RuntimeError: relocation R_X86_64_32S against `.bss' can not be used when making a shared object in _linux_shared

shizukanaskytree · January 15, 2019, 1:36am

I follow setup and installation steps according to
https://docs.tvm.ai/install/from_source.html

Outline

System
config.cmake
Related code
Similar errors

System

Linux x86-64

4.15.0-43-generic

Ubuntu 16.04

config.cmake

My config.cmake is

set(USE_CUDA /usr/local/cuda-10.0)
...
set(USE_LLVM /home/wxf/llvm_download/clang+llvm-7.0.0-x86_64-linux-gnu-ubuntu-16.04/bin/llvm-config)

#--------------------------------------------------------------------
#  Template custom cmake configuration for compiling
#
#  This file is used to override the build options in build.
#  If you want to change the configuration, please use the following
#  steps. Assume you are on the root directory. First copy the this
#  file so that any local changes will be ignored by git
#
#  $ mkdir build
#  $ cp cmake/config.cmake build
#
#  Next modify the according entries, and then compile by
#
#  $ cd build
#  $ cmake ..
#
#  Then buld in parallel with 8 threads
#
#  $ make -j8
#--------------------------------------------------------------------

#---------------------------------------------
# Backend runtimes.
#---------------------------------------------

# Whether enable CUDA during compile,
#
# Possible values:
# - ON: enable CUDA with cmake's auto search
# - OFF: disbale CUDA
# - /path/to/cuda: use specific path to cuda toolkit
set(USE_CUDA /usr/local/cuda-10.0)

# Whether enable ROCM runtime
#
# Possible values:
# - ON: enable ROCM with cmake's auto search
# - OFF: disbale ROCM
# - /path/to/rocm: use specific path to rocm
set(USE_ROCM OFF)

# Whether enable SDAccel runtime
set(USE_SDACCEL OFF)

# Whether enable Intel FPGA SDK for OpenCL (AOCL) runtime
set(USE_AOCL OFF)

# Whether enable OpenCL runtime
set(USE_OPENCL OFF)

# Whether enable Metal runtime
set(USE_METAL OFF)

# Whether enable Vulkan runtime
#
# Possible values:
# - ON: enable Vulkan with cmake's auto search
# - OFF: disbale vulkan
# - /path/to/vulkan-sdk: use specific path to vulkan-sdk
set(USE_VULKAN OFF)

# Whether enable OpenGL runtime
set(USE_OPENGL OFF)

# Whether to enable SGX runtime
#
# Possible values for USE_SGX:
# - /path/to/sgxsdk: path to Intel SGX SDK
# - OFF: disable SGX
#
# SGX_MODE := HW|SIM
set(USE_SGX OFF)
set(SGX_MODE "SIM")
set(RUST_SGX_SDK "/path/to/rust-sgx-sdk")

# Whether enable RPC runtime
set(USE_RPC ON)

# Whether embed stackvm into the runtime
set(USE_STACKVM_RUNTIME OFF)

# Whether enable tiny embedded graph runtime.
set(USE_GRAPH_RUNTIME ON)

# Whether enable additional graph debug functions
set(USE_GRAPH_RUNTIME_DEBUG OFF)

# Whether build with LLVM support
# Requires LLVM version >= 4.0
#
# Possible values:
# - ON: enable llvm with cmake's find search
# - OFF: disbale llvm
# - /path/to/llvm-config: enable specific LLVM when multiple llvm-dev is available.
set(USE_LLVM /home/wxf/llvm_download/clang+llvm-7.0.0-x86_64-linux-gnu-ubuntu-16.04/bin/llvm-config)

#---------------------------------------------
# Contrib libraries
#---------------------------------------------
# Whether use BLAS, choices: openblas, mkl, atlas, apple
set(USE_BLAS none)

# /path/to/mkl: mkl root path when use mkl blas library
# set(USE_MKL_PATH /opt/intel/mkl) for UNIX
# set(USE_MKL_PATH ../IntelSWTools/compilers_and_libraries_2018/windows/mkl) for WIN32
set(USE_MKL_PATH none)

