I think it might also be a good idea to design a DSL for deep learning. I heard that Google wants to develop a special compiler for TensorFlow.
tclarke
They also have developed a specific chip for TensorFlow called a Tensor Processing Unit. This chip was used in the famous AlphaGo machine. This is more on the hardware side than software, but it goes to show how specific needs must sometimes be met with very specific domain-specific tools.
ferozenaina
Apart from the syntax, why can't we just modify the C++ compiler and optimize it for a particular application? If working with meshes, we could optimize it for parallelization and to access the 'stencil' of a node. Where is the line separating a customized compiler and a DSL?
I think it might also be a good idea to design a DSL for deep learning. I heard that Google wants to develop a special compiler for TensorFlow.
They also have developed a specific chip for TensorFlow called a Tensor Processing Unit. This chip was used in the famous AlphaGo machine. This is more on the hardware side than software, but it goes to show how specific needs must sometimes be met with very specific domain-specific tools.
Apart from the syntax, why can't we just modify the C++ compiler and optimize it for a particular application? If working with meshes, we could optimize it for parallelization and to access the 'stencil' of a node. Where is the line separating a customized compiler and a DSL?
An example I'm looking at is this slide