Speed up.

[JosePineiro]

For each 16 bytes:

• Change 16 sums (+) for increments (++). Increment are faster then sum.
• Erase 1 subtract
• Erase (1 256) and (0256). Perhaps the compiler optimize this operations.

In precalc table Change one mult for one rotare right. (Only 16536 operations per execution.)

In my test speed up 8%

[force-net]

Hello Jose I've tested your code, it is slower ~5% About your changes:

• Increments can be faster than sum, but with this increments CPU unable to perform operations in parallel (every operation depends on previous)
• It seems, you've mean addition, it's removing does not give serious increase
• Compiler precalculate constants, so, 1 256, 8 32, 2 2 2 2 2 2 2 * 2 or 256 - it is same code, also 0 + x, also compiles to x

Precalc table can be faster, but it is hard to catch difference because it is calculates only 1 time per application run.

[JosePineiro]

Sorry,

I do not think in parallel excecution. I my machine (AMD four cores) this change give a fast code. I suspected that the compiler precalculate constants. I do not see speed up with the change.

But I think that the change in precalc table are faster. You can test it precalculating the table a thousand times in each execution.

[force-net]

Can you provide a difference between your code and current code? Currently I did not have any good AMD computer for tests.

Precalc table may be slightly faster (but it hard question, jit can optimize both variants to the same code) but there are only one calculation per app lifecycle, so, it some CPU cycles that do not do any observable performance.