MacBook Pro M1 Max doesn't give correct number of threads with --threads=auto

[navidcy]

I get different results for Threads.nthreads() on a Macbook Pro M1 Max using the latest version of julia and the tagged v1.7.1.

In particular, with latest Julia version I get 6! But when I look at System Info I see 8?

For julia#master at e0a4b77

navid:~/ $ /Users/navid/julia/julia --threads=auto
               _
   _       _ _(_)_     |  Documentation: https://docs.julialang.org
  (_)     | (_) (_)    |
   _ _   _| |_  __ _   |  Type "?" for help, "]?" for Pkg help.
  | | | | | | |/ _` |  |
  | | |_| | | | (_| |  |  Version 1.8.0-DEV.1455 (2022-02-06)
 _/ |\__'_|_|_|\__'_|  |  Commit e0a4b7727c (0 days old master)
|__/                   |

julia> versioninfo()
Julia Version 1.8.0-DEV.1455
Commit e0a4b7727c (2022-02-06 12:55 UTC)
Platform Info:
  OS: macOS (arm64-apple-darwin21.3.0)
  CPU: Apple M1 Max
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-13.0.0 (ORCJIT, cyclone)
Environment:
  JULIA_EDITOR = vim

julia> Threads.nthreads()
6

while on Julia v1.7

navid:~/ $ /Users/navid/julia-1.7/julia --threads=auto
               _
   _       _ _(_)_     |  Documentation: https://docs.julialang.org
  (_)     | (_) (_)    |
   _ _   _| |_  __ _   |  Type "?" for help, "]?" for Pkg help.
  | | | | | | |/ _` |  |
  | | |_| | | | (_| |  |  Version 1.7.1 (2021-12-22)
 _/ |\__'_|_|_|\__'_|  |
|__/                   |

julia> versioninfo()
Julia Version 1.7.1
Commit ac5cc99908 (2021-12-22 19:35 UTC)
Platform Info:
  OS: macOS (arm64-apple-darwin21.3.0)
  CPU: Apple M1 Max
  WORD_SIZE: 64
  LIBM: libopenlibm
  LLVM: libLLVM-12.0.1 (ORCJIT, cyclone)
Environment:
  JULIA_EDITOR = vim

julia> Threads.nthreads()
10

Probably the fix is related to this bit?

julia/src/sys.c

Lines 598 to 630 in 546a774

	// Apple's M1 processor is a big.LITTLE style processor, with 4x "performance"
	// cores, and 4x "efficiency" cores. Because Julia expects to be able to run
	// things like heavy linear algebra workloads on all cores, it's best for us
	// to only spawn as many threads as there are performance cores. Once macOS
	// 12 is released, we'll be able to query the multiple "perf levels" of the
	// cores of a CPU (see this PR [0] to pytorch/cpuinfo for an example) but
	// until it's released, we will just recognize the M1 by its CPU family
	// identifier, then subtract how many efficiency cores we know it has.
	JL_DLLEXPORT int jl_cpu_threads(void) JL_NOTSAFEPOINT
	{
	#if defined(HW_AVAILCPU) && defined(HW_NCPU)
	size_t len = 4;
	int32_t count;
	int nm[2] = {CTL_HW, HW_AVAILCPU};
	sysctl(nm, 2, &count, &len, NULL, 0);
	if (count < 1) {
	nm[1] = HW_NCPU;
	sysctl(nm, 2, &count, &len, NULL, 0);
	if (count < 1) { count = 1; }
	}
	#if defined(__APPLE__) && defined(_CPU_AARCH64_)
	// Manually subtract efficiency cores for Apple's big.LITTLE cores
	int32_t family = 0;
	len = 4;
	sysctlbyname("hw.cpufamily", &family, &len, NULL, 0);
	if (family >= 1 && count > 1) {
	if (family == CPUFAMILY_ARM_FIRESTORM_ICESTORM) {
	// We know the Apple M1 has 4 efficiency cores, so subtract them out.
	count -= 4;
	}
	}

@glwagner, @sandreza

[mkitti]

See #42099
Ping @staticfloat

No, Picard, there seem to be only 2.

[mkitti]

Since reference [0] got cut off, here it is again.

[0] https://github.com/pytorch/cpuinfo/blob/8ab2db2d405436f1014ed603021545b3b6b6f1ae/src/arm/mach/init.c#L161-L163

[navidcy]

@mkitti, that's relevant. But on Apple M1 Max processor one needs to subtract only 2 and not 4, as done in

julia/src/sys.c

Line 628 in 546a774

count -= 4;

Right?

[mkitti]

Right. See this comment: #42099 (comment) .

There appears to be a way to query this information rather than hard code it.

[gbaraldi]

@navidcy Could you kindly build the following C script and say what it prints for you.

#include <sys/sysctl.h>
#include <stdio.h>

int main()
{
    char str[7];
    size_t len = 7;
    sysctlbyname("kern.osrelease", str, &len, NULL, 0);
    printf("%s\n", str);
    if (str[0] > 1 && str[1] > 0)
    {
        len = 4;
        int ncpus;
        sysctlbyname("hw.perflevel0.physicalcpu", &ncpus, &len, NULL, 0);
        printf("%d\n", ncpus);
    }
}

[navidcy]

Sure, but I'd like some help with C.

How do I build this?

[gbaraldi]

It should be enough to put it into a file called test.c somewhere in your computer, navigate to that place with your terminal and do clang test.c -o test after that do ./test and it should print two lines.

[navidcy]

21.3.0
8

[gbaraldi]

Thanks for the help. I will open a PR that will make it set the correct number of threads.