Metal fill kernel

[ivarflakstad]

Still uses blit command encoder for zeroes and all u8 values.

[Narsil]

Is there any benefit to using a CPU alloc and simply doing a memcpy ? Do you have benchmarks for it ?

Also keep the blit for zeros all the way. Remove the trait follow the rest of the crates convention.
Don't add any dependencies, they are not necessary.

[ivarflakstad]

It's a tiny bit faster, yeah 😉

fill_metal_u8/iter      time:   [101.33 µs 101.64 µs 101.98 µs]
                        thrpt:  [153.22 GiB/s 153.73 GiB/s 154.20 GiB/s]
                 change:
                        time:   [-96.183% -96.117% -96.019%] (p = 0.00 < 0.05)
                        thrpt:  [+2411.8% +2475.3% +2520.1%]

fill_metal_f32/iter     time:   [409.20 µs 437.94 µs 467.82 µs]
                        thrpt:  [133.60 GiB/s 142.71 GiB/s 152.74 GiB/s]
                 change:
                        time:   [-97.242% -97.113% -96.969%] (p = 0.00 < 0.05)
                        thrpt:  [+3199.1% +3364.0% +3525.2%]
                        Performance has improved.

Also keep the blit for zeros all the way.

That's what I'm doing 👍

[ivarflakstad]

Don't add any dependencies, they are not necessary.

If we want to be able to encode f16 and bf16 values into the kernel we kinda need it.
Alternatively we could send in the kernel name and the bits as u16 I guess?

[Narsil]

LGTM