Merge changes Iee153445,Iee274471 am: 79df15ea88 am: 10f298fc41 am: 7cb5001398t_frc_odp_330442040t_frc_odp_330442000t_frc_ase_330444010android-wear-13.0.0-gpl_r1android-vts-13.0_r8android-vts-13.0_r7android-vts-13.0_r6android-vts-13.0_r5android-vts-13.0_r4android-vts-13.0_r3android-vts-13.0_r2android-t-qpr3-beta-3-gplandroid-t-qpr3-beta-1-gplandroid-t-qpr2-beta-3-gplandroid-t-qpr2-beta-2-gplandroid-t-qpr1-beta-3-gplandroid-t-qpr1-beta-1-gplandroid-cts-13.0_r8android-cts-13.0_r7android-cts-13.0_r6android-cts-13.0_r5android-cts-13.0_r4android-cts-13.0_r3android-cts-13.0_r2android-13.0.0_r83android-13.0.0_r82android-13.0.0_r81android-13.0.0_r80android-13.0.0_r79android-13.0.0_r78android-13.0.0_r77android-13.0.0_r76android-13.0.0_r75android-13.0.0_r74android-13.0.0_r73android-13.0.0_r72android-13.0.0_r71android-13.0.0_r70android-13.0.0_r69android-13.0.0_r68android-13.0.0_r67android-13.0.0_r66android-13.0.0_r65android-13.0.0_r64android-13.0.0_r63android-13.0.0_r62android-13.0.0_r61android-13.0.0_r60android-13.0.0_r59android-13.0.0_r58android-13.0.0_r57android-13.0.0_r56android-13.0.0_r55android-13.0.0_r54android-13.0.0_r53android-13.0.0_r52android-13.0.0_r51android-13.0.0_r50android-13.0.0_r49android-13.0.0_r48android-13.0.0_r47android-13.0.0_r46android-13.0.0_r45android-13.0.0_r44android-13.0.0_r43android-13.0.0_r42android-13.0.0_r41android-13.0.0_r40android-13.0.0_r39android-13.0.0_r38android-13.0.0_r37android-13.0.0_r36android-13.0.0_r35android-13.0.0_r34android-13.0.0_r33android-13.0.0_r32android-13.0.0_r30android-13.0.0_r29android-13.0.0_r28android-13.0.0_r27android-13.0.0_r24android-13.0.0_r23android-13.0.0_r22android-13.0.0_r21android-13.0.0_r20android-13.0.0_r19android-13.0.0_r18android-13.0.0_r17android-13.0.0_r16aml_go_odp_330912000aml_go_ads_330915100aml_go_ads_330915000aml_go_ads_330913000android13-tests-releaseandroid13-tests-devandroid13-qpr3-s9-releaseandroid13-qpr3-s8-releaseandroid13-qpr3-s7-releaseandroid13-qpr3-s6-releaseandroid13-qpr3-s5-releaseandroid13-qpr3-s4-releaseandroid13-qpr3-s3-releaseandroid13-qpr3-s2-releaseandroid13-qpr3-s14-releaseandroid13-qpr3-s13-releaseandroid13-qpr3-s12-releaseandroid13-qpr3-s11-releaseandroid13-qpr3-s10-releaseandroid13-qpr3-s1-releaseandroid13-qpr3-releaseandroid13-qpr3-c-s8-releaseandroid13-qpr3-c-s7-releaseandroid13-qpr3-c-s6-releaseandroid13-qpr3-c-s5-releaseandroid13-qpr3-c-s4-releaseandroid13-qpr3-c-s3-releaseandroid13-qpr3-c-s2-releaseandroid13-qpr3-c-s12-releaseandroid13-qpr3-c-s11-releaseandroid13-qpr3-c-s10-releaseandroid13-qpr3-c-s1-releaseandroid13-qpr2-s9-releaseandroid13-qpr2-s8-releaseandroid13-qpr2-s7-releaseandroid13-qpr2-s6-releaseandroid13-qpr2-s5-releaseandroid13-qpr2-s3-releaseandroid13-qpr2-s2-releaseandroid13-qpr2-s12-releaseandroid13-qpr2-s11-releaseandroid13-qpr2-s10-releaseandroid13-qpr2-s1-releaseandroid13-qpr2-releaseandroid13-qpr2-b-s1-releaseandroid13-qpr1-s8-releaseandroid13-qpr1-s7-releaseandroid13-qpr1-s6-releaseandroid13-qpr1-s5-releaseandroid13-qpr1-s4-releaseandroid13-qpr1-s3-releaseandroid13-qpr1-s2-releaseandroid13-qpr1-s1-releaseandroid13-qpr1-releaseandroid13-mainline-go-adservices-releaseandroid13-frc-odp-releaseandroid13-devandroid13-d4-s2-releaseandroid13-d4-s1-releaseandroid13-d4-releaseandroid13-d3-s1-releaseandroid13-d2-releaseandroid-wear-13.0.0-gpl_r1

