Merge.

3 files changed, 40 insertions, 13 deletions

diff --git a/unsupported/test/CMakeLists.txt b/unsupported/test/CMakeLists.txt
index c6a92fe73..c088df1c1 100644
--- a/unsupported/test/CMakeLists.txt
+++ b/unsupported/test/CMakeLists.txt
@@ -190,6 +190,10 @@ if(CUDA_FOUND AND EIGEN_TEST_CUDA)
   if("${CMAKE_CXX_COMPILER_ID}" STREQUAL "Clang")
     set(CUDA_NVCC_FLAGS "-ccbin /usr/bin/clang" CACHE STRING "nvcc flags" FORCE)
   endif()
+  if(EIGEN_TEST_CUDA_CLANG)
+   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -std=c++11 --cuda-gpu-arch=sm_${EIGEN_CUDA_COMPUTE_ARCH}")
+  endif()
+
   set(CUDA_NVCC_FLAGS "-std=c++11 --relaxed-constexpr -arch compute_${EIGEN_CUDA_COMPUTE_ARCH} -Xcudafe \"--display_error_number\"")
   cuda_include_directories("${CMAKE_CURRENT_BINARY_DIR}" "${CUDA_TOOLKIT_ROOT_DIR}/include")
   set(EIGEN_ADD_TEST_FILENAME_EXTENSION "cu")
@@ -206,10 +210,7 @@ if(CUDA_FOUND AND EIGEN_TEST_CUDA)
     ei_add_test(cxx11_tensor_random_cuda)
   endif()
 
-  # Operations other that casting of half floats are only supported starting with arch 5.3
-  if (${EIGEN_CUDA_COMPUTE_ARCH} GREATER 52)
-    ei_add_test(cxx11_tensor_of_float16_cuda)
-  endif()
+  ei_add_test(cxx11_tensor_of_float16_cuda)
 
   unset(EIGEN_ADD_TEST_FILENAME_EXTENSION)
 endif()
diff --git a/unsupported/test/NonLinearOptimization.cpp b/unsupported/test/NonLinearOptimization.cpp
index 724ea7b5b..6a5ed057f 100644
--- a/unsupported/test/NonLinearOptimization.cpp
+++ b/unsupported/test/NonLinearOptimization.cpp
@@ -14,6 +14,9 @@
 
 using std::sqrt;
 
+// tolerance for chekcing number of iterations
+#define LM_EVAL_COUNT_TOL 4/3
+
 int fcn_chkder(const VectorXd &x, VectorXd &fvec, MatrixXd &fjac, int iflag)
 {
     /*      subroutine fcn for chkder example. */
@@ -1023,7 +1026,8 @@ void testNistLanczos1(void)
   VERIFY_IS_EQUAL(lm.njev, 72);
   // check norm^2
   std::cout.precision(30);
-  VERIFY_IS_APPROX(lm.fvec.squaredNorm(), 1.4290986055242372e-25);  // should be 1.4307867721E-25, but nist results are on 128-bit floats
+  std::cout << lm.fvec.squaredNorm() << "\n";
+  VERIFY(lm.fvec.squaredNorm() <= 1.4307867721E-25);
   // check x
   VERIFY_IS_APPROX(x[0], 9.5100000027E-02);
   VERIFY_IS_APPROX(x[1], 1.0000000001E+00);
@@ -1044,7 +1048,7 @@ void testNistLanczos1(void)
   VERIFY_IS_EQUAL(lm.nfev, 9);
   VERIFY_IS_EQUAL(lm.njev, 8);
   // check norm^2
-  VERIFY_IS_APPROX(lm.fvec.squaredNorm(), 1.430571737783119393e-25);  // should be 1.4307867721E-25, but nist results are on 128-bit floats
+  VERIFY(lm.fvec.squaredNorm() <= 1.4307867721E-25);
   // check x
   VERIFY_IS_APPROX(x[0], 9.5100000027E-02);
   VERIFY_IS_APPROX(x[1], 1.0000000001E+00);
@@ -1354,8 +1358,12 @@ void testNistMGH17(void)
   
   // check return value
   VERIFY_IS_EQUAL(info, 2); 
-  VERIFY(lm.nfev < 650);  // 602
-  VERIFY(lm.njev < 600);  // 545
+  ++g_test_level;
+  VERIFY_IS_EQUAL(lm.nfev, 602);  // 602
+  VERIFY_IS_EQUAL(lm.njev, 545);  // 545
+  --g_test_level;
+  VERIFY(lm.nfev < 602 * LM_EVAL_COUNT_TOL);
+  VERIFY(lm.njev < 545 * LM_EVAL_COUNT_TOL);
 
   /*
    * Second try
diff --git a/unsupported/test/levenberg_marquardt.cpp b/unsupported/test/levenberg_marquardt.cpp
index a2bdb99e4..6dc17bd17 100644
--- a/unsupported/test/levenberg_marquardt.cpp
+++ b/unsupported/test/levenberg_marquardt.cpp
@@ -23,6 +23,9 @@
 
 using std::sqrt;
 
