GPU MODE Lecture 10: Build a Prod Ready CUDA library

Lecture #10 explores the challenges of real-time CUDA development, focusing on building a C++ library for efficient GPU communication and task scheduling to abstract away CUDA complexities for non-expert users while maintaining performance.

This post is part of the following series:

• GPU MODE Lecture Notes: My notes from the GPU MODE reading group lectures run by Andreas Kopf and Mark Saroufim.

• Introduction
  
• Relevant CUDA Concepts
• The CUDA Ninja’s Motivation: Why Create Libraries?
• Automatic TV: Understanding the Application and its Challenges
• Library Challenges: Abstraction vs. Performance
• Use Case 1: GPU Communication Manager
  
• Use Case 2: CUDA Layers for Image Preprocessing
• Conclusion and Future Directions

Resource Links:

• YouTube Recording: Lecture 10: Build a Prod Ready CUDA library
• Slides: Lecture 10: Building production ready CUDA libraries

Introduction

Speaker Introduction

• Oscar Amoros Huguet: Professional CUDA developer and active community member.
• Background:
  • Started with OpenCL for processing 3D volumes (wave propagation simulation and medical imaging).
  • Worked with Barcelona Supercomputing Center on OmpSs (cluster programming abstraction).
  • Taught OpenSphere Encoder and computer science topics at the University of Barcelona.
  • Co-authored papers on OpenCL and related topics.
  • Created “Simple OpenCL” library (linked by Khronos Group).
  • Ported Boost Compute’s OpenCL code to HSAIL (GPU assembly).
  • Currently works at Mediapro on Automatic TV, a soft real-time C++ and CUDA-based application.

Lecture Structure

• Review of Relevant CUDA Concepts: GPU communication, streams, latency hiding, kernel fusion.
• The CUDA Ninja’s Motivation for Library Creation: Addressing practical challenges in a production environment.
• Use Cases:
  • GPU Communication Manager: Optimizing data transfer between different memory spaces.
  • CUDA Layers for Image Preprocessing: Accelerating common image processing operations for tasks like neural network inference.
• Open Source Project: Discussion of an open-source CUDA library project.

Relevant CUDA Concepts

• Focus: Host-side optimization for real-world applications.
• Tools: NVIDIA Nsight Systems for profiling and identifying bottlenecks.
• Key Concepts:
  • GPU Memory Copies (cudaMemcpyAsync): Data transfer between CPU and GPU or between GPUs.
  • CUDA Streams: Enabling asynchronous execution of kernels and memory copies.
  • Latencies and Latency Hiding: Overlapping computation and data transfer to minimize idle time.
  • Kernel Fusion (Vertical and Horizontal): Combining multiple kernels to reduce overhead.
  • Memory-Bound Kernels: Performance limited by memory bandwidth.
  • Compute-Bound Kernels: Performance limited by computational capacity.

The CUDA Ninja’s Motivation: Why Create Libraries?

• Initial Expectations (Ideal Scenario):
  • Bug-free, well-tested CPU code.
  • Unit tests for performance benchmarking and validation.
• Reality in a Startup Environment:
  • Collaboration with Non-CUDA Experts: Computer vision specialists, C++ application developers, etc.
  • Need for Real-Time Prototyping: Enabling quick iteration and testing of computer vision algorithms.
  • Performance Bottlenecks in Production: Handling larger datasets and multiple cameras demanded further optimization beyond initial prototypes.
  • QA and Performance Measurement: Establishing performance benchmarks and cost goals.
• The Need for Automation: Repetitive optimization tasks and the desire to empower non-CUDA programmers led to the idea of creating libraries and abstractions.

