Detect the number of performance cores in the M1 macs via the new macos12 API

[gbaraldi]

Should fix #44067. The logic is a bit convoluted right now due to older mac versions not having the API.

[navidcy]

I tried this on a MacBook Pro M1 Max and it works.

navid:gbaraldi_julia/ (add-ncpus-apple) $ /Users/navid/gbaraldi_julia/julia --threads=auto                                                                                                                         [13:59:59]
               _
   _       _ _(_)_     |  Documentation: https://docs.julialang.org
  (_)     | (_) (_)    |
   _ _   _| |_  __ _   |  Type "?" for help, "]?" for Pkg help.
  | | | | | | |/ _` |  |
  | | |_| | | | (_| |  |  Version 1.8.0-DEV.1459 (2022-02-08)
 _/ |\__'_|_|_|\__'_|  |  add-ncpus-apple/304e20b1b6 (fork: 109 commits, 19 days)
|__/                   |

julia> versioninfo()
Julia Version 1.8.0-DEV.1459
Commit 304e20b1b6 (2022-02-08 02:45 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()
8

[vchuravy]

If someone can also see how to do that for Intel on Alderlake (on Linux/Windows). That might make @chriselrod happy

[giordano]

This is macOS-specific, but now we can run also Linux on this CPU 🙂 Do we know a similar API for Linux? I get 8 virtual cores in Asahi Linux
(sorry for the screenshot, can't boot this system right now)

Edit: well, I guess this is the same question as the one above 😄

[staticfloat]

Just shooting in the dark here, but can you look around in e.g. /sys/devices/system/cpu/cpu*/topology/ to see if there's anything interesting?

[giordano]

Difference between cpu0 and cpu3 (same "family"):

$ diff -ur /sys/devices/system/cpu/cpu{0,3}/topology/ 2>| /dev/null
diff -ur /sys/devices/system/cpu/cpu0/topology/cluster_cpus /sys/devices/system/cpu/cpu3/topology/cluster_cpus
--- /sys/devices/system/cpu/cpu0/topology/cluster_cpus  2022-03-27 18:30:03.818188313 +0100
+++ /sys/devices/system/cpu/cpu3/topology/cluster_cpus  2022-03-27 18:32:38.189584324 +0100
@@ -1 +1 @@
-01
+08
diff -ur /sys/devices/system/cpu/cpu0/topology/cluster_cpus_list /sys/devices/system/cpu/cpu3/topology/cluster_cpus_list
--- /sys/devices/system/cpu/cpu0/topology/cluster_cpus_list     2022-03-27 18:30:03.818188313 +0100
+++ /sys/devices/system/cpu/cpu3/topology/cluster_cpus_list     2022-03-27 18:32:38.189584324 +0100
@@ -1 +1 @@
-0
+3
diff -ur /sys/devices/system/cpu/cpu0/topology/core_cpus /sys/devices/system/cpu/cpu3/topology/core_cpus
--- /sys/devices/system/cpu/cpu0/topology/core_cpus     2022-03-27 18:30:03.818188313 +0100
+++ /sys/devices/system/cpu/cpu3/topology/core_cpus     2022-03-27 18:32:38.189584324 +0100
@@ -1 +1 @@
-01
+08
diff -ur /sys/devices/system/cpu/cpu0/topology/core_cpus_list /sys/devices/system/cpu/cpu3/topology/core_cpus_list
--- /sys/devices/system/cpu/cpu0/topology/core_cpus_list        2022-03-27 18:30:03.818188313 +0100
+++ /sys/devices/system/cpu/cpu3/topology/core_cpus_list        2022-03-27 18:32:38.189584324 +0100
@@ -1 +1 @@
-0
+3
diff -ur /sys/devices/system/cpu/cpu0/topology/core_id /sys/devices/system/cpu/cpu3/topology/core_id
--- /sys/devices/system/cpu/cpu0/topology/core_id       2022-03-27 18:30:03.818188313 +0100
+++ /sys/devices/system/cpu/cpu3/topology/core_id       2022-03-27 18:31:44.204678091 +0100
@@ -1 +1 @@
-0
+3
diff -ur /sys/devices/system/cpu/cpu0/topology/thread_siblings /sys/devices/system/cpu/cpu3/topology/thread_siblings
--- /sys/devices/system/cpu/cpu0/topology/thread_siblings       2022-03-27 18:30:03.818188313 +0100
+++ /sys/devices/system/cpu/cpu3/topology/thread_siblings       2022-03-27 18:32:38.190584322 +0100
@@ -1 +1 @@
-01
+08
diff -ur /sys/devices/system/cpu/cpu0/topology/thread_siblings_list /sys/devices/system/cpu/cpu3/topology/thread_siblings_list
--- /sys/devices/system/cpu/cpu0/topology/thread_siblings_list  2022-03-27 18:30:03.818188313 +0100
+++ /sys/devices/system/cpu/cpu3/topology/thread_siblings_list  2022-03-27 18:32:38.190584322 +0100
@@ -1 +1 @@
-0
+3