# Whether use contrib.random in runtime
set(USE_RANDOM OFF)

# Whether use NNPack
set(USE_NNPACK OFF)

# Whether use CuDNN
set(USE_CUDNN OFF)

# Whether use cuBLAS
set(USE_CUBLAS OFF)

# Whether use MIOpen
set(USE_MIOPEN OFF)

# Whether use MPS
set(USE_MPS OFF)

# Whether use rocBlas
set(USE_ROCBLAS OFF)

# Whether use contrib sort
set(USE_SORT OFF)

# Build ANTLR parser for Relay text format
set(USE_ANTLR OFF)

errors

When I run a simple test from:

https://docs.tvm.ai/tutorials/nnvm_quick_start.html#sphx-glr-tutorials-nnvm-quick-start-py

I get some errors in the end:

Symbol Outputs:
	output[0]=softmax(0)
Variable:data
Variable:bn_data_gamma
Variable:bn_data_beta
Variable:bn_data_moving_mean
Variable:bn_data_moving_var
--------------------
Op:batch_norm, Name=bn_data
Inputs:
	arg[0]=data(0) version=0
	arg[1]=bn_data_gamma(0) version=0
	arg[2]=bn_data_beta(0) version=0
	arg[3]=bn_data_moving_mean(0) version=1
	arg[4]=bn_data_moving_var(0) version=1
Attrs:
	epsilon=2e-05
	scale=False
Variable:conv0_weight
--------------------
...                           # normal output 
--------------------
Op:dense, Name=fc1
Inputs:
	arg[0]=flatten0(0)
	arg[1]=fc1_weight(0) version=0
	arg[2]=fc1_bias(0) version=0
Attrs:
	units=1000
--------------------
Op:softmax, Name=softmax
Inputs:
	arg[0]=fc1(0)

[0.00089283 0.00103331 0.0009094  0.00102275 0.00108751 0.00106737
 0.00106262 0.00095838 0.00110792 0.00113151]



['deploy_graph.json', 'deploy_lib.tar', 'deploy_param.params']
Traceback (most recent call last):
  File "00_nnvm_quick_start.py", line 132, in <module>
    loaded_lib = tvm.module.load(path_lib)
  File "/home/wxf/tvm_prj/tvm/python/tvm/module.py", line 241, in load
    _cc.create_shared(path + ".so", files)
  File "/home/wxf/tvm_prj/tvm/python/tvm/contrib/cc.py", line 33, in create_shared
    _linux_shared(output, objects, options, cc)
  File "/home/wxf/tvm_prj/tvm/python/tvm/contrib/cc.py", line 58, in _linux_shared
    raise RuntimeError(msg)
RuntimeError: Compilation error:
/usr/bin/ld: /tmp/tmpqgibc8e4/lib.o: relocation R_X86_64_32S against `.bss' can not be used when making a shared object; recompile with -fPIC
/tmp/tmpqgibc8e4/lib.o: error adding symbols: Bad value
collect2: error: ld returned 1 exit status

Related code

Related code is in /tvm/python/tvm/contrib/cc.py

https://docs.python.org/3/library/subprocess.html#popen-constructor

...
 12 def create_shared(output,
 13                   objects,
 14                   options=None,
 15                   cc="g++"):
 16     """Create shared library.
 17  
 18     Parameters
 19     ----------
 20     output : str
 21         The target shared library.
 22  
 23     objects : list
 24         List of object files.
 25  
 26     options : list
 27         The list of additional options string.
 28  
 29     cc : str, optional
 30         The compile string.
 31     """
 32     if sys.platform == "darwin" or sys.platform.startswith('linux'):
 33         _linux_shared(output, objects, options, cc)
 34     elif sys.platform == "win32":
 35         _windows_shared(output, objects, options)
 36     else:
 37         raise ValueError("Unsupported platform")
...
 40 def _linux_shared(output, objects, options, cc="g++"):
 41     cmd = [cc]
 42     cmd += ["-shared", "-fPIC"]
 43     if sys.platform == "darwin":
 44         cmd += ["-undefined", "dynamic_lookup"]
 45     cmd += ["-o", output]
 46     if isinstance(objects, str):
 47         cmd += [objects]
 48     else:
 49         cmd += objects
 50     if options:
 51         cmd += options
 52     proc = subprocess.Popen(
 53         cmd, stdout=subprocess.PIPE, stderr=subprocess.STDOUT)
 54     (out, _) = proc.communicate()
 55     if proc.returncode != 0:
 56         msg = "Compilation error:\n"
 57         msg += py_str(out)
 58         raise RuntimeError(msg)

