Exploring Static Code Generation and SIMD-Acceleration for Machine Learning on RISC-V
Talkat RISC-V Forum: Developer Tools and Tool Chains 2021 · on Wednesday, 2021-06-02· at 18:05· Virtual·
summary
Presents two static code generators—one based on TensorFlow Lite for Microcontrollers and one based on TVM—that compile ML models into static executable code for RISC-V microcontrollers to eliminate runtime and memory overhead of dynamic model runtimes. Describes RISC-V CMSIS-NN kernel implementations that use the RISC-V P (packed/SIMD) and V (vector) extensions to accelerate inner loops, and reports evaluation results on TinyMLPerf using the ETISS simulator showing benefits of static code generation and SIMD-accelerated kernels.
AI-written summary · about
chapters (8)
- 0:00Introduction to ML on RISC-V
- 3:04Static code generation process
- 6:05SIMD and vector extensions
- 8:00Open source toolchain support
- 10:04Reference implementations and benchmarks
- 11:17Performance comparison results
- 13:12Runtime overhead analysis
- 15:15Future frameworks and optimizations
related talks
28mAn Efficient Implementation of TensorFlow Lite for RISC-V Vectors
10mSupport TVM QNN Flow on RISC-V with SIMD Computation
17mEmpower ML Framework on RISC-V by Backend Graph Compilers
40mHands-on TinyML Deployment on RISC-V: Building an Optimized Edge AI Pipeline
13mUnleashing ML Processing Power Through RISC-V Vectors: Applications, Implementation and...
40mHands-on TinyML Deployment on RISC-V: Building an Optimized Edge AI Pipeline - Corrected Audio
10mLightning Talk: Software Development for ML and RISC-V Vector Acceler...
36mPorting ML to RISC‑V: How Not to Lose Performance
