Clad is a source-transformation automatic differentiation (AD) library for C++,
implemented as a plugin for the Clang compiler.
Clad enables automatic differentiation (AD) for C++. It is based on LLVM compiler infrastructure and is a plugin for Clang compiler. Clad is based on source code transformation. Given C++ source code of a mathematical function, it can automatically generate C++ code for computing derivatives of the function. It supports both forward-mode and reverse-mode AD.Clad has extensive coverage of modern C++ features and a robust fallback and recovery system in place.
Clad provides five API functions:
clad::differentiateto use forward-mode AD.clad::gradientto use reverse-mode AD.clad::hessianto compute Hessian matrix using a combination of forward-mode and reverse-mode AD.clad::jacobianto compute Jacobian matrix using reverse-mode AD.clad::estimate-errorto compute the floating-point error of the given program using reverse-mode AD.
API functions are used to label an existing function for differentiation.
Both functions return a functor object containing the generated derivative which can be called via .execute method, which forwards provided arguments to the generated derivative function.
For a guide on compiling your clad-based programs, look here.
For a function f of several inputs and single (scalar) output, forward mode AD can be used to compute (or, in case of Clad, create a function) a directional derivative of f with respect to a single specified input variable. Derivative function created by the forward-mode AD is guaranteed to have at most a constant factor (around 2-3) more arithmetical operations compared to the original function.
clad::differentiate(f, ARGS) takes 2 arguments:
fis a pointer to a function or a method to be differentiatedARGSis either:
- a single numerical literal indicating an index of independent variable (e.g.
0forx,1fory) - a string literal with the name of independent variable (as stated in the definition of
f, e.g."x"or"y"), and if the variable is an array the index needs to be specified, e.g."arr[1]"
Generated derivative function has the same signature as the original function f, however its return value is the value of the derivative. Example:
#include "clad/Differentiator/Differentiator.h"
#include <iostream>
double f(double x, double y) { return x * y; }
int main() {
// Call clad to generate the derivative of f wrt x.
auto f_dx = clad::differentiate(f, "x");
// Execute the generated derivative function.
std::cout << f_dx.execute(/*x=*/3, /*y=*/4) << std::endl;
// Dump the generated derivative code to standard output.
f_dx.dump();
}Reverse-mode AD allows computing the gradient of f using at most a constant factor (around 4) more arithmetical operations compared to the original function. While its constant factor and memory overhead is higher than that of the forward-mode, it is independent of the number of inputs. E.g. for a function having N inputs and consisting of T arithmetical operations, computing its gradient takes a single execution of the reverse-mode AD and around 4*T operations, while it would take N executions of the forward-mode, this requiring up to N*3*T operations.
clad::gradient(f, /*optional*/ ARGS) takes 1 or 2 arguments:
fis a pointer to a function or a method to be differentiatedARGSis either:
- not provided, then
fis differentiated w.r.t. its every argument - a string literal with comma-separated names of independent variables (e.g.
"x"or"y"or"x, y"or"y, x")
Since a vector of derivatives must be returned from a function generated by the reverse mode, its signature is slightly different. The generated function has void return type and same input arguments. The function has additional n arguments (where n refers to the number of arguments whose gradient was requested) of type clad::array_ref<T>, where T is the return type of f. Each clad::array_ref variable stores the derivative of the elements as they appear in the orignal function signature. clad::array_ref is a special non-scalar type that allows clad to store the derivatives of scalar and no-scalar values through the same interface. The caller is responsible for allocating and zeroing-out the gradient storage. Example:
auto f_grad = clad::gradient(f);
double dx = 0, dy = 0;
// After this call, dx and dy will store the derivatives of x and y respectively.
f_grad.execute(x, y, &dx, &dy);
std::cout << "dx: " << dx << ' ' << "dy: " << dy << std::endl;
// Same effect as before.
auto f_dx_dy = clad::gradient(f, "x, y");
auto f_dy_dx = clad::gradient(f, "y, x");
// The same effect can be achieved by using an array instead of individual variables.
