Re: ARM NEON Intrinsics guide?

[Kyrill Tkachov <kyrylo dot tkachov at foss dot arm dot com>]

Hi,

On 11/05/16 07:03, Jeffrey Walton wrote:
> Hi Everyone,
>
> I'm working on a port of some software to ARM v7, v7-a and v8-a, which
> means SIMD coprocessor/NEON is available.
>
> The performance increases when the data types are 32-bit using
> int32x4_t vectors, but its not quite what I expected. As a
> measurement, Clang produces code near the theoretical throughput, and
> its about 2 cycles per byte faster than GCC. But GCC NEON is better
> than the C/C++ measurements.
>
> I'm not seeing the performance gain I expected when the data types are
> 64-bit using int64x2_t vectors. GCC produces code that's nowhere near
> Clang or the theoretical limit. In fact, its slower than the C/C++
> code.

What does the generated code look like for GCC vs Clang?
Is it just instruction scheduling differences or is there
more register movement, stack spilling etc?
A bug report with a standalone example would be very helpful.

> I feel like I'm missing something with respect to using GCC and NEON
> intrinsics. For example, maybe there's a technique or pattern to help
> GCC minimizes stalls.


> For completeness, here are the CFLAGS I am using:
>
>      ARMv7:
>          -DNDEBUG -O3 -march=armv7-a -mfpu=neon
>          -DNDEBUG -O3 -mcpu=cortex-a7 -mfpu=neon
>
>      ARMv8:
>          -DNDEBUG -O3 -O3 -march=armv8-a+crc -mcpu=cortex-a53
>          -DNDEBUG -O3 -O3 -march=armv8-a+crc -mcpu=cortex-a53 -mfpu=neon
>
> Is there a GCC guide for NEON intinsics?

As has been mentioned on this thread already,
http://infocenter.arm.com/help/topic/com.arm.doc.ihi0073a/IHI0073A_arm_neon_intrinsics_ref.pdf
is a list of the intrinsics and how they map down to NEON instructions, thought it's
more of a reference rather than a user guide.

If you can isolate a standalone example where GCC NEON intrinsics perform poorly it can you
please file a bug report with the testcase.

As an aside, I notice your command-line options are sub-optimal.
If you're targeting a Cortex-A7  you want to use -mfpu=neon-vfpv4 rather than just -mfpu=neon.
This will give you access to the vfma instructions.
Whereas if you're targeting ARMv8-A on a Cortex-A53 you'll want to use -mfpu=neon-fp-armv8
to enable the ARMv8 floating-point an NEON instructions.

Cheers,
Kyrill