@stdlib/randomand for most of
Tensorflow.js. So far
tfjsis great in terms of API but still several times slower than
numpynative C++ bindings. Is there any benchmark carrying out comparisons between
@alvinsunyixiao Yes, indeed there are benchmarks! For example, https://github.com/stdlib-js/stdlib/blob/b4d11aac1aa74441653f0ba195a184fd5cc51de0/lib/node_modules/%40stdlib/blas/ext/base/dcusum/benchmark/python/numpy/benchmark.py
Unfortuntely, atm, these benchmarks are scattered across the code base, being localized with their respective package, and need to be run locally. We’ve yet to add tooling for aggregating all the benchmarks across the project.
If there are particular features you need, we can work to prioritize those. Certainly some of the lower level building blocks.
a*x + yelement-wise: https://github.com/stdlib-js/stdlib/tree/8542ec63ba38530cdbe61d8ec86a09ff370c95e7/lib/node_modules/%40stdlib/blas/base/daxpy
ndarrayAPI wrappers, as done in the second link. If these would be useful, we can easily prioritize them this week.
stdlibto compile against a particular BLAS library (e.g., OpenBLAS). This allows for an order of magnitude speed-up against reference C implementations.
requirestdlib packages (e.g.,
@stdlib/math/base/special/erf), what version of
@stdlib/stdlibare you using?
@SindujaRajadurai Ah! I know what you did. You’ve installed the wrong package. You did
npm install stdlib
npm install @stdlib/stdlib
This project is published as a scoped package.