add support for Hexagon HVX

[androm3da]

Add support for HVX, omitting support for 128-lane for now.

[programmerjake]

quite a lot of the library assumes the max lane count is 64, at the very least basically all the tests that only test up to 64 lanes should also test more than 64 lanes. also the bitmasks code uses u64 so is restricted to at most 64 lanes.

[programmerjake]

getting all the tests to pass with 128 lanes is going to be a lot of work and be somewhat controversial (though we do want to increase the limit to more than 128 at some point), so if you want this PR to have a good chance of getting merged quickly, I'd suggest leaving out the change to 128 lanes and maybe just have the 1 byte x 128 lanes hexagon types be commented out for now with a fixme to re-enable them when we get around to adding support for >64 lanes.

[androm3da]

somewhat controversial

Would it be a good idea for me to create an MCP?

getting merged quickly

I'm not in a hurry.

leaving out the change to 128 lanes and maybe just have the 1 byte x 128 lanes hexagon types be commented out for now with a fixme to re-enable them when we get around to adding support for >64 lanes.

Sounds good to me: will do.

[androm3da]

BTW: not sure if I understood the goal of the checklist but 3e0f44c Add comprehensive edge case tests for SIMD operations was intended to handle that.

[programmerjake]

somewhat controversial

Would it be a good idea for me to create an MCP?

It's probably not necessary.

BTW: not sure if I understood the goal of the checklist but 3e0f44c Add comprehensive edge case tests for SIMD operations was intended to handle that.

the checklist is more for if you were adding new functions like sin or ilog10 which are non-trivial, you're adding new types where the actual conversion code is pretty trivial so testing the conversions isn't that important.

[androm3da]

the checklist is more for if you were adding new functions like sin or ilog10 which are non-trivial, you're adding new types where the actual conversion code is pretty trivial so testing the conversions isn't that important.

Ok so should I drop 3e0f44c ? Seems like it's not necessary.

[programmerjake]

the checklist is more for if you were adding new functions like sin or ilog10 which are non-trivial, you're adding new types where the actual conversion code is pretty trivial so testing the conversions isn't that important.

Ok so should I drop 3e0f44c ? Seems like it's not necessary.

it isn't necessary for this PR.

[androm3da]

Sorry: I misunderstood how this crate gets integrated - I will need to land rust-lang/stdarch#1999 first and then re-evaluate the design of HVX in portable-simd after that. Closing.