Skip to content

Error compiling on ARM with -mfpu=neon: can’t find a register in class ‘LO_REGS’ while reloading ‘asm’ #5303

New issue
New issue
Closed
Closed
Error compiling on ARM with -mfpu=neon: can’t find a register in class ‘LO_REGS’ while reloading ‘asm’#5303
Assignees
maciekcc
Labels
type:build/installBuild and install issues
@dprylipko

Description

@dprylipko
dprylipko
opened on Oct 31, 2016

Hi,

I am trying to compile TF for an ARM device and to exploit NEON-related optimizations.
Basically, I follow these instructions for PI (my device is a Sabre Board, not PI): https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/makefile

but with a bit different flags, namely:
make -f tensorflow/contrib/makefile/Makefile OPTFLAGS="-Ofast -mfpu=neon -funsafe-math-optimizations -ftree-vectorize" HOST_OS=LINUX

because my CPU does not support vfpv4:

cat /proc/cpuinfo
processor	: 0
model name	: ARMv7 Processor rev 10 (v7l)
BogoMIPS	: 7.54
Features	: half thumb fastmult vfp edsp neon vfpv3 tls vfpd32
CPU implementer	: 0x41
CPU architecture: 7
CPU variant	: 0x2
CPU part	: 0xc09
CPU revision	: 10

The error I got:

/root/tensorflow_exp/tensorflow/contrib/makefile/downloads/gemmlowp/meta/streams_arm_32.h: In static member function ‘static void gemmlowp::meta::GemmExecutorPackLHS::ExecuteDispatch3D(const P&) [with P = gemmlowp::meta::GemmParams<unsigned char, int, gemmlowp::meta::ColumnMajorWithSum, gemmlowp::meta::RowMajorWithSum, gemmlowp::meta::QuantizedStaticPreprocessedAsInt32, gemmlowp::meta::RowMajor>; int m = 1; int n = 8; int k = 8; int m_leftovers = 0; int n_leftovers = 7; int k_leftovers = 4]’:
/root/tensorflow_exp/tensorflow/contrib/makefile/downloads/gemmlowp/meta/streams_arm_32.h:4211:59: error: can’t find a register in class ‘LO_REGS’ while reloading ‘asm’
         "d25", "d26", "d27", "d28", "d29", "cc", "memory");
                                                           ^
/root/tensorflow_exp/tensorflow/contrib/makefile/downloads/gemmlowp/meta/streams_arm_32.h:4211:59: error: ‘asm’ operand has impossible constraints
make: *** [/root/tensorflow_exp/tensorflow/contrib/makefile/gen/obj/tensorflow/core/kernels/meta_support.o] Error 1

As far as I understand, it complains about too little registers, which wonders me, since ARM has 16 registers. If I remove -mfpu=neon flag, everything works like a charm.

I would greatly appreciated for any suggestions.

Environment info

Ubuntu 14.04.4 LTS
Linux sabresd 4.1.15-1.0.0+g3924425 #1 SMP PREEMPT Sun Mar 13 14:09:51 CST 2016 armv7l armv7l armv7l GNU/Linux

Installed version of CUDA and cuDNN:
No CUDA installed
TF commit hash: 1fcd6d1294564066c6f92b121a3aaf4ed186dc1a

Activity

aselle

aselle commented on Oct 31, 2016

@aselle
aselle
on Oct 31, 2016
Contributor

This looks like a problem compiling gemmlowp. So you should perhaps file a bug on http://github.com/google/gemmlowp/issues ... On their github README, they say that this is sometimes caused by having an insufficient compiler.

aselle
added
stat:awaiting responseStatus - Awaiting response from author
on Oct 31, 2016
aselle

aselle commented on Oct 31, 2016

@aselle
aselle
on Oct 31, 2016
Contributor

@maciekcc, could you please comment on this?

aselle
added
stat:awaiting tensorflowerStatus - Awaiting response from tensorflower
on Oct 31, 2016
maciekcc

maciekcc commented on Oct 31, 2016

@maciekcc
maciekcc
on Oct 31, 2016

I'm on it. What compiler are you using? Some quick things you might check: try adding -fomit-frame-pointer if it's not enabled (it saves one extra register), or remove '-mthumb' (-mno-thumb) if you are using that.

dprylipko

dprylipko commented on Nov 1, 2016

@dprylipko
dprylipko
on Nov 1, 2016
Author

@aselle : Yes, it's gemmlowp code.
@maciekcc:

