Rebase to latest.

13 files changed, 48 insertions, 37 deletions

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
index 392acf302..cca716d6f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorBase.h
@@ -129,6 +129,12 @@ class TensorBase<Derived, ReadOnlyAccessors>
     }
 
     EIGEN_DEVICE_FUNC
+    EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_digamma_op<Scalar>, const Derived>
+    digamma() const {
+      return unaryExpr(internal::scalar_digamma_op<Scalar>());
+    }
+
+    EIGEN_DEVICE_FUNC
     EIGEN_STRONG_INLINE const TensorCwiseUnaryOp<internal::scalar_erf_op<Scalar>, const Derived>
     erf() const {
       return unaryExpr(internal::scalar_erf_op<Scalar>());
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
index 624e814e2..1adb68894 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContraction.h
@@ -378,7 +378,7 @@ struct TensorContractionEvaluatorBase
   }
 
   template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
-  void evalGemv(Scalar* buffer) const {
+  EIGEN_DEVICE_FUNC void evalGemv(Scalar* buffer) const {
     const Index rows = m_i_size;
     const Index cols = m_k_size;
 
@@ -516,7 +516,7 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
       Base(op, device) { }
 
   template <bool lhs_inner_dim_contiguous, bool rhs_inner_dim_contiguous, bool rhs_inner_dim_reordered, int Alignment>
-  void evalProduct(Scalar* buffer) const {
+  EIGEN_DEVICE_FUNC void evalProduct(Scalar* buffer) const {
     if (this->m_j_size == 1) {
       this->template evalGemv<lhs_inner_dim_contiguous, rhs_inner_dim_contiguous, rhs_inner_dim_reordered, Alignment>(buffer);
       return;
@@ -582,10 +582,8 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
 
     OutputMapper output(buffer, m);
 
-    typedef typename internal::gemm_blocking_space<ColMajor, LhsScalar, RhsScalar, Dynamic, Dynamic, Dynamic> BlockingType;
-
     // Sizes of the blocks to load in cache. See the Goto paper for details.
-    BlockingType blocking(m, n, k, 1, true);
+    internal::TensorContractionBlocking<LhsMapper, RhsMapper, Index, internal::ShardByCol> blocking(k, m, n, 1);
     const Index kc = blocking.kc();
     const Index mc = numext::mini(m, blocking.mc());
     const Index nc = numext::mini(n, blocking.nc());
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h
index 78ed5038f..3d3f6904f 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionBlocking.h
@@ -28,7 +28,7 @@ class TensorContractionBlocking {
   typedef typename LhsMapper::Scalar LhsScalar;
   typedef typename RhsMapper::Scalar RhsScalar;
 
-  TensorContractionBlocking(Index k, Index m, Index n, Index num_threads = 1) :
+  EIGEN_DEVICE_FUNC TensorContractionBlocking(Index k, Index m, Index n, Index num_threads = 1) :
       kc_(k), mc_(m), nc_(n)
   {
     if (ShardingType == ShardByCol) {
@@ -41,9 +41,9 @@ class TensorContractionBlocking {
     }
   }
 
-  EIGEN_ALWAYS_INLINE Index kc() const { return kc_; }
-  EIGEN_ALWAYS_INLINE Index mc() const { return mc_; }
-  EIGEN_ALWAYS_INLINE Index nc() const { return nc_; }
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Index kc() const { return kc_; }
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Index mc() const { return mc_; }
+  EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE Index nc() const { return nc_; }
 
  private:
   Index kc_;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionMapper.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionMapper.h
index 9b6d18090..63c8ae126 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionMapper.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionMapper.h
@@ -426,15 +426,16 @@ class TensorContractionSubMapper {
 };
 
 
-template<typename Scalar, typename Index, int side,
+template<typename Scalar_, typename Index, int side,
          typename Tensor,
          typename nocontract_t, typename contract_t,
          int packet_size,
          bool inner_dim_contiguous, bool inner_dim_reordered, int Alignment>
 class TensorContractionInputMapper
-  : public BaseTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, packet_size, inner_dim_contiguous, inner_dim_reordered, Alignment> {
+  : public BaseTensorContractionMapper<Scalar_, Index, side, Tensor, nocontract_t, contract_t, packet_size, inner_dim_contiguous, inner_dim_reordered, Alignment> {
 
  public:
+  typedef Scalar_ Scalar;
   typedef BaseTensorContractionMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, packet_size, inner_dim_contiguous, inner_dim_reordered, Alignment> Base;
   typedef TensorContractionSubMapper<Scalar, Index, side, Tensor, nocontract_t, contract_t, packet_size, inner_dim_contiguous, inner_dim_reordered, Alignment> SubMapper;
   typedef SubMapper VectorMapper;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
index 576bea295..51a3b9490 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.h
@@ -176,10 +176,10 @@ struct TensorEvaluator<const TensorContractionOp<Indices, LeftArgType, RightArgT
 
