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

Fixes for the forced evaluation of tensor expressions

Browse Source 
More tests
This commit is contained in:
Benoit Steiner
2014-10-02 10:39:36 -07:00
parent 5cc23199be
commit 8b2afe33a1
4 changed files with 110 additions and 8 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
unsupported/Eigen/CXX11/src/Tensor/TensorForcedEval.h
View File

@@ -87,31 +87,28 @@ struct TensorEvaluator<const TensorForcedEvalOp<ArgType>, Device>
enum {
IsAligned = true,
PacketAccess = true,
PacketAccess = (internal::packet_traits<Scalar>::size > 1),
};
EIGEN_DEVICE_FUNC TensorEvaluator(const XprType& op, const Device& device)
: m_impl(op.expression(), device), m_op(op.expression()), m_device(device), m_buffer(NULL)
{ }
EIGEN_DEVICE_FUNC ~TensorEvaluator() {
eigen_assert(!m_buffer);
}
typedef typename XprType::Index Index;
typedef typename XprType::CoeffReturnType CoeffReturnType;
typedef typename XprType::PacketReturnType PacketReturnType;
EIGEN_DEVICE_FUNC const Dimensions& dimensions() const { return m_impl.dimensions(); }
EIGEN_STRONG_INLINE void evalSubExprsIfNeeded() {
m_impl.evalSubExprsIfNeeded();
EIGEN_STRONG_INLINE bool evalSubExprsIfNeeded(Scalar*) {
m_impl.evalSubExprsIfNeeded(NULL);
m_buffer = (Scalar*)m_device.allocate(m_impl.dimensions().TotalSize() * sizeof(Scalar));
typedef TensorEvalToOp<const ArgType> EvalTo;
EvalTo evalToTmp(m_buffer, m_op);
internal::TensorExecutor<const EvalTo, Device, TensorEvaluator<ArgType, Device>::PacketAccess>::run(evalToTmp, m_device);
m_impl.cleanup();
return true;
}
EIGEN_STRONG_INLINE void cleanup() {
m_device.deallocate(m_buffer);
@@ -129,6 +126,8 @@ struct TensorEvaluator<const TensorForcedEvalOp<ArgType>, Device>
return internal::ploadt<Packet, LoadMode>(m_buffer + index);
}
Scalar* data() const { return m_buffer; }
private:
TensorEvaluator<ArgType, Device> m_impl;
const ArgType m_op;