Code speedup
Closed this issue · 1 comments
There's a lot of low-hanging fruit in the optimization of the code. Most of the simulation time is spent on dict-related operations, such as __hash__ and __eq__ in particular. Re-writing some of the often-called and dictionary-heavy functions will greatly improve the runtime.
See #60 for WIP.
Currently, Zigzag has the concept of MemoryOperands (eg I1, I2) and LayerOperands (I, W). The idea behind this distinction is that users can create layers with different names for the layer operands. To map the layer operands to the memory operands, each layer has a MemoryOperandLinks instance to map layer to memory operands.
In practice, only the same two layer and memory operands occur, and the MemoryOperandLinks is always the exact same, namely the default. I don't know any use case of having a different layer-to-memory-operand mapping.
Dealing with this variable mapping from layer to memory operand costs a lot of execution time and complexity. There are many, many occurrences of this translation. The code could be significantly sped up by removing MemoryOperandLinks, merging the concept of memory and layer operands into one, and hardcoding the operands (e.g. layer.operand_1 or layer.weight_op ).
Are there any drawbacks in this approach? Am I missing use cases where a non-default ``MemoryOperandLinks` is needed?