Just so folks have the heads up, here is the discussion of the blog post on 
reddit: http://www.reddit.com/r/rust/comments/1igvye/vision_for_rust_simd/

~Brendan

On 16/07/2013, at 2:27 PM, Jens Nockert <[email protected]> wrote:

> Hello rust-dev!
> 
> I implemented OpenCL-style[0] accessors for SIMD types in Rust[1], the code 
> quality isn't near /merge-worthy yet, but I wanted some input.
> 
> For some code-examples, go to 
> https://github.com/jensnockert/rust/tree/simd/src/test/run-pass/simd-test and 
> check it out, you should be able to get the idea of how they work.
> 
> Note that I didn't add any actual syntax for vector types yet (since it would 
> be highly controversial and I don't know what would be the best option), so I 
> just added a simd!(name: T * n) syntax extension that declares a new type 
> that maps down to a LLVM <n * T>.
> 
> My preference for syntax right now would be simd!(n x T) if I can get that to 
> parse, or simd!(T, ..n). And then you would declare a type with type f32x4 = 
> simd(4 x f32); and it would magically work. Another option would be some 
> variant of the [T, ..n] syntax used for fixed-length vectors.
> 
> Introducing a new t, ty_simd_vec(t, uint), instead of using the current 
> #[simd] struct { … }, is yet another thing that is controversial about the 
> patch and this needs a bit of explanation of the problem with #[simd] struct 
> { … }.
> 
> To be able to make these accessors work, you unfortunately need to be able to 
> generate anonymous types in the compiler, x.even for example (take the 
> even-indexed elements of a vector) may be a type that is undeclared. And if 
> you want to be able to pretty-print that, you need to be able to generate a 
> type without a name, which makes #[simd] struct { … } impossible.
> 
> You could just pre-declare all possible options up to 256-bit long, which 
> probably would only be a hundred types or so, but would feel a bit silly.
> 
> There are also other operations that could generate (possibly) unnamed types, 
> like a == b, which should generate a i1-vector, or a shufflevector intrinsic 
> that could generate vectors of any length.
> 
> Ps. I didn't think of #[simd] (T, T, T, T) &c. before implementing (sanxiyn 
> gave me that idea), but I still think that is probably a worse idea than 
> adding SIMD as an additional type with actual syntax.
> 
> [0]: http://www.khronos.org/files/opencl-quick-reference-card.pdf, page 2, 
> "Vector Component Addressing"
> [1]: https://github.com/jensnockert/rust/tree/simd
> _______________________________________________
> Rust-dev mailing list
> [email protected]
> https://mail.mozilla.org/listinfo/rust-dev

_______________________________________________
Rust-dev mailing list
[email protected]
https://mail.mozilla.org/listinfo/rust-dev

Reply via email to