learning-chip/symboliclib

SymbolicLib

This repository contains an implementation of algorithms presented in

Sparsity-Specific Code Optimization using Expression Trees
Philipp Herholz, Xuan Tang, Teseo Schneider, Shoaib Kamil, Daniele Panozzo, Olga Sorkine-Hornung
In ACM Transactions on Graphics (2022)
PDF, Project page

Compiling

Compiling from scratch requires CMake and a recent version of XCode on Mac and Visual Studio 2019 on Windows.

On MacOS, compiling should be as simple as

git clone https://github.com/PhHerholz/SymbolicLib
cd SymbolicLib && mkdir build && cd build
cmake .. 
make -j4

The Symbolic type

A symbolic expression is represented by instances of the class Sym::Symbolic. We start by creating two variables. Variables have two parameters, a variable id and a variable group.

Symbolic a(0, 0);
Symbolic b(1, 0);

Symbolic instances can be combined using mathematical operations.

Symbolic c = a + b * sqrt(a * b);

Evaluating the requires concrete values for a and b. The second argument to evaluate defines these values for variable group 0.

cout << evaluate(c, vector<double>{1., 2.}) << endl; // 3.82843

The expression can be differentiated with respect to a set of variables.

auto dc = differentiate(c, vector<Symbolic>{a, b});
   
cout << evaluate(dc[0], vector<double>{1., 2.}) << endl; // 2.41421
cout << evaluate(dc[1], vector<double>{1., 2.}) << endl; // 2.12132

Basic library usage

The library can be used to generate optimized code evaluating a sparse expression. First we load a regular sparse matrix.

 Eigen::SparseMatrix<double> A;
 Eigen::loadMarket(A, "../../data/sphere.mtx");

The function makeSymbolic builds a copy of the sparse matrix A and replaces each value with a variable of group 0.

Eigen::SparseMatrix<Symbolic> AS = makeSymbolic(A, 0);

The symbolic matrix BS stores symbolic expressions for each entry.

Eigen::SparseMatrix<Symbolic> BS = AS.transpose() * AS + AS;

We want to compile a program that evaluates the expression. Compute unit defines such a program; the first parameter requests a vectorized program using 256 AVX2 registers holding 4 doubles. NumThreads(8) defines a parallelized implementation using 8 threads. The next two parameters define the input variables of the program AS and the expression to be evaluated BS.

ComputeUnit<double> unit(Device(VecWidth(4), NumThreads(8)), AS, BS);

Compile, link and execute the program using the numeric values contained in A.

unit.compile().execute(A);

Retrieve the result and compare to a reference solution.

Eigen::SparseMatrix<double> B = A.transpose() * A + A;
Eigen::SparseMatrix<double> B2 = B;

unit.getResults(B2);

cout << "difference: " << (B - B2).norm() << endl;

Generating a program for AMD HIP devices just requires setting the UseHIP device parameter.

ComputeUnit<double> unitHIP(Device(UseHIP(), ThreadsPerBlock(128)), AS, BS);
unitHIP.compile().execute(A).getResults(B2);

Cuda devices can be used in the same way.

ComputeUnit<double> unitCuda(Device(UseCuda(), ThreadsPerBlock(128)), AS, BS);
unitCuda.compile().execute(A).getResults(B2);

The full example can be found here.