Package does not use hardware acceleration

[jakobnissen]

Python's hashlib.sha256(data).hexdigest() is 10x faster than SHA.jl's sha256(data) on my computer:

In [8]: %timeit hashlib.sha256(data).hexdigest()
53.8 µs ± 6.88 ns per loop (mean ± std. dev. of 7 runs, 10,000 loops each)

julia> @benchmark sha256(data)
BenchmarkTools.Trial: 8884 samples with 1 evaluation.
 Range (min … max):  514.644 μs … 913.010 μs  ┊ GC (min … max): 0.00% … 0.00%
 Time  (median):     518.341 μs               ┊ GC (median):    0.00%
 Time  (mean ± σ):   561.483 μs ±  99.395 μs  ┊ GC (mean ± σ):  0.00% ± 0.00%

This is presumably (haven't checked though) because Python's underlying C library uses SHA-specific CPU instructions, where as this package does not.
This might not be possible to fix in current versions of Julia, but given that we see a 10x performance difference, this is something Julia needs to support eventually (and probably will). When that time comes, this package should be updated.

This issue was discovered by @jonalm

[inkydragon]

Python is currently migrating some of the hash algorithms to use the HACL* project, and the rest of the algorithms use the OpensSSL implementation.

python/cpython#99108

Changed in version 3.9: Hashlib now uses SHA3 and SHAKE from OpenSSL if it provides it.

Changed in version 3.12: For any of the MD5, SHA1, SHA2, or SHA3 algorithms that the linked OpenSSL does not provide we fall back to a verified implementation from the HACL* project.
https://docs.python.org/3/library/hashlib.html#hash-algorithms

HACL* is a formally verified library of modern cryptographic algorithms written in a subset of F* called Low* and compiled to C using a compiler called KaRaMeL.

HACL* is released in pure C, so we can wrap it and test the performance.

---

I'd prefer to keep the pure julia implementation and then have the fast impl. as a separate package, maybe called HashlibAcc.jl. The default is to use the pure julia impl., and if users need a faster impl., they can import the accelerated package.

The main concern is that introducing a third-party dependency may increase the maintenance burden.

[jakobnissen]

Would it not be best of both worlds to use hardware acceleration in pure Julia?