Deep Dive: Accelerating Neural Networks using RVV and Open Standard Software
summary
Ends-to-end workflow that compiles SYCL kernels to an RVV-enabled LLVM backend, injects low-level driver abstractions and a Codeplay vectorizer, and runs a ResNet50 inference demo on a two-core Spike simulator model. Build uses computeCpp front end, SYCL runtime, and libraries SYCL-DNN and Eigen; kernels are auto-vectorized from extracted parallel_for code to RVV instructions, offloaded via DMA to on-chip buffers, and produce profiling traces showing per-core cycles, memory reads/writes, and successful end-to-end inference.
AI-written summary · about
chapters (9)
- 0:00Introduction and context
- 1:45SYCL programming model
- 4:21Ecosystem and benefits
- 6:34VGG demo and compilation
- 8:48Under the hood: Compilation
- 11:59Kernel optimization steps
- 15:15Offloading and vectorization
- 18:57VGG demo and profiling
- 21:17Benefits and Q&A
related talks
20mThe Journey to Enabling RVV Solutions with SYCL and the OneAPI Ecosystem
24mOpen Acceleration for RISC-V – Portability, Performance and Partners...
16mAccelerate HPC and AI Applications with RVV Auto-Vectorization
25mOpenCL support by NNVM & TVM
19mBuilding Tool Chains for RISC-V AI Accelerators
13mKeynote: Taking RISC-V Intelligence Everywhere
24mMoving to RISC-V Vector: A Practical Journey of AI Operator Optimization
10mSupport TVM QNN Flow on RISC-V with SIMD Computation