In this post, I will introduce how to solve a parallel computation task using Intel Threading Building Blocks<\/a>.<\/p>\n In the deep learning platform, given inputs contains several thousand images, we want to analyze the data path of a certain deep learning model. The analysis part of each image is identical, for example, we want to collect the minimum and maximum values of a certain layer in the model. Say that we only care about the output of the last Softmax layer which is basically a 1-D array. So the problem is we want to collect the minimum and maximum value (two float numbers) of that layer for all the image inputs.<\/p>\n Since the behavior of each image is similar, the intuition is to utilize multiple threads to do parallel computations. We can easily create threads in C++ 14 thanks to the thread library. However, it makes no sense that we create 5000 threads (actually the overhead of creating and deleting those threads will ruin the performance, see this reference<\/a>).<\/p>\n We can manually maintain a thread pool. We firstly create $MAX_THREADS$ (a typical value is the number of cores) threads and every time a thread finished, we start a new thread until all the images have been processed.<\/p>\n It is not a good idea to build wheels from scratch. There are two most popular libraries that implements thread pools, Intel Thread Building Blocks(TBB)<\/a> and Microsoft Parallel Patterns Library(PPL)<\/a>.<\/p>\n After research, PPL is not good for multi-platform development. Thus, we choose TBB.<\/p>\n There are lots of bugs before I make it work. I only present the succeed way for you and for my later review.<\/p>\nProblem<\/h2>\n
Settings<\/h2>\n
\n
Research<\/h2>\n
Install and Build<\/h2>\n
\n
sudo apt-get install libtbb-dev<\/code><\/pre>\n