Difference between cpu3 and cpu4 (different "families"):

$ diff -ur /sys/devices/system/cpu/cpu{3,4}/topology/ 2>| /dev/null
diff -ur /sys/devices/system/cpu/cpu3/topology/cluster_cpus /sys/devices/system/cpu/cpu4/topology/cluster_cpus
--- /sys/devices/system/cpu/cpu3/topology/cluster_cpus  2022-03-27 18:32:38.189584324 +0100
+++ /sys/devices/system/cpu/cpu4/topology/cluster_cpus  2022-03-27 18:32:38.189584324 +0100
@@ -1 +1 @@
-08
+10
diff -ur /sys/devices/system/cpu/cpu3/topology/cluster_cpus_list /sys/devices/system/cpu/cpu4/topology/cluster_cpus_list
--- /sys/devices/system/cpu/cpu3/topology/cluster_cpus_list     2022-03-27 18:32:38.189584324 +0100
+++ /sys/devices/system/cpu/cpu4/topology/cluster_cpus_list     2022-03-27 18:32:38.189584324 +0100
@@ -1 +1 @@
-3
+4
diff -ur /sys/devices/system/cpu/cpu3/topology/core_cpus /sys/devices/system/cpu/cpu4/topology/core_cpus
--- /sys/devices/system/cpu/cpu3/topology/core_cpus     2022-03-27 18:32:38.189584324 +0100
+++ /sys/devices/system/cpu/cpu4/topology/core_cpus     2022-03-27 18:32:38.189584324 +0100
@@ -1 +1 @@
-08
+10
diff -ur /sys/devices/system/cpu/cpu3/topology/core_cpus_list /sys/devices/system/cpu/cpu4/topology/core_cpus_list
--- /sys/devices/system/cpu/cpu3/topology/core_cpus_list        2022-03-27 18:32:38.189584324 +0100
+++ /sys/devices/system/cpu/cpu4/topology/core_cpus_list        2022-03-27 18:32:38.189584324 +0100
@@ -1 +1 @@
-3
+4
diff -ur /sys/devices/system/cpu/cpu3/topology/core_id /sys/devices/system/cpu/cpu4/topology/core_id
--- /sys/devices/system/cpu/cpu3/topology/core_id       2022-03-27 18:31:44.204678091 +0100
+++ /sys/devices/system/cpu/cpu4/topology/core_id       2022-03-27 18:31:44.204678091 +0100
@@ -1 +1 @@
-3
+0
diff -ur /sys/devices/system/cpu/cpu3/topology/core_siblings /sys/devices/system/cpu/cpu4/topology/core_siblings
--- /sys/devices/system/cpu/cpu3/topology/core_siblings 2022-03-27 18:32:38.189584324 +0100
+++ /sys/devices/system/cpu/cpu4/topology/core_siblings 2022-03-27 18:32:38.189584324 +0100
@@ -1 +1 @@
-0f
+f0
diff -ur /sys/devices/system/cpu/cpu3/topology/core_siblings_list /sys/devices/system/cpu/cpu4/topology/core_siblings_list
--- /sys/devices/system/cpu/cpu3/topology/core_siblings_list    2022-03-27 18:32:38.189584324 +0100
+++ /sys/devices/system/cpu/cpu4/topology/core_siblings_list    2022-03-27 18:32:38.189584324 +0100
@@ -1 +1 @@
-0-3
+4-7
diff -ur /sys/devices/system/cpu/cpu3/topology/package_cpus /sys/devices/system/cpu/cpu4/topology/package_cpus
--- /sys/devices/system/cpu/cpu3/topology/package_cpus  2022-03-27 18:32:38.190584322 +0100
+++ /sys/devices/system/cpu/cpu4/topology/package_cpus  2022-03-27 18:32:38.190584322 +0100
@@ -1 +1 @@
-0f
+f0
diff -ur /sys/devices/system/cpu/cpu3/topology/package_cpus_list /sys/devices/system/cpu/cpu4/topology/package_cpus_list
--- /sys/devices/system/cpu/cpu3/topology/package_cpus_list     2022-03-27 18:32:38.190584322 +0100
+++ /sys/devices/system/cpu/cpu4/topology/package_cpus_list     2022-03-27 18:32:38.190584322 +0100
@@ -1 +1 @@
-0-3
+4-7
diff -ur /sys/devices/system/cpu/cpu3/topology/physical_package_id /sys/devices/system/cpu/cpu4/topology/physical_package_id
--- /sys/devices/system/cpu/cpu3/topology/physical_package_id   2022-03-27 18:32:05.788637996 +0100
+++ /sys/devices/system/cpu/cpu4/topology/physical_package_id   2022-03-27 18:32:05.788637996 +0100
@@ -1 +1 @@
-0
+1
diff -ur /sys/devices/system/cpu/cpu3/topology/thread_siblings /sys/devices/system/cpu/cpu4/topology/thread_siblings
--- /sys/devices/system/cpu/cpu3/topology/thread_siblings       2022-03-27 18:32:38.190584322 +0100
+++ /sys/devices/system/cpu/cpu4/topology/thread_siblings       2022-03-27 18:32:38.190584322 +0100
@@ -1 +1 @@
-08
+10
diff -ur /sys/devices/system/cpu/cpu3/topology/thread_siblings_list /sys/devices/system/cpu/cpu4/topology/thread_siblings_list
--- /sys/devices/system/cpu/cpu3/topology/thread_siblings_list  2022-03-27 18:32:38.190584322 +0100
+++ /sys/devices/system/cpu/cpu4/topology/thread_siblings_list  2022-03-27 18:32:38.190584322 +0100
@@ -1 +1 @@
-3
+4

