Upgraded the bundled Eigen to 3.7

src/eigen/Eigen/Core

@@ -53,9 +53,9 @@
     #endif
 
     #define EIGEN_DEVICE_FUNC __host__ __device__
-    // We need math_functions.hpp to ensure that that EIGEN_USING_STD_MATH macro
+    // We need cuda_runtime.h to ensure that that EIGEN_USING_STD_MATH macro
     // works properly on the device side
-    #include <math_functions.hpp>
+    #include <cuda_runtime.h>
   #else
     #define EIGEN_DEVICE_FUNC
   #endif

src/eigen/Eigen/src/Cholesky/LDLT.h

@@ -305,7 +305,8 @@ template<> struct ldlt_inplace<Lower>
     if (size <= 1)
     {
       transpositions.setIdentity();
-      if (numext::real(mat.coeff(0,0)) > static_cast<RealScalar>(0) ) sign = PositiveSemiDef;
+      if(size==0) sign = ZeroSign;
+      else if (numext::real(mat.coeff(0,0)) > static_cast<RealScalar>(0) ) sign = PositiveSemiDef;
       else if (numext::real(mat.coeff(0,0)) < static_cast<RealScalar>(0)) sign = NegativeSemiDef;
       else sign = ZeroSign;
       return true;

src/eigen/Eigen/src/Core/Array.h

@@ -153,8 +153,6 @@ class Array
       : Base(std::move(other))
     {
       Base::_check_template_params();
-      if (RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic)
-        Base::_set_noalias(other);
     }
     EIGEN_DEVICE_FUNC
     Array& operator=(Array&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_assignable<Scalar>::value)

src/eigen/Eigen/src/Core/ConditionEstimator.h

@@ -160,7 +160,7 @@ rcond_estimate_helper(typename Decomposition::RealScalar matrix_norm, const Deco
 {
   typedef typename Decomposition::RealScalar RealScalar;
   eigen_assert(dec.rows() == dec.cols());
-  if (dec.rows() == 0)              return RealScalar(1);
+  if (dec.rows() == 0)              return NumTraits<RealScalar>::infinity();
   if (matrix_norm == RealScalar(0)) return RealScalar(0);
   if (dec.rows() == 1)              return RealScalar(1);
   const RealScalar inverse_matrix_norm = rcond_invmatrix_L1_norm_estimate(dec);

src/eigen/Eigen/src/Core/MapBase.h

@@ -43,6 +43,7 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
     enum {
       RowsAtCompileTime = internal::traits<Derived>::RowsAtCompileTime,
       ColsAtCompileTime = internal::traits<Derived>::ColsAtCompileTime,
+      InnerStrideAtCompileTime = internal::traits<Derived>::InnerStrideAtCompileTime,
       SizeAtCompileTime = Base::SizeAtCompileTime
     };
 
@@ -187,8 +188,11 @@ template<typename Derived> class MapBase<Derived, ReadOnlyAccessors>
     void checkSanity(typename internal::enable_if<(internal::traits<T>::Alignment>0),void*>::type = 0) const
     {
 #if EIGEN_MAX_ALIGN_BYTES>0
+      // innerStride() is not set yet when this function is called, so we optimistically assume the lowest plausible value:
+      const Index minInnerStride = InnerStrideAtCompileTime == Dynamic ? 1 : Index(InnerStrideAtCompileTime);
+      EIGEN_ONLY_USED_FOR_DEBUG(minInnerStride);
       eigen_assert((   ((internal::UIntPtr(m_data) % internal::traits<Derived>::Alignment) == 0)
-                    || (cols() * rows() * innerStride() * sizeof(Scalar)) < internal::traits<Derived>::Alignment ) && "data is not aligned");
+                    || (cols() * rows() * minInnerStride * sizeof(Scalar)) < internal::traits<Derived>::Alignment ) && "data is not aligned");
 #endif
     }
 