     // compute block sizes (which depend on number of threads)
     const Index num_threads = this->m_device.numThreads();
-    Index mc = m;
-    Index nc = n;
-    Index kc = k;
-    internal::computeProductBlockingSizes<LhsScalar,RhsScalar,1>(kc, mc, nc, num_threads);
+    internal::TensorContractionBlocking<LhsMapper, RhsMapper, Index, internal::ShardByCol> blocking(k, m, n, num_threads);
+    Index mc = blocking.mc();
+    Index nc = blocking.nc();
+    Index kc = blocking.kc();
     eigen_assert(mc <= m);
     eigen_assert(nc <= n);
     eigen_assert(kc <= k);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
index 367a152a0..67c797802 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
@@ -21,7 +21,7 @@ namespace Eigen {
   */
 namespace internal {
 
-template <typename Index, typename InputDims, size_t NumKernelDims, int Layout>
+template <typename Index, typename InputDims, int NumKernelDims, int Layout>
 class IndexMapper {
  public:
   IndexMapper(const InputDims& input_dims, const array<Index, NumKernelDims>& kernel_dims,
@@ -123,7 +123,7 @@ class IndexMapper {
       }
       inputIndex += p * m_inputStrides[NumKernelDims];
     } else {
-      int limit = 0;
+      std::ptrdiff_t limit = 0;
       if (NumKernelDims < NumDims) {
         limit = NumDims - NumKernelDims - 1;
       }
@@ -147,7 +147,7 @@ class IndexMapper {
       }
       outputIndex += p * m_outputStrides[NumKernelDims];
     } else {
-      int limit = 0;
+      std::ptrdiff_t limit = 0;
       if (NumKernelDims < NumDims) {
         limit = NumDims - NumKernelDims - 1;
       }
@@ -206,7 +206,7 @@ class IndexMapper {
   }
 
  private:
-  static const size_t NumDims = internal::array_size<InputDims>::value;
+  static const int NumDims = internal::array_size<InputDims>::value;
   array<Index, NumDims> m_inputStrides;
   array<Index, NumDims> m_outputStrides;
   array<Index, NumDims> m_cudaInputStrides;
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h
index 5abdc489b..e684ab8f7 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorDeviceCuda.h
@@ -109,10 +109,12 @@ class CudaStreamDevice : public StreamInterface {
 struct GpuDevice {
   // The StreamInterface is not owned: the caller is
   // responsible for its initialization and eventual destruction.
-  explicit GpuDevice(const StreamInterface* stream) : stream_(stream) {
+  explicit GpuDevice(const StreamInterface* stream) : stream_(stream), max_blocks_(INT_MAX) {
+    eigen_assert(stream);
+  }
+  explicit GpuDevice(const StreamInterface* stream, int num_blocks) : stream_(stream), max_blocks_(num_blocks) {
     eigen_assert(stream);
   }
-
   // TODO(bsteiner): This is an internal API, we should not expose it.
   EIGEN_STRONG_INLINE const cudaStream_t& stream() const {
     return stream_->stream();
@@ -246,6 +248,10 @@ struct GpuDevice {
 #endif
   }
 
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE int maxBlocks() const {
+    return max_blocks_;
+  }
+
   // This function checks if the CUDA runtime recorded an error for the
   // underlying stream device.
   inline bool ok() const {
@@ -259,7 +265,7 @@ struct GpuDevice {
 
  private:
   const StreamInterface* stream_;
-
+  int max_blocks_;
 };
 
 #ifndef __CUDA_ARCH__
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h
index e7daf7304..bd83d5de8 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvalTo.h
@@ -136,7 +136,7 @@ struct TensorEvaluator<const TensorEvalToOp<ArgType>, Device>
   }
 
   template<int LoadMode>
-  EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
   {
     return internal::ploadt<Packet, LoadMode>(m_buffer + index);
   }
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
index d2ab70f2b..df15c6204 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
@@ -220,7 +220,7 @@ EIGEN_DEVICE_FUNC inline void TensorExecutor<Expression, GpuDevice, false>::run(
   if (needs_assign)
   {
     const int block_size = device.maxCudaThreadsPerBlock();
-    const int max_blocks = device.getNumCudaMultiProcessors() * device.maxCudaThreadsPerMultiProcessor() / block_size;
+    const int max_blocks = numext::maxi<int>(device.maxBlocks(), device.getNumCudaMultiProcessors() * device.maxCudaThreadsPerMultiProcessor() / block_size);
     const Index size = array_prod(evaluator.dimensions());
     // Create a least one block to ensure we won't crash if we're called with tensors of size 0.
     const int num_blocks = numext::maxi<int>(numext::mini<int>(max_blocks, (size + block_size - 1) / block_size), 1);
@@ -239,7 +239,7 @@ EIGEN_DEVICE_FUNC inline void TensorExecutor<Expression, GpuDevice, true>::run(c
   if (needs_assign)
   {
     const int block_size = device.maxCudaThreadsPerBlock();
-    const int max_blocks = device.getNumCudaMultiProcessors() * device.maxCudaThreadsPerMultiProcessor() / block_size;
+    const int max_blocks = numext::maxi<int>(device.maxBlocks(), device.getNumCudaMultiProcessors() * device.maxCudaThreadsPerMultiProcessor() / block_size);
     const Index size = array_prod(evaluator.dimensions());
     // Create a least one block to ensure we won't crash if we're called with tensors of size 0.
     const int num_blocks = numext::maxi<int>(numext::mini<int>(max_blocks, (size + block_size - 1) / block_size), 1);
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
index c9b0b2f28..58b864787 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
@@ -106,7 +106,6 @@ struct TensorEvaluator<const TensorForcedEvalOp<ArgType>, Device>
   EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_impl.dimensions(); }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType*) {
-    m_impl.evalSubExprsIfNeeded(NULL);
     const Index numValues = m_impl.dimensions().TotalSize();
     m_buffer = (CoeffReturnType*)m_device.allocate(numValues * sizeof(CoeffReturnType));
     // Should initialize the memory in case we're dealing with non POD types.
@@ -119,7 +118,6 @@ struct TensorEvaluator<const TensorForcedEvalOp<ArgType>, Device>
     EvalTo evalToTmp(m_buffer, m_op);
     const bool PacketAccess = internal::IsVectorizable<Device, const ArgType>::value;
     internal::TensorExecutor<const EvalTo, Device, PacketAccess>::run(evalToTmp, m_device);
-    m_impl.cleanup();
     return true;
   }
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void cleanup() {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
index 09ee0c2c6..22aea5ea4 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
@@ -345,8 +345,8 @@ template <typename Self, typename Op, typename Device>
 struct InnerReducer {
   static const bool HasOptimizedImplementation = false;
 
