Generating OpenCL/CUDA source code from C++ expressions in VexCL

on YouTube ↗58m

" summary

Library generates OpenCL/CUDA kernels JIT from C++ expression templates and generates OpenCL/CUDA kernels JIT that replace temporaries with fused computations, uses Boost.Proto to build and traverse expression trees, and caches compiled kernels in local statics to avoid repeated compile overhead. Built device-agnostic backends (OpenCL and CUDA), supports random generators and symbolic recording, runs ensemble ODE solves, and reports roughly 1.5× speed vs some vendor kernels for large problems and up to ~10×–20× when full fusion removes global-memory temporaries.

AI-written summary · about

" chapters (12)

1. 0:00VexCL and GPU frameworks
2. 3:14What is VexCL
3. 8:56Generating OpenCL/CUDA kernels
4. 10:00Expression templates for CPU
5. 17:46Extending to GPU code
6. 26:41Boost.Proto for expressions
7. 34:26Defining terminals in VexCL
8. 37:38Lazy functions and variables
9. 41:19Performance gains with kernels
10. 47:48OpenCL CPU vs GPU performance
11. 50:23Why OpenCL and CUDA
12. 57:56Q&A

" related talks

• 01h 01mTake Advantage of All the MIPS - SYCL & C++Michael S Wong, Nevin Liber, Ronan Keryell, Hugh Delaney, Jakub Chlanda · CppCon 2022
• 36mCUDA Kernels with C++Michael Gopshtein · NDC TechTown 2018
• 01h 32mHeterogeneous Programming in C++ todayMichael S Wong · ACCU 2018
• 56m“Towards Heterogeneous Programming in C++"Gordon Brown, Michael Wong · CppCon 2016
• 37mC++ for multi-accelerator, multi-vendor systemsGuy Tamir · Meeting C++ 2022
• 01h 09mC++20 on Xilinx FPGA with SYCL for VitisRonan Keryell · CppCon 2021
• 01h 00mGPU Accelerated Computing & Optimizations on Cross-Vendor Graphics Cards with Vulkan & KomputeAlejandro Saucedo · CppCon 2022
• 01h 07mHeterogeneous Modern C++ with SYCLMichael S Wong, Nevin Liber, Tom Deakin, Gordon Brown · CppCon 2021