ProdEnvMatA

[npiroozan]

Good Afternoon,

I hope you are doing well.

I noticed something interesting in the Tensorboard trace using se_e2_a as the descriptor. After tuning OMP, TF_Intra_OP, and TF_Inter_OP Parallelism a remaining bottleneck is ProdEnvMatA. Attached please find the respective tensorboard trace.

This was run on a CPU node with 8 workers and a local batch size of 16. As one can see, all tf_computes are held up initially until ProdEnvMatA is complete. May I please ask 2 questions regarding this op:

1. What is the function of this Op.
2. Is there any way to reduce its compute time?

Thank you very much!

[wanghan-iapcm]

@denghuilu Please take a look a the question. Thanks!

[njzjz]

This Op computes the environmental matrix. See eq. 5 in this page.

I strongly agree that it can be further optimized.

[npiroozan]

I see, thank you. Perhaps it can be parallelized in the same manner as ProdForceSeA?