CAUSED BY:

(out, _) = proc.communicate()

RuntimeError: Compilation error:
/usr/bin/ld: /tmp/tmpwl0pygoi/lib.o: relocation R_X86_64_32S against `.bss' can not be used when making a shared object; recompile with -fPIC
/tmp/tmpwl0pygoi/lib.o: error adding symbols: Bad value
collect2: error: ld returned 1 exit status

Similar situations in other tutorial examples

When I run another tutorial example, I also get similar error information.
https://docs.tvm.ai/tutorials/autotvm/tune_nnvm_cuda.html

DEBUG:autotvm:No: 217	GFLOPS: 0.00/0.00	result: MeasureResult(costs=(RuntimeError('Except caught from RPC call: TVMCall CFunc Error:

Traceback (most recent call last):

  File "/home/wxf/tvm_prj/tvm/python/tvm/_ffi/_ctypes/function.py", line 55, in cfun

    rv = local_pyfunc(*pyargs)

  File "/home/wxf/tvm_prj/tvm/python/tvm/rpc/server.py", line 50, in load_module

    m = _load_module(path)

  File "/home/wxf/tvm_prj/tvm/python/tvm/module.py", line 241, in load

    _cc.create_shared(path + ".so", files)

  File "/home/wxf/tvm_prj/tvm/python/tvm/contrib/cc.py", line 33, in create_shared

    _linux_shared(output, objects, options, cc)

  File "/home/wxf/tvm_prj/tvm/python/tvm/contrib/cc.py", line 58, in _linux_shared

    raise RuntimeError(msg)

RuntimeError: Compilation error:

/usr/bin/ld: /tmp/tmpgqcx3hzt/lib.o: relocation R_X86_64_32S against `.bss\' can not be used when making a shared object; recompile with -fPIC

/tmp/tmpgqcx3hzt/lib.o: error adding symbols: Bad value

collect2: error: ld returned 1 exit status



',),), error_no=4, all_cost=7.192363262176514, timestamp=1547500859.3370917)	[('tile_b', [16, 1, 1, 1]), ('tile_y', [1, 128, 2, 2]), ('tile_x', [2, 1, 2, 4]), ('tile_rc', [64, 8]), ('auto_unroll_max_step', 1500), ('unroll_explicit', 0)],winograd,None,183331

Another example code shows error:

<class 'tvm.tensor.Tensor'>
-----GPU code-----
extern "C" __global__ void myadd_kernel0( float* __restrict__ C,  float* __restrict__ A,  float* __restrict__ B, int n) {
  if (((int)blockIdx.x) < (n / 64)) {
    C[((((int)blockIdx.x) * 64) + ((int)threadIdx.x))] = (A[((((int)blockIdx.x) * 64) + ((int)threadIdx.x))] + B[((((int)blockIdx.x) * 64) + ((int)threadIdx.x))]);
  } else {
    if ((((int)blockIdx.x) * 64) < (n - ((int)threadIdx.x))) {
      C[((((int)blockIdx.x) * 64) + ((int)threadIdx.x))] = (A[((((int)blockIdx.x) * 64) + ((int)threadIdx.x))] + B[((((int)blockIdx.x) * 64) + ((int)threadIdx.x))]);
    }
  }
}


Traceback (most recent call last):
  File "01_get_started.py", line 214, in <module>
    cc.create_shared(temp.relpath("myadd.so"), [temp.relpath("myadd.o")])
  File "/home/wxf/tvm_prj/tvm/python/tvm/contrib/cc.py", line 33, in create_shared
    _linux_shared(output, objects, options, cc)
  File "/home/wxf/tvm_prj/tvm/python/tvm/contrib/cc.py", line 58, in _linux_shared
    raise RuntimeError(msg)
RuntimeError: Compilation error:
/usr/bin/ld: /tmp/tmpfijjmtj5/myadd.o: relocation R_X86_64_32S against `.bss' can not be used when making a shared object; recompile with -fPIC
/tmp/tmpfijjmtj5/myadd.o: error adding symbols: Bad value
collect2: error: ld returned 1 exit status

The hint shows recompile with -fPIC, but I think the Makefile has already set the -fPIC flag.

