experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Rely on the tensor cost model to figure out the best strategy to parallelize

Browse Source 
the transpose operation instead of using a hardcoded cutoff point.
Change: 122193123
This commit is contained in:
Benoit Steiner 2016-05-12 12:48:25 -08:00 committed by TensorFlower Gardener
parent b082c4f921
commit 313408ba1f
1 changed files with 1 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

9
tensorflow/core/kernels/transpose_functor_cpu.cc
View File

@ -56,14 +56,7 @@ void TransposeUsingEigen(const Device& d, const Tensor& in,
auto y = typename TTypes<T, NDIMS>::Tensor( auto y = typename TTypes<T, NDIMS>::Tensor(
reinterpret_cast<T*>(const_cast<char*>(out->tensor_data().data())), reinterpret_cast<T*>(const_cast<char*>(out->tensor_data().data())),
out->shape().AsEigenDSizes<NDIMS>()); out->shape().AsEigenDSizes<NDIMS>());
auto nelem = in.NumElements(); y.device(d) = x.shuffle(p);
static const int64 kInlineThreshold = 131072;
if (nelem * sizeof(T) < kInlineThreshold) {
// Don't bother multi-threaded transpose if 'in' is small.
y = x.shuffle(p);
} else {
y.device(d) = x.shuffle(p);
}
} }
} // end namespace internal } // end namespace internal