How to make use of neon co processor for openssl?

Rakshith rao

I want to use the neon coprocessor. I have to compile openssl library for neon and use the library with my application. To enable neon, the procedure is described here. There it is given that, a set of commands has to be given to the system to ennable neon. But I din't get when and how to use it. Also, Is this auto vectorization possible with openssl? Any idea?

over 8 years ago

0 Biser Gatchev-XID over 8 years ago

TI__Guru**** 393215 points

Hi,

I will forward this to the SW team.

0 Georgi Karpov over 8 years ago

Expert 1185 points

Hi,

To enable NEON/VFP you must use corresponding co-processor commands for your compiler, as is described in the link.
Set of commands can be added to the kernel code in init/main.c using asm() .

Vectorizable Loops for Auto-vectorization are shown here:
gcc.gnu.org/.../vectorization.html

So, if openssl code corresponds to these examples, Auto-vectorization will be applied.

BR,
Georgi

0 Rakshith rao over 8 years ago in reply to Georgi Karpov

Intellectual 390 points

Thank you. This is for auto vectorization. Is there any implementation/support for neon intrinsic code for openssl?

0 Rakshith rao over 8 years ago in reply to Georgi Karpov

Intellectual 390 points

Hi,
While building the kernel with those co processor command I got the error message "selected processor does not support ARM mode `fmxr FPEXC,r2'" What does that mean? Could you please clarify it? Moreover are those assembly commands necessary?

0 Rakshith rao over 8 years ago in reply to Georgi Karpov

Intellectual 390 points

I tried those commands to enable neon co processor. But application is not giving any better performance though I compiled openssl for neon with essential flags.

0 Georgi Karpov over 8 years ago in reply to Rakshith rao

Expert 1185 points

Hi,

can you describe how you test your openssl performance?

Georgi

0 Rakshith rao over 8 years ago in reply to Georgi Karpov

Intellectual 390 points

I compiled openssl for neon. Then I wrote a sample code to encrypt and decrypt a file. The time taken for both the operations are same to that of normal openssl (which was not compiled for neon).

0 Georgi Karpov over 8 years ago in reply to Rakshith rao

Expert 1185 points

Have you tried it with other applications with Auto-vectorization, to see if there is effect after enabling neon?

0 Rakshith rao over 8 years ago in reply to Georgi Karpov

Intellectual 390 points

I wrote a sample program which has a huge loop operation. I compiled the program with "-O3 -mfpu=neon -ftree=vectorize" and with no compiler directives. In the first case, the time taken was 1/3rd of the later. But I think it is because of -O3 command(compiler optimization) rather than the neon optimization(SIMD optimization).
By the way, to enable neon, I have to get into the supervisor/privileged mode. I think, it is not happening right now. So I am looking into it.

0 Joe G. over 8 years ago in reply to Rakshith rao

TI__Mastermind 25686 points

Some good advice on Neon here:
processors.wiki.ti.com/.../Cortex_A8

Also a biquad example and screenshot here:
e2e.ti.com/.../1539592

0 Rakshith rao over 8 years ago in reply to Joe G.

Intellectual 390 points

The first link I had referred before. To executed those commands. Especially the command "FMXR FPEXC, r0" needs privileged mode to be executed.

Processors

Processors forum

How to make use of neon co processor for openssl?