How to identify which assembly file is being executed when cgemm is called from bench_blas.py?

[rajith-ravikumar-mcw]

There are several files located in the OpenBLAS/kernel/arm64/ directory that begin with the prefix "cgemm":

• cgemm_kernel_4x4.S
• cgemm_kernel_8x4.S
• cgemm_kernel_8x4_cortexa53.c
• cgemm_kernel_8x4_thunderx2t99.S
• cgemm_kernel_sve_v1x4.S
• cgemm_ncopy_sve_v1.c
• cgemm_tcopy_sve_v1.c

The script bench_blas.py calls functions for the input types c, d, s, and z, using input sizes of 100 and 1000.

I am looking to optimize these specific functions' implementations.

[martin-frbg]

The cpu on which the benchmark is being run determines which kernel will be used, through the CGEMMKERNEL entry in kernel/arm64/KERNEL.target (where target is one of the supported cpu names from TargetList.txt, like ARMV8, ARMV8SVE or CORTEXA57, etc). (These KERNEL files use Makefile syntax and are evaluated at build time)
If you are using a multi-cpu (DYNAMIC_ARCH=1) build, you can set the environment variable OPENBLAS_VERBOSE=2 to tell you
what cpu target was autodetected by the library
If the entry uses variable names CGEMM_UNROLL_M and/or CGEMM_UNROLL_N, you need to look up their values in the toplevel param.h first - this is where the "4x4" or "8x4" etc come from. There is no difference in CGEMM kernel choice between different sizes, but small sizes may be computed in single-threaded mode (see file interface/gemm.c for the threshold).

For SGEMM and DGEMM, there is now an additional set of "small" kernels that gets invoked directly, without passing through the block matrix setup code in driver/level3 first

[rajith-ravikumar-mcw]

Hi Martin,

Thanks for the clear explanation. I have added an example for reference to this thread.

For example:
Lets take cgemm test being executed on an armv8 machine

OpenBLAS/kernel/arm64/KERNEL.ARMV8 contains the definition for CGEMMKERNEL

as

CGEMMKERNEL = cgemm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N).S

The definitions for $(CGEMM_UNROLL_M) and $(CGEMM_UNROLL_N) are present in

OpenBLAS/param.h

For this example:
/* Common ARMv8 parameters */ #if defined(ARMV8) #define CGEMM_DEFAULT_UNROLL_M 8 #define CGEMM_DEFAULT_UNROLL_N 4

So the kernel that is executed is
OpenBLAS/kernel/arm64/cgemm_kernel_8x4.S

[rajith-ravikumar-mcw]

Hi Martin,

Quick question: If I modify cgemm_kernel_8x4.S, build OpenBLAS, and use the new openblas.dll and openblas.lib to generate the pyd file with Meson for benchmarking , will the updated kernel be used?

[martin-frbg]

I would expect so - but to be sure, please do a complete rebuild (i.e. run make clean or equivalent between builds) as the build file structure is a bit complicated and just rebuilding/replacing a single object may not always work as expected.

[Mugundanmcw]

@martin-frbg I could see a tool codespeed.io that could provide execution details upto kernal execution, Is there any equivalent tool that could to the same as the tool I mentioned?
Thanks

[martin-frbg]

codspeed.io is what we use for the benchmark that runs on every PR, but I am almost certain that this service runs on Linux where one can easily use perf to get a performance analysis down to function level (and individual lines in the machine code of each). I do not know of any Windows tool that would do the same