double result2[2] = {};
f_dy_dx.execute(x, y, /*dx=*/&result2[0], /*dy=*/&result2[1]);
// note that the derivatives are mapped to the "result" indices in the same order as they were specified in the argument:
std::cout << "dy: " << result2[0] << ' ' << "dx: " << result2[1] << std::endl;Clad can produce the hessian matrix of a function using its forward and reverse mode capabilities. Its interface is similar to reverse mode but differs when arrays are involved. It returns the matrix as a flattened vector in row major format.
clad::hessian(f, /*optional*/ ARGS) takes 1 or 2 arguments:
fis a pointer to a function or a method to be differentiatedARGSis either:- not provided, then
fis differentiated w.r.t. its every argument except in the case of arrays where it needs to be provided - a string literal with comma-separated names of independent variables (e.g.
"x"or"y"or"x, y"or"y, x"or in case of arrays"x[0:2]")
- not provided, then
The generated function has void return type and same input arguments. The function has an additional argument of
type clad::array_ref<T>, where T is the return type of f. This clad::array_ref variable stores the hessian
matrix. The caller is responsible for allocating and zeroing-out the hessian storage. Example:
#include "clad/Differentiator/Differentiator.h"
#include <iostream>
double f(double x, double y) { return x * y; }
double g(double x, double y[2]) { return x * y[0] * y[1]; }
int main() {
// Since we are differentiating variables that are not arrays the interface
// is same as in reverse mode
auto f_hess = clad::hessian(f);
// The size of the resultant matrix should be the square of the
// number of independent variables
double mat_f[4] = {0};
clad::array_ref<double> mat_f_ref(mat_f, 4);
// Execute the hessian function
f_hess.execute(/*x=*/3, /*y=*/4, mat_f_ref);
std::cout << "[" << mat_f_ref[0] << ", " << mat_f_ref[1] << "\n "
<< mat_f_ref[2] << ", " << mat_f_ref[3] << "]";
// When arrays are involved the array indexes that are to be differentiated needs to be specified
// even if the whole array is being differentiated
auto g_hess = clad::hessian(g, "x, y[0:1]");
// The rest of the steps are the same.
}Clad can produce the jacobian of a function using its reverse mode. It returns the jacobian matrix as a flattened vector in row major format.
clad::jacobian(f, /*optional*/ ARGS) takes 1 or 2 arguments:
fis a pointer to a function or a method to be differentiatedARGSis either:- not provided, then
fis differentiated w.r.t. its every argument - a string literal with comma-separated names of independent variables (e.g.
"x"or"y"or"x, y"or"y, x")
- not provided, then
The generated function has void return type and same input arguments. The function has an additional argument of
type T *, where T is the pointee type of the output (the last variable) of f. This variable stores the jacobian
matrix. The caller is responsible for allocating and zeroing-out the jacobian storage. Example:
#include "clad/Differentiator/Differentiator.h"
#include <iostream>
void h(double a, double b double output[]) {
output[0] = a * a * a;
output[1] = a * a * a + b * b * b;
output[2] = 2 * (a + b);
}
int main() {
// This sets all the input variables (i.e a and b) as independent variables
auto h_jac = clad::jacobian(h);
// The jacobian matrix size should be the number of
// independent variables * the number of outputs of the original function
// In this case it is 2 * 3 = 6
double jac[6] = {0};
double output[3] = {0};
h_jac.execute(/*a=*/3, /*b=*/4, output, jac);
std::cout << jac[0] << " " << jac[1] << std::endl
<< jac[2] << " " << jac[3] << std::endl
<< jac[4] << " " << jac[5] << std::endl;
}Clad is capable of annotating a given function with floating point error estimation code using the reverse mode of AD. An interface similar to clad::gradient(f) is provided as follows:
clad::estimate_error(f) takes 1 argument:
fis a pointer to the function or method to be annotated with floating point error estimation code.