Also, looking at /proc/cpuinfo I noticed that CPUs from the two different "families" have different "CPU part" value:

$ grep -A6 'processor.*:.*[34]' /proc/cpuinfo
processor       : 3
BogoMIPS        : 48.00
Features        : fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm jscvt fcma lrcpc dcpop sha3 asimddp sha512 asimdfhm dit uscat ilrcpc flagm ssbs sb paca pacg dcpodp flagm2 frint
CPU implementer : 0x61
CPU architecture: 8
CPU variant     : 0x1
CPU part        : 0x022
--
processor       : 4
BogoMIPS        : 48.00
Features        : fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm jscvt fcma lrcpc dcpop sha3 asimddp sha512 asimdfhm dit uscat ilrcpc flagm ssbs sb paca pacg dcpodp flagm2 frint
CPU implementer : 0x61
CPU architecture: 8
CPU variant     : 0x1
CPU part        : 0x023

I'd also note that the "About this System" window in Asahi Linux correctly identifies the two different CPUs (4 firestorm and 4 icestorm):

So I'm moderately sure there is a way to programmatically distinguish the two CPUs (I'd probably need to dig out the code of this window), although I'm not sure there is a macOS-like way to query the "performance" ones only, and hardcoding this logic isn't very scalable.

[yuyichao]

So I'm moderately sure there is a way to programmatically distinguish the two CPUs

julia/src/processor_arm.cpp

Line 784 in 19eb307

// /sys/devices/system/cpu/cpu<n>/regs/identification/midr_el1 reader

[chritkhalil]

I'm currently running the newly released Julia version 1.9.1 on an m2 mac air, and I'm facing the same issue.

After reading through this thread about the detection of the number of performance cores in m1 CPUs, I wonder if the same patch may need to be applied to the m2 chips or if there's a different problem. I believe that BLAS.get_num_threads() should return 8 on an 8-core system, unless its intently restricted to use of some cores due to the nature of the architecture (maybe?).

Testing with 1.7.1 and 1.8.0-beta1~x64, BLAS.get_num_threads() returns 8

[gbaraldi]

So, from measurements, using all cores lead to slower performance than limiting to only performance cores, it also made the system very sluggish. I might be wrong but the 8 core m2 has 4 performance and 4 efficiency which means we tell blas to use 4 cores.

[staticfloat]

That is correct, we default to the number of performance cores, as the efficiency cores provide very little benefit to Julia, but create a large performance issue for the rest of the system. You can see how many performance cores you have by opening Activity Monitor and then clicking Window -> CPU History, which gives you a nice core-by-core activity graph, and annotates which are efficiency and performance. As an example, my M1 Pro has 8 performance cores: