NEON detection on aarch64 / TX1

[dusty-nv]

Torch is failing to build on ARMv8 (Tegra X1) under Ubuntu 16.04 aarch64.
During configuration:
-- Could not find hardware support for NEON on this machine.

Then while building:

[ 46%] Building C object lib/TH/CMakeFiles/TH.dir/THVector.c.o
In file included from /tmp/luarocks_torch-scm-1-6670/torch7/lib/TH/THVector.c:2:0:
/tmp/luarocks_torch-scm-1-6670/torch7/lib/TH/generic/THVectorDispatch.c: In function ‘THByteVector_vectorDispatchInit’:
/tmp/luarocks_torch-scm-1-6670/torch7/lib/TH/generic/simd/simd.h:60:3: error: impossible constraint in ‘asm’
   asm volatile ( "cpuid\n\t"
   ^
/tmp/luarocks_torch-scm-1-6670/torch7/lib/TH/generic/simd/simd.h:60:3: error: impossible constraint in ‘asm’
   asm volatile ( "cpuid\n\t"
   ^
lib/TH/CMakeFiles/TH.dir/build.make:350: recipe for target 'lib/TH/CMakeFiles/TH.dir/THVector.c.o' failed

It looks like FindARM.cmake searches /proc/cpuinfo for 'neon'. However, ARMv8 output is the following:

Features        : fp asimd aes pmull sha1 sha2 crc32
CPU implementer : 0x41
CPU architecture: 8
CPU variant     : 0x1
CPU part        : 0xd07
CPU revision    : 1
Hardware        : jetson_tx1

It appears like other [1][2] ARMv8 platforms may also list asimd in leui of neon in their /proc/cpuinfo.

[dusty-nv]

Note: after manually patching FindARM.cmake to set NEON_TRUE and omitting -mfpu=neon (as mentioned here) from CMakeLists, I encounter the same errors from issue 763. Anyone know a workaround for now?

[soumith]

@dusty-nv #763 is now fixed in master

[soumith]

if you can send a patch to FindARM.cmake to work for armv8, it would be much appreciated.

[dusty-nv]

Thanks Soumith. Try these patches below for FindARM.cmake and lib/TH/CMakeLists are a bit crude, because I still define NEON cflags directive on ARMv8 (but without -mfpu=neon) so it compiles the rest of the code normally:

FindARM-patch.txt
CMakeLists-patch.txt

Let me know if/when you commit to master so I can re-build from master.

[rtarquini]

Did a pull today 10/14/2016. There is still are problems on the X1 build.

(1) /pkg/torch/lib/TH/CMakeLists.txt
FIND_PACKAGE(ARM) - does not result in NEON_FOUND.
i.e the default case does not set -D__NEON___
- I'm only building for the X1/64 bit, so I just always define this
(2) /pkg/torch/lib/TH/vector/NEON.c
Assembly instructions are invalid for ARM64.
- reverted to generic 'C' functions from previous ThVector.c
I can post this file if interested.
(3) Build processes are being killed. (Low memory?)

Not sure how to resolve #3? Ideas?

-Rich

[rtarquini]

Replacing the ../luarocks make with ../luarocks install resolves the memory issues with the problematic rocks such as cutorch. The rock specs in the distribution need to be refreshed.

[soumith]

@rtarquini fixed via torch/cutorch@ac40c05

[dusty-nv]

@rtarquini can you post your ThVector.c? I'm getting invalid assembly instructions too after forcing NEON_FOUND with latest pull.

[rtarquini]

ThVector.c includes ./vector/NEON.c. That is the file with the mods.

NEON.c.txt

[dusty-nv]

Thanks @rtarquini, I got master building on aarch64 again with the file. On TX1 it still requires swap and enough disk space in /tmp for me with jopts=1, when cutorch begins compiling large templates in THCTensorMathPointwise.cu.

[Atcold]

Installation on TX1 still breaks (fixed with the two patches and the new NEON.c) and it's taking ages (it's been 4 hours that I've been trying to install cutorch). Installing an older version does take a minute or two.
For the swap, I just plugged in a SSD as "additional RAM".
This is a nightmare, especially with deadlines coming up too soon.

[hit1001]

While trying out on aarch64 platform i am still getting these errors.

[Atcold]

@hit1001, you need to cd into torch7 repo, and patch <CMakeLists-patch.txt, replace lib/TH/vector/NEON.c with NEON.c, cd into torch7/lib/TH/cmake/ and run patch <FindARM-patch.txt.
You will also need to create a swap file. See this for a nice tutorial about the how.

[soumith]

@Atcold if you hav ea solution to this, can you please send in a PR? I dont have an AArch64 to test out, and hence was unable to fix it myself.

[Atcold]

Oh, OK. I didn't know @soumith.
The problem is that it takes one whole night to install cutorch. I'm not sure this is a feasible solution. Moreover, one needs an external SSD drive (swapping on USB is a suicide).

[hit1001]

@Atcold, i am trying to install torch/distro. so the files you mentioned lies in /pkg/torch/TH ..
The thing are torch7 and cutorch additional packages apart from those in directory pkg? after patching it still doesnt build up for me.
If yes, how do i install them? @dusty-nv :_ Do i follow the CMakePrebuild.sh for torch7?

[Atcold]

@hit1001, this is my ~/torch/pkg/torch repo on my TX1.

You can simply apply this patch.txt.
Let me know if you need further assistance.

Here is my current setup.

[soumith]

@Atcold i would appreciate it if you send a PR so that folks dont need to manually apply this patch. I cant create the PR, as I cannot test the ARM64 side.
Thanks.

[Atcold]

@soumith, I am not the author of the code. I've just used @dusty-nv and @rtarquini work. Shall I send a PR citing the authors? I am also not aware if this breaks anything else.
Please, instruct.

[soumith]

yes, please do. we will catch other issues in contbuild