add accelerate provider

[astraw]

Hi, I managed to get Apple's Accelerate.framework packaged up and published to crates.io. It would be create if you could merge this and make a new release. Thanks!

[IvanUkhov]

@astraw, I really appreciate these four pull requests! I’ve been planning to create this provider for ages. It’ll be interesting to see how Accelerate compares with OpenBLAS.

Regards,
Ivan

[IvanUkhov]

Here are some superficial benchmarks:

Accelerate:

multiply_left_0010   ...         429 ns/iter (+/- 341)
multiply_left_0100   ...     204,151 ns/iter (+/- 173,491)
multiply_left_1000   ... 114,508,598 ns/iter (+/- 10,434,712)
multiply_right_0010  ...         428 ns/iter (+/- 312)
multiply_right_0100  ...     211,328 ns/iter (+/- 50,826)
multiply_right_1000  ... 114,822,024 ns/iter (+/- 45,520,929)
symmetric_eigen_0010 ...      21,009 ns/iter (+/- 17,384)
symmetric_eigen_0100 ...   2,710,091 ns/iter (+/- 170,804)
symmetric_eigen_1000 ... 643,218,416 ns/iter (+/- 81,163,491)

OpenBLAS:

multiply_left_0010   ...         637 ns/iter (+/- 254)
multiply_left_0100   ...     136,785 ns/iter (+/- 101,942)
multiply_left_1000   ... 105,401,081 ns/iter (+/- 36,737,045)
multiply_right_0010  ...         583 ns/iter (+/- 101)
multiply_right_0100  ...     121,236 ns/iter (+/- 27,124)
multiply_right_1000  ... 104,822,927 ns/iter (+/- 16,554,278)
symmetric_eigen_0010 ...      39,096 ns/iter (+/- 45,000)
symmetric_eigen_0100 ...  11,619,836 ns/iter (+/- 1,533,766)
symmetric_eigen_1000 ... 918,963,320 ns/iter (+/- 140,549,251)

Looks like Accelerate is good at computing eigendecompositions.

Regards,
Ivan