Hi,

I see in the multi-threaded implementation of the tensor contraction, that contractions where n==1 have parallelisation disabled because "the cost model is not tuned" (see https://bitbucket.org/eigen/eigen/commits/256b51362a2d9321c4ff2110ecf23cb8ef16a836#Lunsupported/Eigen/CXX11/src/Tensor/TensorContractionThreadPool.hT210).

I was wondering if this is still necessary - it's been 6 months since that commit and my Tensorflow code is 2x faster with the check commented out.

Regards,
Tim