Why Does -march=native Cause Crashes on My Zen2 Chip?

  • Thread starter Vanadium 50
  • Start date
In summary, using the -march=native flag on a Zen2 chip can cause crashes due to the software being optimized for a different architecture. This can result in instructions that are not compatible with the chip, leading to unexpected and potentially harmful crashes. Additionally, the flag may not always provide a performance boost on Zen2 chips and can cause issues with compatibility and stability. It is important to carefully consider the use of this flag and to test for any potential crashes before implementing it in production.
  • #1
Vanadium 50
Staff Emeritus
Science Advisor
Education Advisor
2023 Award
35,005
21,671
I don't understand the march/mtune behavior.

This is Linux GCC, on a Zen2 chip.

Mtune seems to do nothing much. OK, sometimes your code is as tuned as its going to get out of the box. I am compiling with -O3 and -Ofast, so maybe it is so optimized there is little to tune.

Default march works fine. -march=znver1 works fine, but again, no faster. OK, again, maybe there's little to be done. -march=native and -march=znver2 should do the same thing. I guess they do. They both hang at pthread_create.

This is not really a problem - I don't really need to squeeze the last bin of performance out of the code - but it sure seems mysterious.

gcc -v gives

Code: 
Using built-in specs.
COLLECT_GCC=gcc
COLLECT_LTO_WRAPPER=/usr/libexec/gcc/x86_64-redhat-linux/11/lto-wrapper
OFFLOAD_TARGET_NAMES=nvptx-none
OFFLOAD_TARGET_DEFAULT=1
Target: x86_64-redhat-linux
Configured with: ../configure --enable-bootstrap --enable-host-pie --enable-host-bind-now --enable-languages=c,c++,fortran,lto --prefix=/usr --mandir=/usr/share/man --infodir=/usr/share/info --with-bugurl=http://bugs.almalinux.org/ --enable-shared --enable-threads=posix --enable-checking=release --with-system-zlib --enable-__cxa_atexit --disable-libunwind-exceptions --enable-gnu-unique-object --enable-linker-build-id --with-gcc-major-version-only --enable-plugin --enable-initfini-array --without-isl --enable-multilib --with-linker-hash-style=gnu --enable-offload-targets=nvptx-none --without-cuda-driver --enable-gnu-indirect-function --enable-cet --with-tune=generic --with-arch_64=x86-64-v2 --with-arch_32=x86-64 --build=x86_64-redhat-linux --with-build-config=bootstrap-lto --enable-link-serialization=1
Thread model: posix
Supported LTO compression algorithms: zlib zstd
gcc version 11.4.1 20231218 (Red Hat 11.4.1-3) (GCC)
 
Computer science news on Phys.org
  • #2
Vanadium 50 said:
Mtune seems to do nothing much. OK, sometimes your code is as tuned as its going to get out of the box. I am compiling with -O3 and -Ofast, so maybe it is so optimized there is little to tune.
If you leave out those optimization options, does mtune seem to do more?
 
  • #3
I have not played with that. I am more interested in my march crashes. The setting march=native should never generate code that the CPU cannot handle.

The way I got into this rabbit hole was noticing that the block of code that takes the longest has some multiply-and-adds. This was to see if the compiler could speed it up by using FMA instructions. What could be simpler than throwing a compiler switch?
 

Hot Threads

Recent Insights

Back
Top