Digital Mars - digitalmars.D.learn

↑ ↓ ← → "dominik" <asd asd.com> writes:

I'm working on my vector and matrix math library, and now I'm prototyping my 
SIMD optimizations, obviously I would like data alignment.. fact is I have 
no clue how to do it ( I vaguely remember align directive)

for example something like (mixed C++ and D! :) ), what would I need for 
pragma pack equivalent in D then for a struct?

#pragma pack(16)
struct Vector4d {
  union {
   struct { float x = 0.0f, y = 0.0f, z = 0.0f, w = 0.0f; };
   struct { float r, g, b, a; };
   struct { float cell[4]; };
   Repeat!(float, 4) tuple;
  };
...
}

also somewhere down the road in the code I'd like equivalent of declspec 
align too, for example something like

Vector4d opAdd( ref Vector4d v ) {

version(NORMAL) {
    return Vector4d( x + v.x, y + v.y, z + v.z, w + v.w );
}
else version(SIMD) {
    __declspec(align(16)) Vector4d v4_ret;
    asm {
        mov esi, this
        mov edi, v
        movaps xmm0, [esi]
        addps xmm0, [edi]
        movaps v4_ret, xmm0
    }
    return v4_ret;
}
}

or something like that, its illustrative only :)

Apr 13 2008

↑ ↓ ← → Don <nospam nospam.com> writes:

dominik wrote:
 I'm working on my vector and matrix math library, and now I'm prototyping my 
 SIMD optimizations, obviously I would like data alignment.. fact is I have 


Good news!

 no clue how to do it ( I vaguely remember align directive)


 
 for example something like (mixed C++ and D! :) ), what would I need for 
 pragma pack equivalent in D then for a struct?
 
 #pragma pack(16)
 struct Vector4d {




align(16) struct Vector4d {


 also somewhere down the road in the code I'd like equivalent of declspec 
 align too, for example something like
 
     __declspec(align(16)) Vector4d v4_ret;
     asm {
         mov esi, this
         mov edi, v


asm {
    align 16;
    mov esi, this;
    mov edi, v;
}


Bad news!

One problem is that the stack (and hence all local variables) are only 
aligned to 32 bytes. This means you can only use SSE aligned loads 
(movaps, movapd) on static and heap variables. (This is something that D 
really should fix while it has the chance).

Apr 16 2008

↑ ↓ ← → "Lionello Lunesu" <lionello lunesu.remove.com> writes:

 One problem is that the stack (and hence all local variables) are only 
 aligned to 32 bytes. This means you can only use SSE aligned loads 
 (movaps, movapd) on static and heap variables. (This is something that D 
 really should fix while it has the chance).


bits?

Apr 16 2008

↑ ↓ ← → Don Clugston <nospam nospam.com> writes:

Lionello Lunesu wrote:
 
 One problem is that the stack (and hence all local variables) are only 
 aligned to 32 bytes. This means you can only use SSE aligned loads 
 (movaps, movapd) on static and heap variables. (This is something that 
 D really should fix while it has the chance).


 bits?


Um actually the '32 bytes' is completely wrong. You need to have 128 bit 
alignment = 16 bytes, but currently DMD only lets you have 64 bit 
alignment for locals. Since the D calling convention is different from 
C, D could actually define that 128 bit alignment is preserved for 
function calls (ie, stack pushes/pops between function calls must always 
be a multiple of 16 bytes). Extern(C) and extern(Windows) functions will 
need a stack alignment fixup at the start of them.

Apr 16 2008

digitalmars.D.learn - struct alignments