src/eigen/Eigen/src/Core/MathFunctions.h

@@ -616,21 +616,28 @@ template<typename Scalar>
 struct random_default_impl<Scalar, false, true>
 {
   static inline Scalar run(const Scalar& x, const Scalar& y)
-  { 
-    typedef typename conditional<NumTraits<Scalar>::IsSigned,std::ptrdiff_t,std::size_t>::type ScalarX;
-    if(y<x)
+  {
+    if (y <= x)
       return x;
-    // the following difference might overflow on a 32 bits system,
-    // but since y>=x the result converted to an unsigned long is still correct.
-    std::size_t range = ScalarX(y)-ScalarX(x);
-    std::size_t offset = 0;
-    // rejection sampling
-    std::size_t divisor = 1;
-    std::size_t multiplier = 1;
-    if(range<RAND_MAX) divisor = (std::size_t(RAND_MAX)+1)/(range+1);
-    else               multiplier = 1 + range/(std::size_t(RAND_MAX)+1);
+    // ScalarU is the unsigned counterpart of Scalar, possibly Scalar itself.
+    typedef typename make_unsigned<Scalar>::type ScalarU;
+    // ScalarX is the widest of ScalarU and unsigned int.
+    // We'll deal only with ScalarX and unsigned int below thus avoiding signed
+    // types and arithmetic and signed overflows (which are undefined behavior).
+    typedef typename conditional<(ScalarU(-1) > unsigned(-1)), ScalarU, unsigned>::type ScalarX;
+    // The following difference doesn't overflow, provided our integer types are two's
+    // complement and have the same number of padding bits in signed and unsigned variants.
+    // This is the case in most modern implementations of C++.
+    ScalarX range = ScalarX(y) - ScalarX(x);
+    ScalarX offset = 0;
+    ScalarX divisor = 1;
+    ScalarX multiplier = 1;
+    const unsigned rand_max = RAND_MAX;
+    if (range <= rand_max) divisor = (rand_max + 1) / (range + 1);
+    else                   multiplier = 1 + range / (rand_max + 1);
+    // Rejection sampling.
     do {
-      offset = (std::size_t(std::rand()) * multiplier) / divisor;
+      offset = (unsigned(std::rand()) * multiplier) / divisor;
     } while (offset > range);
     return Scalar(ScalarX(x) + offset);
   }
@@ -1006,7 +1013,8 @@ inline int log2(int x)
 
 /** \returns the square root of \a x.
   *
-  * It is essentially equivalent to \code using std::sqrt; return sqrt(x); \endcode,
+  * It is essentially equivalent to
+  * \code using std::sqrt; return sqrt(x); \endcode
   * but slightly faster for float/double and some compilers (e.g., gcc), thanks to
   * specializations when SSE is enabled.
   *

src/eigen/Eigen/src/Core/Matrix.h

@@ -274,8 +274,6 @@ class Matrix
       : Base(std::move(other))
     {
       Base::_check_template_params();
-      if (RowsAtCompileTime!=Dynamic && ColsAtCompileTime!=Dynamic)
-        Base::_set_noalias(other);
     }
     EIGEN_DEVICE_FUNC
     Matrix& operator=(Matrix&& other) EIGEN_NOEXCEPT_IF(std::is_nothrow_move_assignable<Scalar>::value)

src/eigen/Eigen/src/Core/MatrixBase.h

@@ -444,16 +444,24 @@ template<typename Derived> class MatrixBase
 ///////// MatrixFunctions module /////////
 
     typedef typename internal::stem_function<Scalar>::type StemFunction;
-    const MatrixExponentialReturnValue<Derived> exp() const;
+#define EIGEN_MATRIX_FUNCTION(ReturnType, Name, Description) \
+    /** \returns an expression of the matrix Description of \c *this. \brief This function requires the <a href="unsupported/group__MatrixFunctions__Module.html"> unsupported MatrixFunctions module</a>. To compute the coefficient-wise Description use ArrayBase::##Name . */ \
+    const ReturnType<Derived> Name() const;
+#define EIGEN_MATRIX_FUNCTION_1(ReturnType, Name, Description, Argument) \
+    /** \returns an expression of the matrix Description of \c *this. \brief This function requires the <a href="unsupported/group__MatrixFunctions__Module.html"> unsupported MatrixFunctions module</a>. To compute the coefficient-wise Description use ArrayBase::##Name . */ \
+    const ReturnType<Derived> Name(Argument) const;
+
+    EIGEN_MATRIX_FUNCTION(MatrixExponentialReturnValue, exp, exponential)
+    /** \brief Helper function for the <a href="unsupported/group__MatrixFunctions__Module.html"> unsupported MatrixFunctions module</a>.*/
     const MatrixFunctionReturnValue<Derived> matrixFunction(StemFunction f) const;
-    const MatrixFunctionReturnValue<Derived> cosh() const;
-    const MatrixFunctionReturnValue<Derived> sinh() const;
-    const MatrixFunctionReturnValue<Derived> cos() const;
-    const MatrixFunctionReturnValue<Derived> sin() const;
-    const MatrixSquareRootReturnValue<Derived> sqrt() const;
-    const MatrixLogarithmReturnValue<Derived> log() const;
-    const MatrixPowerReturnValue<Derived> pow(const RealScalar& p) const;
-    const MatrixComplexPowerReturnValue<Derived> pow(const std::complex<RealScalar>& p) const;
+    EIGEN_MATRIX_FUNCTION(MatrixFunctionReturnValue, cosh, hyperbolic cosine)
+    EIGEN_MATRIX_FUNCTION(MatrixFunctionReturnValue, sinh, hyperbolic sine)
+    EIGEN_MATRIX_FUNCTION(MatrixFunctionReturnValue, cos, cosine)
+    EIGEN_MATRIX_FUNCTION(MatrixFunctionReturnValue, sin, sine)
+    EIGEN_MATRIX_FUNCTION(MatrixSquareRootReturnValue, sqrt, square root)
+    EIGEN_MATRIX_FUNCTION(MatrixLogarithmReturnValue, log, logarithm)
+    EIGEN_MATRIX_FUNCTION_1(MatrixPowerReturnValue,        pow, power to \c p, const RealScalar& p)
+    EIGEN_MATRIX_FUNCTION_1(MatrixComplexPowerReturnValue, pow, power to \c p, const std::complex<RealScalar>& p)
 
