GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-06-04 18:54:00 +08:00
Code Issues Packages Projects Releases Wiki Activity

Optimize evaluation strategy for TensorSlicingOp and TensorChippingOp

Browse Source
This commit is contained in:
Eugene Zhulenev 2019-06-25 15:41:37 -07:00
parent ba506d5bd2
commit 229db81572
2 changed files with 30 additions and 19 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

45
unsupported/Eigen/CXX11/src/Tensor/TensorChipping.h
View File

@ -145,10 +145,18 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
// Alignment can't be guaranteed at compile time since it depends on the // Alignment can't be guaranteed at compile time since it depends on the
// slice offsets. // slice offsets.
IsAligned = false, IsAligned = false,
Layout = TensorEvaluator<ArgType, Device>::Layout,
PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess, PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
BlockAccess = TensorEvaluator<ArgType, Device>::BlockAccess, BlockAccess = TensorEvaluator<ArgType, Device>::BlockAccess,
PreferBlockAccess = true, // Chipping of outer-most dimension is a trivial operation, because we can
Layout = TensorEvaluator<ArgType, Device>::Layout, // read and write directly from the underlying tensor using single offset.
IsOuterChipping = (static_cast<int>(Layout) == ColMajor && DimId == NumInputDims - 1) ||
(static_cast<int>(Layout) == RowMajor && DimId == 0),
// 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,
CoordAccess = false, // to be implemented CoordAccess = false, // to be implemented
RawAccess = false RawAccess = false
}; };
@ -230,8 +238,7 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE) EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
eigen_assert(index+PacketSize-1 < dimensions().TotalSize()); eigen_assert(index+PacketSize-1 < dimensions().TotalSize());
if ((static_cast<int>(Layout) == static_cast<int>(ColMajor) && m_dim.actualDim() == 0) || if (IsInnerChipping) {
(static_cast<int>(Layout) == static_cast<int>(RowMajor) && m_dim.actualDim() == NumInputDims-1)) {
// m_stride is equal to 1, so let's avoid the integer division. // m_stride is equal to 1, so let's avoid the integer division.
eigen_assert(m_stride == 1); eigen_assert(m_stride == 1);
Index inputIndex = index * m_inputStride + m_inputOffset; Index inputIndex = index * m_inputStride + m_inputOffset;
@ -242,8 +249,7 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
} }
PacketReturnType rslt = internal::pload<PacketReturnType>(values); PacketReturnType rslt = internal::pload<PacketReturnType>(values);
return rslt; return rslt;
} else if ((static_cast<int>(Layout) == static_cast<int>(ColMajor) && m_dim.actualDim() == NumInputDims - 1) || } else if (IsOuterChipping) {
(static_cast<int>(Layout) == static_cast<int>(RowMajor) && m_dim.actualDim() == 0)) {
// m_stride is always greater than index, so let's avoid the integer division. // m_stride is always greater than index, so let's avoid the integer division.
eigen_assert(m_stride > index); eigen_assert(m_stride > index);
return m_impl.template packet<LoadMode>(index + m_inputOffset); return m_impl.template packet<LoadMode>(index + m_inputOffset);
@ -345,9 +351,7 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Eigen::internal::traits<XprType>::PointerType data() const { EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE typename Eigen::internal::traits<XprType>::PointerType data() const {
CoeffReturnType* result = const_cast<CoeffReturnType*>(m_impl.data()); CoeffReturnType* result = const_cast<CoeffReturnType*>(m_impl.data());
if (((static_cast<int>(Layout) == static_cast<int>(ColMajor) && m_dim.actualDim() == NumDims) || if (IsOuterChipping && result) {
(static_cast<int>(Layout) == static_cast<int>(RowMajor) && m_dim.actualDim() == 0)) &&
result) {
return result + m_inputOffset; return result + m_inputOffset;
} else { } else {
return NULL; return NULL;
@ -370,13 +374,11 @@ struct TensorEvaluator<const TensorChippingOp<DimId, ArgType>, Device>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE Index srcCoeff(Index index) const
{ {
Index inputIndex; Index inputIndex;
if ((static_cast<int>(Layout) == static_cast<int>(ColMajor) && m_dim.actualDim() == 0) || if (IsInnerChipping) {
(static_cast<int>(Layout) == static_cast<int>(RowMajor) && m_dim.actualDim() == NumInputDims - 1)) {
