Use L-BFGS-B Fortran library for native logistic regression benchmark (#8)

This PR swaps out the native logistic regression benchmark's solver for the same one used in SciPy. L-BFGS-B is wrapped in a minimal DAAL optimization_solver class and directly set as the solver for DAAL's logistic regression. This brings native performance for this benchmark to match our optimized scikit-learn performance for a sample binary classification problem with 100k samples and 1k features.

Because we must now directly call BLAS and LAPACK functions instead of using DAAL's internal BLAS and LAPACK implementations, we add a dependency on MKL as well.

Author: bibikar

LICENSES_bundled

@@ -0,0 +1,17 @@
+scikit-learn_bench bundles parts of the compatibly-licensed libraries
+listed below.
+
+Name: CLI11
+Files: native/include/CLI11.hpp
+License: BSD license
+  For details, see https://github.com/CLIUtils/CLI11
+
+Name: L-BFGS-B
+Files: native/lbfgsb/*.f
+License: BSD license
+  For details, see scipy/optimize/lbfgsb/README
+
+Name: scipy
+Files: native/lbfgsb/linpack.f
+License: 3-clause BSD
+  For details, see scipy/optimize/lbfgsb/README

README.md

@@ -5,7 +5,7 @@ Python*. See benchmark results [here](https://intelpython.github.io/scikit-learn
 
 ## Prerequisites
 - python and scikit-learn to run python versions
-- `icc` and `daal` to compile and run native benchmarks
+- `icc`, `ifort`, `mkl`, `daal` to compile and run native benchmarks
 
 ## Automatically build and run
 - Run `make`. This will generate data, compile benchmarks, and run them.

native/Makefile

@@ -5,13 +5,15 @@
 BENCHMARKS += distances kmeans linear ridge pca \
 			  two_class_svm multi_class_svm log_reg_lbfgs \
 			  decision_forest_regr decision_forest_clsf
+FOBJ = $(addprefix lbfgsb/,lbfgsb.o linpack.o timer.o)
 CXXSRCS = $(addsuffix _bench.cpp,$(BENCHMARKS))
 
 CXX = icc
 CXXFLAGS += -m64 -fPIC -fp-model strict -O3 -fomit-frame-pointer \
 		  -xSSE4.2 -axCORE-AVX2,COMMON-AVX512
 CXXFLAGS += -std=c++14 -g
-LDFLAGS += -ldaal_core -ldaal_thread -Wl,-rpath,$(CONDA_PREFIX)/lib
+LDFLAGS += -ltbb -lstdc++ -lpthread -lm -ldaal_core -ldaal_thread \
+		   -Wl,-rpath,$(CONDA_PREFIX)/lib
 CXXINCLUDE += include
 
 ifneq ($(CONDA_PREFIX),)
@@ -26,10 +28,19 @@ all: $(addprefix bin/,$(BENCHMARKS))
 bin:
 	mkdir -p bin
 
+bin/log_reg_lbfgs: log_reg_lbfgs_bench.cpp $(FOBJ) | bin
+	$(CXX) $^ $(CXXINCLUDE) $(CXXFLAGS) $(LDFLAGS) -mkl=parallel \
+		-lmkl_rt -lifcore -limf -o $@
+
+
 bin/%: %_bench.cpp | bin
 	$(CXX) $< $(CXXINCLUDE) $(CXXFLAGS) $(LDFLAGS) -o $@
 
 
+$(FOBJ):
+	make -C lbfgsb
+
+
 clean:
 	rm -rf bin/
 

native/lbfgsb/Makefile

@@ -0,0 +1,10 @@
+FC = ifort
+FFLAGS = -g -O3 -m64 -fPIC -fp-model strict -xSSE4.2 -fomit-frame-pointer \
+		 -axCORE-AVX2,COMMON-AVX512 -mkl
+FSRC = lbfgsb.f linpack.f timer.f
+
+all: $(FSRC:.f=.o)
+
+%.o: %.f
+	$(FC) $(FFLAGS) -c $< -o $@
+

native/lbfgsb/README

@@ -0,0 +1,31 @@
+L-BFGS-B
+========
+
+This directory contains version 3.0 of L-BFGS-B in Fortran, available
+at http://users.iems.northwestern.edu/~nocedal/lbfgsb.html
+under the terms of the 3-clause BSD license. Fortran files `lbfgsb.f`,
+`linpack.f`, and `timer.f` come from this library.
+
+`linpack.f` was modified to include calls to LAPACK as in scipy v0.10.0.
+Scipy is also licensed under the BSD license.
+
+`lbfgsb.h` contains a C header to call the Fortran setulb_ routine.
+`lbfgsb_daal.h` contains a DAAL class wrapping the L-BFGS-B library for use
+as a DAAL iterative solver.
+
+The authors of the library ask that we quote at least one of the references
+below.
+
+References
+----------
+
+1.  R. H. Byrd, P. Lu and J. Nocedal. A Limited Memory Algorithm for Bound
+    Constrained Optimization, (1995), SIAM Journal on Scientific and
+    Statistical Computing , 16, 5, pp. 1190-1208.
+2.  C. Zhu, R. H. Byrd and J. Nocedal. L-BFGS-B: Algorithm 778: L-BFGS-B,
+    FORTRAN routines for large scale bound constrained optimization (1997),
+    ACM Transactions on Mathematical Software, Vol 23, Num. 4, pp. 550 - 560.
+3.  J.L. Morales and J. Nocedal. L-BFGS-B: Remark on Algorithm 778: L-BFGS-B,
+    FORTRAN routines for large scale bound constrained optimization (2011),
+    to appear in ACM Transactions on Mathematical Software.
+