devtools/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2026-04-10 11:34:33 +08:00
Code Issues Packages Projects Releases Wiki Activity

Eigen cost model part 1. This implements a basic recursive framework to estimate the cost of evaluating tensor expressions.

Browse Source
This commit is contained in:
Rasmus Munk Larsen
2016-04-14 13:57:35 -07:00
parent 3551dea887
commit 235e83aba6
28 changed files with 695 additions and 185 deletions
Download Patch File Download Diff File Expand all files Collapse all files

46
unsupported/Eigen/CXX11/src/Tensor/TensorConvolution.h
View File

@@ -297,6 +297,11 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
typedef typename XprType::Index Index;
typedef DSizes<Index, NumDims> Dimensions;
typedef typename XprType::Scalar Scalar;
typedef typename XprType::CoeffReturnType CoeffReturnType;
typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
enum {
IsAligned = TensorEvaluator<InputArgType, Device>::IsAligned & TensorEvaluator<KernelArgType, Device>::IsAligned,
PacketAccess = TensorEvaluator<InputArgType, Device>::PacketAccess & TensorEvaluator<KernelArgType, Device>::PacketAccess,
@@ -367,10 +372,6 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
}
}
typedef typename XprType::Scalar Scalar;
typedef typename XprType::CoeffReturnType CoeffReturnType;
typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) {
@@ -405,7 +406,6 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
template<int LoadMode>
EIGEN_DEVICE_FUNC PacketReturnType packet(const Index index) const
{
const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
Index indices[2] = {index, index+PacketSize-1};
Index startInputs[2] = {0, 0};
if (static_cast<int>(Layout) == static_cast<int>(ColMajor)) {
@@ -448,6 +448,23 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
}
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
costPerCoeff(bool vectorized) const {
const double kernel_size = m_kernelImpl.dimensions().TotalSize();
// We ignore the use of fused multiply-add.
const double convolve_compute_cost =
TensorOpCost::AddCost<Scalar>() + TensorOpCost::MulCost<Scalar>();
const double firstIndex_compute_cost =
NumDims *
(2 * TensorOpCost::AddCost<Index>() + 2 * TensorOpCost::MulCost<Index>() +
TensorOpCost::DivCost<Index>());
return TensorOpCost(0, 0, firstIndex_compute_cost, vectorized, PacketSize) +
kernel_size * (m_inputImpl.costPerCoeff(vectorized) +
m_kernelImpl.costPerCoeff(vectorized) +
TensorOpCost(0, 0, convolve_compute_cost, vectorized,
PacketSize));
}
EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
private:
@@ -1044,6 +1061,25 @@ struct TensorEvaluator<const TensorConvolutionOp<Indices, InputArgType, KernelAr
return internal::ploadt<PacketReturnType, LoadMode>(m_buf+index);
}
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost
costPerCoeff(bool vectorized) const {
// TODO(rmlarsen): FIXME: For now, this is just a copy of the CPU cost
// model.
const double kernel_size = m_kernelImpl.dimensions().TotalSize();
// We ignore the use of fused multiply-add.
const double convolve_compute_cost =
TensorOpCost::AddCost<Scalar>() + TensorOpCost::MulCost<Scalar>();
const double firstIndex_compute_cost =
NumDims *
(2 * TensorOpCost::AddCost<Index>() + 2 * TensorOpCost::MulCost<Index>() +
TensorOpCost::DivCost<Index>());
return TensorOpCost(0, 0, firstIndex_compute_cost, vectorized, PacketSize) +
kernel_size * (m_inputImpl.costPerCoeff(vectorized) +
m_kernelImpl.costPerCoeff(vectorized) +
TensorOpCost(0, 0, convolve_compute_cost, vectorized,
PacketSize));
}
private:
// No assignment (copies are needed by the kernels)
TensorEvaluator& operator = (const TensorEvaluator&);