01_vector_add test compile error. Just want to confirm can this test be well supported.

[FatJhon]

Describe the bug

user guide test vector_add test, just fail to make_so,get error information as follows:
ret = subprocess.check_call(cc_cmd)
File "/usr/local/lib/python3.10/subprocess.py", line 369, in check_call
raise CalledProcessError(retcode, cmd)
subprocess.CalledProcessError: Command '['/opt/rh/gcc-toolset-11/root/usr/bin/gcc', '/tmp/tmp59kzdlsc/kernel.s', '-O3', '-shared', '-fPIC', '-Wno-psabi', '-o', '/tmp/tmp59kzdlsc/kernel.cpython-310-x86_64-linux-gnu.so', '-lm', '-lTritonCPURuntime', '-lsleef', '-lgcc', '-L/home/workspace/torch-triton/triton-cpu/python/triton/_C', '-L/usr/local/lib', '-I/usr/local/include', '-I/tmp/tmp59kzdlsc', '-I/usr/local/include/python3.10', '-Wl,-rpath', PosixPath('/home/workspace/torch-triton/triton-cpu/python/triton/_C'), '-Wl,-rpath', '/usr/local/lib', '-g']' returned non-zero exit status 1.

Environment details

use the newest master && cpu backend test

[krishagarwal]

I found the same issue. Looking further up in the error log, I saw the following message from the assembler:

/tmp/tmpfne6g4g3/kernel.s: Assembler messages: /tmp/tmpfne6g4g3/kernel.s:4363: Error: junk at end of line, first unrecognized character is `"'

I manually printed the assembly source from python (TRITON_KERNEL_DUMP=1 TRITON_DUMP_DIR=<some directory> wasn't working) and saw that the line causing the error is:

.file	1 "<path>/triton-cpu/python/tutorials" "01-vector-add.py"

I guess the assembler needs this path to be one string rather than split into two, meaning it should be

.file	1 "<path>/triton-cpu/python/tutorials/01-vector-add.py"

For now, as a really dirty fix, I inserted the following at line 93 in python/triton/runtime/build.py:

with open(src, "r") as f:
    src_txt = f.read()
src_txt = src_txt.replace('" "', "/")
with open(src, "w") as f:
    f.write(src_txt)

Basically, this will manually combine the path into one string in the assembly source before compiling. It seems to be working fine for me so far, although I would obviously hope a more proper fix is implemented.

[ienkovich]

I've seen such a problem previously. It was caused by as not recognizing .file directive and we've fixed it by adding -g compilation option. Initially, we used -gdwarf-5, but later moved to -g (IIRC -gdwarf-5 caused failure on older toolchains). So I suspect the issue might be related to the binary toolchain version on your system. What is our toolchain version? Can you successfully compile the original asm file with your gcc/as providing -gdwarf-5 options (asm compilation options are added at python/triton/runtime/build.py:92)?

[krishagarwal]

gcc version 9.4.0, as version 2.34. The issue persists even after using -gdwarf-5.

[ienkovich]

In our flows we use GCC 11, Binutils 2.38 or newer, so upgrading should resolve the issue.

A possible workaround is to use TRITON_DISABLE_LINE_INFO=1. This should remove all .file directives you have a problem with.