Compilers leverage the knowledge of computational graph to fuse operators mainly to eliminate memory loading operations. A common example is fusing batchnorm into the preceding convolution layer. Batch Normalization can be implemented as a matrix-vector operation, which can then be implemented as a 1×1 convolution.
Operation Fusion
Creator
Creator
Seonglae ChoCreated
Created
2023 Oct 29 7:20Editor
Editor
Seonglae ChoEdited
Edited
2023 Oct 29 8:7Refs
Refs
Kernel Fusion