Merge.

4 files changed, 51 insertions, 7 deletions

diff --git a/CMakeLists.txt b/CMakeLists.txt
index 5255e9600..48c0a6367 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -469,6 +469,8 @@ if(BUILD_TESTING)
   else()
     add_subdirectory(test EXCLUDE_FROM_ALL)
   endif()
+
+  add_subdirectory(failtest)
 endif()
 
 if(EIGEN_LEAVE_TEST_IN_ALL_TARGET)
@@ -519,8 +521,6 @@ message(STATUS "")
 message(STATUS "Configured Eigen ${EIGEN_VERSION_NUMBER}")
 message(STATUS "")
 
-add_subdirectory(failtest)
-
 string(TOLOWER "${CMAKE_GENERATOR}" cmake_generator_tolower)
 if(cmake_generator_tolower MATCHES "makefile")
   message(STATUS "Some things you can do now:")
@@ -537,8 +537,10 @@ if(cmake_generator_tolower MATCHES "makefile")
   message(STATUS "              |   Or:")
   message(STATUS "              |     cmake . -DINCLUDE_INSTALL_DIR=yourdir")
   message(STATUS "make doc      | Generate the API documentation, requires Doxygen & LaTeX")
-  message(STATUS "make check    | Build and run the unit-tests. Read this page:")
-  message(STATUS "              |   http://eigen.tuxfamily.org/index.php?title=Tests")
+  if(BUILD_TESTING)
+    message(STATUS "make check    | Build and run the unit-tests. Read this page:")
+    message(STATUS "              |   http://eigen.tuxfamily.org/index.php?title=Tests")
+  endif()
   message(STATUS "make blas     | Build BLAS library (not the same thing as Eigen)")
   message(STATUS "make uninstall| Removes files installed by make install")
   message(STATUS "--------------+--------------------------------------------------------------")
diff --git a/Eigen/src/Core/arch/AVX512/Complex.h b/Eigen/src/Core/arch/AVX512/Complex.h
index b7e68d2ab..f2034a713 100644
--- a/Eigen/src/Core/arch/AVX512/Complex.h
+++ b/Eigen/src/Core/arch/AVX512/Complex.h
@@ -406,12 +406,40 @@ template<> EIGEN_STRONG_INLINE Packet4cd pcplxflip<Packet4cd>(const Packet4cd& x
 
 EIGEN_DEVICE_FUNC inline void
 ptranspose(PacketBlock<Packet8cf,4>& kernel) {
-  ptranspose(reinterpret_cast<PacketBlock<Packet8d,4>&>(kernel));
+  PacketBlock<Packet8d,4> pb;
+  
+  pb.packet[0] = _mm512_castps_pd(kernel.packet[0].v);
+  pb.packet[1] = _mm512_castps_pd(kernel.packet[1].v);
+  pb.packet[2] = _mm512_castps_pd(kernel.packet[2].v);
+  pb.packet[3] = _mm512_castps_pd(kernel.packet[3].v);
+  ptranspose(pb);
+  kernel.packet[0].v = _mm512_castpd_ps(pb.packet[0]);
+  kernel.packet[1].v = _mm512_castpd_ps(pb.packet[1]);
+  kernel.packet[2].v = _mm512_castpd_ps(pb.packet[2]);
+  kernel.packet[3].v = _mm512_castpd_ps(pb.packet[3]);
 }
 
