From aa33446dace833fbf06632e586c80119b3d8ac11 Mon Sep 17 00:00:00 2001
From: Benoit Steiner <benoit.steiner.goog@gmail.com>
Date: Thu, 9 Jun 2016 08:22:27 -0700
Subject: Improved support for vectorization of 16-bit floats

---
 .../Eigen/CXX11/src/Tensor/TensorFunctors.h        | 86 ++++++++++++++++++++++
 unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h    |  5 ++
 .../Eigen/CXX11/src/Tensor/TensorReductionCuda.h   |  8 +-
 3 files changed, 95 insertions(+), 4 deletions(-)

(limited to 'unsupported/Eigen')

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
index 3dd32e9d1..bf52e490f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorFunctors.h
@@ -84,6 +84,14 @@ struct functor_traits<scalar_sigmoid_op<T> > {
 };
 
 
+template<typename Reducer, typename Device>
+struct reducer_traits {
+  enum {
+    Cost = 1,
+    PacketAccess = false
+  };
+};
+
 // Standard reduction functors
 template <typename T> struct SumReducer
 {
@@ -119,6 +127,15 @@ template <typename T> struct SumReducer
   }
 };
 
+template <typename T, typename Device>
+struct reducer_traits<SumReducer<T>, Device> {
+  enum {
+    Cost = NumTraits<T>::AddCost,
+    PacketAccess = PacketType<T, Device>::type::HasAdd
+  };
+};
+
+
 template <typename T> struct MeanReducer
 {
   static const bool PacketAccess = packet_traits<T>::HasAdd && !NumTraits<T>::IsInteger;
@@ -162,6 +179,15 @@ template <typename T> struct MeanReducer
     DenseIndex packetCount_;
 };
 
+template <typename T, typename Device>
+struct reducer_traits<MeanReducer<T>, Device> {
+  enum {
+    Cost = NumTraits<T>::AddCost,
+    PacketAccess = PacketType<T, Device>::type::HasAdd
+  };
+};
+
+
 template <typename T> struct MaxReducer
 {
   static const bool PacketAccess = packet_traits<T>::HasMax;
@@ -195,6 +221,15 @@ template <typename T> struct MaxReducer
   }
 };
 
+template <typename T, typename Device>
+struct reducer_traits<MaxReducer<T>, Device> {
+  enum {
+    Cost = NumTraits<T>::AddCost,
+    PacketAccess = PacketType<T, Device>::type::HasMax
+  };
+};
+
+
 template <typename T> struct MinReducer
 {
   static const bool PacketAccess = packet_traits<T>::HasMin;
@@ -228,6 +263,14 @@ template <typename T> struct MinReducer
   }
 };
 
+template <typename T, typename Device>
+struct reducer_traits<MinReducer<T>, Device> {
+  enum {
+    Cost = NumTraits<T>::AddCost,
+    PacketAccess = PacketType<T, Device>::type::HasMin
+  };
+};
+
 
 template <typename T> struct ProdReducer
 {
@@ -263,6 +306,14 @@ template <typename T> struct ProdReducer
   }
 };
 
+template <typename T, typename Device>
+struct reducer_traits<ProdReducer<T>, Device> {
+  enum {
+    Cost = NumTraits<T>::MulCost,
+    PacketAccess = PacketType<T, Device>::type::HasMul
+  };
+};
+
 
 struct AndReducer
 {
@@ -280,6 +331,15 @@ struct AndReducer
   }
 };
 
+template <typename Device>
+struct reducer_traits<AndReducer, Device> {
+  enum {
+    Cost = 1,
+    PacketAccess = false
+  };
+};
+
+
 struct OrReducer {
   static const bool PacketAccess = false;
   static const bool IsStateful = false;
@@ -295,6 +355,15 @@ struct OrReducer {
   }
 };
 
