www.digitalmars.com         C & C++   DMDScript  

digitalmars.D.learn - How do I write __simd(void16*, void16) ?

reply Etienne <etcimon gmail.com> writes:
I can't seem to find this function anywhere: __simd(void16*, void16)

The mangling seems to go through to dmd's glue.lib

This is for SSE2 operations:

MOVDQU => void _mm_storeu_si128 ( __m128i *p, __m128i a)
MOVDQU => __m128i _mm_loadu_si128 ( __m128i *p)

Would I have to write this with ASM?
Oct 08 2014
next sibling parent reply Benjamin Thaut <code benjamin-thaut.de> writes:
Am 08.10.2014 20:56, schrieb Etienne:

 I can't seem to find this function anywhere: __simd(void16*, void16)

 The mangling seems to go through to dmd's glue.lib

 This is for SSE2 operations:

 MOVDQU => void _mm_storeu_si128 ( __m128i *p, __m128i a)
 MOVDQU => __m128i _mm_loadu_si128 ( __m128i *p)

 Would I have to write this with ASM?


I strongly advise to not use core.simd at this point. It is in a 
horribly broken state and generates code that is far from efficient. If 
you use core.simd your code will be slower then if you use normal 
floating point math. Also afaik core.simd is currently only supported by 
dmd meaning that yuo have to rewrite the code in case you want to go to 
ldc or gdc. If you need simd with dmd, write inline assembly. If you 
need simd with the other two compilers, use the gcc intrinsics, they 
work on both compilers.

Kind Regards
Benjamin Thaut
Oct 08 2014
parent reply Etienne <etcimon gmail.com> writes:
On 2014-10-08 3:04 PM, Benjamin Thaut wrote:

 I strongly advise to not use core.simd at this point. It is in a
 horribly broken state and generates code that is far from efficient. If


I think I'll have to re-write the xmmintrin.h functions I need as string 
mixins to inline the assembly. Is that supposed to be compatible with 
LDC/GDC anyways?
Oct 08 2014
parent reply Benjamin Thaut <code benjamin-thaut.de> writes:
Am 08.10.2014 21:12, schrieb Etienne:

 On 2014-10-08 3:04 PM, Benjamin Thaut wrote:

 I strongly advise to not use core.simd at this point. It is in a
 horribly broken state and generates code that is far from efficient. If


 I think I'll have to re-write the xmmintrin.h functions I need as string
 mixins to inline the assembly. Is that supposed to be compatible with
 LDC/GDC anyways?


I know that GDC stopped supporting D style inline asm a while ago. If 
you need inline asm with GDC you have to use the gcc style inline 
assembly. I don't know about ldc though. But generally you want to use 
the official intrinsics with gdc and ldc because they won't perform any 
optimizations on inline assembly.

Kind Regards
Benjamin Thaut
Oct 09 2014
parent reply Etienne <etcimon gmail.com> writes:
On 2014-10-09 2:32 PM, Benjamin Thaut wrote:

 I know that GDC stopped supporting D style inline asm a while ago. If
 you need inline asm with GDC you have to use the gcc style inline
 assembly. I don't know about ldc though. But generally you want to use
 the official intrinsics with gdc and ldc because they won't perform any
 optimizations on inline assembly.

 Kind Regards
 Benjamin Thaut


Any idea where I can find the headers in D for it?
Oct 09 2014
parent reply Benjamin Thaut <code benjamin-thaut.de> writes:
Am 09.10.2014 21:04, schrieb Etienne:

 On 2014-10-09 2:32 PM, Benjamin Thaut wrote:

 I know that GDC stopped supporting D style inline asm a while ago. If
 you need inline asm with GDC you have to use the gcc style inline
 assembly. I don't know about ldc though. But generally you want to use
 the official intrinsics with gdc and ldc because they won't perform any
 optimizations on inline assembly.

 Kind Regards
 Benjamin Thaut


 Any idea where I can find the headers in D for it?


