|
|
|
|
|
|
by carterschonwald
4578 days ago
|
|
|
indeed. The generalized stream fusion work is about lifting pointwise operations to their SIMD vectorized analogue. Likewise, stream fusion isn't always a win! Stream fusion and its siblings are great for pointwise operations, but aren't always a good idea for computations with heavy reuse, like nested convolutions or performant matrix multipy |
|
|