Tech Exeter Conference: Intro to GPU Development in Python

Writing code for GPUs has come a long way over the last few years and it is now easier than ever to get started. You can even do it in Python! This talk will cover setting up your Python environment for GPU development. How coding for GPUs differs from CPUs, and the kind of problems GPUs excel at solving. We will dive into some real examples using Numba and also touch on a suite of Python Data Science tools called RAPIDS.

Session takeaways

* You don't need to learn C++ to develop on GPUs
* GPUs are useful for more than just machine learning
* hardware accelerators like GPUs are going to be more important than ever in order to scale our current workloads

Transcript

1. Jacob Tomlinson | Tech Exeter Conference 2020
   Intro to GPU development
   in Python
   

   View Slide

2. View Slide

3. 3
   RAPIDS Github
   https://github.com/rapidsai
   

   View Slide

4. 4
   GPU-Accelerated ETL
   The Average Data Scientist Spends 90+% of Their
   Time in ETL as Opposed to Training Models
   

   View Slide

5. 5
   Lightning-fast performance on real-world use cases
   Up to 350x faster queries; Hours to Seconds!
   TPCx-BB is a data science benchmark consisting of 30 end-to-end
   queries representing real-world ETL and Machine Learning workflows,
   involving both structured and unstructured data. It can be run at
   multiple “Scale Factors”.
   ▸ SF1 - 1GB
   ▸ SF1K - 1 TB
   ▸ SF10K - 10 TB
   RAPIDS results at SF1K (2 DGX A100s) and SF10K (16 DGX A100s) show
   GPUs provide dramatic cost and time-savings for small scale and
   large-scale data analytics problems
   ▸ SF1K 37.1x average speed-up
   ▸ SF10K 19.5x average speed-up (7x Normalized for Cost)
   

   View Slide

6. 6
   Dask
   DEPLOYABLE
   ▸ HPC: SLURM, PBS, LSF, SGE
   ▸ Cloud: Kubernetes
   ▸ Hadoop/Spark: Yarn
   PYDATA NATIVE
   ▸ Easy Migration: Built on top of NumPy, Pandas Scikit-Learn, etc
   ▸ Easy Training: With the same APIs
   ▸ Trusted: With the same developer community
   EASY SCALABILITY
   ▸ Easy to install and use on a laptop
   ▸ Scales out to thousand node clusters
   POPULAR
   ▸ Most Common parallelism framework today in the PyData and SciPy community
   

   View Slide

7. 7
   What are GPUs?
   

   View Slide

8. 8
   Gaming Hardware
   For pwning n00bs
   

   View Slide

9. 9
   Mysterious Machine Learning Hardware
   For things like GauGAN
   Semantic Image Synthesis
   with Spatially-Adaptive
   Normalization
   Taesung Park, Ming-Yu Liu,
   Ting-Chun Wang,
   Jun-Yan Zhu
   arXiv:1903.07291 [cs.CV]
   

   View Slide

10. 10
    CPU GPU
    

    View Slide

11. 11
    https://youtu.be/-P28LKWTzrI
    

    View Slide

12. 12
    GPU vs CPU
    https://docs.nvidia.com/cuda/cuda-c-programming-guide/
    

    View Slide

13. 13
    Using a GPU is like using two computers
    Icons made by Freepik from Flaticon
    Network
    SSH
    VNC
    RD
    SCP
    FTP
    SFTP
    Robocopy
    

    View Slide

14. 14
    Using a GPU is like using two computers
    Icons made by Freepik from Flaticon
    PCI
    CUDA
    

    View Slide

15. 15
    What is CUDA?
    

    View Slide

16. 16
    CUDA
    

    View Slide

17. 17
    I don’t write C/C++
    

    View Slide

18. 18
    What does CUDA do?
    Construct GPU code with CUDA C/C++ language extensions
    Copy data from RAM to GPU
    Copy compiled code to GPU
    Execute code
    Copy data from GPU to RAM
    How do we run stuff on the GPU?
    

    View Slide

19. 19
    Let’s do it in Python
    

    View Slide

20. View Slide

21. View Slide

22. 22
    Writing a Kernel
    Differences between a kernel and a function
    ● A kernel cannot return anything, it must instead modify memory
    ● A kernel must specify its thread hierarchy (threads and blocks)
    A kernel is a GPU function
    

    View Slide

23. 23
    Threads, blocks, grids and warps
    https://docs.nvidia.com/cuda/cuda-c-programming-guide/
    

    View Slide

24. 24
    What?
    Rules of thumb for threads per block:
    ● Should be a round multiple of the warp size (32)
    ● A good place to start is 128-512 but benchmarking is required to
    determine the optimal value.
    

    View Slide

25. 25
    Imports
    

    View Slide

26. 26
    Data arrays
    

    View Slide

27. 27
    Example kernel
    

    View Slide

28. 28
    Running the kernel
    

    View Slide

29. 29
    Absolute positions
    

    View Slide

30. 30
    Thread and block positions
    

    View Slide

31. 31
    Example kernel (again)
    

    View Slide

32. 32
    How did the GPU update our numpy array?
    If you call a Numba CUDA kernel with data that isn’t on the GPU it will be
    copied to the GPU before running the kernel and copied back after.
    This isn’t always ideal as copying data can waste time.
    

    View Slide

33. 33
    Create a GPU array
    

    View Slide

34. 34
    Simplified position kernel
    

    View Slide

35. 35
    GPU Array
    

    View Slide

36. 36
    Copy to host
    

    View Slide

37. 37
    Higher level APIs
    

    View Slide

38. View Slide

39. 39
    cuDF
    

    View Slide

40. 40
    cuDF cuIO
    Analytics
    Data Preparation Visualization
    Model Training
    cuML
    Machine Learning
    cuGraph
    Graph Analytics
    PyTorch,
    TensorFlow, MxNet
    Deep Learning
    cuxfilter, pyViz,
    plotly
    Visualization
    Dask
    GPU Memory
    RAPIDS
    End-to-End GPU Accelerated Data Science
    

    View Slide

41. 41
    Interoperability for the Win
    DLPack and __cuda_array_interface__
    mpi4py
    

    View Slide

42. 42
    __cuda_array_interface__
    

    View Slide

43. 43
    Recap
    

    View Slide

44. 44
    Recap
    GPUs run the same function (kernel) many times in parallel
    When being called each function gets a unique index
    CUDA/C++ is used to write kernels, but high level languages like Python can also compile to it
    Memory must be copied between the CPU (host) and GPU (device)
    Many familiar Python APIs have GPU accelerated implementations to abstract all this away
    Takeaways on GPU computing
    

    View Slide

45. THANK YOU
    Jacob Tomlinson
    @_jacobtomlinson
    

    View Slide