Created by: jczaja
Max Sequence Pool is filling index Tensor that is utilized in relevant grad op. However this is not needed when is_test is set to true eg. we are performing inference/scoring.
Notes:
- After optimization operator is ~10% faster (on CAPI test_seq_conv1 model, one thread).
- To take advantage of this optimization is_test has to be set to true eg. model that CAPI is fetching may need to be modified to have such an attribute.
- Changes are introducing extension of sequence_pool API with is_test
- Pooling ops (both native PaddlePaddle and MKL-DNN one) should undergo similar optimizations to get performance boost.