mirror/eigen
2
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-03-07 18:27:40 +08:00
Code Issues Packages Projects Releases Wiki Activity

Fixed compilation error

Browse Source
This commit is contained in:
Benoit Steiner 2016-05-17 12:54:18 -07:00
parent ebf6ada5ee
commit 997c335970
1 changed files with 17 additions and 2 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

19
unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h
View File

@ -93,10 +93,16 @@ struct EvalRange {
evaluator.evalScalar(i);
}
}
static Index alignBlockSize(Index size) {
return size;
}
};
template <typename Evaluator, typename Index>
struct EvalRange<Evaluator, Index, true> {
static const int PacketSize = unpacket_traits<typename Evaluator::PacketReturnType>::size;
static void run(Evaluator* evaluator_in, const Index first, const Index last) {
Evaluator evaluator = *evaluator_in;
eigen_assert(last >= first);
@ -122,6 +128,15 @@ struct EvalRange<Evaluator, Index, true> {
evaluator.evalScalar(i);
}
}
static Index alignBlockSize(Index size) {
// Align block size to packet size and account for unrolling in run above.
if (size >= 16 * PacketSize) {
return (size + 4 * PacketSize - 1) & ~(4 * PacketSize - 1);
}
// Aligning to 4 * PacketSize would increase block size by more than 25%.
return (size + PacketSize - 1) & ~(PacketSize - 1);
}
};
template <typename Expression, bool Vectorizable>
@ -139,9 +154,9 @@ class TensorExecutor<Expression, ThreadPoolDevice, Vectorizable> {
const Index size = array_prod(evaluator.dimensions());
#if defined(EIGEN_USE_NONBLOCKING_THREAD_POOL) && defined(EIGEN_USE_COST_MODEL)
device.parallelFor(size, evaluator.costPerCoeff(Vectorizable),
EvalRange::alignBlockSize,
EvalRange<Evaluator, Index, Vectorizable>::alignBlockSize,
[&evaluator](Index first, Index last) {
EvalRange::run(&evaluator, first, last);
EvalRange<Evaluator, Index, Vectorizable>::run(&evaluator, first, last);
});
#else
size_t num_threads = device.numThreads();