3 ms·
I thought that NumPy is already supposed to use "BLAS and company" internally and that BLAS is already as optimized as possible. What am I missing or misunderst
by krastanov 11y ago
I thought that NumPy is already supposed to use "BLAS and company" internally and that BLAS is already as optimized as possible. What am I missing or misunderstanding?
- Veedrac 11y agoI'm not sure what the BLAS story in PyPy is, but manual looping over elements is not going to work like that, and SIMD will help significantly there. If you're only doing broadcasted operations, there's not much draw in PyPy over CPython.
- SixSigma 11y agoThat you don't need Numpy or even python. This is about the Vm JIT using simd if it can
- srean 11y agoEven when you use BLAS calls there are some inefficiencies that remain. There can be (i) no optimization across BLAS calls, sometimes these can help a lot (ii) the loops or code containing the BLAS call can benefit significantly from this