[WIP] Add test for DNRM2 bug on Neoverse

[martin-frbg]

#3654

[larsoner]

Hooray Travis!

0.01s$ ./gesdd
Note: The following floating-point exceptions are signalling: IEEE_INVALID_FLAG IEEE_OVERFLOW_FLAG IEEE_UNDERFLOW_FLAG
STOP Error while computing the SVD! 
 ** On entry to DLASCL parameter number  4 had an illegal value
 ** On entry to DLASCL parameter number  4 had an illegal value

+1 for adding this as a proper test case to the test suite if possible/reasonable. (I am not sure how this would best be done...)

So far in #3654 I tested locally that KERNEL.VORTEX works when it's:

include $(KERNELDIR)/KERNEL.NEOVERSEN1
DNRM2KERNEL = ../arm/nrm2.c
SNRM2KERNEL = ../arm/nrm2.c

I saw in this comment you mention using nrm2.S -- I'm not sure if this is the one you meant, but locally for me using:

include $(KERNELDIR)/KERNEL.NEOVERSEN1
DNRM2KERNEL = ../arm64/nrm2.S
SNRM2KERNEL = ../arm64/nrm2.S

It also appears to work!

So +1 from me to changing the target to use ../arm64/nrm2.S

[larsoner]

... I see you pushed 62a21e3 while I was typing. Hopefully I'm faster than Travis on this one:

el_LN\" -DNO_AFFINITY -I.. -DDOUBLE  -UCOMPLEX -c -DTRMMKERNEL -DDOUBLE -UCOMPLEX -DLEFT -UTRANSA ../kernel/arm64/dtrmm_kernel_8x4.S -o dtrmm_kernel_LN.o
../kernel/arm64/dznrm2_thunderx2t99.c:87:3: error: invalid operand for instruction
        "       fmov     x6, 1.0 \n" 
         ^
<inline asm>:15:15: note: instantiated into assembly here
        fmov     x6, 1.0 
                     ^
../kernel/arm64/dznrm2_thunderx2t99.c:87:3: error: invalid operand for instruction
        "       fmov     x6, 1.0 \n" 
         ^
<inline asm>:15:15: note: instantiated into assembly here
        fmov     x6, 1.0 
                     ^

[larsoner]

FYI you can also push with [skip appveyor] [skip actions] in your commit if you want to save the CIs some cycles. You can also manually add a skipper for Azure to allow [skip azp] if you want like SciPy has at some point:

https://github.com/scipy/scipy/blob/962c3ddfd018b2ba1ba99c4aed70bc565a3693a7/azure-pipelines.yml#L35-L53

[martin-frbg]

Thx. Too hot here today - I'll probably move the initial testing to my new ARM9 tablet before this gets too embarassing.

[larsoner]

FWIW I'm happy to test whatever you push -- or ignore if you'd rather wait for Travis :) I'm just happy you're taking the time to look!

[larsoner]

@martin-frbg WDYT about for now just using

include $(KERNELDIR)/KERNEL.NEOVERSEN1
DNRM2KERNEL = ../arm64/nrm2.S
SNRM2KERNEL = ../arm64/nrm2.S

? It would remove any time pressure from getting this fix out in the wild, especially if it were cut as a minor bugfix release