器尚未支持 f16/f64 算术、warp shuffle、寄存器溢出等，且仅支持 sm_80 架构。

查看 README 依据

Every number in this README is reproducible; see [REPRODUCING.md](REPRODUCING.md) and [docs/STATUS.md](docs/STATUS.md) for the exact command behind each one. > **Honesty up front.** OpenAssembler does not replace `ptxas` end to end: it > still uses a `ptxas`-built cubin as the ELF container (symbol table, > relocations, constant banks) and only replaces the *code* inside it — see > [How it works](#how-it-works). It also doesn't cover all of PTX yet: no > f16/f64 arithmetic, no warp shuffles, no register spilling. See > [Gaps](#gaps) and [docs/STATUS.md](docs/STATUS.md) for the honest list.

arch-independent; the encoder and latency model are sm_80 (Ampere) only. The living version of this list, with what's measured vs. hypothesized, is [docs/STATUS.md § Gaps and next steps](docs/STATUS.md#gaps-and-next-steps). --- ## Repository layout ``` openassembler/ the package: backend/ (compiler), isa/ (encoder), emu/ (SASS emulator), container/ (ELF splicing), ptx/ (parser + reference interpreter), opt/ (div-by-const),