I think a good starting point would be Manu's std.simd module. I don't 
know if he is still working on it, but a old version can be found here:

https://github.com/TurkeyMan/simd/blob/master/std/simd.d

If you have further questions you might be well advised to ask him: 
turkeyman gmail.com

You can also find the druntime versions of ldc and gdc on github. For 
example:

https://github.com/ldc-developers/druntime/blob/ldc/src/ldc/simd.di
https://github.com/D-Programming-GDC/GDC/blob/master/libphobos/libdruntime/gcc/builtins.d

Unforunately the gcc.buildints module seems to be generated during 
compilation of gdc, so you might want to get a binary version or compile 
it yourself to see the module.

Kind Regards
Benjamin Thaut
Oct 09 2014
next sibling parent reply "David Nadlinger" <code klickverbot.at> writes:
On Thursday, 9 October 2014 at 20:29:44 UTC, Benjamin Thaut wrote:

 Unforunately the gcc.buildints module seems to be generated 
 during compilation of gdc, so you might want to get a binary 
 version or compile it yourself to see the module.


By the way, LDC has ldc.gccbuiltins_x86 too. LLVM doesn't export 
all the GCC-style intrinsics, though, if they are easily 
representable in normal LLVM IR (thus ldc.simd).

Daivd
Oct 09 2014
parent reply Etienne <etcimon gmail.com> writes:
On 2014-10-09 5:05 PM, David Nadlinger wrote:

 On Thursday, 9 October 2014 at 20:29:44 UTC, Benjamin Thaut wrote:

 Unforunately the gcc.buildints module seems to be generated during
 compilation of gdc, so you might want to get a binary version or
 compile it yourself to see the module.


 By the way, LDC has ldc.gccbuiltins_x86 too. LLVM doesn't export all the
 GCC-style intrinsics, though, if they are easily representable in normal
 LLVM IR (thus ldc.simd).

 Daivd


That's very helpful, the problem remains that the API is unfamiliar. I 
think most of the time, simd code will only need to be translated from 
basic function calls, it would've been nice to have equivalents :-p
Oct 09 2014
parent Etienne Cimon <etcimon gmail.com> writes:
On 2014-10-09 17:32, Etienne wrote:

 That's very helpful, the problem remains that the API is unfamiliar. I
 think most of the time, simd code will only need to be translated from
 basic function calls, it would've been nice to have equivalents :-p


Sorry, I think I had a bad understanding. I found out through a github 
issue that you need to use

pragma(LDC_intrinsic, "llvm.*") [function declaration]

https://github.com/ldc-developers/ldc/issues/627

And the possible gcc-style intrinsics are defined here:

https://www.opensource.apple.com/source/clamav/clamav-158/clamav.Bin/clamav-0.98/libclamav/c++/llvm/include/llvm/Intrinsics.gen

This really begs for at binding hat works with all compilers.
Oct 09 2014
prev sibling parent reply Etienne <etcimon gmail.com> writes:
On 2014-10-09 4:29 PM, Benjamin Thaut wrote:

 I think a good starting point would be Manu's std.simd module. I don't
 know if he is still working on it, but a old version can be found here:

 https://github.com/TurkeyMan/simd/blob/master/std/simd.d


That's a great reference! I can do a lot from that. I wish it wasn't an 
EDSL, makes it really hard to translate the simd code to D.


 You can also find the druntime versions of ldc and gdc on github. For
 example:

 https://github.com/ldc-developers/druntime/blob/ldc/src/ldc/simd.di
 https://github.com/D-Programming-GDC/GDC/blob/master/libphobos/libdruntime/gcc/builtins.d


 Unforunately the gcc.buildints module seems to be generated during
 compilation of gdc, so you might want to get a binary version or compile
 it yourself to see the module.


OK, thanks !
Oct 09 2014
parent reply ketmar via Digitalmars-d-learn <digitalmars-d-learn puremagic.com> writes:
On Thu, 09 Oct 2014 17:19:05 -0400
Etienne via Digitalmars-d-learn <digitalmars-d-learn puremagic.com>
wrote:


 Unforunately the gcc.buildints module seems to be generated during
 compilation of gdc, so you might want to get a binary version or
 compile it yourself to see the module.


