Can't generate instructions for ARMv7 NEON with GCC 7.4.1

[zambony]

I'm new to SIMD usage, so I'm a bit unsure if I've missed something.

I am trying to cross-compile an application from my x86 linux desktop to an SoC using ARMv7, but Highway's target detection seems to be using scalar code (from what I see in the assembly output).

The SoC's info, from cat /proc/cpuinfo:

processor       : 0
model name      : ARMv7 Processor rev 0 (v7l)
BogoMIPS        : 666.66
Features        : half thumb fastmult vfp edsp neon vfpv3 tls vfpd32
CPU implementer : 0x41
CPU architecture: 7
CPU variant     : 0x3
CPU part        : 0xc09
CPU revision    : 0

processor       : 1
model name      : ARMv7 Processor rev 0 (v7l)
BogoMIPS        : 666.66
Features        : half thumb fastmult vfp edsp neon vfpv3 tls vfpd32
CPU implementer : 0x41
CPU architecture: 7
CPU variant     : 0x3
CPU part        : 0xc09
CPU revision    : 0

Hardware        : Xilinx Zynq Platform
Revision        : 0003

I'm using this for my CMake to enable NEON

target_compile_options(application PRIVATE -march=armv7-a -mfpu=neon -mfloat-abi=hard)

My compiler:

arm-linux-gnueabihf-g++ (Linaro GCC 7.4-2019.02) 7.4.1 20181213 [linaro-7.4-2019.02 revision 56ec6f6b99cc167ff0c2f8e1a2eed33b1edc85d4]
Copyright (C) 2017 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

I have enabled -DHWY_CMAKE_ARM7:BOOL=ON, too.

However, HWY_NAMESPACE seems to refer to N_SCALAR, as in the assembly.

simd::N_SCALAR::myFunction(...)

The performance of the SIMD code is also worse than the hand-written scalar version, so I am pretty sure it's not working correctly.

Have I missed some setup? Is this a limitation of the old version of GCC I'm using?

[johnplatts]

The HWY_NEON_WITHOUT_AES target on Armv7 requires support for VFPv4 in addition to NEON.

[zambony]

Oh, darn. Is this a limitation of Highway? As in, could I use the intrinsics manually? I'm trying to operate on a large dataset of doubles, so I was hoping to somehow leverage SIMD. If I have to write it by hand I guess I'll do it.

[jan-wassenberg]

You could try removing the check for HWCAP_VFPv4 in targets.cc and seeing if it works for your application.
I think the main things missing in VFPv3 are fp16 and FMA support. FMA is quite important, and I have seen VFPv3 so rarely, that I doubt we'd remove the requirement. Also, ArmV7 does not have vectorized f64 instructions, which is exactly what you want - I think the SoC is too old for this use case :)

[zambony]

Also, ArmV7 does not have vectorized f64 instructions, which is exactly what you want - I think the SoC is too old for this use case :)

I see :( I'll have to come up with a different solution then. Thanks for the responses.