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

33
unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h
View File

@@ -411,6 +411,9 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
typedef typename XprType::Scalar Scalar;
typedef TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device> Self;
static const bool InputPacketAccess = TensorEvaluator<ArgType, Device>::PacketAccess;
typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
static const int PacketSize = internal::unpacket_traits<PacketReturnType>::size;
enum {
IsAligned = false,
@@ -495,9 +498,6 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE const Dimensions& dimensions() const { return m_dimensions; }
typedef typename internal::remove_const<typename XprType::CoeffReturnType>::type CoeffReturnType;
typedef typename PacketType<CoeffReturnType, Device>::type PacketReturnType;
EIGEN_STRONG_INLINE EIGEN_DEVICE_FUNC bool evalSubExprsIfNeeded(CoeffReturnType* data) {
m_impl.evalSubExprsIfNeeded(NULL);
@@ -584,16 +584,15 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
template<int LoadMode>
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE PacketReturnType packet(Index index) const
{
const int packetSize = internal::unpacket_traits<PacketReturnType>::size;
EIGEN_STATIC_ASSERT(packetSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE)
eigen_assert(index + packetSize - 1 < dimensions().TotalSize());
EIGEN_STATIC_ASSERT(PacketSize > 1, YOU_MADE_A_PROGRAMMING_MISTAKE)
eigen_assert(index + PacketSize - 1 < dimensions().TotalSize());
EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[packetSize];
EIGEN_ALIGN_MAX typename internal::remove_const<CoeffReturnType>::type values[PacketSize];
if (ReducingInnerMostDims) {
const Index num_values_to_reduce =
(static_cast<int>(Layout) == static_cast<int>(ColMajor)) ? m_preservedStrides[0] : m_preservedStrides[NumPreservedStrides - 1];
const Index firstIndex = firstInput(index);
for (Index i = 0; i < packetSize; ++i) {
for (Index i = 0; i < PacketSize; ++i) {
Op reducer(m_reducer);
values[i] = internal::InnerMostDimReducer<Self, Op>::reduce(*this, firstIndex + i * num_values_to_reduce,
num_values_to_reduce, reducer);
@@ -602,18 +601,18 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
const Index firstIndex = firstInput(index);
const int innermost_dim = (static_cast<int>(Layout) == static_cast<int>(ColMajor)) ? 0 : NumOutputDims - 1;
// TBD: extend this the the n innermost dimensions that we preserve.
if (((firstIndex % m_dimensions[innermost_dim]) + packetSize - 1) < m_dimensions[innermost_dim]) {
if (((firstIndex % m_dimensions[innermost_dim]) + PacketSize - 1) < m_dimensions[innermost_dim]) {
Op reducer(m_reducer);
typename Self::PacketReturnType accum = reducer.template initializePacket<typename Self::PacketReturnType>();
internal::InnerMostDimPreserver<NumReducedDims-1, Self, Op>::reduce(*this, firstIndex, reducer, &accum);
return reducer.finalizePacket(accum);
} else {
for (int i = 0; i < packetSize; ++i) {
for (int i = 0; i < PacketSize; ++i) {
values[i] = coeff(index + i);
}
}
} else {
for (int i = 0; i < packetSize; ++i) {
for (int i = 0; i < PacketSize; ++i) {
values[i] = coeff(index + i);
}
}
@@ -621,6 +620,18 @@ struct TensorEvaluator<const TensorReductionOp<Op, Dims, ArgType>, Device>
return rslt;
}
// Must be called after evalSubExprsIfNeeded().
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE TensorOpCost costPerCoeff(bool vectorized) const {
if (RunningFullReduction && m_result) {
return TensorOpCost(sizeof(CoeffReturnType), 0, 0, vectorized, PacketSize);
} else {
const Index num_values_to_reduce = internal::array_prod(m_reducedDims);
const double compute_cost = num_values_to_reduce * internal::functor_traits<Op>::Cost;
return m_impl.costPerCoeff(vectorized) * num_values_to_reduce +
TensorOpCost(0, 0, compute_cost, vectorized, PacketSize);
}
}
EIGEN_DEVICE_FUNC Scalar* data() const { return NULL; }
private: