How to use core.simd?

[jkm]

I'm trying out latest alpha release (LDC 0.13.0 alpha 2) from https://github.com/ldc-developers/ldc/releases on Linux x86_64. I'd like to compare ldc to dmd but the simple program

void main(string[] args)                                                        
{                                                                               
    import core.simd;                                                           
    float4 a = 1, b = 2;                                                        
    float4 c = __simd(XMM.ADDSS, a, b);                                         
}

does not compile

test.d(5): Error: undefined identifier __simd
test.d(5): Error: undefined identifier XMM

I suppose I must be doing something fundamentally wrong since I found posts where core.simd works with ldc. Please enlighten me.

[redstar]

Quite easy: the instructions from core.simd are not supported by LDC. Instead, you additionally import ldc.simd. Then you have 3 new functions which operates on vectors. shufflevector, extractelement and insertelement (and some other stuff). Other functions are not available at the D level - they are produced by the different optimization passes of LLVM.
The advantage of this approach is that the D code is CPU-independent. The disadvantage is that it is easy to write sub-optimal code (e.g. if the target CPU does not support vector instructions).

[jkm]

I like that vectorization is left to the compiler but at least for dmd is was decided that it is suboptimal (see http://dlang.org/simd.html). I suppose you are going to add more functions to ldc.simd. Is there a way (in ldc) to detect for which CPU the code is being compiled?

[redstar]

No, the DMD way of doing vector instruction will not be added to LDC. It is more or less impossible to implement this in LLVM. (See it the positive way: LDC can generate avx or altivec instructions if available...)

There is no way to detect the CPU during compilation. You have to specify the cpu on the command line (e.g. -mcpu=krait). You can then also define an identifier for the cpu if you need it.

[jkm]

What do you mean by

define an identifier for the cpu if you need it

?
How do I do that?

[redstar]

E.g. ldc2 -mcpu=corei7 -d-version=COREI7.
In D you can then use version(COREI7) { ... }

[dnadlinger]

You can, however, also use a large portion of the GCC SIMD builtins. Some elementary insertions/shuffles are not offered as builtins though (because LLVM does not expose them). For those, use ldc.simd.

[jkm]

How do use GCC SIMD builtins? Can you give an example?

[dnadlinger]

Just import ldc.gccbuiltins_x86 (or whatever platform you are targeting), then use the various builtins like __builtin_ia32_ptestz128 as usual.

[redstar]

I close this issue because there is nothing to fix. If needed the discussion can continue in the forum (http://forum.dlang.org/group/digitalmars.D.ldc).

[dnadlinger]

We should probably offer a tutorial or a few examples of how to write SIMD code with LDC. Maybe @jerro can point us to a few public examples of his that could be linked on a wiki page or such.