The function signature of the generated code is the same as from clad::gradient(f) with the exception that it has an extra argument at the end of type double&, which returns the total floating point error in the function by reference. For a user function double f(double, double) example usage is described below:
// Generate the floating point error estimation code for 'f'.
auto df = clad::estimate_error(f);
// Print the generated code to standard output.
df.dump();
// Declare the necessary variables.
double x, y, d_x, d_y, final_error = 0;
// Finally call execute on the generated code.
df.execute(x, y, &d_x, &d_y, final_error);
// After this, 'final_error' contains the floating point error in function 'f'.The above example generates the the error code using an in-built taylor approximation model. However, clad is capable of using any user defined custom model, for information on how to use you own custom model, please visit this demo.
More detail on the APIs can be found under clad's user documentation.
xeus-cling provides a Jupyter kernel for C++ with the help of the C++ interpreter Cling and the native implementation of the Jupyter protocol xeus. Within the xeus-cling framework, Clad can enable automatic differentiation (AD) such that users can automatically generate C++ code for their computation of derivatives of their functions.
To set up your environment, use:
mamba create -n xeus-clad -c conda-forge clad xeus-cling jupyterlab
conda activate xeus-clad
Next, running jupyter notebook will show 3 new kernels for C++ 11/14/17 with Clad attached.
Try out a Clad tutorial interactively in your browser through binder:
Since Clad is a Clang plugin, it must be properly attached when Clang compiler is invoked. First, the plugin must be built to get libclad.so (or .dylib). To compile SourceFile.cpp with Clad enabled use:
clang -cc1 -x c++ -std=c++11 -load /full/path/to/lib/clad.so -plugin clad SourceFile.cpp
To compile using clang-9, use:
clang-9 -I /full/path/to/include/ -x c++ -std=c++11 -fplugin=/full/path/to/lib/clad.so SourceFile.cpp -o sourcefile -lstdc++ -lm
To save the Clad generated derivative code to Derivatives.cpp add:
-Xclang -plugin-arg-clad -Xclang -fgenerate-source-file
To print the Clad generated derivative add:
-Xclang -plugin-arg-clad -Xclang -fdump-derived-fn
At the moment, LLVM/Clang 5.0.x - 14.0.4 are supported.
Clad is available using conda:
conda install -c conda-forge cladIf you have already added conda-forge as a channel, the -c conda-forge is unnecessary. Adding the channel is recommended because it ensures that all of your packages use compatible versions:
conda config --add channels conda-forge
conda update --all#sudo apt install clang-11 libclang-11-dev llvm-11-tools llvm-11-dev
sudo bash -c "$(wget -O - https://apt.llvm.org/llvm.sh)"
sudo -H pip install lit
git clone https://github.com/vgvassilev/clad.git clad
mkdir build_dir inst; cd build_dir
cmake ../clad -DClang_DIR=/usr/lib/llvm-11 -DLLVM_DIR=/usr/lib/llvm-11 -DCMAKE_INSTALL_PREFIX=../inst -DLLVM_EXTERNAL_LIT="``which lit``"
make && make install
brew install llvm@12
brew install python
python -m pip install lit
git clone https://github.com/vgvassilev/clad.git clad
mkdir build_dir inst; cd build_dir
cmake ../clad -DLLVM_DIR=/usr/local/Cellar/llvm/12.0.0_1/lib/cmake/llvm -DClang_DIR=/usr/local/Cellar/llvm/12.0.0_1/lib/cmake/clang -DCMAKE_INSTALL_PREFIX=../inst -DLLVM_EXTERNAL_LIT="`which lit`"
make && make install
make check-clad
sudo -H pip install lit
git clone https://github.com/llvm/llvm-project.git src
cd src; git chekout llvmorg-13.0.0
cd /tools
git clone https://github.com/vgvassilev/clad.git clad
cd ../../../
mkdir obj inst
cd obj
cmake -S ../src/llvm -DLLVM_ENABLE_PROJECTS="clang" -DCMAKE_BUILD_TYPE="Debug" -DLLVM_TARGETS_TO_BUILD=host -DCMAKE_INSTALL_PREFIX=../inst
make && make install
Clad is based on compile-time analysis and transformation of C++ abstract syntax tree (Clang AST). This means that Clad must be able to see the body of a function to differentiate it (e.g. if a function is defined in an external library there is no way for Clad to get its AST).
