diff options
author | Benoit Steiner <benoit.steiner.goog@gmail.com> | 2016-04-25 09:16:08 -0700 |
---|---|---|
committer | Benoit Steiner <benoit.steiner.goog@gmail.com> | 2016-04-25 09:16:08 -0700 |
commit | fd9401f2601fe8ed16c1a386f217da02dc8bfd2f (patch) | |
tree | 66d74b165e8faf3564c2e2f888214a2482c458c3 /unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h | |
parent | 5c372d19e38fc5d96a559b253cdaba0145d9b468 (diff) |
Refined the cost of the striding operation.
Diffstat (limited to 'unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h')
-rw-r--r-- | unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h | 16 |
1 files changed, 11 insertions, 5 deletions
diff --git a/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h b/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h index 23248c626..52b7d216a 100644 --- a/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h +++ b/unsupported/Eigen/CXX11/src/Tensor/TensorStriding.h @@ -209,11 +209,17 @@ struct TensorEvaluator<const TensorStridingOp<Strides, ArgType>, Device> } EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const { - const double compute_cost = NumDims * (2 * TensorOpCost::AddCost<Index>() + - 2 * TensorOpCost::MulCost<Index>() + - TensorOpCost::DivCost<Index>()); - return m_impl.costPerCoeff(vectorized) + - TensorOpCost(0, 0, compute_cost, false /* vectorized */, PacketSize); + double compute_cost = (NumDims - 1) * (TensorOpCost::AddCost<Index>() + + TensorOpCost::MulCost<Index>() + + TensorOpCost::DivCost<Index>()) + + TensorOpCost::MulCost<Index>(); + if (vectorized) { + compute_cost *= 2; // packet() computes two indices + } + const int innerDim = (static_cast<int>(Layout) == static_cast<int>(ColMajor)) ? 0 : (NumDims - 1); + return m_impl.costPerCoeff(vectorized && m_inputStrides[innerDim] == 1) + + // Computation is not vectorized per se, but it is done once per packet. + TensorOpCost(0, 0, compute_cost, vectorized, PacketSize); } EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; } |