Module reduce 

High-performance CPU memory reduction kernels.

Reductions are bandwidth-bound (read N elements, write 1 scalar). Single-threaded loops with LLVM auto-vectorization already saturate the memory bus. Rayon is not used.

Functions

max_all

Executes a global max reduction, finding the largest single value in the tensor.

mean_all

Executes a global mean reduction, calculating the average of all elements.

sum_all

Executes a global sum reduction, collapsing the entire tensor into a single scalar value.