Noeda / htensorizer

Mikko's dumb optimizing Tensor library

Mikko's stupid optimizing Tensor Library

yes good tensors mm yes

Goals

• An interpreted Haskell backend
• A Vulkan backend
• Enough features to run some neural networks
• Competetive performance compared to torch and some graph-based tensor libraries
• Should work on NVidia cards, CPU and M1.
• If there's a Vulkan backend, would it work through MoltenVk on Apple?
• If I find a way to test on AMD card, I will test on AMD card
• Idiomatic-Haskell type API on top of the "program" API I started