Support TVM QNN Flow on RISC-V with SIMD Computation

on YouTube ↗10m

" summary

Compiles quantized (int8) models with TVM QNN flow and replaces convolution kernels with LLVM intrinsics mapped to RISC-V Superword and Subword SIMD, using TVM tensorization to emit vectorized primitives. Runs on Spike and an NDIS cycle-accurate simulator, benchmarks MobileNet v1/v2 and Inception v3/v4, measures instruction-count reductions of 1.79–7.58x versus FP32, and evaluates accuracy on 1k ImageNet images showing acceptable loss. Implementation modifies TVM schedules and LLVM backend to target RISC‑V P extension SIMD.

AI-written summary · about

" related talks

• 17mExploring Static Code Generation and SIMD-Acceleration for Machine Learning on RISC-VRafael Stahl · RISC-V Forum: Developer Tools and Tool Chains 2021
• 28mAn Efficient Implementation of TensorFlow Lite for RISC-V VectorsMostafa Hagog · RISC-V Forum: Vector & Machine Learning 2021
• 24mMoving to RISC-V Vector: A Practical Journey of AI Operator OptimizationGuodong Xu · RISC-V Summit North America 2025
• 16mAccelerate HPC and AI Applications with RVV Auto-VectorizationHualin Wu · RISC-V Summit North America 2023
• 13mKeynote: Taking RISC-V Intelligence EverywhereCharlie Su · RISC-V Summit North America 2023
• 01h 45mPorting SIMD to RISC-V Vectors | Corrected AudioDaniel Thompson · RISC-V Summit Europe 2026 Developer Workshops
• 40mHands-on TinyML Deployment on RISC-V: Building an Optimized Edge AI Pipeline - Corrected AudioYueh-Feng Lee, Niraj Dengale · RISC-V Summit Europe 2026 Developer Workshops
• 19mRISC-V: Changing the Way AI/ML Accelerators and Computing Infrastructure Are BuiltDavid Chen · RISC-V Summit North America 2024