AADC Prototype Library

Overview

That library is a prototype implementation of automatic adjoint differentiation (AAD) approach developed by MatLogica. Please check website for more information: http://www.matlogica.com/

As an example, we attempt to differentiate code recorded inside function mySuperModelFunc() and calculate derivatives using:

• finite-differences method
• AAD method using double scalar function
• AAD method using AVX256 or AVX512 vector

Build and Run

Library was tested to compile with GCC, CLang & Intel compilers on Linux platform and MSVC on Windows.

On Windows, just open solution file aadc.sln using Visual Studio.

On Linux, run the following commands to build the project:

• mkdir build
• cd build
• cmake ..
• make
• run binary ./example

Regenerate AAD function

Pre-generated AAD code is included in the project. If you wish to make changes to primal function and regenerate AAD code, you will need:

• uncomment line #define REGENERATE_AAD_FUNCTION inside example.cpp file
• run command make
• run command ./example
• comment out line #define REGENERATE_AAD_FUNCTION
• run command make && ./example

That will update auto-generated AAD function and compile + link updated definition inside "example" binary file.

Sample Output

Primal function & FD derivatives:
f(0.5,1) = 406.892
dx = -10.2952
da = -0.793778

Compiled function & AAD derivatives (scalar case):
f(0.5,1) = 406.892
dx = -10.2904
da = -0.793772

Integral of Primal function & FD derivative:
f() = 451.175
da  = -0.717925
time= 1650860

Compiled function & AAD derivatives (AVX256)
f (0.5,1) = 406.892
dx = -10.2904
da = -0.793772

Integral[0,1] of AVX AAD function using Monte-Carlo & 1 thread:
f() = 451.175
da  = -0.718056
time= 676434

Integral[0,1] of AVX AAD function using Monte-Carlo & 4 threads:
f() = 451.175
da  = -0.718056
time= 177143