Note: Clad currently differentiates types such as int/char/boolean as any real type (such as float, double, etc.) would be differentiated. Users should keep in mind that Clad does not warn against lossy casts, which on differentiation may result in incorrect derivatives.
Note: If for any reason clad is unable to algorithmically differentiate a function, it automatically switches to numerically differentiating the same. To disable this behavior, please compile your programs with the -DCLAD_NO_NUM_DIFF flag. The numerical differentiation functionality can also be used standalone over a wide range of function signatures with minimal user intervention. This presentation provides more information on what can be numerically differentiated. For a comprehensive demo on using custom user defined types with numerical differentiation, you can check out this demo.
Sometimes Clad may be unable to differentiate your function (e.g. if its definition is in a library and source code is not available). Alternatively, an efficient/more numerically stable expression for derivatives may be know. In such cases, it is useful to be able to specify a custom derivatives for your function.
Clad supports that functionality by allowing to specify your own derivatives in namespace custom_derivatives. For a function named FNAME you can specify:
- a custom derivative w.r.t
I-th argument by defining a functionFNAME_dargIinsidenamespace custom_derivatives - a custom gradient w.r.t every argument by defining a function
FNAME_gradinsidenamespace custom_derivatives
When Clad will encounter a function FNAME, it will first do a lookup inside the custom_derivatives namespace to try to find a suitable custom function, and only if none is found will proceed to automatically derive it.
Example:
- Suppose that you have a function
my_pow(x, y)which computesxto the power ofy. However, Clad is not able to differentiatemy_pow's body (e.g. it calls an external library or uses some non-differentiable approximation):
double my_pow(double x, double y) { // something non-differentiable here... }However, you know analytical formulas of its derivatives, and you can easily specify custom derivatives:
namespace custom_derivatives {
double my_pow_darg0(double x, double y) { return y * my_pow(x, y - 1); }
double my_pow_darg1(dobule x, double y) { return my_pow(x, y) * std::log(x); }
}You can also specify a custom gradient:
namespace custom_derivatives {
void my_pow_grad(double x, double y, array_ref<double> _d_x, array_ref<double> _d_y) {
double t = my_pow(x, y - 1);
*_d_x = y * t;
*_d_y = x * t * std::log(x);
}
}Whenever Clad will encounter my_pow inside differentiated function, it will find and use provided custom functions instead of attempting to differentiate it.
Note: Clad provides custom derivatives for some mathematical functions from <cmath> inside clad/Differentiator/BuiltinDerivatives.h.
Details on custom derivatives, other supported C++ syntax (already supported or in-progress) and further resources can be found over at clad's user documentation.
% Peer-Reviewed Publication
%
% 16th International workshop on Advanced Computing and Analysis Techniques
% in physics research (ACAT), 1-5 September, 2014, Prague, The Czech Republic
%
@inproceedings{Vassilev_Clad,
author = {Vassilev,V. and Vassilev,M. and Penev,A. and Moneta,L. and Ilieva,V.},
title = {{Clad -- Automatic Differentiation Using Clang and LLVM}},
journal = {Journal of Physics: Conference Series},
year = 2015,
month = {may},
volume = {608},
number = {1},
pages = {012055},
doi = {10.1088/1742-6596/608/1/012055},
url = {https://iopscience.iop.org/article/10.1088/1742-6596/608/1/012055/pdf},
publisher = {{IOP} Publishing}
}Founder of the project is Vassil Vassilev as part of his research interests and vision. He holds the exclusive copyright and other related rights, described in Copyright.txt.
clad is an open source project, licensed by GNU LESSER GENERAL PUBLIC LICENSE (see License.txt). If there is module with different that LGPL license it will be explicitly stated in the License.txt in the module's source code folder.
Please see License.txt for further information.
We are always looking for improvements to the tool, as such open-source developers are greatly appreciated! If you are interested in getting started with contributing to clad, make sure you checkout our contribution guide.