# gcc --version
gcc (Ubuntu/Linaro 4.8.5-2ubuntu1~14.04.1) 4.8.5

Did you mean -mno-thumb-interwork? Thanks for the hint, but I tried this:

make -f tensorflow/contrib/makefile/Makefile OPTFLAGS="-Ofast -mfpu=neon -funsafe-math-optimizations -ftree-vectorize -mno-thumb-interwork -fomit-frame-pointer" HOST_OS=LINUX

with the same result...

aselle
removed
stat:awaiting responseStatus - Awaiting response from author
on Nov 2, 2016
dprylipko

dprylipko commented on Nov 3, 2016

@dprylipko
dprylipko
on Nov 3, 2016
Author

I tried to compile TF on RPi 3 model B:

make -f tensorflow/contrib/makefile/Makefile HOST_OS=PI TARGET=PI OPTFLAGS="-Os -mfpu=neon-vfpv4 -funsafe-math-optimizations -ftree-vectorize" CXX=g++-4.8

It worked fine, except some minor issues with the Makefile.

maciekcc
self-assigned this
on Nov 3, 2016
maciekcc
removed
stat:awaiting tensorflowerStatus - Awaiting response from tensorflower
on Nov 3, 2016
aselle
added
type:build/installBuild and install issues
on Nov 3, 2016
samjabrahams

samjabrahams commented on Jan 16, 2017

@samjabrahams
samjabrahams
on Jan 16, 2017
Contributor

Currently updating my TensorFlow on Raspberry Pi guide and compiling version 0.12 (and hopefully 1.0.0-alpha afterward); just wanted to note that adding --copt="-fomit-frame-pointer" as a flag to bazel build let me get past the gemmlowp compilation steps.

Here's hoping everything else is peachy!

dprylipko

dprylipko commented on Jan 16, 2017

@dprylipko
dprylipko
on Jan 16, 2017
Author

I managed to build TF on RPi as well, I can't do that on Sabre Board though.

samjabrahams

samjabrahams commented on Jan 16, 2017

@samjabrahams
samjabrahams
on Jan 16, 2017
Contributor

I should have been more specific- my last comment was for compiling TensorFlow through Bazel instead of using the Makefile.

pczekalski

pczekalski commented on Apr 22, 2017

@pczekalski
pczekalski
on Apr 22, 2017

Same problem on Orange Pi+ 2E. Did you find the solution?

8 remaining items

itsmeolivia
closed this as completedon Jun 16, 2017
eric-ibarra

eric-ibarra commented on Jun 22, 2017

@eric-ibarra
eric-ibarra
on Jun 22, 2017

I'm still having this issue compiling with gcc-4.8.4. It works if I compile without the neon optimization, however, I believe there is significant performance to be gained using the neon optimization with eigen

godardt

godardt commented on Aug 2, 2017

@godardt
godardt
on Aug 2, 2017 · edited by godardt

Any updates on this? I'm still having the issue with TF 1.3.0 rc1:

/tensorflow/core/kernels/_objs/quantized_ops/tensorflow/core/kernels/meta_support.pic.o): com.google.devtools.build.lib.shell.BadExitStatusException: Process exited with status 1.
In file included from external/gemmlowp/meta/streams.h:293:0,                                                                            from external/gemmlowp/meta/quantized_mul_kernels.h:22,
                 from ./tensorflow/core/kernels/meta_support.h:21,                                                                       from tensorflow/core/kernels/meta_support.cc:18:                                                       external/gemmlowp/meta/streams_arm_32.h: In static member function 'static void gemmlowp::meta::GemmExecutorPackLHS::ExecuteDispatch3D(const P&) [with P = gemmlowp::meta::GemmParams<unsigned char, int, gemmlowp::meta::ColumnMajorWithSum, gemmlowp::meta::RowMajorWithSum, gemmlowp::meta::QuantizedStaticPreprocessedAsInt32, gemmlowp::meta::RowMajor>; int m = 1; int n = 8; int k = 8; int m_leftovers = 0; int n_leftovers = 7; int k_leftovers = 4]':
external/gemmlowp/meta/streams_arm_32.h:4211:59: error: can't find a register in class 'LO_REGS' while reloading 'asm'
         "d25", "d26", "d27", "d28", "d29", "cc", "memory");
                                                           ^
external/gemmlowp/meta/streams_arm_32.h:4211:59: error: 'asm' operand has impossible constraints
Target //tensorflow:libtensorflow.so failed to build
INFO: Elapsed time: 17200.207s, Critical Path: 601.22s

with:

$ gcc --version
gcc (Ubuntu/Linaro 4.8.5-4ubuntu2) 4.8.5

and:

