Pascal

Pascal is a software framework written in C that enables hardware-enhanced Bayesian Learning. It contains modules that can carry out operations on tensors, similar to NumPy Arrays, PyTorch Tensors, and other popular tools. Pascal can also carry out reverse-mode automatic differentiation using the pascal-autodiff module.

Requirements

Pascal can run using several backends. By default, Pascal uses clapack, which is included in this repository. Therefore you do not need any other requirements, other than a C compiler.

Pascal uses either a cblas or a libgsl backend to carry out some operations. Thus, these should be installed. The Makefile includes flags that should point to standard locations. If these don't work, please updated the INCLUDE_FLAGS, LDFLAGS and FRAMEWORK_FLAGS variables in the configs/default.conf as needed. Alternatively, you can create a new configuration file (.conf) in the configs folder and update the CONFIGURATION variable in the Makefile (note that if your new configuration is configs/new-config.conf, CONFIGURATION should be set to new-config).

CBLAS and GSL on MacOS

On MacOS, these are included in the Accelerate framework. This is installed with the XCode command-line tools, by running:

xcode-select --install

Currently, default.conf assumes that libgsl is installed using homebrew by running:

brew install gsl

If you install it using other means, please update the INCLUDE_FLAGS and LDFLAGS in the correct if-else block.

CBLAS and GSL on Linux

To be written

Build

To use this package, simply clone this repository. This can be done as:

git clone --recursive https://github.com/physical-computation/pascal.git

Note

The --recursive is important since this repository uses submodules. This also means that if you use Pascal as a submodule to another repository, you need to run the following after adding the Pascal submodule.

git submodule update --init --recursive

Then you can build it using

make

This creates a library file in libs/

To test, run:

make test

If everything worked well, this will print:

tests/load_data/test; tests/tensor_integration/test; tests/tensor/test; tests/tensor_random_sample/test; tests/tensor_utils/test; tests/autodiff/test; tests/autodiff_primitives/test; tests/tensor_broadcast/test; tests/tensor_iterate/test;
Running tests in load_data:
Completed:       3/3 passed in 596µs
Running tests in pascal_tensor_integrations:
Completed:       15/15 passed in 192µs
Running tests in tensor:
Completed:       48/48 passed in 372µs
Running tests in pascal_tensor_random_sample:
Completed:       2/2 passed in 3µs
Running tests in pascal_tensor_utils:
Completed:       8/8 passed in 10µs
Running tests in autodiff:
Completed:       15/15 passed in 273µs
Running tests in autodiff_primitives:
Completed:       19/19 passed in 106µs
Running tests in pascal_tensor_broadcast:
Completed:       3/3 passed in 12µs
Running tests in pascal_tensor_iterate:
Completed:       6/6 passed in 8µs

Examples and Benchmarks

Note

Currently, the Makefiles in examples/ and benchmarks/ are written to work with macOS only.

A basic implementation of a tensor operation using Pascal is

The examples and benchmarks use Python to run some analyses and baselines. It is recommended that you use a virtual environment. Install the required packages using

pip install -r requirements.txt

Please see the examples/ folder. It includes a few simple examples of using the functionality that Pascal offers. The Makefiles in these folders currently only support macOS. However, they should be a useful guide for other platforms: you can think of them as compiling your application while including libpascal.a (which gets created when you run make).

Running examples

All examples can be executed by running the following from the root directory of this project:

make examples

Individual examples can be executed using the following syntax:

make eg-<example>

<example> can be replaced by the names of the folders inside of examples/, such as nn-simple and bnn-simple.

Running benchmarks

We currently compare some of our tensor operations against NumPy. To run these benchmarks, run

make benchmarks

Documentation

There is a set of incomplete documentation in docs/. If you load docs/html/index.html locally, you can navigate this documentation on your browser.

Furthermore, documentation comments can be found on some of the function in include/pascal.h.