// m_stride is equal to 1, so let's avoid the integer division. // m_stride is equal to 1, so let's avoid the integer division.
eigen_assert(m_stride == 1); eigen_assert(m_stride == 1);
inputIndex = index * m_inputStride + m_inputOffset; inputIndex = index * m_inputStride + m_inputOffset;
} else if ((static_cast<int>(Layout) == static_cast<int>(ColMajor) && m_dim.actualDim() == NumInputDims - 1) || } else if (IsOuterChipping) {
(static_cast<int>(Layout) == static_cast<int>(RowMajor) && m_dim.actualDim() == 0)) {
// m_stride is always greater than index, so let's avoid the integer // m_stride is always greater than index, so let's avoid the integer
// division. // division.
eigen_assert(m_stride > index); eigen_assert(m_stride > index);
@ -425,7 +427,16 @@ struct TensorEvaluator<TensorChippingOp<DimId, ArgType>, Device>
PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess, PacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess,
BlockAccess = TensorEvaluator<ArgType, Device>::BlockAccess, BlockAccess = TensorEvaluator<ArgType, Device>::BlockAccess,
Layout = TensorEvaluator<ArgType, Device>::Layout, Layout = TensorEvaluator<ArgType, Device>::Layout,
RawAccess = false RawAccess = false,
// Chipping of outer-most dimension is a trivial operation, because we can
// read and write directly from the underlying tensor using single offset.
IsOuterChipping =
(static_cast<int>(Layout) == ColMajor && DimId == NumInputDims - 1) ||
(static_cast<int>(Layout) == RowMajor && DimId == 0),
// Chipping inner-most dimension.
IsInnerChipping =
(static_cast<int>(Layout) == ColMajor && DimId == 0) ||
(static_cast<int>(Layout) == RowMajor && DimId == NumInputDims - 1),
}; };
typedef typename internal::remove_const<Scalar>::type ScalarNoConst; typedef typename internal::remove_const<Scalar>::type ScalarNoConst;
@ -449,8 +460,7 @@ struct TensorEvaluator<TensorChippingOp<DimId, ArgType>, Device>
{ {
EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE) EIGEN_STATIC_ASSERT((PacketSize > 1), YOU_MADE_A_PROGRAMMING_MISTAKE)
if ((static_cast<int>(this->Layout) == static_cast<int>(ColMajor) && this->m_dim.actualDim() == 0) || if (IsInnerChipping) {
(static_cast<int>(this->Layout) == static_cast<int>(RowMajor) && this->m_dim.actualDim() == NumInputDims-1)) {
// m_stride is equal to 1, so let's avoid the integer division. // m_stride is equal to 1, so let's avoid the integer division.
eigen_assert(this->m_stride == 1); eigen_assert(this->m_stride == 1);
EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize]; EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
@ -460,8 +470,7 @@ struct TensorEvaluator<TensorChippingOp<DimId, ArgType>, Device>
this->m_impl.coeffRef(inputIndex) = values[i]; this->m_impl.coeffRef(inputIndex) = values[i];
inputIndex += this->m_inputStride; inputIndex += this->m_inputStride;
} }
} else if ((static_cast<int>(this->Layout) == static_cast<int>(ColMajor) && this->m_dim.actualDim() == NumInputDims-1) || } else if (IsOuterChipping) {
(static_cast<int>(this->Layout) == static_cast<int>(RowMajor) && this->m_dim.actualDim() == 0)) {
// m_stride is always greater than index, so let's avoid the integer division. // m_stride is always greater than index, so let's avoid the integer division.
eigen_assert(this->m_stride > index); eigen_assert(this->m_stride > index);
this->m_impl.template writePacket<StoreMode>(index + this->m_inputOffset, x); this->m_impl.template writePacket<StoreMode>(index + this->m_inputOffset, x);

4
unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h
View File

@ -577,7 +577,9 @@ struct TensorEvaluator<const TensorSlicingOp<StartIndices, Sizes, ArgType>, Devi
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType* data) { EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(CoeffReturnType* data) {
m_impl.evalSubExprsIfNeeded(NULL); m_impl.evalSubExprsIfNeeded(NULL);
if (!NumTraits<typename internal::remove_const<Scalar>::type>::RequireInitialization && data && m_impl.data()) { if (!NumTraits<typename internal::remove_const<Scalar>::type>::RequireInitialization
&& data && m_impl.data()
&& !BlockAccess) {
Index contiguous_values = 1; Index contiguous_values = 1;
if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) { if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
for (int i = 0; i < NumDims; ++i) { for (int i = 0; i < NumDims; ++i) {