From 13c3327f5cf829fd9d04a2ab46861e722cd74ca0 Mon Sep 17 00:00:00 2001
From: Eugene Zhulenev <ezhulenev@google.com>
Date: Tue, 12 Nov 2019 10:12:28 -0800
Subject: Remove legacy block evaluation support

---
 .../Eigen/CXX11/src/Tensor/TensorChipping.h        | 68 ----------------------
 1 file changed, 68 deletions(-)

(limited to 'unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h')

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
index 32d6960bf..098110217 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
@@ -148,7 +148,6 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
     IsAligned         = false,
     Layout            = TensorEvaluator<ArgType, Device>::Layout,
     PacketAccess      = TensorEvaluator<ArgType, Device>::PacketAccess,
-    BlockAccess       = TensorEvaluator<ArgType, Device>::BlockAccess,
     BlockAccessV2     = TensorEvaluator<ArgType, Device>::BlockAccessV2,
     // Chipping of outer-most dimension is a trivial operation, because we can
     // read and write directly from the underlying tensor using single offset.
@@ -167,11 +166,6 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
 
   typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
 
-  typedef internal::TensorBlock<ScalarNoConst, Index, NumInputDims, Layout>
-      InputTensorBlock;
-  typedef internal::TensorBlock<ScalarNoConst, Index, NumDims, Layout>
-      OutputTensorBlock;
-
   //===- Tensor block evaluation strategy (see TensorBlock.h) -------------===//
   typedef internal::TensorBlockDescriptor<NumDims, Index> TensorBlockDesc;
   typedef internal::TensorBlockScratchAllocator<Device> TensorBlockScratch;
@@ -218,20 +212,6 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
     }
     m_inputStride *= input_dims[m_dim.actualDim()];
     m_inputOffset = m_stride * op.offset();
-
-    if (BlockAccess) {
-      if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
-        m_inputStrides[0] = 1;
-        for (int i = 1; i < NumInputDims; ++i) {
-          m_inputStrides[i] = m_inputStrides[i - 1] * input_dims[i - 1];
-        }
-      } else {
-        m_inputStrides[NumInputDims - 1] = 1;
-        for (int i = NumInputDims - 2; i >= 0; --i) {
-          m_inputStrides[i] = m_inputStrides[i + 1] * input_dims[i + 1];
-        }
-      }
-    }
   }
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
@@ -323,52 +303,6 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
     m_impl.getResourceRequirements(resources);
   }
 
-  // TODO(andydavis) Reduce the overhead of this function (experiment with
-  // using a fixed block size).
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void block(
-      OutputTensorBlock* output_block) const {
-    // Calculate input block sizes.
-    const DSizes<Index, NumDims>& output_block_sizes =
-        output_block->block_sizes();
-    const DSizes<Index, NumDims>& output_block_strides =
-        output_block->block_strides();
-    const Index chip_dim = m_dim.actualDim();
-    DSizes<Index, NumInputDims> input_block_sizes;
-    DSizes<Index, NumInputDims> input_block_strides;
-    for (Index i = 0; i < NumInputDims; ++i) {
-      if (i < chip_dim) {
-        input_block_sizes[i] = output_block_sizes[i];
-        input_block_strides[i] = output_block_strides[i];
-      } else if (i > chip_dim) {
-        input_block_sizes[i] = output_block_sizes[i - 1];
-        input_block_strides[i] = output_block_strides[i - 1];
-      } else {
-        input_block_sizes[i] = 1;
-      }
-    }
-    // Fix up input_block_stride for chip dimension.
-    if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
-      if (chip_dim == 0) {
-        input_block_strides[chip_dim] = 1;
-      } else {
-        input_block_strides[chip_dim] =
-            input_block_strides[chip_dim - 1] * input_block_sizes[chip_dim - 1];
-      }
-    } else {
-      if (chip_dim == NumInputDims - 1) {
-        input_block_strides[chip_dim] = 1;
-      } else {
-        input_block_strides[chip_dim] =
-            input_block_strides[chip_dim + 1] * input_block_sizes[chip_dim + 1];
-      }
-    }
-    // Instantiate and read input block from input tensor.
-    InputTensorBlock input_block(srcCoeff(output_block->first_coeff_index()),
-                                 input_block_sizes, input_block_strides,
-                                 m_inputStrides, output_block->data());
-    m_impl.block(&input_block);
-  }
-
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorBlockV2
   blockV2(TensorBlockDesc& desc, TensorBlockScratch& scratch,
           bool root_of_expr_ast = false) const {
@@ -482,7 +416,6 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
   Index m_stride;
   Index m_inputOffset;
   Index m_inputStride;
-  DSizes<Index, NumInputDims> m_inputStrides;
   TensorEvaluator<ArgType, Device> m_impl;
   const internal::DimensionId<DimId> m_dim;
   const Device EIGEN_DEVICE_REF m_device;
@@ -508,7 +441,6 @@ struct TensorEvaluator<TensorChippingOp<DimId, ArgType>, Device>
   enum {
     IsAligned     = false,
     PacketAccess  = TensorEvaluator<ArgType, Device>::PacketAccess,
-    BlockAccess   = TensorEvaluator<ArgType, Device>::BlockAccess,
     BlockAccessV2 = TensorEvaluator<ArgType, Device>::RawAccess,
     Layout        = TensorEvaluator<ArgType, Device>::Layout,
     RawAccess     = false
-- 
cgit v1.2.3