39

AMD, Intel, Red Hat, and SUSE have defined a set of "architecture levels" for x86-64 CPUs. For example x86-64-v2 means that a CPU support not only the basic x86-64 instructions set, but also other instructions like SSE4.2, SSSE3 or POPCNT.

How can I check which architecture levels are supported by my CPU?

8 Answers 8

41

This is based on gioele’s answer; the whole script might as well be written in AWK:

#!/usr/bin/awk -f

BEGIN {
    while (!/flags/) if (getline < "/proc/cpuinfo" != 1) exit 1
    if (/lm/&&/cmov/&&/cx8/&&/fpu/&&/fxsr/&&/mmx/&&/syscall/&&/sse2/) level = 1
    if (level == 1 && /cx16/&&/lahf/&&/popcnt/&&/sse4_1/&&/sse4_2/&&/ssse3/) level = 2
    if (level == 2 && /avx/&&/avx2/&&/bmi1/&&/bmi2/&&/f16c/&&/fma/&&/abm/&&/movbe/&&/xsave/) level = 3
    if (level == 3 && /avx512f/&&/avx512bw/&&/avx512cd/&&/avx512dq/&&/avx512vl/) level = 4
    if (level > 0) { print "CPU supports x86-64-v" level; exit level + 1 }
    exit 1
}

This also checks for the baseline (“level 1” here), only outputs the highest supported level, and exits with an exit code matching the first unsupported level.

6
  • 3
    Some of the checking of earlier levels is redundant, but not a bad idea I guess. In real hardware, SSE4.2 for example already implies support for all previous (Intel) SSE versions (but not AMD SSE4a). In a virtual machine CPUID is virtualized so it's theoretically possible to indicate SSSE3 support without SSE3. Only in a software emulator would could you make SSE3 instructions fault while SSSE3 instructions didn't. (BTW, you omitted /sse3/.) The de-facto standard is that runtime CPU dispatching only needs to check the highest SSE feature flag it depends on. Jan 27, 2021 at 19:02
  • 1
    There are other de-facto standards like SSE4.2 implying popcnt, but that's good to check explicitly. And other non-SIMD extensions like BMI1 are fully independent of SIMD (although since some BMI1/2 instructions use VEX encoding, they're normally only found on CPUs that support AVX. And unfortunately Intel even disables BMI1/2 on their Pentium/Celeron CPUs, perhaps as a way of fully disabling AVX.). Jan 27, 2021 at 19:08
  • 1
    BTW, level 2 = Nehalem and current Silvermont, and current-gen Pentium/Celeron. Also AMD Bulldozer family since even Excavator doesn't have BMI2, only AVX2 and FMA3. Level 3 = Haswell (and Zen), and includes most of the really good stuff. MacOS apparently can make fat binaries with baseline x86-64 and Haswell feature-level, allowing usage of BMI2 efficient shift instructions all over the place, and of AVX everywhere. Level 4 = -march=skylake-avx512. Jan 27, 2021 at 19:12
  • 1
    @PeterCordes yes, there are a number of deficiencies and redundancies here (in particular, I should check full fields instead of using regexes, since for example /lm/ will match anything containing those characters). I followed the exhaustive level definitions as used in the first answer (that’s where /ssse3/ without /sse3/ came from), even though as you say many of them are redundant. (I’ve been following the discussions leading up to the definition of these levels.) Jan 27, 2021 at 19:29
  • 1
    TBH this was more an exercise in showing that all the checks could be done in AWK instead of a mixture of AWK ans shell, rather than coming up with the best level checker ;-). Jan 27, 2021 at 19:32
28

Originally copied from https://gitlab.archlinux.org/archlinux/rfcs/-/merge_requests/2/diffs

With glibc 2.33 or later (Arch Linux, Debian 12, Ubuntu 21.04, Fedora 34, etc.), or patched glibc (RHEL 8), you can see what architecture is supported by your CPU by running:

$ /lib/ld-linux-x86-64.so.2 --help

Subdirectories of glibc-hwcaps directories, in priority order:
  x86-64-v4
  x86-64-v3 (supported, searched)
  x86-64-v2 (supported, searched)

On Debian derivatives the path is different, you need to run /lib64/ld-linux-x86-64.so.2 --help.

1
  • 5
    The AMD64 ABI remarks that /lib64/ld-linux-x86-64.so.2 is the standard place (on Linux) for the program interpreter.
    – Sam Morris
    Jan 17, 2023 at 11:50
12

On Linux, one can check the CPU capabilities reported by /proc/cpuinfo against the requirements described in the x86-psABI documentation.

The following script automates that process (the exit code is the number of the first non-supported architecture level).