Automatic TV: Understanding the Application and its Challenges

• Automatic TV: A software for automated sports recording that tracks the game and provides dynamic camera views.
• Functionality:
  • Uses multiple fixed 4K cameras as input.
  • Stitches camera feeds to create a seamless panoramic view.
  • Employs AI (computer vision and neural networks) for player tracking, object detection, and camera switching.
  • Generates a Full HD output video with enhanced image quality and features (scoreboard overlays, etc.).
• Performance Challenges:
  • Real-time/Soft Real-time Processing: Maintaining a smooth output video with minimal delay.
  • Multi-GPU Support: Distributing the processing workload across up to three GPUs for increased throughput.
  • Efficient GPU Communication: Minimizing the overhead of data transfer between GPUs and the CPU.

Library Challenges: Abstraction vs. Performance

• Key Challenges:
  • Performance: Achieving optimal performance comparable to hand-tuned CUDA code.
  • Abstraction: Hiding CUDA complexities and providing an easy-to-use interface for non-CUDA programmers.
  • Balancing Abstraction and Performance: Finding the right level of abstraction that simplifies usage without significantly sacrificing performance.
  • User Requirements: Understanding the target users’ needs and preferences (level of CUDA knowledge, performance expectations, etc.).
  • API Design: Creating an intuitive and familiar API that integrates seamlessly with existing workflows.

Use Case 1: GPU Communication Manager

Problem Definition: Efficient Data Transfer

• Goal: Create a system for efficient data transfer (e.g., images) between different memory spaces (CPU, GPU, CPU pinned memory).
• Memory Spaces:
  • CPU Memory: Standard system RAM.
  • GPU Memory: Dedicated memory on the graphics card.
  • CPU Pinned Memory: A special type of CPU memory that can be directly accessed by the GPU, eliminating the need for an intermediate copy.
• CUDA memcpy: The standard CUDA function for transferring data between memory spaces.

Requirements and Considerations

• Zero Allocation During Runtime: Allocate all necessary memory upfront to avoid blocking CPU threads during real-time processing.
• Handling Same Memory Space Transfers: Optimize for cases where the source and destination memory spaces are the same, avoiding unnecessary copies.
• Minimizing CPU Thread Blocking: Ensure the CPU thread responsible for scheduling CUDA work remains responsive and is not blocked by memory operations.

Memory Space Combinations and Copy Strategies

Source Memory Space	Destination Memory Space	Desired Copies	CUDA memcpy Strategy
CPU	CPU	0	No copy
CPU Pinned	CPU Pinned	0	No copy
GPU	GPU	0	No copy
CPU	CPU Pinned	1	cudaMemcpyHostToHost
CPU Pinned	CPU	1	cudaMemcpyHostToHost
CPU	GPU	2	cudaMemcpyHostToHost (CPU to CPU Pinned) followed by cudaMemcpyHostToDevice
GPU	CPU	2	cudaMemcpyDeviceToHost followed by cudaMemcpyHostToHost (CPU Pinned to CPU)
CPU Pinned	GPU	1	cudaMemcpyHostToDevice
GPU	CPU Pinned	1	cudaMemcpyDeviceToHost
GPU	GPU (Same Device)	0	No copy
GPU	GPU (Different Devices, Peer-to-Peer)	1	cudaMemcpyPeerAsync (if supported)
GPU	GPU (Different Devices, No Peer-to-Peer)	2	cudaMemcpyDeviceToHost followed by cudaMemcpyHostToDevice

• Peer-to-Peer Communication: Direct data transfer between GPU memories without involving the CPU, achievable through NVLink or PCI Express (if supported by the hardware and drivers).
• Manual Pinned Memory Allocation: Bypassing CUDA runtime’s automatic pinned memory allocation to prevent CPU thread blocking.

Overcoming Sequential Data Transfer Bottlenecks

• Problem: Sequential memory copies and kernel launches introduce significant delays in a multi-GPU pipeline.
• Solution: Introduce delays (buffers) and utilize CUDA streams to enable parallel execution of kernels and memory transfers.

Producer-Consumer Model

• Traditional CPU-Based Model:
  • Producer: A thread that generates data and writes it to a shared buffer.
  • Consumer: A thread that reads and processes data from the shared buffer.
  • Buffer: Manages access to the shared data, ensuring thread safety.
• Key Features:
  • Task Parallelism: Producer and consumer can operate concurrently.
  • Variable Buffer Size: Adjusts to fluctuations in producer and consumer execution times.

