Book Image

Hands-On GPU Programming with Python and CUDA

By : Dr. Brian Tuomanen
Book Image

Hands-On GPU Programming with Python and CUDA

By: Dr. Brian Tuomanen

Overview of this book

Hands-On GPU Programming with Python and CUDA hits the ground running: you’ll start by learning how to apply Amdahl’s Law, use a code profiler to identify bottlenecks in your Python code, and set up an appropriate GPU programming environment. You’ll then see how to “query” the GPU’s features and copy arrays of data to and from the GPU’s own memory. As you make your way through the book, you’ll launch code directly onto the GPU and write full blown GPU kernels and device functions in CUDA C. You’ll get to grips with profiling GPU code effectively and fully test and debug your code using Nsight IDE. Next, you’ll explore some of the more well-known NVIDIA libraries, such as cuFFT and cuBLAS. With a solid background in place, you will now apply your new-found knowledge to develop your very own GPU-based deep neural network from scratch. You’ll then explore advanced topics, such as warp shuffling, dynamic parallelism, and PTX assembly. In the final chapter, you’ll see some topics and applications related to GPU programming that you may wish to pursue, including AI, graphics, and blockchain. By the end of this book, you will be able to apply GPU programming to problems related to data science and high-performance computing.
Table of Contents (15 chapters)

Questions

  1. In the atomic operations example, try changing the grid size from 1 to 2 before the kernel is launched while leaving the total block size at 100. If this gives you the wrong output for add_out (anything other than 200), then why is it wrong, considering that atomicExch is thread-safe?
  2. In the atomic operations example, try removing __syncthreads, and then run the kernel over the original parameters of grid size 1 and block size 100. If this gives you the wrong output for add_out (anything other than 100), then why is it wrong, considering that atomicExch is thread-safe?
  3. Why do we not have to use __syncthreads to synchronize over a block of size 32 or less?

  1. We saw that sum_ker is around five times faster than PyCUDA's sum operation for random-valued arrays of length 640,000 (10000*2*32). If you try adding a zero to the end of this number (that is, multiply it...