#!/bin/sh -eu

flags=$(cat /proc/cpuinfo | grep flags | head -n 1 | cut -d: -f2)

supports_v2='awk "/cx16/&&/lahf/&&/popcnt/&&/sse4_1/&&/sse4_2/&&/ssse3/ {found=1} END {exit !found}"'
supports_v3='awk "/avx/&&/avx2/&&/bmi1/&&/bmi2/&&/f16c/&&/fma/&&/abm/&&/movbe/&&/xsave/ {found=1} END {exit !found}"'
supports_v4='awk "/avx512f/&&/avx512bw/&&/avx512cd/&&/avx512dq/&&/avx512vl/ {found=1} END {exit !found}"'

echo "$flags" | eval $supports_v2 || exit 2 && echo "CPU supports x86-64-v2"
echo "$flags" | eval $supports_v3 || exit 3 && echo "CPU supports x86-64-v3"
echo "$flags" | eval $supports_v4 || exit 4 && echo "CPU supports x86-64-v4"
4
  • 2
    Instead of using a variable and evaling it, you could have used a function
    – muru
    Jan 27, 2021 at 8:48
  • As an FYI, my old AMD FX-6100 supports v2, but not v3 or v4.
    – RonJohn
    Jan 27, 2021 at 18:47
  • 1
    @RonJohn: Yup, even Bulldozer-family is only "level 2", even though Excavator has AVX2 and FMA. It's missing BMI2 and movbe. (Piledriver / Steamroller have AVX1 and FMA; Bulldozer has AVX1 and FMA4 but not FMA3; Intel pulled the rug out from under AMD as late as they could. See Stop the instruction set war on Agner Fog's blog.) To be fair, having another level with AVX but not BMI2 would be of limited value, and BMI2 is quite nice for Intel CPUs: variable-count shifts with SHLX/SHRX are 1 uop instead of 3, and can use any reg instead of CL Jan 27, 2021 at 19:16
  • 1
    Level 3 = Haswell and Zen1. Level 4 = -march=skylake-avx512. Jan 27, 2021 at 19:16
6

Here's a shell script to determine the x86_64 CPU architecture level on Linux. It's compatible with BusyBox. With the option -v, it shows which flags you're missing to reach the next level. See What do the flags in /proc/cpuinfo mean? for an explanation of the flags.

#!/bin/sh
set -e

verbose=
while getopts v OPTLET; do
  case "$OPTLET" in
    v) verbose=1;;
    \?) exit 2;;
  esac
done

flags=$(grep '^flags\b' </proc/cpuinfo | head -n 1)
flags=" ${flags#*:} "

has_flags () {
  for flag; do
    case "$flags" in
      *" $flag "*) :;;
      *)
        if [ -n "$verbose" ]; then
          echo >&2 "Missing $flag for the next level"
        fi
        return 1;;
    esac
  done
}

determine_level () {
  level=0
  has_flags lm cmov cx8 fpu fxsr mmx syscall sse2 || return 0
  level=1
  has_flags cx16 lahf_lm popcnt sse4_1 sse4_2 ssse3 || return 0
  level=2
  has_flags avx avx2 bmi1 bmi2 f16c fma abm movbe xsave || return 0
  level=3
  has_flags avx512f avx512bw avx512cd avx512dq avx512vl || return 0
  level=4
}

determine_level
echo "$level"

(Acknowledgement: I reused the list of flags from Stephen Kitt's answer which in turns builds on gioele's answer.)

3
  • .@Gilles, thanks for this. I created a 'x86-64-level' tool (<github.com/HenrikBengtsson/x86-64-level/>) that originated from your script here. I did it to give it a home, make it downloadable, and to be able to add a README with more details. I'd like to add a FOSS license to it; do you have a preference?
    – HenrikB
    Dec 17, 2022 at 3:49
  • @HenrikB Anything posted on this site is already open source. But if you prefer a different license: I hereby allow anyone to make a derivative work based on the code in unix.stackexchange.com/a/631320 and license it (or not) however they please. Dec 17, 2022 at 11:22
  • Thank you @Gilles. I've added the 'CC BY-SA 4.0' license per the default here. Also, I've listed you as one of the authors - I hope that's okay. Also, if you want to be listed with anything else that your StackOverflow user name, please let me know.
    – HenrikB
    Dec 18, 2022 at 0:06
1

I've created an x86-64-level tool based on the suggestions here. Examples:

$ x86-64-level
3

$ level=$(x86-64-level)
$ echo "x86-64-v${level}"
x86-64-v3

## Output an explanation to stderr
$ x86-64-level --verbose
Identified x86-64-v3, because x86-64-v4 requires 'avx512f', which
is not supported by this CPU [Intel(R) Core(TM) i7-8650U CPU @ 1.90GHz]
3

If you want to assert that the current machine supports a certain x86-64 level in a shell script, add the following one-line gatekeeper;

x86-64-level --assert=4 || exit 1

This will be silent if the host supports x86-64-v4, otherwise it'll output:

The CPU [Intel(R) Core(TM) i7-8650U CPU @ 1.90GHz] on this host ('dev2')
supports x86-64-v3, which is less than the required x86-64-v4

and exit with exit value 1.

