Scribd Logo
Upload

Welcome to Scribd!

  • 97 views

    PyCUDA AH PDF

    Uploaded by

    jackops
    This document provides an introduction to using PyCUDA to run CUDA programs in Python. It demonstrates adding two matrices in parallel on a GPU. The tutorial shows how to set up a Google Colab notebook with GPU acceleration, import PyCUDA, allocate memory on the GPU, copy matrices from CPU to GPU, write a CUDA kernel to add the matrices in parallel, call the kernel to perform the calculation on the GPU, extract the result back to the CPU, and compare the result to adding the matrices sequentially on the CPU. The advantages of parallel programming include faster execution of repeated computations, efficient data handling, and not requiring as powerful machines as CPUs.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd

    PyCUDA AH PDF

    Uploaded by

    jackops
    97 views16 pages
    This document provides an introduction to using PyCUDA to run CUDA programs in Python. It demonstrates adding two matrices in parallel on a GPU. The tutorial shows how to set up a Google Colab notebook with GPU acceleration, import PyCUDA, allocate memory on the GPU, copy matrices from CPU to GPU, write a CUDA kernel to add the matrices in parallel, call the kernel to perform the calculation on the GPU, extract the result back to the CPU, and compare the result to adding the matrices sequentially on the CPU. The advantages of parallel programming include faster execution of repeated computations, efficient data handling, and not requiring as powerful machines as CPUs.

    Original Title

    PyCUDA-AH.pdf

    Copyright

    © © All Rights Reserved

    Available Formats

    PDF, TXT or read online from Scribd

    Did you find this document useful?

    Is this content inappropriate?

    This document provides an introduction to using PyCUDA to run CUDA programs in Python. It demonstrates adding two matrices in parallel on a GPU. The tutorial shows how to set up a Google Colab notebook with GPU acceleration, import PyCUDA, allocate memory on the GPU, copy matrices from CPU to GPU, write a CUDA kernel to add the matrices in parallel, call the kernel to perform the calculation on the GPU, extract the result back to the CPU, and compare the result to adding the matrices sequentially on the CPU. The advantages of parallel programming include faster execution of repeated computations, efficient data handling, and not requiring as powerful machines as CPUs.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    97 views16 pages

    PyCUDA AH PDF

    Uploaded by

    jackops
    This document provides an introduction to using PyCUDA to run CUDA programs in Python. It demonstrates adding two matrices in parallel on a GPU. The tutorial shows how to set up a Google Colab notebook with GPU acceleration, import PyCUDA, allocate memory on the GPU, copy matrices from CPU to GPU, write a CUDA kernel to add the matrices in parallel, call the kernel to perform the calculation on the GPU, extract the result back to the CPU, and compare the result to adding the matrices sequentially on the CPU. The advantages of parallel programming include faster execution of repeated computations, efficient data handling, and not requiring as powerful machines as CPUs.

    Copyright:

    © All Rights Reserved
    Download as PDF, TXT or read online from Scribd
    Download as pdf or txt
    of 16

    PyCUDA

    Tutorial
    Introduction
    As we may have already seen, NVIDIA GPUs use CUDA cores to perform calculations. In this segment, we
    will be using Python to run CUDA programs.

    The difference between running C and Python is that Python has a lot more QOL especially in regards to
    data visualization and scientific computing.
    Aim
    We will demonstrate one of the simplest applications of parallel programming. Adding of two matrices.
    Note that each element in the matrix will be added in parallel i.e. in a single step. We will forgo the
    classical for loop in this case. We will start with 2 2D matrices with an arbitrary size in numpy and then
    will proceed to add them.
    Getting started
    To start our virtual workspace, we will use Google Colaboratory, address at:

    https://colab.research.google.com/

    Google’s cloud service will provide this to us for free (note that you will require a gmail account)
    Setup

    1. Top left, File-> new Python 3 notebook


    2. Runtime -> Hardware accelerator -> GPU
    3. Setup workspace with a new code cell

    Since pycuda is not a native library in colab we need an additional line before importing the libraries.

    !pip install pycuda.

    Run the code segment first before proceeding (at the left, a play button)
    Building
    Wait for a bit while pycuda is being installed. After the build finishes, we are ready to proceed.

    Firstly we need to import the basic libraries.


    Declaring
    We declare our matrix as follows

    Of course, we are free to declare as we wish, but this will create a standard random number matrix,
    repeat for a “b” matrix. Note that the cloud probably does not support double precision accuracy so to
    avoid difficulties, we pre-convert to float32 or single precision.
    GPU details
    As we may have seen earlier, but GPU’s don’t support data in them, that is to say, we can’t say that
    > accessing GPU
    > GPU variable a
    >a=3

    We will need to allocate memory and then create a copy from the device and access it via pointers and
    the run calculations on that.
    We do this as follows:

    As the memory allocation step


    Then proceed with:

    Which essentially copies a to a_gpu

    htod stands for “host to device” that is, from your system to the GPU, we will be making the
    reverse of this command to extract data from our GPU.
    Writing the function
    Interestingly enough, the function body has to be The equivalence in this is such that all a[idx] are
    written in C being executed simultaneously.

    Linearization of the 2D matrix will give us the


    following relation between idx and (x,y) of an
    index.

    PS : don’t forget the “;”


    a a+1 a+2 a+3 a+4 Array indexing.

    threadIdx.x refers to x
    a+5 a+(5+1) a+(5+2) a+(5+3) a+(5+4) threadIdx.y refers to y

    blockDim.x refers to
    a+(5*2) a+(5*2+1) . . . dimension in x which
    in our case is 5
    Hence, we can create
    . . . . . an equivalent 1D
    array of size 25x1
    with the following
    . . . . a+(5*4+4) index notation.
    = a+24

    s e
    Calling the function

    Have to create a separate variable to get This will extract the function from
    the function as follows “SourceModule” and execute it on the
    GPU copies of a and b
    Extracting the
    result

    This step is pretty straightforward. For


    simplicity, we will create a different matrix
    and copy data from device to host using

    dtoh

    Which stands for “device to host”


    Result

    We should get some output in the form of


    matrices

    As an exercise, we could check the check


    the result without the GPU in action as a+b

    And compare the results


    Advantages of parallel programming
    1. Faster execution of repeated computation
    2. Efficient handling of data
    3. Does not need high power machines in comparison to CPU’s

    Note: As we may already have seen, but parallel programming needs to be applied in very specific
    circumstances. More specifically, it is best suited for work where a computation is not affected by its next
    step.

    Ex. repeated derivatives as opposed to complete derivative calculation.


    Thank You

    You might also like