[SYCL] This PR adds the minimum modifications to the Eigen unsupported module required to run it on devices supporting SYCL.

* Abstracting the pointer type so that both SYCL memory and pointer can be captured.
* Converting SYCL virtual pointer to SYCL device memory in Eigen evaluator class.
* Binding SYCL placeholder accessor to command group handler by using bind method in Eigen evaluator node.
* Adding SYCL macro for controlling loop unrolling.
* Modifying the TensorDeviceSycl.h and SYCL executor method to adopt the above changes.

1 files changed, 13 insertions, 10 deletions

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
index 4292fe0c2..47db839db 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
@@ -89,6 +89,8 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
   typedef typename XprType::CoeffReturnType CoeffReturnType;
   typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
   static const int PacketSize = PacketType<CoeffReturnType, Device>::size;
+  typedef StorageMemory<CoeffReturnType, Device> Storage;
+  typedef typename Storage::Type EvaluatorPointerType;
 
 
   enum {
@@ -103,9 +105,6 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorEvaluator(const XprType& op, const Device& device)
       : m_impl(op.expression(), device)
-#ifdef EIGEN_USE_SYCL
-      , m_patch_dims(op.patch_dims())
-#endif
   {
     Index num_patches = 1;
     const typename TensorEvaluator<ArgType, Device>::Dimensions& input_dims = m_impl.dimensions();
@@ -149,7 +148,7 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
 
   EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
 
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar* /*data*/) {
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(EvaluatorPointerType /*data*/) {
     m_impl.evalSubExprsIfNeeded(NULL);
     return true;
   }
@@ -167,6 +166,7 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
     Index patchOffset = index - patchIndex * m_outputStrides[output_stride_index];
     Index inputIndex = 0;
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      EIGEN_UNROLL_LOOP
       for (int i = NumDims - 2; i > 0; --i) {
         const Index patchIdx = patchIndex / m_patchStrides[i];
         patchIndex -= patchIdx * m_patchStrides[i];
@@ -175,6 +175,7 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
         inputIndex += (patchIdx + offsetIdx) * m_inputStrides[i];
       }
     } else {
+      EIGEN_UNROLL_LOOP
       for (int i = 0; i < NumDims - 2; ++i) {
         const Index patchIdx = patchIndex / m_patchStrides[i];
         patchIndex -= patchIdx * m_patchStrides[i];
@@ -202,6 +203,7 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
 
     Index inputIndices[2] = {0, 0};
     if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
+      EIGEN_UNROLL_LOOP
       for (int i = NumDims - 2; i > 0; --i) {
         const Index patchIdx[2] = {patchIndices[0] / m_patchStrides[i],
                                    patchIndices[1] / m_patchStrides[i]};
@@ -217,6 +219,7 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
         inputIndices[1] += (patchIdx[1] + offsetIdx[1]) * m_inputStrides[i];
       }
     } else {
+      EIGEN_UNROLL_LOOP
       for (int i = 0; i < NumDims - 2; ++i) {
         const Index patchIdx[2] = {patchIndices[0] / m_patchStrides[i],
                                    patchIndices[1] / m_patchStrides[i]};
@@ -243,6 +246,7 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
       EIGEN_ALIGN_MAX CoeffReturnType values[PacketSize];
       values[0] = m_impl.coeff(inputIndices[0]);
       values[PacketSize-1] = m_impl.coeff(inputIndices[1]);
+      EIGEN_UNROLL_LOOP
       for (int i = 1; i < PacketSize-1; ++i) {
         values[i] = coeff(index+i);
       }
@@ -259,11 +263,13 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
            TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
   }
 
-  EIGEN_DEVICE_FUNC typename Eigen::internal::traits<XprType>::PointerType data() const { return NULL; }
+  EIGEN_DEVICE_FUNC EvaluatorPointerType data() const { return NULL; }
 
 #ifdef EIGEN_USE_SYCL
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const TensorEvaluator<ArgType, Device>& impl() const { return m_impl; }
-  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const PatchDim& functor() const { return m_patch_dims; }
+  // binding placeholder accessors to a command group handler for SYCL
+  EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE void bind(cl::sycl::handler &cgh) const {
+    m_impl.bind(cgh); 
+  }
 #endif
 
  protected:
@@ -274,9 +280,6 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
 
   TensorEvaluator<ArgType, Device> m_impl;
 
-#ifdef EIGEN_USE_SYCL
-  const PatchDim m_patch_dims;
-#endif
 };
 
 } // end namespace Eigen