Created by: zhangting2020

PR types

Performance optimization

PR changes

OPs

Describe

IndexList in Eigen is used to encode a set of Tensor dimensions/indices. The indices in the list can be known at compile time or at runtime. A mix of static and dynamic indices can also be provided if needed. The tensor code will attempt to take advantage of the indices that are known at compile time to optimize the code it generates. Using IndexList instead of arrays of indices can speed up CPU and GPU performance.

Note: This functionality requires a c++11 compliant compiler. If the compiler is older we need to use arrays of indices instead. That's why the EIGEN_HAS_INDEX_LIST is used in the code.

Performance

CPU：

op	input shape	before	after	speed up
instance_norm	[1, 64, 128, 128]	41.2712 ms	3.21222 ms	13x
instance_norm_grad	[1, 64, 128, 128]	149.14 ms	11.949 ms	12x
instance_norm	[1, 128, 64, 64]	20.6193 ms	1.61026 ms	13x
instance_norm_grad	[1, 128, 64, 64]	74.4767 ms	5.19748 ms	14x
instance_norm	[1, 256, 32, 32]	10.308 ms	0.821658 ms	13x
instance_norm_grad	[1, 256, 32, 32]	37.1751 ms	2.60926 ms	14x