How to fix the problem? Thanks in advance!

srkreddy1238 · January 15, 2019, 5:48am

Suggest to use a bit latest Ubuntu if possible.

How about gcc version (try with GCC-5.x)?

OFShare · January 22, 2019, 9:46am

@shizukanaskytree have you solved this problem?

I recently encounter the same problem as you post.

I just follow the tutorial:
1.https://docs.tvm.ai/tutorials/nnvm_quick_start.html#sphx-glr-tutorials-nnvm-quick-start-py
2.https://docs.tvm.ai/tutorials/get_started.html#sphx-glr-tutorials-get-started-py

everything goes well until I execute this line code "cc.create_shared(temp.relpath(“myadd.so”), [temp.relpath(“myadd.o”)]) " ,it seems produce the .so, link error:

“/usr/bin/x86_64-linux-gnu-ld: /tmp/tmphk243coy/myadd.o: relocation R_X86_64_32S against `.bss’ can not be used when making a shared object; recompile with -fPIC”

but in ‘create_shared’ function it already have default ‘-fPIC’ optional.

System:
ubuntu 18.04
cuda-9.0
gcc version 6.5.0
LLVM 7.0.0(Pre-Built Binaries:Ubuntu 16.04 (.sig) )
cmake version 3.10.2

as @srkreddy1238 suggested, I will try gcc-g.x later.

denis0x0D · January 22, 2019, 1:48pm

@shizukanaskytree this is interesting error, because myadd.o generates from assembler generated by llvm from tvm with position independent code

github.com

dmlc/tvm/blob/master/src/codegen/llvm/llvm_common.cc#L146


}


std::string err;
const llvm::Target* target =
    llvm::TargetRegistry::lookupTarget(target_triple, err);
if (target == nullptr) {
  CHECK(allow_null) << err << " target_triple=" << target_triple;
  return nullptr;
}
llvm::TargetMachine* tm = target->createTargetMachine(
    target_triple, mcpu, mattr, opt, llvm::Reloc::PIC_);
return std::unique_ptr<llvm::TargetMachine>(tm);
}


}  // namespace codegen
}  // namespace tvm
#endif  // TVM_LLVM_VERSION

But it in your case, it seems that, the code were generated with llvm::Reloc::Static.
Could you check the generated code, in case pic code all calls should go through plt https://docs.oracle.com/cd/E19120-01/open.solaris/819-0690/feowg/index.html

print(fadd.get_source("asm"))



python get_started.py > dump.s
grep GOTPCREL dump.s | wc -l

or yor can save directrly myadd.o

readelf -r myadd.o | grep GOTPCREL | wc -l

OFShare · January 22, 2019, 1:56pm

I have solved this problem.Those are my steps:

tvm install from source ,you must notice LLVM version.

“Since LLVM takes long time to build from source, you can download pre-built version of LLVM from LLVM Download Page.”

when you download pre-built version of LLVM,you should download LLVM correspond to you local environment of gcc and g++ version.

For example,I used download the LLVM 7.0.0,but my local ubuntu18.04 is gcc 6.x and g++6.x, so I got this error.However,when I changed to LLVM 6.0.0, everything goes well.

System:
ubuntu 18.04
cuda-9.0
gcc version 6.5.0
g++ version 6.5.0
LLVM 6.0.0(Pre-Built Binaries:Ubuntu 16.04 (.sig) )
cmake version 3.10.2

good luck.

shizukanaskytree · January 22, 2019, 7:36pm

I find the solution from this post.

CMAKE_C_COMPILER
CMAKE_CXX_COMPILER

need to be explicitly set.