Created by: tpatejko

MKLDNN supports operator fusion either through special primitives (now deprecated convolution_relu_forward), or through special post_ops API introduced in MKLDNN v0.12.

According to MKLDNN release notes, the following fusion scenarios are supported:

• forward convolution with eltwise for inference and training
• convolution with sum for inference
• batch normalization with eltwise for training

Does Fluid implement any fusion passes of existing GPU or CPU operators?

Some graph frameworks implement operator fusion using compiler passes for graph matching and graph rewriting. Does Fluid implement such infrastructure?

We would like to dive a bit deeper into graph compilation in Fluid. Where can we start?