Beyond CUDA: GPU Accelerated Computing on Cross-Vendor Graphics Cards with Vulk...
Talkat KubeCon + CloudNativeCon + Open Source Summit China 2021 · on Thursday, 2021-12-09· at 11:20· Open Source Summit 演讲厅, 虚拟会议 (Virtual)·
summary
A cross‑vendor GPU compute stack is demonstrated using Vulkan SDK and the Kompute framework with both Python and C++ bindings, running on AMD, Qualcomm and NVIDIA hardware; a simple multiply kernel returns 2 4 6 and a logistic‑regression shader trains for 100 iterations using micro‑batching (example batch size 5) to update weights on the device. The presenter contrasts 500–2000 lines vs dozens of code, shows pre‑recorded sequences, async submission and concurrent queues, and confirms the examples execute and return learned parameters.
AI-written summary · about
chapters (17)
- 0:00Intro
- 0:11Hello, my name is Alejandro
- 0:49Why Parallel Processing?
- 3:07Parallel Processing: Options
- 4:50Vulkan C++ SDK Disadvantages.
- 10:54Enter Kompute The General Purpose Vulkan Compute Framework.
- 12:35Vulkan Kompute: Components
- 18:34The Hello World of ML
- 18:56ML Example Intuition
- 20:03Kompute Logic to Set Up
- 22:15LR Shader Logic
- 23:32Kompute Logic: Create Tensors
- 24:16Kompute Logic: Init Tensors
- 24:34Kompute Logic: Main Sequence
- 25:17Kompute Logic: "Learn" LR Params
- 26:19Kompute Logic: Print LR Params
- 27:35High level Roadmap
related talks
18mCompiling Strassen-like Matrix Multiplication Algorithms to Fast CUDA Kernels
27mVendor-Neutral GPU Programming in Chapel
08mGPU Programming with Triton Kernels
05mLightning Talk: GPU Programming with C++ and Triton
02h 51mCUDA Python Kernel Authoring
25mLeveraging AI Ecosystem for Portable and Sustainable GPU Kernels in HPC
42mOpen Source is Better Together- GPU Python Libraries Unite
53mIntroduction to C++ AMP (GPGPU Computing)