Fix NVCC+ICC issues.

NVCC does not understand `__forceinline`, so we need to use `inline`
when compiling for GPU.

ICC specializes `std::complex` operators for `float` and `double`
by default, which cannot be used on device and conflict with Eigen's
workaround in CUDA/Complex.h.  This can be prevented by defining
`_OVERRIDE_COMPLEX_SPECIALIZATION_` before including `<complex>`.
Added this define to the tests and to `Eigen/Core`, but this will
not work if the user includes `<complex>` before `<Eigen/Core>`.

ICC also seems to generate a duplicate `Map` symbol in
`PlainObjectBase`:
```
error: "Map" has already been declared in the current scope
  static ConstMapType Map(const Scalar *data)

```
I tracked this down to `friend class Eigen::Map`.  Putting the `friend`
statements at the bottom of the class seems to resolve this issue.

Fixes #2180

5 files changed, 35 insertions, 13 deletions

diff --git a/Eigen/Core b/Eigen/Core
index 1a60dcba4..5921e15f9 100644
--- a/Eigen/Core
+++ b/Eigen/Core
@@ -40,6 +40,13 @@
   #pragma GCC optimize ("-fno-ipa-cp-clone")
 #endif
 
+// Prevent ICC from specializing std::complex operators that silently fail
+// on device. This allows us to use our own device-compatible specializations
+// instead.
+#if defined(EIGEN_COMP_ICC) && defined(EIGEN_GPU_COMPILE_PHASE) \
+    && !defined(_OVERRIDE_COMPLEX_SPECIALIZATION_)
+#define _OVERRIDE_COMPLEX_SPECIALIZATION_ 1
+#endif
 #include <complex>
 
 // this include file manages BLAS and MKL related macros
diff --git a/Eigen/src/Core/PlainObjectBase.h b/Eigen/src/Core/PlainObjectBase.h
index ca5b5ee1d..e53ca1b16 100644
--- a/Eigen/src/Core/PlainObjectBase.h
+++ b/Eigen/src/Core/PlainObjectBase.h
@@ -118,16 +118,8 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
     using Base::IsVectorAtCompileTime;
     using Base::Flags;
 
-    template<typename PlainObjectType, int MapOptions, typename StrideType> friend class Eigen::Map;
-    friend  class Eigen::Map<Derived, Unaligned>;
     typedef Eigen::Map<Derived, Unaligned>  MapType;
-    friend  class Eigen::Map<const Derived, Unaligned>;
     typedef const Eigen::Map<const Derived, Unaligned> ConstMapType;
-#if EIGEN_MAX_ALIGN_BYTES>0
-    // for EIGEN_MAX_ALIGN_BYTES==0, AlignedMax==Unaligned, and many compilers generate warnings for friend-ing a class twice.
-    friend  class Eigen::Map<Derived, AlignedMax>;
-    friend  class Eigen::Map<const Derived, AlignedMax>;
-#endif
     typedef Eigen::Map<Derived, AlignedMax> AlignedMapType;
     typedef const Eigen::Map<const Derived, AlignedMax> ConstAlignedMapType;
     template<typename StrideType> struct StridedMapType { typedef Eigen::Map<Derived, Unaligned, StrideType> type; };
@@ -989,6 +981,17 @@ class PlainObjectBase : public internal::dense_xpr_base<Derived>::type
 
     enum { IsPlainObjectBase = 1 };
 #endif
+  public:
+    // These apparently need to be down here for nvcc+icc to prevent duplicate
+    // Map symbol.
+    template<typename PlainObjectType, int MapOptions, typename StrideType> friend class Eigen::Map;
+    friend class Eigen::Map<Derived, Unaligned>;
+    friend class Eigen::Map<const Derived, Unaligned>;
+#if EIGEN_MAX_ALIGN_BYTES>0
+    // for EIGEN_MAX_ALIGN_BYTES==0, AlignedMax==Unaligned, and many compilers generate warnings for friend-ing a class twice.
+    friend class Eigen::Map<Derived, AlignedMax>;
+    friend class Eigen::Map<const Derived, AlignedMax>;
+#endif
 };
 
 namespace internal {
diff --git a/Eigen/src/Core/arch/CUDA/Complex.h b/Eigen/src/Core/arch/CUDA/Complex.h
index caf3fe74b..b1618e567 100644
--- a/Eigen/src/Core/arch/CUDA/Complex.h
+++ b/Eigen/src/Core/arch/CUDA/Complex.h
@@ -12,9 +12,6 @@
 #define EIGEN_COMPLEX_CUDA_H
 
 // clang-format off
-
-#if defined(EIGEN_CUDACC) && defined(EIGEN_GPU_COMPILE_PHASE)
-
 // Many std::complex methods such as operator+, operator-, operator* and
 // operator/ are not constexpr. Due to this, GCC and older versions of clang do
 // not treat them as device functions and thus Eigen functors making use of
@@ -22,6 +19,17 @@
 // operators and functors for complex types when building for CUDA to enable
 // their use on-device.
 
+#if defined(EIGEN_CUDACC) && defined(EIGEN_GPU_COMPILE_PHASE)
+    
+// ICC already specializes std::complex<float> and std::complex<double>
+// operators, preventing us from making them device functions here.
+// This will lead to silent runtime errors if the operators are used on device.
+//
+// To allow std::complex operator use on device, define _OVERRIDE_COMPLEX_SPECIALIZATION_
+// prior to first inclusion of <complex>.  This prevents ICC from adding
+// its own specializations, so our custom ones below can be used instead.
+#if !(defined(EIGEN_COMP_ICC) && defined(_USE_COMPLEX_SPECIALIZATION_))
+
 // Import Eigen's internal operator specializations.
 #define EIGEN_USING_STD_COMPLEX_OPERATORS           \
   using Eigen::complex_operator_detail::operator+;  \
@@ -244,6 +252,8 @@ EIGEN_USING_STD_COMPLEX_OPERATORS
 }  // namespace internal
 }  // namespace Eigen
 
-#endif
+#endif  // !(EIGEN_COMP_ICC && _USE_COMPLEX_SPECIALIZATION_)
+
+#endif  // EIGEN_CUDACC && EIGEN_GPU_COMPILE_PHASE
 
 #endif  // EIGEN_COMPLEX_CUDA_H
diff --git a/Eigen/src/Core/util/Macros.h b/Eigen/src/Core/util/Macros.h
index 00f3690d1..bdc0de0ea 100644
--- a/Eigen/src/Core/util/Macros.h
+++ b/Eigen/src/Core/util/Macros.h
@@ -905,7 +905,7 @@
 // but it still doesn't use GCC's always_inline. This is useful in (common) situations where MSVC needs forceinline
 // but GCC is still doing fine with just inline.
 #ifndef EIGEN_STRONG_INLINE
-#if EIGEN_COMP_MSVC || EIGEN_COMP_ICC
+#if (EIGEN_COMP_MSVC || EIGEN_COMP_ICC) && !defined(EIGEN_GPUCC)
 #define EIGEN_STRONG_INLINE __forceinline
 #else
 #define EIGEN_STRONG_INLINE inline
diff --git a/test/main.h b/test/main.h
index cf061730f..3e80c9f7d 100644
--- a/test/main.h
+++ b/test/main.h
@@ -40,6 +40,8 @@
 // definitions.
 #include <limits>
 #include <algorithm>
+// Disable ICC's std::complex operator specializations so we can use our own.
+#define _OVERRIDE_COMPLEX_SPECIALIZATION_ 1
 #include <complex>
 #include <deque>
 #include <queue>