DYNAMIC_ARCH woes

[dnoan]

Recently, OpenBLAS added support for VORTEXM4. I attempted to build OpenBLAS with support for both VORTEX and VORTEXM4 in order to achieve optimal performance across Apple M-series chips.

I built the library using CMake with the following options:

-DDYNAMIC_ARCH=ON  
-DDYNAMIC_LIST="VORTEX;VORTEXM4"  
-DTARGET=VORTEX

Expected behavior

• The common code should be compiled for VORTEX.
• On Apple M1/M2/M3 systems, the VORTEX code path should be selected at runtime.
• On Apple M4/M5 systems, the VORTEXM4 code path should be selected at runtime.

Observed behavior (macOS / Apple Silicon)

To verify runtime dispatch, I set OPENBLAS_VERBOSE=2 and ran tests on two systems:

sysctl "machdep.cpu.brand_string"
machdep.cpu.brand_string: Apple M2 Max
Core: armv8

sysctl "machdep.cpu.brand_string"
machdep.cpu.brand_string: Apple M4 Pro
Core: armv8

In both cases, OpenBLAS reported ARMV8 rather than selecting VORTEX or VORTEXM4. Performance was also lower than expected.

Then I decided to check how this works on other platforms.

Observed behavior (AArch64)

I then tested on several AArch64 systems using the following settings in Makefile.rule:

TARGET=ARMV8
DYNAMIC_ARCH=1

OpenBLAS did not report any specific core type at runtime.

Observed behavior (x86_64 Linux)

Makefile.rule settings:

TARGET=NEHALEM
DYNAMIC_ARCH=1

I observed the following:

Model name: AMD EPYC 7301 16-Core Processor
Core: Zen

Model name: AMD Ryzen 5 5500
Core: Zen

Model name: AMD Ryzen Threadripper PRO 7995WX 96-Cores
Core: Cooperlake

Model name: AMD Ryzen 9 9950X3D 16-Core Processor
Core: Cooperlake

Zen was used for cores without AVX512 and Cooperlake for cores with AVX512.

Observed behavior (Windows on ARM)

I also attempted to build on Windows on ARM with:

-DDYNAMIC_ARCH=ON  
-DDYNAMIC_LIST="NEOVERSEN1;CORTEXX1"  
-DTARGET=ARMV8

This resulted in a compilation error:

C:\OpenBLAS-0.3.32\driver\others\dynamic_arm64.c(41,10): fatal error: 'strings.h' file not found
   41 | #include <strings.h>
      |          ^~~~~~~~~~~

Summary

• Runtime CPU detection does not appear to select VORTEX/VORTEXM4 on Apple Silicon.
• AArch64 builds do not report a detected core type.
• Cooperlake code path is used on ZEN 4/5.
• Windows on ARM build fails due to missing <strings.h>.

Any guidance on whether this behavior is expected (or if I am misconfiguring the build) would be appreciated.

[dnoan]

Additional information:

I used the openblas_get_corename() call to get the core type used on aarch64. It always returns ARMV8, even on Neoverse CPUs. For example, I have this Neoverse N1 CPU:

head /proc/cpuinfo
processor       : 0
BogoMIPS        : 50.00
Features        : fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm lrcpc dcpop asimddp ssbs
CPU implementer : 0x41
CPU architecture: 8
CPU variant     : 0x3
CPU part        : 0xd0c
CPU revision    : 1

lscpu | grep "Model name:"
Model name:                              Neoverse-N1

openblas_get_corename() returns the correct core on x64, for example SkylakeX on Xeon Silver 4114.

[martin-frbg]

Possibly DYNAMIC_LIST does not work correctly with the VORTEX types, I'll try to look later today. But DYNAMIC_ARCH should definitely work on aarch64 in general, and report the correct coretype at least on non-Mac systems.
The WoA build issue is that LLVM does not install the strings.h header - I am preparing a PR for your older issue that simply spells out the prototype for strncasecmp directly.

Cooperlake for Zen5 is by design, Zen4 uses SkylakeX (unless compiler or os preclude the use of AVX512, in which case both fall back to Zen that is essentially Haswell)

[martin-frbg]

VORTEX doesn't currently exist for the purpose of DYNAMIC_ARCH (gets mapped to NeoverseN1 if that is available). VORTEXM4 does exist and recognition in DYNAMIC_ARCH depends entirely on the hw.optional.arm.FEAT_SME flag so should work for any M4 or M5 - but if there is no SME support in the compiler toolchain, it too gets mapped to NeoverseN1 if available. With no fallback to NeoverseN1 due to the restricted list, everything drops to ArmV8