The x86-64-level tool is a standalone Bash script that's available at https://github.com/HenrikBengtsson/x86-64-level.

1

One way is to use the Function Multiversioning feature in GCC, write a test program, and see what version of the function (dependent on your CPU arch) will it pick.

The foo function from the program below will create multiple symbols in the binary, and the "best" version will be picked at runtime

$ nm a.out | grep foo
0000000000402236 T _Z3foov
000000000040224c T _Z3foov.arch_x86_64
0000000000402257 T _Z3foov.arch_x86_64_v2
0000000000402262 T _Z3foov.arch_x86_64_v3
000000000040226d T _Z3foov.arch_x86_64_v4
0000000000402290 W _Z3foov.resolver
0000000000402241 T _Z3foov.sse4.2
0000000000402290 i _Z7_Z3foovv
// multiversioning.c

#include <stdio.h>

__attribute__ ((target ("default")))
const char* foo () { return "default"; }

__attribute__ ((target ("sse4.2")))
const char* foo () { return "sse4.2"; }

__attribute__ ((target ("arch=x86-64")))
const char* foo () { return "x86-64-v1"; }

__attribute__ ((target ("arch=x86-64-v2")))
const char* foo () { return "x86-64-v2"; }

__attribute__ ((target ("arch=x86-64-v3")))
const char* foo () { return "x86-64-v3"; }

__attribute__ ((target ("arch=x86-64-v4")))
const char* foo () { return "x86-64-v4"; }

int main ()
{
  printf("%s\n", foo());
  return 0;
}

On my laptop, this prints

$ g++ multiversioning.c 
$ ./a.out 
x86-64-v3

Note that the use of g++ is intentional here.

If I used gcc to compile, it would fail with error: redefinition of ‘foo’.

1

The single thing that worked for me is to use gcc with its __builtin_cpu_supports feature.  Since I invoked it in msys it is likely to work on Windows too.  Can be done with C++ too.

// test_cpu.c
#ifndef __GNUC__
#error "You must use gnu"
#endif

#include <stdio.h>
int main() {
    if (__builtin_cpu_supports("x86-64-v4"))
        puts("v=4");
    else if (__builtin_cpu_supports("x86-64-v3"))
        puts("v=3");
    else if (__builtin_cpu_supports("x86-64-v2"))
        puts("v=2");
    else
        puts("v=1");
}

Usage:

$ gcc /test_cpu.c -o /test_cpu

$ /test_cpu
v=3
1

On more modern Fedora / Red Hat systems do this:

$ /usr/lib64/ld-linux-x86-64.so.2 --help
Usage: /usr/lib64/ld-linux-x86-64.so.2 [OPTION]... EXECUTABLE-FILE [ARGS-FOR-PROGRAM...]
You have invoked 'ld.so', the program interpreter for dynamically-linked
ELF programs.  Usually, the program interpreter is invoked automatically
when a dynamically-linked executable is started.

You may invoke the program interpreter program directly from the command
line to load and run an ELF executable file; this is like executing that
file itself, but always uses the program interpreter you invoked,
instead of the program interpreter specified in the executable file you
run.  Invoking the program interpreter directly provides access to
additional diagnostics, and changing the dynamic linker behavior without
setting environment variables (which would be inherited by subprocesses).

  --list                list all dependencies and how they are resolved
  --verify              verify that given object really is a dynamically linked
                        object we can handle
  --inhibit-cache       Do not use /etc/ld.so.cache
  --library-path PATH   use given PATH instead of content of the environment
                        variable LD_LIBRARY_PATH
  --glibc-hwcaps-prepend LIST
                        search glibc-hwcaps subdirectories in LIST
  --glibc-hwcaps-mask LIST
                        only search built-in subdirectories if in LIST
  --inhibit-rpath LIST  ignore RUNPATH and RPATH information in object names
                        in LIST
  --audit LIST          use objects named in LIST as auditors
  --preload LIST        preload objects named in LIST
  --argv0 STRING        set argv[0] to STRING before running
  --list-tunables       list all tunables with minimum and maximum values
  --list-diagnostics    list diagnostics information
  --help                display this help and exit
  --version             output version information and exit

This program interpreter self-identifies as: /lib64/ld-linux-x86-64.so.2

Shared library search path:
  (libraries located via /etc/ld.so.cache)
  /lib64 (system search path)
  /usr/lib64 (system search path)

Subdirectories of glibc-hwcaps directories, in priority order:
  x86-64-v4
  x86-64-v3
  x86-64-v2 (supported, searched)

You must log in to answer this question.

Not the answer you're looking for? Browse other questions tagged .