+// tolerance for chekcing number of iterations
+#define LM_EVAL_COUNT_TOL 4/3
+
 struct lmder_functor : DenseFunctor<double>
 {
     lmder_functor(void): DenseFunctor<double>(3,15) {}
@@ -631,7 +634,7 @@ void testNistLanczos1(void)
   VERIFY_IS_EQUAL(lm.nfev(), 79);
   VERIFY_IS_EQUAL(lm.njev(), 72);
   // check norm^2
-//   VERIFY_IS_APPROX(lm.fvec().squaredNorm(), 1.430899764097e-25);  // should be 1.4307867721E-25, but nist results are on 128-bit floats
+  VERIFY(lm.fvec().squaredNorm() <= 1.4307867721E-25);
   // check x
   VERIFY_IS_APPROX(x[0], 9.5100000027E-02);
   VERIFY_IS_APPROX(x[1], 1.0000000001E+00);
@@ -652,7 +655,7 @@ void testNistLanczos1(void)
   VERIFY_IS_EQUAL(lm.nfev(), 9);
   VERIFY_IS_EQUAL(lm.njev(), 8);
   // check norm^2
-//   VERIFY_IS_APPROX(lm.fvec().squaredNorm(), 1.428595533845e-25);  // should be 1.4307867721E-25, but nist results are on 128-bit floats
+  VERIFY(lm.fvec().squaredNorm() <= 1.4307867721E-25);
   // check x
   VERIFY_IS_APPROX(x[0], 9.5100000027E-02);
   VERIFY_IS_APPROX(x[1], 1.0000000001E+00);
@@ -789,7 +792,8 @@ void testNistMGH10(void)
   MGH10_functor functor;
   LevenbergMarquardt<MGH10_functor> lm(functor);
   info = lm.minimize(x);
-  VERIFY_IS_EQUAL(info, LevenbergMarquardtSpace::RelativeErrorTooSmall);
+  VERIFY_IS_EQUAL(info, LevenbergMarquardtSpace::RelativeReductionTooSmall);
+  // was: VERIFY_IS_EQUAL(info, 1);
 
   // check norm^2
   VERIFY_IS_APPROX(lm.fvec().squaredNorm(), 8.7945855171E+01);
@@ -799,9 +803,13 @@ void testNistMGH10(void)
   VERIFY_IS_APPROX(x[2], 3.4522363462E+02);
   
   // check return value
-  //VERIFY_IS_EQUAL(info, 1);
+
+  ++g_test_level;
   VERIFY_IS_EQUAL(lm.nfev(), 284 );
   VERIFY_IS_EQUAL(lm.njev(), 249 );
+  --g_test_level;
+  VERIFY(lm.nfev() < 284 * LM_EVAL_COUNT_TOL);
+  VERIFY(lm.njev() < 249 * LM_EVAL_COUNT_TOL);
 
   /*
    * Second try
@@ -809,7 +817,10 @@ void testNistMGH10(void)
   x<< 0.02, 4000., 250.;
   // do the computation
   info = lm.minimize(x);
+  ++g_test_level;
   VERIFY_IS_EQUAL(info, LevenbergMarquardtSpace::RelativeReductionTooSmall);
+  // was: VERIFY_IS_EQUAL(info, 1);
+  --g_test_level;
 
   // check norm^2
   VERIFY_IS_APPROX(lm.fvec().squaredNorm(), 8.7945855171E+01);
@@ -819,9 +830,12 @@ void testNistMGH10(void)
   VERIFY_IS_APPROX(x[2], 3.4522363462E+02);
   
   // check return value
-  //VERIFY_IS_EQUAL(info, 1);
+  ++g_test_level;
   VERIFY_IS_EQUAL(lm.nfev(), 126);
   VERIFY_IS_EQUAL(lm.njev(), 116);
+  --g_test_level;
+  VERIFY(lm.nfev() < 126 * LM_EVAL_COUNT_TOL);
+  VERIFY(lm.njev() < 116 * LM_EVAL_COUNT_TOL);
 }
 
 
@@ -896,8 +910,12 @@ void testNistBoxBOD(void)
 
   // check return value
   VERIFY_IS_EQUAL(info, 1); 
+  ++g_test_level;
   VERIFY_IS_EQUAL(lm.nfev(), 16 );
   VERIFY_IS_EQUAL(lm.njev(), 15 );
+  --g_test_level;
+  VERIFY(lm.nfev() < 16 * LM_EVAL_COUNT_TOL);
+  VERIFY(lm.njev() < 15 * LM_EVAL_COUNT_TOL);
   // check norm^2
   VERIFY_IS_APPROX(lm.fvec().squaredNorm(), 1.1680088766E+03);
   // check x