MOD: optimize DGEMM of large matrices on cortex A53 & A55

[wjc404]

No description provided.

[wjc404]

Test programs:
OpenBLAS-test.tar.gz
(src/test_kernel.cpp for the correctness of kernel function, src/test_dgemm.cpp for the result and performance of DGEMM)

[wjc404]

I need some help for figuring out the syntax error in inline asm:) The asm block can be complied with NDK-r23b and gcc-7, but fails in the CI job.

[martin-frbg]

Sorry I do not see why it does not like the local labels here. As I read the error messages, it seems to think that jumping forward to 3: in the "blt 3f; beq 2f; .align 4; 1:\n\t" of line 84 would cross into a different loop section, and anything behind the "blt3f" on that line seems to get ignored so label 1: does not get defined at all(?) The only thing that stands out to me is the ".align 4" (and I remember earlier Mac assemblers being peculiar in their handlng of .align, making it necessary to use .p2align instead).

[martin-frbg]

Looks like the Apple Clang assembler simply wants to see the labels on a separate line (cf my test in #3452) - go figure...

[wjc404]

Thank you very much @martin-frbg :)