 EIGEN_DEVICE_FUNC inline void
 ptranspose(PacketBlock<Packet8cf,8>& kernel) {
-  ptranspose(reinterpret_cast<PacketBlock<Packet8d,8>&>(kernel));
+  PacketBlock<Packet8d,8> pb;
+  
+  pb.packet[0] = _mm512_castps_pd(kernel.packet[0].v);
+  pb.packet[1] = _mm512_castps_pd(kernel.packet[1].v);
+  pb.packet[2] = _mm512_castps_pd(kernel.packet[2].v);
+  pb.packet[3] = _mm512_castps_pd(kernel.packet[3].v);
+  pb.packet[4] = _mm512_castps_pd(kernel.packet[4].v);
+  pb.packet[5] = _mm512_castps_pd(kernel.packet[5].v);
+  pb.packet[6] = _mm512_castps_pd(kernel.packet[6].v);
+  pb.packet[7] = _mm512_castps_pd(kernel.packet[7].v);
+  ptranspose(pb);
+  kernel.packet[0].v = _mm512_castpd_ps(pb.packet[0]);
+  kernel.packet[1].v = _mm512_castpd_ps(pb.packet[1]);
+  kernel.packet[2].v = _mm512_castpd_ps(pb.packet[2]);
+  kernel.packet[3].v = _mm512_castpd_ps(pb.packet[3]);
+  kernel.packet[4].v = _mm512_castpd_ps(pb.packet[4]);
+  kernel.packet[5].v = _mm512_castpd_ps(pb.packet[5]);
+  kernel.packet[6].v = _mm512_castpd_ps(pb.packet[6]);
+  kernel.packet[7].v = _mm512_castpd_ps(pb.packet[7]);
 }
 
 EIGEN_DEVICE_FUNC inline void
diff --git a/Eigen/src/Core/arch/GPU/PacketMathHalf.h b/Eigen/src/Core/arch/GPU/PacketMathHalf.h
index 00e40d40b..85a32a18d 100644
--- a/Eigen/src/Core/arch/GPU/PacketMathHalf.h
+++ b/Eigen/src/Core/arch/GPU/PacketMathHalf.h
@@ -143,7 +143,11 @@ template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pabs<half2>(const half2&
   return result;
 }
 
+<<<<<<< working copy
 template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 ptrue<half2>(const half2& a) {
+=======
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE half2 pones<half2>(const half2& a) {
+>>>>>>> merge rev
   half2 result;
   *(reinterpret_cast<unsigned*>(&(result))) = 0xffffffffu;
 }  
@@ -648,8 +652,13 @@ template<> EIGEN_STRONG_INLINE Packet16h pnot(const Packet16h& a) {
   Packet16h r; r.x = _mm256_xor_si256(a.x, pcmp_eq(a.x, a.x)); return r;
 }
 
+<<<<<<< working copy
 template<> EIGEN_STRONG_INLINE Packet16h ptrue(const Packet16h& a) {
   Packet16h r; r.x = Packet8i(ptrue(a.x)); return r;
+=======
+template<> EIGEN_STRONG_INLINE Packet16h pones(const Packet16h& a) {
+  Packet16h r; r.x = Packet8i(pones(a.x)); return r;
+>>>>>>> merge rev
 }
 
 template<> EIGEN_STRONG_INLINE Packet16h por(const Packet16h& a,const Packet16h& b) {
@@ -1097,7 +1106,11 @@ EIGEN_STRONG_INLINE Packet8h float2half(const Packet8f& a) {
 #endif
 }
 
+<<<<<<< working copy
 template<> EIGEN_STRONG_INLINE Packet8h ptrue(const Packet8h& a) {
+=======
+template<> EIGEN_STRONG_INLINE Packet8h pones(const Packet8h& a) {
+>>>>>>> merge rev
   Packet8h r; r.x = _mm_cmpeq_epi32(a.x, a.x); return r;
 }
 
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
index 9666bf167..d68409e26 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
@@ -890,7 +890,8 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
 
       TENSOR_CONTRACTION_DISPATCH(
           this->template evalGemmPartialWithoutOutputKernel, Alignment,
-          (buf, begin, end, /*num_threads=*/num_blocks));
+          (buf, begin, end,
+           /*num_threads=*/internal::convert_index<int>(num_blocks)));
 
       // Check if it was the last task in l0 range.
       const Index l0_index = block_idx / l0_size;