mkl CI config

.ci/ci.yml

@@ -132,13 +132,24 @@ jobs:
     ccacheDir: '$(build.binariesDirectory)/ccache'
     ccacheArtifactName: 'ccache-macos'
     ccacheArchive: '$(Build.ArtifactStagingDirectory)/ccache-$(Build.BuildId).tar'
+    commonMacOSCMakeFlags: '-DBUILD_EXAMPLES=OFF -DBUNDLE_JSON=OFF -DBUNDLE_NLOPT=OFF -DENABLE_TESTING=ON -DENABLE_COVERAGE=OFF -DBUILD_META_EXAMPLES=OFF'
 
   pool:
     vmImage: xcode9-macos10.13
 
   steps:
   - template: ./macos-steps.yml
 
+  strategy:
+    maxParallel: 2
+    matrix:
+      ACCELERATE:
+        use_mkl: false
+        cmakeOptions: '$(commonMacOSCMakeFlags)'
+      MKL:
+        use_mkl: true
+        cmakeOptions: '$(commonMacOSCMakeFlags) -DBLAS_LIBRARIES=$(CONDA)/lib/libmkl_core.dylib -DLAPACK_LIBRARIES=$(CONDA)/lib/libmkl_core.dylib -DBLA_VENDOR=Intel'
+
 - job: windows_libshogun
   displayName: Windows
   timeoutInMinutes: 120

.ci/macos-steps.yml

@@ -11,7 +11,7 @@ steps:
 - template: ./download-extract.yml
   parameters:
     name: 'CCache'
-    artifactName: $(ccacheArtifactName)
+    artifactName: $(ccacheArtifactName)-$(use_mkl)
     destinationFolder: $(ccacheDir)
     downloadPath: $(build.binariesDirectory)
 
@@ -20,10 +20,14 @@ steps:
     brew install ccache pkg-config arpack eigen glpk hdf5 json-c lapack lzo nlopt snappy xz
   displayName: Install dependencies
 
+- script: sudo $(CONDA)/bin/conda install mkl-devel -y -q
+  condition: eq(variables['use_mkl'], 'true')
+  displayName: Install MKL
+
 - task: CMake@1
   displayName: Setup
   inputs:
-    cmakeArgs: -DBUILD_EXAMPLES=OFF -DBUNDLE_JSON=OFF -DBUNDLE_NLOPT=OFF -DENABLE_TESTING=ON -DENABLE_COVERAGE=OFF -DBUILD_META_EXAMPLES=OFF ..
+    cmakeArgs: $(cmakeOptions) ..
 
 - script: make -j4
   displayName: Build
@@ -40,4 +44,4 @@ steps:
     name: 'CCache'
     dir: $(ccacheDir)
     archive: $(ccacheArchive)
-    artifactName: $(ccacheArtifactName)
+    artifactName: $(ccacheArtifactName)-$(use_mkl)

cmake/ShogunFindLAPACK.cmake

@@ -38,7 +38,7 @@ IF (LAPACK_FOUND)
         endif()
       endif()
     endif()
-  ELSEIF("${LAPACK_LIBRARIES}" MATCHES ".*/.*mkl_.*")
+  ELSEIF(EXISTS "${LAPACK_LIBRARIES}" AND "${LAPACK_LIBRARIES}" MATCHES ".*/.*mkl_.*")
     # in case MKL is available enable Eigen to use it.
     # for more fine grained control and details see:
     # https://eigen.tuxfamily.org/dox/TopicUsingIntelMKL.html
@@ -66,14 +66,15 @@ IF (LAPACK_FOUND)
     ENDIF()
 
