Propagate block evaluation preference through rvalue tensor expressions

2025-08-11 19:29:02 +08:00 · 2019-10-17 11:17:33 -07:00 · 2019-10-17 11:17:33 -07:00 · df0e8b8137
commit df0e8b8137
parent 0d2a14ce11
10 changed files with 51 additions and 41 deletions
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorArgMax.h
@ -90,7 +90,7 @@ struct TensorEvaluator<const TensorIndexTupleOp<ArgType>, Device>
    PacketAccess = /*TensorEvaluator<ArgType, Device>::PacketAccess*/ false,
    BlockAccess = false,
    BlockAccessV2 = false,
-    PreferBlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
    Layout = TensorEvaluator<ArgType, Device>::Layout,
    CoordAccess = false,  // to be implemented
    RawAccess = false
@ -232,7 +232,7 @@ struct TensorEvaluator<const TensorTupleReducerOp<ReduceOp, Dims, ArgType>, Devi
    PacketAccess      = /*TensorEvaluator<ArgType, Device>::PacketAccess*/ false,
    BlockAccess       = false,
    BlockAccessV2     = false,
-    PreferBlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
    Layout            = TensorEvaluator<const TensorReductionOp<ReduceOp, Dims, const TensorIndexTupleOp<ArgType> >, Device>::Layout,
    CoordAccess       = false,  // to be implemented
    RawAccess         = false
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
@ -157,8 +157,10 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
    // Chipping inner-most dimension.
    IsInnerChipping   = (static_cast<int>(Layout) == ColMajor && DimId == 0) ||
                        (static_cast<int>(Layout) == RowMajor && DimId == NumInputDims - 1),
-    // Do not choose block access if chipping is trivial.
-    PreferBlockAccess = !IsOuterChipping,
+    // Prefer block access if the underlying expression prefers it, otherwise
+    // only if chipping is not trivial.
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess ||
+                        !IsOuterChipping,
    CoordAccess       = false,  // to be implemented
    RawAccess         = false
  };
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorConcatenation.h
@ -123,10 +123,12 @@ struct TensorEvaluator<const TensorConcatenationOp<Axis, LeftArgType, RightArgTy
  typedef typename Storage::Type EvaluatorPointerType;
  enum {
    IsAligned         = false,
-    PacketAccess = TensorEvaluator<LeftArgType, Device>::PacketAccess & TensorEvaluator<RightArgType, Device>::PacketAccess,
+    PacketAccess      = TensorEvaluator<LeftArgType, Device>::PacketAccess &&
+                        TensorEvaluator<RightArgType, Device>::PacketAccess,
    BlockAccess       = false,
    BlockAccessV2     = false,
-    PreferBlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<LeftArgType, Device>::PreferBlockAccess ||
+                        TensorEvaluator<RightArgType, Device>::PreferBlockAccess,
    Layout            = TensorEvaluator<LeftArgType, Device>::Layout,
    RawAccess         = false
  };
@ -321,10 +323,12 @@ template<typename Axis, typename LeftArgType, typename RightArgType, typename De
  typedef typename Base::Dimensions Dimensions;
  enum {
    IsAligned         = false,
-    PacketAccess = TensorEvaluator<LeftArgType, Device>::PacketAccess & TensorEvaluator<RightArgType, Device>::PacketAccess,
+    PacketAccess      = TensorEvaluator<LeftArgType, Device>::PacketAccess &&
+                        TensorEvaluator<RightArgType, Device>::PacketAccess,
    BlockAccess       = false,
    BlockAccessV2     = false,
-    PreferBlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<LeftArgType, Device>::PreferBlockAccess ||
+                        TensorEvaluator<RightArgType, Device>::PreferBlockAccess,
    Layout            = TensorEvaluator<LeftArgType, Device>::Layout,
    RawAccess         = false
  };
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h
@ -764,11 +764,15 @@ struct TensorEvaluator<const TensorCwiseTernaryOp<TernaryOp, Arg1Type, Arg2Type,

  enum {
    IsAligned = TensorEvaluator<Arg1Type, Device>::IsAligned & TensorEvaluator<Arg2Type, Device>::IsAligned & TensorEvaluator<Arg3Type, Device>::IsAligned,
-    PacketAccess = TensorEvaluator<Arg1Type, Device>::PacketAccess & TensorEvaluator<Arg2Type, Device>::PacketAccess & TensorEvaluator<Arg3Type, Device>::PacketAccess &
+    PacketAccess      = TensorEvaluator<Arg1Type, Device>::PacketAccess &&
