[Fix] Compiling Neural Network models with JITCPUTensor especially when reverse mode is signifcantly slow

hikettei / cl-waffe2

[Experimental] Graph and Tensor Abstraction for Deep Learning all in Common Lisp

https://hikettei.github.io/cl-waffe2/

MIT License

133 stars 5 forks source link

[Fix] Compiling Neural Network models with JITCPUTensor especially when reverse mode is signifcantly slow #53

Closed hikettei closed 1 year ago

hikettei commented 1 year ago

[Fix] Compiling time of simple MLP models with reverse mode.
[Added] the new package cl-waffe2/benchmark

JITCPUTensor has a lot of room to improve: In fact, it doesn't use parallel for large-scale matrices, and, Inability to support complex Views. mathematical functions (e.g.: sin/cos/tan) do not use SIMD. (tensor-ptr tensor) function should be much more improved... should be done at top-level , and at least pinning is (to avoid being moved by GC) must.