As part of my GSoC project, I wanted to use SIMD intrinsics. But I could not find a way to write cross-compiler code. That includes core.simd, but also intel-intrinsics.
Specifically intel-intrinsics is a great library, but it could not be used in GDC (at least I did not find a way) and it seems LDC focused.

So, I wrote a mini SIMD module [1]. This is not very general as it was targeted only for the Dmem* utilities. But then I thought, why not make such a module in the druntime?

In the link, you can see for example a simple store function. Also, it shows
the best I got as far as how to answer the question "Is SIMD supported?" across
compilers.

[1] https://github.com/dlang/druntime/pull/2687/files#diff-c2fcd73761ae6659ef91245ce1195b6dR70

On Friday, 2 August 2019 at 14:56:58 UTC, Stefanos Baziotis wrote:
> Specifically intel-intrinsics is a great library, but it could not be used in GDC (at least I did not find a way) and it seems LDC focused.
>

Hello,

The reason is that no one stepped up to add this support for GDC intel-intrinsics, but it is on the roadmap for next year. I can't contribute to D OSS all day, since I need to put out products made with D too ; so users of intel-intrinsics have to be patient or contribute themselves.

Do not underestimate the sheer amount of work needed to make "cross-compiler SIMD module".

On Friday, 2 August 2019 at 20:38:37 UTC, Guillaume Piolat wrote:
>
> Hello,
>
> The reason is that no one stepped up to add this support for GDC intel-intrinsics, but it is on the roadmap for next year. I can't contribute to D OSS all day, since I need to put out products made with D too ;
>

Hello Guillaume,

Don't take this as criticism to intel-intrinsics. On the contrary, I mention it here and there as it is the best SIMD library for D I know (apart from core.simd).

However, I feel like this should be part of druntime core and not an external
library since core.simd already exists but is not cross-compiler.

> Do not underestimate the sheer amount of work needed to make "cross-compiler SIMD module".

I agree, although I probably don't know all the hurdles since I have not made one.
I proposed the idea and I'm willing to put time on it, if
the community wants it. It doesn't necessarily mean that I will be able to do it myself.

I invite people who have more experience (like you) to give advice and / or help.

> so users of intel-intrinsics have to be patient or contribute themselves.

If the community is not positive on that, then my plan is to contribute
to intel-intrinsics.

Best,
Stefanos

On Friday, 2 August 2019 at 14:56:58 UTC, Stefanos Baziotis wrote:
> But then I thought, why not make such a module in the druntime?

Johan suggests implementing some small core.simd subset (load/storeUnaligned, prefetch) in the LDC and GDC druntime flavours instead:
https://github.com/ldc-developers/ldc/issues/3121

On Saturday, 3 August 2019 at 00:13:09 UTC, kinke wrote:
>
> Johan suggests implementing some small core.simd subset (load/storeUnaligned, prefetch) in the LDC and GDC druntime flavours instead:
> https://github.com/ldc-developers/ldc/issues/3121

Yes, that could be done as well.

On Friday, 2 August 2019 at 20:59:04 UTC, Stefanos Baziotis wrote:
>
> However, I feel like this should be part of druntime core and not an external
> library since core.simd already exists but is not cross-compiler.


Well yes what you did makes a lot of sense, since it's druntime that will have to implement memcpy/memset/etc. So there is indeed a need for something in druntime.

(intel-intrinsics is Boost-licensed and one can use any part of it, but it's not like is has been "approved" in terms of DUB popularity (which I feel is a necessary first step, along with GDC support)).

On Saturday, 3 August 2019 at 10:01:37 UTC, Guillaume Piolat wrote:
> On Friday, 2 August 2019 at 20:59:04 UTC, Stefanos Baziotis
>
> Well yes what you did makes a lot of sense, since it's druntime that will have to implement memcpy/memset/etc. So there is indeed a need for something in druntime.
>

Apart from memset, memcpy etc., I think that if we are to have a simd
module, that makes sense to be cross-compiler. On the other hand, people seem to
like the idea to add them separately on LDC and GDC versions. I don't know.

> (intel-intrinsics is Boost-licensed and one can use any part of it, but it's not like is has been "approved" in terms of DUB popularity (which I feel is a necessary first step, along with GDC support)).

Well, it seems like a sane way to do SIMD. I hope it goes well.

Stefanos