Wei LIU (Xiaoze LU)

University of Chinese Academy of Sciences

Ocean University of China

NIO Autonomous Driving Technology Co., Ltd.

• Establish CUDA operator optimization methodology for Computer Vision inference tasks(eg. Transformer, Self Attention, Cross Attention, Grid Sample, RoiAlign, BEV2BEV, NMS, CropAndResize)
• Establish CUDA operator optimization methodology for Lidar inference tasks(eg. 3D Sparse Convolution, Voxel Generator, Mapping, Freespace, RangeView)
• Optimize Deep Learning Operator with CUDA to accelerate the training performance
• Optimize CPU Algorithm with SIMD technology to accelerate its performance
• Implement VLIW optimization to accelerate Image Processing tasks
• Design and implement profiling tools for PyTorch system with cupti and nvtx to guide the layer by layer optimization
• Design and implement high performance communication middleware
• Design decentrialized discovery mechanism for communication middleware

TuSimple Inc.

• Optimize Deep Learning operator to accelerate the Deep Learning Inference system with GPU equiped device
• Design and implement Publisher/Subscriber communication semantic on GPU equipped system to improve the transination performance for large memory object
• Introduce MPS(Multi-Process Service) to autonomous driving environment to improve GPU utilization with multi-processes loosely couppled architecture
• Hook CUDA runtime API to fulfill software GPU virtualization and Deep Learning performance improvement
• Accelerate Binary Neutral Network with SIMD(SSE/AVX) instructions on CPU
• Design and implement high performance, high robustic, high reliability communication middleware
• Design and implement transport layer for communication middleware with shared pointer, shared memory, socket mechanism
• Design a publisher-centric lifecycle mangment method for shared memory and notification mechanism
• Design and implement weak-centrialized discovery layer for communication middleware with ETCD as the discovery mechanism
• Optimize and re-implement serialization method for message type system in communication middleware
• Integrate performance metrics and monitor tools into one docker container, to make its' usage more easy

Cute Series(in Chinese)

• Layout for CUTE
• Algebra and Geometry Explantion for CUTE
• Tensor for CUTE
• MMA Abstraction for CUTE
• COPY Abstraction for CUTE
• Simple GEMM Implementation for CUTE
• GEMM Pipeline for CUTE
• Swizzle for CUTE
• High Performance GEMM Implementation for CUTE

CUDA Related Topics(in Chinese)

• Registers are shared between CUDA Core and Tensor Core
• The Advantage of ldmatrix Instruction in CUDA
• The Difference between CUDA Core and Tensor Core in GPU
• How to optimize Transformer Attention

cutlass

• https://github.com/NVIDIA/cutlass

Protocol Buffers

• https://github.com/protocolbuffers/protobuf

ROS2

• https://github.com/ros2

libuv

• https://github.com/libuv/libuv

Jsoncpp

• https://github.com/open-source-parsers/jsoncpp

Google Scholar

• https://scholar.google.com/citations?user=By4g5MUAAAAJ&hl=en