What's the difference between Tiramisu and TACO compiler

[jinderek]

I searched through the internet but can not find the comparison. AFAIK, they are both tensor compiler.

[rbaghdadi]

They are complementary and they attack different problems.

Tiramisu is designed for optimizing dense loops using the polyhedral model. Its strength is in optimizing dense loops.

TACO is designed for generating loops from sparse tensor algebra expressions. So its strength is in generating the right loop and sparse data structure for your input tensor algebra expression.

While Tiramisu has basic support of sparse loops, it is not designed to generate the sparse loops themselves. You can use TACO on top of Tiramisu to do that.

While TACO has basic support of dense loop optimizations, it is not designed to do that (currently). You can use Tiramisu to optimize the dense loops generated by TACO.