Since existing context-reduction works mainly target the context encoding and decoding logic design, and apply a post-context-generation method, there is not much information on how the code is deployed onto the CGRA. For example, how the loop is converted to DFG may significantly influence the...
Ref. [42] uses HRNET as the backbone to acquire high-resolution global features without going through the decoding layer and combines the adaptive spatial pooling (ASP) module to collect and fuse the local information. Based on the attention mechanism, HMANet [43] makes use of the extended ...