Adapting Producer-Consumer for GPU Communication

• Iterative Memory Manager: A specialized buffer designed for GPU communication in iterative applications.

  

  Slide 24

• Key Features:

  • Iteration-Based Synchronization: All asynchronous operations (kernels and copies) are synchronized at the end of each iteration.
  • Ping-Pong Buffers: Utilizes at least two pointers (memory regions) to enable concurrent reading and writing.

• Example:

  • A kernel writes its output to a pointer in the manager.
  • A copy operation transfers data from that pointer to another pointer in a different memory space.
  • Another kernel reads from the second pointer.
  • Pointers are swapped at the end of each iteration to ensure continuous data flow.

Delay Buffers and Optimizing Memory Manager

• Delay Buffer: A buffer that introduces a fixed delay in the data flow.
• Optimization: When a delay buffer is used, the memory manager can directly access data from the delay buffer’s pointer, eliminating the need for an extra copy.
• Example:
  • A kernel writes to a delay buffer.
  • Instead of copying from the kernel’s output to the memory manager, the manager directly reads from the delay buffer after the specified delay.

Provider-Taker Model: An Abstraction for Memory Management

• Motivation: Introduce a new abstraction to simplify the interaction with the memory manager.
• Key Concepts:
  • Taking: Requesting a pointer from the memory manager (the manager allocates and owns the pointer).
  • Providing: Supplying a pointer to the memory manager (the caller allocates and owns the pointer).
• Ownership:
  • Taking: The memory manager owns the pointer.
  • Providing: The caller owns the pointer.
• Difference from Producer-Consumer:
  • Producer-Consumer focuses on data movement.
  • Provider-Taker focuses on pointer ownership and allocation responsibility.

Iterative Memory Manager Interface

• Data Structure: Define a class representing the data to be transferred.

  struct DataInfo {
      int numElements;
    int elemSizeInBytes;
    MemorySpace memSpace;
  };

  • Includes information like width, height, and memory space.

• Producer/Consumer Roles: Define an enum to specify whether the producer and consumer will take or provide pointers:

  enum Actions { ProducerProvides, ProducerTakes, ConsumerProvides, ConsumerTakes};

• Memory Manager Initialization:

  DataInfo producerDataInfo{1024, 4, HostPageable};
  DataInfo consumerDataInfo{1024, 4, Device_1};
  
  Data ptrToProduce(producerDataInfo);
  Data ptrToConsume(consumerDataInfo);
  
  //Initialization
  MemoryManager<ProducerProvides, ConsumerProvides> manager(producerDataInfo, consumerDataInfo);

• Managing Data Transfer:

  int delay = manager.getTotalDelay(); // Query delay generated by the manager
  
  manager.manage(ptrToProduce, ptrToConsume); // Usage

• Understanding Provide and Take:

  • Provide: Pass data structures as arguments to manage(). The caller is responsible for allocating these data structures.
  • Take: The manage() function returns pointers to data structures allocated by the manager.

Memory Manager Configurations

• Four Possible Configurations:
  • Take-Take: Producer and consumer both take pointers from the manager.
  • Provide-Provide: Producer and consumer both provide pointers to the manager.
  • Take-Provide: Producer takes a pointer, consumer provides a pointer.
  • Provide-Take: Producer provides a pointer, consumer takes a pointer.
• Memory Space Awareness: The memory manager automatically handles copies based on the specified memory spaces of the producer and consumer data.
• Zero-Copy Optimization: If the source and destination memory spaces are the same, the manager performs no copy and simply forwards the pointer.

Example: Memory Manager with a Delay Buffer

• Scenario:
  • A kernel writes to a buffer.
  • A delay buffer stores the kernel’s output.
  • Another kernel reads from the delay buffer after a fixed delay.
• Memory Manager Configuration: Take-Provide (producer takes, consumer provides).
• Zero-Copy Optimization: If the delay buffer and the consumer kernel are in the same memory space, no copy is performed.

