Keccak perf

[benaadams]

Changes

• Avoid the copy in and copy out step by operating over same data. Since there are so many fields they spill to stack anyway (and is stackalloc'd array anyway to start), so operate with stack and remove two 200 byte copies, or 50 ulongs of memory shuffling

KeccakF asm

Before

; Total bytes of code 2606

After

; Total bytes of code 1942

Types of changes

What types of changes does your code introduce?

• Optimization

Testing

Requires testing

• No

[LukaszRozmej]

Do you have any benchmarks?

[LukaszRozmej]

Do you have any - even simple benchmark?

[benaadams]

Do you have any - even simple benchmark?

I benchmarked it with Paprika's benchamarks and it was pretty much the same 😢

There are two many varibles and too few registers so it spills to stack quite a lot and is very very sensitive to instruction order and data dependencies formed