Triton-RV

NOTE: This repo is now archived. For the latest updates, please see RuyiAI-Stack/triton-riscv.

Overview

This repo conducts a comparative analysis of runtime auto-tuning in Triton-CPU across multi-core x86 and RISC-V architectures, benchmarking it against compiler-based auto-tuning approaches to systematically evaluate the impact of various tuning parameters.

1. We introduced cross-platform capabilities to the toolchain, enabling seamless support for both x86 and RISC-V architectures.
2. We developed a visualization module that converts benchmark results into CSV format for streamlined data analysis.
3. We bumped this repo to a Triton-CPU version (2fa1c59), which introduces critical enhancements for CPU performance including Sleef math library integration and optimized multi-threading support.

Environmental Setup

1. Clone and Initialize

git clone git@github.com:xlinsist/Triton-RV.git
cd Triton-RV
git submodule update --init

2. Build triton-cpu

Build triton-cpu as suggested here.

3. Modify global configuration

modify global_config.sh before running. For example, switch PLATFORM to rv or x86.

4. Compile, execute, and report the results

Run the all-in-one script, you may see the generated report.xls in the local 'build-<benchmark_name>' directories.

cd scripts
./run_all.sh