Merge pull request #48 from fabinsch/add-benchmark-vectorization

benchmark: document speed up by vectorization

Author: jcarpent

README.md

@@ -16,18 +16,18 @@ Through **ProxSuite**, we aim at offering to the community scalable optimizers w
 
 **ProxSuite** is integrated into the [cvxpy](https://www.cvxpy.org/) modeling language for convex optimization problems.
 
-## Quick install
+## Getting started
 
 **ProxSuite** is distributed on many well-known package managers.
 
-### With <img src="https://www.python.org/static/community_logos/python-logo-inkscape.svg" height="30" style="vertical-align: -1em;">:
+### Quick install with <img src="https://www.python.org/static/community_logos/python-logo-inkscape.svg" height="30" style="vertical-align: -1em;">:
 
 ```bash
    pip install proxsuite
 ```
 This approach is only available on Linux and Mac OS X.
 
-### With <img src="https://s3.amazonaws.com/conda-dev/conda_logo.svg" height="18">:
+### Quick install with <img src="https://s3.amazonaws.com/conda-dev/conda_logo.svg" height="18">:
 
 ```bash
    conda install proxsuite -c conda-forge
@@ -36,8 +36,13 @@ This approach is available on Linux, Windows and Mac OS X.
 
 ### Alternative approaches
 
-Alternative installation procedures are presented in the [Installation Procedure](#installation-procedure) section.
+Installation from source is presented [here](https://github.com/Simple-Robotics/proxsuite/blob/main/doc/5-installation.md).
 
+### Compiling a program
+For the fastest performance use the following command to enable vectorization
+```bash
+g++ -O3 -march=native -DNDEBUG -std=gnu++17 -DPROXSUITE_VECTORIZE examples/benchmark_dense_qp.cpp -o benchmark_dense_qp $(pkg-config --cflags proxsuite)
+```
 ## ProxSuite main features
 
 **Proxsuite** is fast:

doc/1-Overview.md

@@ -48,6 +48,11 @@ You can compile the C++ version by including ProxSuite and Eigen header director
 
 \code g++ -std=c++17 examples/cpp/overview-simple.cpp -o overview-simple $(pkg-config --cflags proxsuite)  \endcode
 
+If you are looking for the fastest performance, use the following flags to use SIMDE in proxsuite and tell your
+compiler to use the corresponding cpu instruction set
+\code g++ -O3 -march=native -DNDEBUG -DPROXSUITE_VECTORIZE -std=c++17 examples/cpp/overview-simple.cpp -o overview-simple $(pkg-config --cflags proxsuite)
+\endcode
+
 Once your code is compiled, you might then run it using
 
 \code ./overview-simple \endcode

examples/cpp/benchmark_dense_qp.cpp

@@ -0,0 +1,94 @@
+/*
+
+Compile this code once with and without vectorization to see performance
+difference.
+
+Use vectorization:
+g++ -O3 -march=native -DNDEBUG -std=gnu++17 -DPROXSUITE_VECTORIZE
+benchmark_dense_qp.cpp -o benchmark_dense_qp $(pkg-config --cflags proxsuite)
+
+Do not use vectorization:
+g++ -DNDEBUG -std=gnu++17 benchmark_dense_qp.cpp -o benchmark_dense_qp
+$(pkg-config --cflags proxsuite)
+
+Comparison of timings on Intel i7-11850H and ubuntu 20.04 using this file:
+
+With vectorization:
+sparsity_factor: 0.1
+Setup Time consumption(dense): 0.000237295s
+Solve Time consumption(dense): 0.000500206s
+sparsity_factor: 0.2
+Setup Time consumption(dense): 0.000465961s
+Solve Time consumption(dense): 0.000903787s
+sparsity_factor: 0.3
+Setup Time consumption(dense): 0.000697931s
+Solve Time consumption(dense): 0.00136976s
+sparsity_factor: 0.4
+Setup Time consumption(dense): 0.000931736s
+Solve Time consumption(dense): 0.00185252s
+
+
+Without vectorization:
+sparsity_factor: 0.1
+Setup Time consumption(dense): 0.0147825s
+Solve Time consumption(dense): 0.0277815s
+sparsity_factor: 0.2
+Setup Time consumption(dense): 0.029592s
+Solve Time consumption(dense): 0.0490869s
+sparsity_factor: 0.3
+Setup Time consumption(dense): 0.0443664s
+Solve Time consumption(dense): 0.0746045s
+sparsity_factor: 0.4
+Setup Time consumption(dense): 0.0592621s
+Solve Time consumption(dense): 0.101507s
+
+*/
+
+#include <proxsuite/proxqp/dense/dense.hpp>
+#include <proxsuite/proxqp/utils/random_qp_problems.hpp>
+
+using namespace proxsuite::proxqp;
+using T = double;
+
+int
+main()
+{
+  double N = 100;
+  double solve_time = 0.0;
+  double setup_time = 0.0;
+
+  dense::isize dim = 100;
+  dense::isize n_eq(dim * 0.5);
+  dense::isize n_in(dim * 0.5);
+
+  for (T sparsity_factor = 0.1; sparsity_factor < 0.5; sparsity_factor += 0.1) {
+    T strong_convexity_factor(1.e-2);
+    dense::Model<T> qp_random = utils::dense_strongly_convex_qp(
+      dim, n_eq, n_in, sparsity_factor, strong_convexity_factor);
+
+    for (int i = 0; i < N; i++) {
+      dense::QP<T> qp(dim, n_eq, n_in);
+      qp.settings.max_iter = 10000;
+      qp.settings.max_iter_in = 1000;
+      qp.settings.eps_abs = 1e-5;
+      qp.settings.eps_rel = 0;
+      qp.init(qp_random.H,
+              qp_random.g,
+              qp_random.A,
+              qp_random.b,
+              qp_random.C,
+              qp_random.l,
+              qp_random.u);
+      qp.solve();
+      solve_time += qp.results.info.solve_time / N;
+      setup_time += qp.results.info.setup_time / N;
+    }
+    std::cout << "sparsity_factor: " << sparsity_factor << std::endl;
+    std::cout << "Setup Time consumption(dense): " << setup_time / 1e6 << "s"
+              << std::endl
+              << "Solve Time consumption(dense): " << solve_time / 1e6 << "s"
+              << std::endl;
+  }
+
+  return 0;
+}