@smake/eigen 1.0.1 → 1.1.0

package/eigen/Eigen/src/Core/arch/Default/Settings.h

@@ -21,7 +21,7 @@
   * it does not correspond to the number of iterations or the number of instructions
   */
 #ifndef EIGEN_UNROLLING_LIMIT
-#define EIGEN_UNROLLING_LIMIT 100
+#define EIGEN_UNROLLING_LIMIT 110
 #endif
 
 /** Defines the threshold between a "small" and a "large" matrix.

package/eigen/Eigen/src/Core/arch/Default/TypeCasting.h

@@ -0,0 +1,120 @@
+// This file is part of Eigen, a lightweight C++ template library
+// for linear algebra.
+//
+// Copyright (C) 2016 Benoit Steiner <benoit.steiner.goog@gmail.com>
+// Copyright (C) 2019 Rasmus Munk Larsen <rmlarsen@google.com>
+//
+// This Source Code Form is subject to the terms of the Mozilla
+// Public License v. 2.0. If a copy of the MPL was not distributed
+// with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
+
+#ifndef EIGEN_GENERIC_TYPE_CASTING_H
+#define EIGEN_GENERIC_TYPE_CASTING_H
+
+namespace Eigen {
+
+namespace internal {
+
+template<>
+struct scalar_cast_op<float, Eigen::half> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
+  typedef Eigen::half result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const float& a) const {
+    #if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
+      (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
+      return __float2half(a);
+    #else
+      return Eigen::half(a);
+    #endif
+  }
+};
+
+template<>
+struct functor_traits<scalar_cast_op<float, Eigen::half> >
+{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
+
+
+template<>
+struct scalar_cast_op<int, Eigen::half> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
+  typedef Eigen::half result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::half operator() (const int& a) const {
+    #if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
+      (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
+      return __float2half(static_cast<float>(a));
+    #else
+      return Eigen::half(static_cast<float>(a));
+    #endif
+  }
+};
+
+template<>
+struct functor_traits<scalar_cast_op<int, Eigen::half> >
+{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
+
+
+template<>
+struct scalar_cast_op<Eigen::half, float> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
+  typedef float result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float operator() (const Eigen::half& a) const {
+    #if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && EIGEN_CUDA_ARCH >= 300) || \
+      (defined(EIGEN_HAS_HIP_FP16) && defined(EIGEN_HIP_DEVICE_COMPILE))
+      return __half2float(a);
+    #else
+      return static_cast<float>(a);
+    #endif
+  }
+};
+
+template<>
+struct functor_traits<scalar_cast_op<Eigen::half, float> >
+{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
+
+
+template<>
+struct scalar_cast_op<float, Eigen::bfloat16> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
+  typedef Eigen::bfloat16 result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::bfloat16 operator() (const float& a) const {
+    return Eigen::bfloat16(a);
+  }
+};
+
+template<>
+struct functor_traits<scalar_cast_op<float, Eigen::bfloat16> >
+{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
+
+
+template<>
+struct scalar_cast_op<int, Eigen::bfloat16> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
+  typedef Eigen::bfloat16 result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Eigen::bfloat16 operator() (const int& a) const {
+    return Eigen::bfloat16(static_cast<float>(a));
+  }
+};
+
+template<>
+struct functor_traits<scalar_cast_op<int, Eigen::bfloat16> >
+{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
+
+
+template<>
+struct scalar_cast_op<Eigen::bfloat16, float> {
+  EIGEN_EMPTY_STRUCT_CTOR(scalar_cast_op)
+  typedef float result_type;
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE float operator() (const Eigen::bfloat16& a) const {
+    return static_cast<float>(a);
+  }
+};
+
+template<>
+struct functor_traits<scalar_cast_op<Eigen::bfloat16, float> >
+{ enum { Cost = NumTraits<float>::AddCost, PacketAccess = false }; };
+
+
+}
+}
+
+#endif  // EIGEN_GENERIC_TYPE_CASTING_H

package/eigen/Eigen/src/Core/arch/{CUDA → GPU}/MathFunctions.h

@@ -7,8 +7,8 @@
 // Public License v. 2.0. If a copy of the MPL was not distributed
 // with this file, You can obtain one at http://mozilla.org/MPL/2.0/.
 
-#ifndef EIGEN_MATH_FUNCTIONS_CUDA_H
-#define EIGEN_MATH_FUNCTIONS_CUDA_H
+#ifndef EIGEN_MATH_FUNCTIONS_GPU_H
+#define EIGEN_MATH_FUNCTIONS_GPU_H
 
 namespace Eigen {
 
@@ -17,7 +17,7 @@ namespace internal {
 // Make sure this is only available when targeting a GPU: we don't want to
 // introduce conflicts between these packet_traits definitions and the ones
 // we'll use on the host side (SSE, AVX, ...)
-#if defined(__CUDACC__) && defined(EIGEN_USE_GPU)
+#if defined(EIGEN_GPUCC) && defined(EIGEN_USE_GPU)
 template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
 float4 plog<float4>(const float4& a)
 {
@@ -56,6 +56,18 @@ double2 pexp<double2>(const double2& a)
   return make_double2(exp(a.x), exp(a.y));
 }
 
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+float4 pexpm1<float4>(const float4& a)
+{
+  return make_float4(expm1f(a.x), expm1f(a.y), expm1f(a.z), expm1f(a.w));
+}
+
+template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
+double2 pexpm1<double2>(const double2& a)
+{
+  return make_double2(expm1(a.x), expm1(a.y));
+}
+
 template<> EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE
 float4 psqrt<float4>(const float4& a)
 {
@@ -88,4 +100,4 @@ double2 prsqrt<double2>(const double2& a)
 
 } // end namespace Eigen
 
-#endif // EIGEN_MATH_FUNCTIONS_CUDA_H
+#endif // EIGEN_MATH_FUNCTIONS_GPU_H