+template <typename Device>
+struct reducer_traits<OrReducer, Device> {
+  enum {
+    Cost = 1,
+    PacketAccess = false
+  };
+};
+
+
 // Argmin/Argmax reducers
 template <typename T> struct ArgMaxTupleReducer
 {
@@ -312,6 +381,15 @@ template <typename T> struct ArgMaxTupleReducer
   }
 };
 
+template <typename T, typename Device>
+struct reducer_traits<ArgMaxTupleReducer<T>, Device> {
+  enum {
+    Cost = NumTraits<T>::AddCost,
+    PacketAccess = false
+  };
+};
+
+
 template <typename T> struct ArgMinTupleReducer
 {
   static const bool PacketAccess = false;
@@ -328,6 +406,14 @@ template <typename T> struct ArgMinTupleReducer
   }
 };
 
+template <typename T, typename Device>
+struct reducer_traits<ArgMinTupleReducer<T>, Device> {
+  enum {
+    Cost = NumTraits<T>::AddCost,
+    PacketAccess = false
+  };
+};
+
 
 // Random number generation
 namespace {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h b/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
index b1645d56f..82a905a65 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMeta.h
@@ -54,6 +54,11 @@ struct PacketType {
 
 // For CUDA packet types when using a GpuDevice
 #if defined(EIGEN_USE_GPU) && defined(__CUDACC__)
+template <>
+  struct PacketType<half, GpuDevice> {
+  typedef half2 type;
+  static const int size = 2;
+ };
 template <>
 struct PacketType<float, GpuDevice> {
   typedef float4 type;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
index e82530955..1b4fdd03f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReductionCuda.h
@@ -331,7 +331,7 @@ struct FullReducer<Self, Op, GpuDevice, Vectorizable> {
   #ifdef EIGEN_HAS_CUDA_FP16
   static const bool HasOptimizedImplementation = !Op::IsStateful &&
       (internal::is_same<typename Self::CoeffReturnType, float>::value ||
-       (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && Op::PacketAccess));
+       (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && reducer_traits<Op, GpuDevice>::PacketAccess));
 #else
   static const bool HasOptimizedImplementation = !Op::IsStateful &&
                                                  internal::is_same<typename Self::CoeffReturnType, float>::value;
@@ -346,7 +346,7 @@ struct FullReducer<Self, Op, GpuDevice, Vectorizable> {
       return;
     }
 
-    FullReductionLauncher<Self, Op, OutputType, Op::PacketAccess>::run(self, reducer, device, output, num_coeffs);
+    FullReductionLauncher<Self, Op, OutputType, reducer_traits<Op, GpuDevice>::PacketAccess>::run(self, reducer, device, output, num_coeffs);
   }
 };
 
@@ -608,7 +608,7 @@ struct InnerReducer<Self, Op, GpuDevice> {
 #ifdef EIGEN_HAS_CUDA_FP16
   static const bool HasOptimizedImplementation = !Op::IsStateful &&
       (internal::is_same<typename Self::CoeffReturnType, float>::value ||
-       (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && Op::PacketAccess));
+       (internal::is_same<typename Self::CoeffReturnType, Eigen::half>::value && reducer_traits<Op, GpuDevice>::PacketAccess));
 #else
   static const bool HasOptimizedImplementation = !Op::IsStateful &&
                                                  internal::is_same<typename Self::CoeffReturnType, float>::value;
@@ -627,7 +627,7 @@ struct InnerReducer<Self, Op, GpuDevice> {
       return true;
     }
 
-    return InnerReductionLauncher<Self, Op, OutputType, Op::PacketAccess>::run(self, reducer, device, output, num_coeffs_to_reduce, num_preserved_vals);
+    return InnerReductionLauncher<Self, Op, OutputType, reducer_traits<Op, GpuDevice>::PacketAccess>::run(self, reducer, device, output, num_coeffs_to_reduce, num_preserved_vals);
   }
 };
 
-- 
cgit v1.2.3