gotta say, having written some scientific computing code, the libraries out there do not always cover the exact operation you need and are not always using the best algo in the literature. i was able to beat the existing ecosystem 6x head to head and thousands of times faster for my use case. ymmv ofc depending on the problem.

that said, it was not easy!!

I worked on some proprietary video/image encoding application. In that context we hand wrote things like colour space conversions, wavelet transforms, arithmetic coders, compression algorithms, even hashing functions, in SIMD and we got better performance than anything off the shelf. We still used some off the shelf code where we could (e.g. Intel's hand written libraries). The thing is that this was the core of our business and our success depended on how performant these pieces were. That was also some time back, maybe today the situation is different. In this sort of situation you should absolutely put in the effort. But that typically accounts for some small % of the overall software you're going to be a user of. This is really just another variation of the premature optimization statement: "We should forget about small efficiencies, say about 97% of the time: premature optimization is the root of all evil. Yet we should not pass up our opportunities in that critical 3%". So if you're in the 3% then by all means go for it (you gotta). But if you're in the 97% it's silly to not use what's out there (other than for fun, learning etc.)