A library for numerical computing. It provides
- Expression templates for vectors, matrices, and higher-dimensional arrays.
- Support for execution policies including SIMD vectorization and task-based parallelization.
- Strongly typed versions of structural matrices (e.g. triangular and symmetric).
- A modern C++ interface to BLAS and LAPACK.
- GCC 7 or later.
- TBB for parallelization. (optional but recommended)
- BLAS and LAPACK libraries (See, for example, MKL or ATLAS (optional)).
- jemalloc. (optional).
git clone https://github.com/rnburn/satyr
cd satyr
mkdir .build && cd .build
cmake -DWITH_TBB=[ON|OFF] \
-DWITH_JEMALLOC=[ON|OFF] \
-DWITH_BLAS_LAPACK=[ON|OFF] \
..
make && make install
// Declare some random vector and matrices.
satyr::vector<double> v(5);
satyr::matrix<double> A(5, 5), B(5, 5);
satyr::symmetric_matrix<double> S(5);
satyr::lower_triangular_matrix<double> L(5);
// The standard arithmetic operators and mathematical functions can be used to
// execute expression templates.
A = B + 2.0 * cos(S);
// Expressions involving only structural matrices are computed in an efficient
// manner that avoids unnecessary work.
S = sqrt(abs(S)); // computes only over a triangular portion of the matrix.
// Additionally execution policies can be applied to parallelize or vectorize
// the computation.
A = sqrt(abs(L)) << satyr::parallel_v << satyr::simd_v;
// For parallelization, you can also specify a grainsize if the cost of
// managing tasks could potentially be more expensive than the computation
// itself.
A += cos(B) - as_diagonal_matrix(v) << satyr::grainsize{
10}; // Don't create tasks with fewer than 10 iterations.
// Additionally, you can declare numerical arrays of arbitrary dimension.
satyr::n_array<float, 3> H(5, 2, 6);
// And all array-like objects support indexing and slicing.
A(0, 0) -= 5;
A += H(satyr::all_v, 1, satyr::range{1, 6});
// There is a wrapper for many BLAS-LAPACK functions.
auto C = product(A, B); // calls gemm.
auto w = left_solve(L, v); // calls trsv.