AfterSessionExercises.txt

(1) In the slides, Sharad shows a simple matmul kernel.

(a) Time it! Do you get the perf you expect? (You certainly should on a v4 chip!)
(b) Normally we want to fuse in activation. Instead of A@B we could compute activation(A@B) where commonly activation is relu. Can you do this? How does it impact perf? How does it compare to jit(relu(A@B))?
(c) Can you make the kernel take arbitrary activations (without paying a perf penalty)?