Timeline Analysis: Achieving Parallelism

• Multi-GPU Pipeline with Delay Buffers:
  • Kernels on different GPUs execute in parallel.
  • Memory transfers are overlapped with computation using CUDA streams.
• Benefits:
  • Maximizes GPU utilization.
  • Reduces overall processing time.

Real-World Application: Automatic TV’s GPU Pipeline

• Automatic TV’s Implementation:

  • GPU 0: Camera input processing, data transfer to CPU.
  • GPU 1: AI processing (computer vision, neural networks).
  • GPU 2: Output processing, data transfer to CPU for encoding.

• NVIDIA Nsight Systems Analysis: Demonstrates efficient parallelism between compute and data transfer operations.

  

  Slide 45

  

  Slide 45

Potential Application: Multi-GPU Neural Network Training

• Proposed Idea: Apply the Iterative Memory Manager concept to multi-GPU neural network training.
• Potential Benefits:
  • Parallelize forward and backward passes across multiple GPUs.
  • Overlap computation with data transfer of intermediate results.
• Challenges:
  • Requires sufficient GPU memory to store multiple copies of intermediate data.
  • Complexity of integrating with existing training frameworks.

Use Case 2: CUDA Layers for Image Preprocessing

• Problem: Accelerate common image preprocessing operations for tasks like neural network inference.

• GitHub Project: cvGPUSpeedup

• OpenCV CUDA vs. Custom CUDA Layer:

  // OpenCV version
  cv::cuda::resize(d_input(crop), d_up, targetRes, 0., 0., cv::INTER_LINEAR, cv_stream);
  d_up.convertTo(d_temp, CV_32FC3, alpha, cv_stream);
  
  cv::cuda::subtract(d_temp, val_sub, d_temp2, cv::noArray(), -1, cv_stream);
  cv::cuda::divide(d_temp2, val_div, d_temp, 1.0, -1, cv_stream);
  
  cv::cuda::split(d_temp, d_output, cv_stream);
  
  // cvGPUSpeedup version
  cv::Scalar val_alpha(alpha, alpha, alpha);
  cvGS::executeOperations(cv_stream,
      cvGS::resize<CV_8UC3, cv::INTER_LINEAR>(d_input(crop), targetRes, 0., 0.),
      cvGS::convertTo<CV_8UC3, CV_32FC3>(),
      cvGS::multiply<CV_32FC3>(val_alpha),
      cvGS::subtract<CV_32FC3>(val_sub),
      cvGS::divide<CV_32FC3>(val_div),
      cvGS::split<CV_32FC3>(d_output));

  • OpenCV CUDA provides individual functions for operations like cropping, resizing, and color conversion.
  • The custom CUDA layer fuses these operations into a single kernel, reducing overhead and improving performance.

• Performance Gains: Significant speedups (e.g., 167x) achieved by fusing operations and optimizing memory access.

• Section cut short due to time constraints for the live-stream.

Conclusion and Future Directions

• Key Takeaways:
  • Libraries can significantly improve CUDA development by optimizing performance and providing abstractions for non-CUDA programmers.
  • The Iterative Memory Manager is a powerful tool for managing data transfer in multi-GPU pipelines.
  • CUDA layers can accelerate common image processing tasks, particularly beneficial for neural network inference.
• Future Work:
  • Explore the application of the Iterative Memory Manager to multi-GPU neural network training.
  • Further develop and refine the open-source CUDA layer library for image preprocessing.
  • Improve the abstraction and usability of the libraries to make them more accessible to a wider audience.

---

About Me:

• I’m Christian Mills, a deep learning consultant specializing in computer vision and practical AI implementations.
• I help clients leverage cutting-edge AI technologies to solve real-world problems.
• Learn more about me or reach out via email at [email protected] to discuss your project.