     IF (ENABLE_EIGEN_LAPACK)
-      FIND_PATH(MKL_INCLUDE_DIR mkl.h)
+      FIND_PATH(MKL_INCLUDE_DIR mkl.h HINTS ${MKL_PATH}/../include/)
       IF(NOT MKL_INCLUDE_DIR)
         MESSAGE(STATUS "Found MKL, but not mkl.h. Make sure that mkl headers are available in order to use MKL as BLAS/Lapack backend for Eigen.")
         SET(ENABLE_EIGEN_LAPACK OFF CACHE BOOL "Use ${ENABLE_EIGEN_LAPACK}" FORCE)
       ELSE()
         MESSAGE(STATUS "Enabling MKL as BLAS/Lapack backend for Eigen.")
         SET(EIGEN_USE_MKL_ALL 1)
         target_include_directories(shogun PUBLIC ${MKL_INCLUDE_DIR})
+        target_include_directories(libshogun PRIVATE ${MKL_INCLUDE_DIR})
         IF (LIBSHOGUN_BUILD_STATIC)
           target_include_directories(shogun-static PUBLIC ${MKL_INCLUDE_DIR})
         ENDIF()

src/shogun/converter/ica/SOBI.cpp

@@ -66,8 +66,11 @@ void CSOBI::fit_dense(CDenseFeatures<float64_t>* features)
 	MatrixXd M0 = cor(EX,int(m_tau[0]));
 	EigenSolver<MatrixXd> eig;
 	eig.compute(M0);
-	MatrixXd SPH = (eig.pseudoEigenvectors() * eig.pseudoEigenvalueMatrix().cwiseSqrt() * eig.pseudoEigenvectors ().transpose()).inverse();
-	MatrixXd spx = SPH*EX;
+	MatrixXd EVMsqrt = eig.pseudoEigenvalueMatrix().cwiseSqrt();
+	MatrixXd SPH = (eig.pseudoEigenvectors() * EVMsqrt *
+	                eig.pseudoEigenvectors().transpose())
+	                   .inverse();
+	MatrixXd spx = SPH * EX;
 
 	// Compute Correlation Matrices
 	index_t * M_dims = SG_MALLOC(index_t, 3);

src/shogun/machine/gp/LogitVGPiecewiseBoundLikelihood.cpp

@@ -502,11 +502,12 @@ void CLogitVGPiecewiseBoundLikelihood::precompute()
 
 	float64_t h_bak = eigen_h(eigen_h.size()-1);
 	eigen_h(eigen_h.size()-1) = 0;
-
 	//bsxfun(@plus, l.^2, v)
-	eigen_l2_plus_s2 = (eigen_s2.replicate(1,eigen_l.rows()).array().transpose().colwise() + eigen_l.array().pow(2)).matrix();
+	VectorXd epl_pow = eigen_l.array().pow(2);
+	VectorXd eph_pow = eigen_h.array().pow(2);
+	eigen_l2_plus_s2 = (eigen_s2.replicate(1,eigen_l.rows()).array().transpose().colwise() + epl_pow.array()).matrix();
 	//bsxfun(@plus, h.^2, v)
-	eigen_h2_plus_s2 = (eigen_s2.replicate(1,eigen_h.rows()).array().transpose().colwise() + eigen_h.array().pow(2)).matrix();
+	eigen_h2_plus_s2 = (eigen_s2.replicate(1,eigen_h.rows()).array().transpose().colwise() + eph_pow.array()).matrix();
 	//pl.*l - ph.*h
 	eigen_weighted_pdf_diff = (eigen_pl.array().colwise() * eigen_l.array() - eigen_ph.array().colwise() * eigen_h.array()).matrix();
 

src/shogun/mathematics/Statistics.cpp

@@ -605,7 +605,7 @@ float64_t CStatistics::log_det_general(const SGMatrix<float64_t> A)
 {
 	Map<MatrixXd> eigen_A(A.matrix, A.num_rows, A.num_cols);
 	REQUIRE(eigen_A.rows()==eigen_A.cols(),
-		"Input matrix should be a sqaure matrix row(%d) col(%d)\n",
+		"Input matrix should be a square matrix row(%d) col(%d)\n",
 		eigen_A.rows(), eigen_A.cols());
 
