unrolled transform!

[jw3126]

Towards better performance. Roughly 3-4 times faster then baseline, still 2-3 times slower then nettle.

[oxinabox]

How do the if statements compare to using a series of loops?
I think the series of for loops should be faster

[jw3126]

These are compile time ifs. No runtime branches left.

[oxinabox]

add @inbounds to all of these?

[jw3126]

The whole body is @inbounds.

[oxinabox]

Or does it propagate correctly from the last line?
I'm never sure about @inbounds propergation

[jw3126]

It just does not propagate into function calls. But I think you are still right it is better style to be explict, if something is refactored into a function.

[oxinabox]

These are compile time ifs. No runtime branches left.

Ah yes, I just realized that.

[oxinabox]

You should add yourself to the LISCENSE.md

[oxinabox]

nettle is hand-optimized assembly.
I'm not sure how bad one can feel about losing to it.
I'm pretty sure it actually manages to not use RAM for anything but the message.
Keeping the A,B,C,D,F all in the CPU registers.

[jw3126]

I read this their baseline is similar to ours and the significant optimizations were unrolling and telling C to put variables into registers. (They do use some assembly, but it does not give much performance.) Not sure if there is a way to control register usage from Julia.

[oxinabox]

thanks, this is great
There might be a way to to register forcing via Core.Intrinsics.llvmcall but I'm not sure.
It can come in another PR if it matters.

MD5 checking isn't a massive bottleneck ever anyway.
Its basically only for legacy file checksums, right?

[jw3126]

Yeah performance is not critical here. My motivation is just curiosity.