search

cinecert / asdcplib

AS-DCP and AS-02 File Access Library
Other
70 stars 54 forks source link

AES: use EVP functions instead of deprecated functions #133

Open ArnaudBienner opened 8 months ago

ArnaudBienner commented 8 months ago

Among other things, they use hardware acceleration (like AES-NI) which brings a significant performance improvement.

From my tests on an Intel CPU, this makes unwrapping of encrypted MXF ~20x faster.

I started to look at the HMAC part too, but it looks like it is already using hardware acceleration and it didn't bring any performance improvement, so I didn't updated that part for now, despite the functions being deprecated too.

Jason-elkins commented 7 months ago

@ArnaudBienner, What is the cpu count for your benchmark?

ArnaudBienner commented 7 months ago

@ArnaudBienner, What is the cpu count for your benchmark?

4 cores, this runs only on one core (either the old version or the new one with one with my patch) as asdcplib is not multithreaded.

Jason-elkins commented 7 months ago

@ArnaudBienner, What is the cpu you are using? (so we can make sure we're using the same instruction set).

ArnaudBienner commented 7 months ago

@ArnaudBienner, What is the cpu you are using? (so we can make sure we're using the same instruction set).

I tried on Intel CPU: in this case openssl uses AES-NI instructions set.

The information about whether or not AES-NI is available for a given processor is available on Intel website. e.g. for i7-7920HQ "Intel® AES New Instructions Yes"

So using a modern Intel CPU, which has AES-NI, you should be able to see a difference in performance between the old and new implementation.

AFAICT all recent Intel CPU now have AES-NI.

Note that it is also possible to deactivate the hardware acceleration in openssl through environment variable. See https://www.openssl.org/docs/manmaster/man3/OPENSSL_ia32cap.html for more details.

For instance, to disable Intel AES-NI you can do

export OPENSSL_ia32cap="~0x200000200000000"

before executing asdcp-unwrap or asdcp-wrap and you will be able to notice a difference with and without the hardware acceleration.

It's likely using openssl EVP functions provides other acceleration for other type of CPU (e.g. ARM) but I didn't test those.