Implementaion of dasum, sasum with AVX2 & AVX512 intrinsic

[xiegengxin]

Implement dasum, sasum with AVX2 & AVX512 intrinsic for x86_64 architecture to improve the performance.

[brada4]

I think you need to force-align presumed jump targets to palign(5) or so , but measure.
What about AMD ZEN and Intel COOPERLAKE ?

[brada4]

AVX2 and AVX512 needs some warmup, otherwise for small samples it is hundreds of cycles unproductive

[xiegengxin]

1. Align to 64 byte or 32 byte.
2. When input array size is small (less than 256), using SSE intrinsic to instead of AVX512/AVX2. The performance won't worse too much.

input size	Optimizaion vs Before
16	0.96x
32	1.03x
64	1.01x
128	0.99x
256	0.99x
512	0.95x
1024	1.16x
2048	1.18x

[brada4]

From excellent resource at agner.org

The first YMM or ZMM instruction takes 150-250 clock cycles -probably to start a power-up process

How does it work with intentionally unaligned input, and not ending so round?