 OK, thanks !


why, ketmar to the rescue! gcc.builtins is empty module. ;-)

actually, importing it works like a trigger, and then programmer has
access to GCC builtins defined in GCC source. one can read GCC
documentation to find more information about 'em.

also, importing this module defines some types. here they are:

  __builtin_va_list  The target's va_list type ))
  __builtin_clong    The D equivalent of the target's C "long" type
  __builtin_culong   The D equivalent of the target's C "unsigned long"
  __builtin_machine_int    Signed word-sized integer
  __builtin_machine_uint   Unsigned word-sized integer
  __builtin_pointer_int    Signed pointer-sized integer
  __builtin_pointer_uint   Unsigned pointer-sized integer
Oct 10 2014
parent reply Etienne <etcimon gmail.com> writes:
On 2014-10-10 4:12 AM, ketmar via Digitalmars-d-learn wrote:

 actually, importing it works like a trigger, and then programmer has
 access to GCC builtins defined in GCC source. one can read GCC
 documentation to find more information about 'em.


Hi ketmar,

Which type would have to be sent to the corresponding functions? I have 
a hard time figuring out how to use the __m128i with the proper 
mangling. Does it use core.simd's Vector!x types there?
Oct 10 2014
parent reply ketmar via Digitalmars-d-learn <digitalmars-d-learn puremagic.com> writes:
On Fri, 10 Oct 2014 08:31:44 -0400
Etienne via Digitalmars-d-learn <digitalmars-d-learn puremagic.com>
wrote:


 Which type would have to be sent to the corresponding functions? I
 have a hard time figuring out how to use the __m128i with the proper=20
 mangling. Does it use core.simd's Vector!x types there?


i know nothing about SIMD, but this compiles:

  import core.simd;
  import gcc.builtins;
  void main () {
    float4 a, b;
    auto tmp =3D __builtin_ia32_mulps(a, b); // a*b
  }

i don't know what the hell this means, but at least it accepts types
from core.simd. ;-) so i assume that other such builtins will accept
other types too.
Oct 10 2014
parent reply Etienne <etcimon gmail.com> writes:
On 2014-10-10 9:01 AM, ketmar via Digitalmars-d-learn wrote:

    import core.simd;
    import gcc.builtins;
    void main () {
      float4 a, b;
      auto tmp = __builtin_ia32_mulps(a, b); // a*b
    }

 i don't know what the hell this means, but at least it accepts types
 from core.simd. ;-) so i assume that other such builtins will accept
 other types too.


Nice! Nobody knows simd but they all know how to make it work. Go figure =)
Oct 10 2014
parent ketmar via Digitalmars-d-learn <digitalmars-d-learn puremagic.com> writes:
On Fri, 10 Oct 2014 09:09:09 -0400
Etienne via Digitalmars-d-learn <digitalmars-d-learn puremagic.com>
wrote:


 On 2014-10-10 9:01 AM, ketmar via Digitalmars-d-learn wrote:

    import core.simd;
    import gcc.builtins;
    void main () {
      float4 a, b;
      auto tmp =3D __builtin_ia32_mulps(a, b); // a*b
    }

 i don't know what the hell this means, but at least it accepts types
 from core.simd. ;-) so i assume that other such builtins will accept
 other types too.


=20
 Nice! Nobody knows simd but they all know how to make it work. Go
 figure =3D)


btw, i'm ready to do tests with 32-bit gdc if necessary. you can
mail/jabber me (the address is the same ;-) if you need.
Oct 10 2014
prev sibling parent "Trass3r" <un known.com> writes:
On Wednesday, 8 October 2014 at 18:56:31 UTC, Etienne wrote:

 I can't seem to find this function anywhere: __simd(void16*, 
 void16)



 MOVDQU => void _mm_storeu_si128 ( __m128i *p, __m128i a)
 MOVDQU => __m128i _mm_loadu_si128 ( __m128i *p)


Is there a module by now that allows to directly write Intel 
intrinsics?
Oct 09 2014