 	PartialPivLU<MatrixXd> lu(eigen_A);

tests/unit/features/iterators/DotIterator_unittest.cc

@@ -92,5 +92,5 @@ TEST(DotIterator, add)
 	}
 
 	for (i = 0; i < vec.vlen; ++i)
-		EXPECT_EQ(vec[i], res[i]);
+		EXPECT_NEAR(vec[i], res[i], 1E-12);
 }

tests/unit/preprocessor/FisherLDA_unittest.cc

@@ -152,26 +152,26 @@ TEST_F(FLDATest, CANVAR_FLDA_Unit_test)
 	float64_t epsilon=0.00000000001;
 
 	// comparing projection outputs against Matlab 'CannonVar' implementation
-	EXPECT_NEAR(+51.3760082238439111, y(0,0), epsilon);
-	EXPECT_NEAR(-56.4869601858493127, y(0,1), epsilon);
-	EXPECT_NEAR(-63.3150724246548862, y(0,2), epsilon);
-	EXPECT_NEAR(+75.1693851132601338, y(0,3), epsilon);
-	EXPECT_NEAR( 0.00000000000000000, y(0,4), epsilon);
-	EXPECT_NEAR(+87.4243337392790067, y(0,5), epsilon);
-	EXPECT_NEAR(-114.092258182259556, y(0,6), epsilon);
-	EXPECT_NEAR(-94.010102920274619,  y(0,7), epsilon);
-	EXPECT_NEAR(-66.5754164410987244, y(0,8), epsilon);
-	EXPECT_NEAR(-109.483855254332568, y(0,9), epsilon);
-	EXPECT_NEAR(+27.5378037199449182, y(0,10), epsilon);
-	EXPECT_NEAR(-158.649815592855759, y(0,11), epsilon);
+	EXPECT_NEAR(51.3760082238439111, std::abs(y(0,0)), epsilon);
+	EXPECT_NEAR(56.4869601858493127, std::abs(y(0,1)), epsilon);
+	EXPECT_NEAR(63.3150724246548862, std::abs(y(0,2)), epsilon);
+	EXPECT_NEAR(75.1693851132601338, std::abs(y(0,3)), epsilon);
+	EXPECT_NEAR(0.00000000000000000, std::abs(y(0,4)), epsilon);
+	EXPECT_NEAR(87.4243337392790067, std::abs(y(0,5)), epsilon);
+	EXPECT_NEAR(114.092258182259556, std::abs(y(0,6)), epsilon);
+	EXPECT_NEAR(94.010102920274619,  std::abs(y(0,7)), epsilon);
+	EXPECT_NEAR(66.5754164410987244, std::abs(y(0,8)), epsilon);
+	EXPECT_NEAR(109.483855254332568, std::abs(y(0,9)), epsilon);
+	EXPECT_NEAR(27.5378037199449182, std::abs(y(0,10)), epsilon);
+	EXPECT_NEAR(158.649815592855759, std::abs(y(0,11)), epsilon);
 
 	// comparing eigenvectors from the transformation_matrix with that from the
 	// 'CannonVar' implementation.
 	SGMatrix<float64_t> transformy=fisherlda.get_transformation_matrix();
 	EXPECT_NEAR(0.576749097670396393, std::abs(transformy[0]), epsilon);
 	EXPECT_NEAR(0.158373354160231516, std::abs(transformy[1]), epsilon);
-	EXPECT_NEAR(0.47693388209865617, std::abs(transformy[2]), epsilon);
-	EXPECT_NEAR(0.64405933820939687, std::abs(transformy[3]), epsilon);
+	EXPECT_NEAR(0.47693388209865617,  std::abs(transformy[2]), epsilon);
+	EXPECT_NEAR(0.64405933820939687,  std::abs(transformy[3]), epsilon);
 
 	SG_UNREF(dense_feat);
 	SG_UNREF(labels);