Original change: https://android-review.googlesource.com/c/platform/external/eigen/+/1999079

Change-Id: I4c76dc5ddc7fb0ae9fc42436f28bd8bf9de50a97

1 files changed, 86 insertions, 4 deletions

diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
index 52b803d7f..246ebe44e 100644
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorForwardDeclarations.h
@@ -12,7 +12,7 @@
 
 namespace Eigen {
 
-// MakePointer class is used as a container of the adress space of the pointer
+// MakePointer class is used as a container of the address space of the pointer
 // on the host and on the device. From the host side it generates the T* pointer
 // and when EIGEN_USE_SYCL is used it construct a buffer with a map_allocator to
 // T* m_data on the host. It is always called on the device.
@@ -20,8 +20,35 @@ namespace Eigen {
 // map_allocator.
 template<typename T> struct MakePointer {
   typedef T* Type;
+  typedef const T* ConstType;
 };
 
+template <typename T>
+EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE T* constCast(const T* data) {
+  return const_cast<T*>(data);
+}
+
+// The StorageMemory class is a container of the device specific pointer
+// used for refering to a Pointer on TensorEvaluator class. While the TensorExpression
+// is a device-agnostic type and need MakePointer class for type conversion,
+// the TensorEvaluator class can be specialized for a device, hence it is possible
+// to construct different types of temproray storage memory in TensorEvaluator
+// for different devices by specializing the following StorageMemory class.
+template<typename T, typename device> struct StorageMemory: MakePointer <T> {};
+
+namespace internal{
+template<typename A, typename B> struct Pointer_type_promotion {
+  static const bool val=false;
+};
+template<typename A> struct Pointer_type_promotion<A, A> {
+  static const bool val = true;
+};
+template<typename A, typename B> struct TypeConversion {
+  typedef A* type;
+};
+}
+
+
 template<typename PlainObjectType, int Options_ = Unaligned, template <class> class MakePointer_ = MakePointer> class TensorMap;
 template<typename Scalar_, int NumIndices_, int Options_ = 0, typename IndexType = DenseIndex> class Tensor;
 template<typename Scalar_, typename Dimensions, int Options_ = 0, typename IndexType = DenseIndex> class TensorFixedSize;
@@ -37,7 +64,7 @@ template<typename Op, typename Dims, typename XprType, template <class> class Ma
 template<typename XprType> class TensorIndexTupleOp;
 template<typename ReduceOp, typename Dims, typename XprType> class TensorTupleReducerOp;
 template<typename Axis, typename LeftXprType, typename RightXprType> class TensorConcatenationOp;
-template<typename Dimensions, typename LeftXprType, typename RightXprType> class TensorContractionOp;
+template<typename Dimensions, typename LeftXprType, typename RightXprType, typename OutputKernelType> class TensorContractionOp;
 template<typename TargetType, typename XprType> class TensorConversionOp;
 template<typename Dimensions, typename InputXprType, typename KernelXprType> class TensorConvolutionOp;
 template<typename FFT, typename XprType, int FFTDataType, int FFTDirection> class TensorFFTOp;
@@ -58,21 +85,50 @@ template<typename Strides, typename XprType> class TensorInflationOp;
 template<typename Generator, typename XprType> class TensorGeneratorOp;
 template<typename LeftXprType, typename RightXprType> class TensorAssignOp;
 template<typename Op, typename XprType> class TensorScanOp;
+template<typename Dims, typename XprType> class TensorTraceOp;
 
