Any usable SIMD implementation?

[Manu via Digitalmars-d]

On 7 April 2016 at 13:27, Walter Bright via Digitalmars-d
<digitalmars-d at puremagic.com> wrote:
> On 4/6/2016 7:43 PM, Manu via Digitalmars-d wrote:
>>>
>>> 1. This has been characterized as a blocker, it is not, as it does not
>>> impede writing code that takes advantage of various SIMD code generation
>>> at
>>> compile time.
>>
>>
>> It's sufficiently blocking that I have not felt like working any
>> further without this feature present. I can't feel like it 'works' or
>> it's 'done', until I can demonstrate this functionality.
>> Perhaps we can call it a psychological blocker, and I am personally
>> highly susceptible to those.
>
>
> I can understand that it might be demotivating for you, but that is not a
> blocker. A blocker has no reasonable workaround. This has a trivial
> workaround:
>
>    gdc -simd=AFX foo.d
>
> becomes:
>
>    gdc -simd=AFX -version=AFX foo.d
>
> It's even simpler if you use a makefile variable:
>
>     FPU=AFX
>
>     gdc -simd=$(FPU) -version=$(FPU)

Sure. I've done this in my own tests. I just never published that
anyone else should do it.


> You also mentioned being blocked (i.e. demotivated) for *years* by this, and
> I assume that may be because we don't care about SIMD support. That would be
> wrong, as I care a lot about it. But I had no idea you were having a problem
> with this, as you did not file any bug reports. Suffering in silence is
> never going to work :-)

There's been threads, but sure, I could have done more to push it along.
Motivation is a complex and not particularly logical emotion, there's
a lot of factors feeding into it.

Not least of which, is that I haven't been working in games for a
while, which means I haven't depended on it for my work. Don't take
that to read I have lost interest in the support, just that the
pressure is reduced.
You'll have noticed that C++ interaction is my recent focus, since
that's directly related to my current day-job, and the path that I
need to solve now to get D into my work.
That's consuming almost 100% of my D-time-allocation... if I could
ever manage to just kick that goal, it might free me up >_< .. I keep
on trying.


>>> 2. I'm not sure these global settings are the best approach, especially
>>> if
>>> one is writing applications that dynamically adjusts based on the CPU the
>>> user is running on.
>>
>>
>> They are necessary to provide a baseline. It is typical when building
>> code that you specify a min-spec. This is what's used by default
>> throughout the application.
>
>
> It is not necessary to do it that way. Call std.cpuid to determine what is
> available at runtime, and issue an error message if not. There is no runtime
> cost to that. In fact, it has to be done ANYWAY, as it isn't user friendly
> to seg fault trying to execute instructions that do not exist.

The author still needs to be able to control at compile-time what
min-spec shall be supported.
I agree the check is valuable, but I think it's an unrelated detail.


>> Runtime selection is not practical in a broad sense. Emitting small
>> fragments of SIMD here and there will probably take a loss if they are
>> all surrounded by a runtime selector. SIMD is all about pipelining,
>> and runtime branches on SIMD version are antithesis to good SIMD
>> usage; they can't be applied for small-scale deployment.
>> In my experience, runtime selection is desirable for large scale
>> instantiations at an outer level of the work loop. I've tried to
>> design this intent in my library, by making each simd API capable of
>> receiving SIMD version information via template arg, and within the
>> library, the version is always passed through to dependent calls.
>> The Idea is, if you follow this pattern; propagating a SIMD version
>> template arg through to your outer function, then you can instantiate
>> your higher-level work function for any number of SIMD feature
>> combinations you feel is appropriate.
>
>
> Doing it at a high level is what I meant, not for each SIMD code fragment.

Sure, so you agree we need a mechanism for the author to tune the
default selection then? Or are you suggesting SSE2 is 'fine' as a
default? (ie, that is what is implied by D_SIMD)


>> Naturally, this process requires a default, otherwise this usage
>> baggage will cloud the API everywhere (rather than in the few cases
>> where a developer specifically wants to make use of it), and many
>> developers in 2015 feel SSE2 is a weak default. I would choose SSE4.1
>> in my applications, xbox developers would choose AVX1, it's very
>> application/target-audience specific, but SSE2 is the only reasonable
>> selection if we are not to accept a hint from the command line.
>
>
> I still don't see how it is a problem to do the switch at a high level.

It's not a problem, that's exactly my design, but it's not a universal solution.

> Heck, you could put the ENTIRE ENGINE inside a template, have a template
> parameter be the instruction set, and instantiate the template for each
> supported instruction set.
>
> Then,
>
>     void app(int simd)() { ... my fabulous app ... }
>
>     int main() {
>       auto fpu = core.cpuid.getfpu();
>       switch (fpu) {
>         case SIMD: app!(SIMD)(); break;
>         case SIMD4: app!(SIMD4)(); break;
>         default: error("unsupported FPU"); exit(1);
>       }
>     }

Sure, I've designed for this specifically, but it's not practical to
wind this all the way to the top of the stack.
Some hot code will make make use of this pattern, but small fragments
that appear throughout the code don't want to have this baggage
applied. They should just work with the developer's deliberately
selected default. It's not worth runtime selection on small
deployments. You will likely end up with numerous helper functions,
which when involved in the runtime-selected loops, would have
different versions generated appropriately, but when these helper
functions appear on their own, they would want to use a sensible
default.

>> I've done it with a template arg because it can be manually
>> propagated, and users can extrapolate the pattern into their outer
>> work functions, which can then easily have multiple versions
>> instantiated for runtime selection.
>> I think it's also important to mangle it into the symbol name for the
>> reasons I mention above.
>
>
> Note that version identifiers are not usable directly as template
> parameters. You'd have to set up a mapping.

I guess you haven't looked at my code, but yes, it's all mapped to
enums used by the templates. The versions would assign a constant used
as the template's default arg.