$cat /proc/cpuinfo
processor       : 0
model name      : ARMv7 Processor rev 4 (v7l)
BogoMIPS        : 76.80
Features        : half thumb fastmult vfp edsp neon vfpv3 tls vfpv4 idiva idivt vfpd32 lpae evtstrm crc32
CPU implementer : 0x41
CPU architecture: 7
CPU variant     : 0x0
CPU part        : 0xd03
CPU revision    : 4

processor       : 1
model name      : ARMv7 Processor rev 4 (v7l)
BogoMIPS        : 76.80
Features        : half thumb fastmult vfp edsp neon vfpv3 tls vfpv4 idiva idivt vfpd32 lpae evtstrm crc32
CPU implementer : 0x41
CPU architecture: 7
CPU variant     : 0x0
CPU part        : 0xd03
CPU revision    : 4

processor       : 2
model name      : ARMv7 Processor rev 4 (v7l)
BogoMIPS        : 76.80
Features        : half thumb fastmult vfp edsp neon vfpv3 tls vfpv4 idiva idivt vfpd32 lpae evtstrm crc32
CPU implementer : 0x41
CPU architecture: 7
CPU variant     : 0x0
CPU part        : 0xd03
CPU revision    : 4

processor       : 3
model name      : ARMv7 Processor rev 4 (v7l)
BogoMIPS        : 76.80
Features        : half thumb fastmult vfp edsp neon vfpv3 tls vfpv4 idiva idivt vfpd32 lpae evtstrm crc32
CPU implementer : 0x41
CPU architecture: 7
CPU variant     : 0x0
CPU part        : 0xd03
CPU revision    : 4

Hardware        : BCM2709
Revision        : a02082
Serial          : 0000000081b86d8e

while compiling with Bazel with:

bazel build -c opt --copt="-mfpu=neon-vfpv4" --copt="-funsafe-math-optimizations" --copt="-ftree-vectorize" --copt="-fomit-frame-pointer" --local_resources 1024,1.0,1.0 --verbose_failures //tensorflow:libtensorflow.so
maciekcc

maciekcc commented on Aug 2, 2017

@maciekcc
maciekcc
on Aug 2, 2017

The fix is already in gemmlowp (since May actually) : google/gemmlowp@9941cad

I will check if it ended up in tensorflow.

godardt

godardt commented on Aug 3, 2017

@godardt
godardt
on Aug 3, 2017 · edited by godardt

@maciekcc Doesn't seems so. The most recent commit from the bazel workspace is this one.

godardt
mentioned this on Aug 4, 2017
zhonghuaxi

zhonghuaxi commented on Aug 8, 2017

@zhonghuaxi
zhonghuaxi
on Aug 8, 2017 · edited by zhonghuaxi

+1

make -f tensorflow/contrib/makefile/Makefile TARGET=PI OPTFLAGS="-O3 -mcpu=cortex-a15 -mfpu=neon-vfpv4 -funsafe-math-optimizations -ftree-vectorize -Llibs" -j100 CXX=/opt/toolchains/gcc-linaro-arm-linux-gnueabihf-4.8-2013.09_linux/bin/arm-linux-gnueabihf-g++

/home/zxi/projects/tensorflow/tensorflow/contrib/makefile/downloads/gemmlowp/meta/streams_arm_32.h: In static member function ‘static void gemmlowp::meta::GemmExecutorPackLHS::ExecuteDispatch3D(const P&) [with P = gemmlowp::meta::GemmParams<unsigned char, int, gemmlowp::meta::ColumnMajorWithSum, gemmlowp::meta::RowMajorWithSum, gemmlowp::meta::QuantizedStaticPreprocessedAsInt32, gemmlowp::meta::RowMajor>; int m = 1; int n = 8; int k = 8; int m_leftovers = 0; int n_leftovers = 7; int k_leftovers = 4]’:
/home/zxi/projects/tensorflow/tensorflow/contrib/makefile/downloads/gemmlowp/meta/streams_arm_32.h:4211:59: error: can’t find a register in class ‘LO_REGS’ while reloading ‘asm’
         "d25", "d26", "d27", "d28", "d29", "cc", "memory");
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

Labels

type:build/installBuild and install issues

Type

No type

Projects

No projects

Milestone

No milestone

Relationships

None yet

    Development

    No branches or pull requests

      Participants

      @aselle@itsmeolivia@godardt@samjabrahams@eric-ibarra

      Issue actions
        Error compiling on ARM with -mfpu=neon: can’t find a register in class ‘LO_REGS’ while reloading ‘asm’ · Issue #5303 · tensorflow/tensorflow