-  static void run(const Self&, Op&, const Device&, typename Self::CoeffReturnType*, typename Self::Index, typename Self::Index) {
-    assert(false && "Not implemented");
+  EIGEN_DEVICE_FUNC static void run(const Self&, Op&, const Device&, typename Self::CoeffReturnType*, typename Self::Index, typename Self::Index) {
+    eigen_assert(false && "Not implemented");
   }
 };
 
@@ -355,8 +355,8 @@ template <typename Self, typename Op, typename Device>
 struct OuterReducer {
   static const bool HasOptimizedImplementation = false;
 
-  static void run(const Self&, Op&, const Device&, typename Self::CoeffReturnType*, typename Self::Index, typename Self::Index) {
-    assert(false && "Not implemented");
+  EIGEN_DEVICE_FUNC static void run(const Self&, Op&, const Device&, typename Self::CoeffReturnType*, typename Self::Index, typename Self::Index) {
+    eigen_assert(false && "Not implemented");
   }
 };
 
@@ -463,7 +463,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
 	  m_outputStrides[i] = m_outputStrides[i - 1] * m_dimensions[i - 1];
 	}
       } else {
-	m_outputStrides[NumOutputDims - 1] = 1;
+	m_outputStrides.back() = 1;
 	for (int i = NumOutputDims - 2; i >= 0; --i) {
 	  m_outputStrides[i] = m_outputStrides[i + 1] * m_dimensions[i + 1];
 	}
@@ -479,7 +479,7 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
 	  input_strides[i] = input_strides[i-1] * input_dims[i-1];
 	}
       } else {
-	input_strides[NumInputDims - 1] = 1;
+	input_strides.back() = 1;
 	for (int i = NumInputDims - 2; i >= 0; --i) {
 	  input_strides[i] = input_strides[i + 1] * input_dims[i + 1];
 	}
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h
index 98631fc7f..ed933b6ac 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStorage.h
@@ -41,7 +41,10 @@ class TensorStorage<T, FixedDimensions, Options_>
  private:
   static const std::size_t Size = FixedDimensions::total_size;
 
-  EIGEN_ALIGN_MAX T m_data[Size];
+  // Allocate an array of size at least one to prevent compiler warnings.
+  static const std::size_t MinSize = max_n_1<Size>::size;
+  EIGEN_ALIGN_MAX T m_data[MinSize];
+
   FixedDimensions m_dimensions;
 
  public:
@@ -105,7 +108,6 @@ class TensorStorage<T, DSizes<IndexType, NumIndices_>, Options_>
 
     EIGEN_DEVICE_FUNC void resize(Index size, const array<Index, NumIndices_>& nbDimensions)
     {
-      eigen_assert(size >= 1);
       const Index currentSz = internal::array_prod(m_dimensions);
       if(size != currentSz)
       {
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorUInt128.h b/unsupported/Eigen/CXX11/src/Tensor/TensorUInt128.h
index 19352eb5e..0d34f7ee6 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorUInt128.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorUInt128.h
@@ -34,11 +34,11 @@ struct TensorUInt128
   LOW low;
 
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
-  TensorUInt128(int x) : high(0), low(x) {
+  TensorUInt128(int32_t x) : high(0), low(x) {
     eigen_assert(x >= 0);
   }
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
-  TensorUInt128(unsigned int x) : high(0), low(x) { }
+  TensorUInt128(uint32_t x) : high(0), low(x) { }
   EIGEN_DEVICE_FUNC EIGEN_ALWAYS_INLINE
   TensorUInt128(long x) : high(0), low(x) {
     eigen_assert(x >= 0);