I'm trying to write (in DMD) the assembler that handles the function :

__m256i _mm256_permute4x64_epi64(__m256i a, in int M);

This translates to vpermq

The closest thing I could find in DMD assembly is VPERMILPS, which is called with:

asm { vpermilps YMM0, YMM1, IMM8; }

However, I cannot find out how to make IMM8 with `in int M`. I can convert int -> short[4] -> ubyte, but I can't move it into imm8

asm { mov imm8, ub; } <-- Fails.


Does anyone have an idea of what this is and how to define it?

Thanks in advance.

On Tuesday, 28 October 2014 at 21:05:15 UTC, Etienne wrote:
> I'm trying to write (in DMD) the assembler that handles the function :
>
> __m256i _mm256_permute4x64_epi64(__m256i a, in int M);
>
> This translates to vpermq
>
> The closest thing I could find in DMD assembly is VPERMILPS, which is called with:
>
> asm { vpermilps YMM0, YMM1, IMM8; }
>
> However, I cannot find out how to make IMM8 with `in int M`. I can convert int -> short[4] -> ubyte, but I can't move it into imm8
>
> asm { mov imm8, ub; } <-- Fails.
>
>
> Does anyone have an idea of what this is and how to define it?

Disclaimer: I only know a little basic X86 assembler, and I'm
just going of the first google result for "vpermilps" [1].

"imm8" is not a register. "imm" stands for "immediate", i.e. a
constant, hard-coded value. E.g.:

asm { vpermilps YMM0, YMM1, 0 /* no idea what would be a
meaningful value */; }

There are variants of vpermilps that work on three registers. The
third register is a XMM/YMM register, too, then. E.g.:

asm { vpermilps YMM0, YMM1, YMM2; }

I think my CPU doesn't have AVX, so I can't test anything beyond
compiling. When running all I get is "Illegal instruction (core
dumped)".

[1] http://www.felixcloutier.com/x86/VPERMILPS.html

> "imm8" is not a register. "imm" stands for "immediate", i.e. a
> constant, hard-coded value. E.g.:
>
> asm { vpermilps YMM0, YMM1, 0 /* no idea what would be a
> meaningful value */; }
>

Oh, well, that makes sense. This means I should use string mixins to insert the actual value.

I couldn't run AVX2 either, I have to work on this blindly until I get my hands on a new Xeon E5-2620 3rd generation processor. :/


> I think my CPU doesn't have AVX, so I can't test anything beyond
compiling. When running all I get is "Illegal instruction (core
dumped)".

My error is vmovdqu, I get:

 Error: AVX vector types not supported

Anyone know what this is? I can't find a reference to it in the dmd compiler code.

Also, 'vpunpcklqdq _ymm, _ymm, _ymm' is undefined in DMD, so I get : avx2.d(23): Error: bad type/size of operands 'vpunpcklqdq'


Looks like abunch of work is needed to get avx2 working...