+                        TensorEvaluator<Arg2Type, Device>::PacketAccess &&
+                        TensorEvaluator<Arg3Type, Device>::PacketAccess &&
                        internal::functor_traits<TernaryOp>::PacketAccess,
    BlockAccess       = false,
    BlockAccessV2     = false,
-    PreferBlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<Arg1Type, Device>::PreferBlockAccess ||
+                        TensorEvaluator<Arg2Type, Device>::PreferBlockAccess ||
+                        TensorEvaluator<Arg3Type, Device>::PreferBlockAccess,
    Layout            = TensorEvaluator<Arg1Type, Device>::Layout,
    CoordAccess       = false,  // to be implemented
    RawAccess         = false
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorLayoutSwap.h
@ -121,7 +121,7 @@ struct TensorEvaluator<const TensorLayoutSwapOp<ArgType>, Device>
    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
    BlockAccess = false,
    BlockAccessV2 = false,
-    PreferBlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
    Layout = (static_cast<int>(TensorEvaluator<ArgType, Device>::Layout) == static_cast<int>(ColMajor)) ? RowMajor : ColMajor,
    CoordAccess = false,  // to be implemented
    RawAccess = TensorEvaluator<ArgType, Device>::RawAccess
@ -201,7 +201,7 @@ template<typename ArgType, typename Device>
    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
    BlockAccess = false,
    BlockAccessV2 = false,
-    PreferBlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
    Layout = (static_cast<int>(TensorEvaluator<ArgType, Device>::Layout) == static_cast<int>(ColMajor)) ? RowMajor : ColMajor,
    CoordAccess = false  // to be implemented
  };
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
@ -1118,7 +1118,7 @@ struct TensorEvaluator<const TensorStridingSlicingOp<StartIndices, StopIndices,
    PacketAccess = false,
    BlockAccess = false,
    BlockAccessV2 = false,
-    PreferBlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
    Layout = TensorEvaluator<ArgType, Device>::Layout,
    RawAccess = false
  };
@ -1300,7 +1300,7 @@ struct TensorEvaluator<TensorStridingSlicingOp<StartIndices, StopIndices, Stride
    PacketAccess = false,
    BlockAccess = false,
    BlockAccessV2 = false,
-    PreferBlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
    Layout = TensorEvaluator<ArgType, Device>::Layout,
    CoordAccess = TensorEvaluator<ArgType, Device>::CoordAccess,
    RawAccess = false
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorPatch.h
@ -98,7 +98,7 @@ struct TensorEvaluator<const TensorPatchOp<PatchDim, ArgType>, Device>
    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
    BlockAccess = false,
    BlockAccessV2 = false,
-    PreferBlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
    Layout = TensorEvaluator<ArgType, Device>::Layout,
    CoordAccess = false,
    RawAccess = false
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h
@ -116,7 +116,7 @@ struct TensorEvaluator<const TensorStridingOp<Strides, ArgType>, Device>
    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
    BlockAccess = false,
    BlockAccessV2 = false,
-    PreferBlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
    Layout = TensorEvaluator<ArgType, Device>::Layout,
    CoordAccess = false,  // to be implemented
    RawAccess = false
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorTrace.h
@ -99,7 +99,7 @@ struct TensorEvaluator<const TensorTraceOp<Dims, ArgType>, Device>
    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
    BlockAccess = false,
    BlockAccessV2 = false,
-    PreferBlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
    Layout = TensorEvaluator<ArgType, Device>::Layout,
    CoordAccess = false,
    RawAccess = false
--- a/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
+++ b/unsupported/Eigen/CXX11/src/Tensor/TensorVolumePatch.h
@ -185,7 +185,7 @@ struct TensorEvaluator<const TensorVolumePatchOp<Planes, Rows, Cols, ArgType>, D
    PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
    BlockAccess = false,
    BlockAccessV2 = false,
-    PreferBlockAccess = false,
+    PreferBlockAccess = TensorEvaluator<ArgType, Device>::PreferBlockAccess,
    Layout = TensorEvaluator<ArgType, Device>::Layout,
    CoordAccess = false,
    RawAccess = false