   protected:
     EIGEN_DEVICE_FUNC MatrixBase() : Base() {}

src/eigen/Eigen/src/Core/SolveTriangular.h

@@ -169,6 +169,9 @@ void TriangularViewImpl<MatrixType,Mode,Dense>::solveInPlace(const MatrixBase<Ot
   OtherDerived& other = _other.const_cast_derived();
   eigen_assert( derived().cols() == derived().rows() && ((Side==OnTheLeft && derived().cols() == other.rows()) || (Side==OnTheRight && derived().cols() == other.cols())) );
   eigen_assert((!(Mode & ZeroDiag)) && bool(Mode & (Upper|Lower)));
+  // If solving for a 0x0 matrix, nothing to do, simply return.
+  if (derived().cols() == 0)
+    return;
 
   enum { copy = (internal::traits<OtherDerived>::Flags & RowMajorBit)  && OtherDerived::IsVectorAtCompileTime && OtherDerived::SizeAtCompileTime!=1};
   typedef typename internal::conditional<copy,

src/eigen/Eigen/src/Core/arch/AVX/PacketMath.h

@@ -159,11 +159,12 @@ template<> EIGEN_STRONG_INLINE Packet8i pdiv<Packet8i>(const Packet8i& /*a*/, co
 
 #ifdef __FMA__
 template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f& b, const Packet8f& c) {
-#if ( EIGEN_COMP_GNUC_STRICT || (EIGEN_COMP_CLANG && (EIGEN_COMP_CLANG<308)) )
-  // clang stupidly generates a vfmadd213ps instruction plus some vmovaps on registers,
-  // and gcc stupidly generates a vfmadd132ps instruction,
-  // so let's enforce it to generate a vfmadd231ps instruction since the most common use case is to accumulate
-  // the result of the product.
+#if ( (EIGEN_COMP_GNUC_STRICT && EIGEN_COMP_GNUC<80) || (EIGEN_COMP_CLANG) )
+  // Clang stupidly generates a vfmadd213ps instruction plus some vmovaps on registers,
+  //  and even register spilling with clang>=6.0 (bug 1637).
+  // Gcc stupidly generates a vfmadd132ps instruction.
+  // So let's enforce it to generate a vfmadd231ps instruction since the most common use
+  //  case is to accumulate the result of the product.
   Packet8f res = c;
   __asm__("vfmadd231ps %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b));
   return res;
@@ -172,7 +173,7 @@ template<> EIGEN_STRONG_INLINE Packet8f pmadd(const Packet8f& a, const Packet8f&
 #endif
 }
 template<> EIGEN_STRONG_INLINE Packet4d pmadd(const Packet4d& a, const Packet4d& b, const Packet4d& c) {
-#if ( EIGEN_COMP_GNUC_STRICT || (EIGEN_COMP_CLANG && (EIGEN_COMP_CLANG<308)) )
+#if ( (EIGEN_COMP_GNUC_STRICT && EIGEN_COMP_GNUC<80) || (EIGEN_COMP_CLANG) )
   // see above
   Packet4d res = c;
   __asm__("vfmadd231pd %[a], %[b], %[c]" : [c] "+x" (res) : [a] "x" (a), [b] "x" (b));