 template<typename CustomUnaryFunc, typename XprType> class TensorCustomUnaryOp;
 template<typename CustomBinaryFunc, typename LhsXprType, typename RhsXprType> class TensorCustomBinaryOp;
 
 template<typename XprType, template <class> class MakePointer_ = MakePointer> class TensorEvalToOp;
-template<typename XprType, template <class> class MakePointer_ = MakePointer> class TensorForcedEvalOp;
+template<typename XprType> class TensorForcedEvalOp;
 
 template<typename ExpressionType, typename DeviceType> class TensorDevice;
+template<typename ExpressionType, typename DeviceType, typename DoneCallback> class TensorAsyncDevice;
 template<typename Derived, typename Device> struct TensorEvaluator;
 
+struct NoOpOutputKernel;
+
 struct DefaultDevice;
 struct ThreadPoolDevice;
 struct GpuDevice;
 struct SyclDevice;
 
+#ifdef EIGEN_USE_SYCL
+
+template <typename T> struct MakeSYCLPointer {
+  typedef Eigen::TensorSycl::internal::RangeAccess<cl::sycl::access::mode::read_write, T> Type;
+};
+
+template <typename T>
+EIGEN_STRONG_INLINE const Eigen::TensorSycl::internal::RangeAccess<cl::sycl::access::mode::read_write, T>&
+constCast(const Eigen::TensorSycl::internal::RangeAccess<cl::sycl::access::mode::read_write, T>& data) {
+  return data;
+}
+
+template <typename T>
+struct StorageMemory<T, SyclDevice> : MakeSYCLPointer<T> {};
+template <typename T>
+struct StorageMemory<T, const SyclDevice> : StorageMemory<T, SyclDevice> {};
+
+namespace TensorSycl {
+namespace internal{
+template <typename Evaluator, typename Op> class GenericNondeterministicReducer;
+}
+}
+#endif
+
+
 enum FFTResultType {
   RealPart = 0,
   ImagPart = 1,
@@ -98,10 +154,36 @@ struct IsVectorizable<GpuDevice, Expression> {
                             TensorEvaluator<Expression, GpuDevice>::IsAligned;
 };
 
+// Tiled evaluation strategy.
+enum TiledEvaluation {
+  Off = 0,    // tiled evaluation is not supported
+  On = 1,     // still work in progress (see TensorBlock.h)
+};
+
+template <typename Device, typename Expression>
+struct IsTileable {
+  // Check that block evaluation is supported and it's a preferred option (at
+  // least one sub-expression has much faster block evaluation, e.g.
+  // broadcasting).
+  static const bool BlockAccess =
+      TensorEvaluator<Expression, Device>::BlockAccess &&
+      TensorEvaluator<Expression, Device>::PreferBlockAccess;
+
+  static const TiledEvaluation value =
+      BlockAccess ? TiledEvaluation::On : TiledEvaluation::Off;
+};
+
 template <typename Expression, typename Device,
-          bool Vectorizable = IsVectorizable<Device, Expression>::value>
+          bool Vectorizable      = IsVectorizable<Device, Expression>::value,
+          TiledEvaluation Tiling = IsTileable<Device, Expression>::value>
 class TensorExecutor;
 
+template <typename Expression, typename Device, typename DoneCallback,
+          bool Vectorizable = IsVectorizable<Device, Expression>::value,
+          TiledEvaluation Tiling = IsTileable<Device, Expression>::value>
+class TensorAsyncExecutor;
+
+
 }  // end namespace internal
 
 }  // end namespace Eigen