www.digitalmars.com         C & C++   DMDScript  

digitalmars.D.learn - XMM Intrinsics

reply Marcio Martins <marcioapm gmail.com> writes:
Hi,

I am building a CRC32C implementation using SSE for D, because I 
couldn't find any readily available :[

However, I am unable to find any documentation regarding which 
SSE instructions are available and how I could use them in D. I 
can see core.simd can't seem to make any use of it.

How would I go about calling _mm_* functions in D in a way that 
is portable between D compilers?

I am currently using LDC so if someone could point me a direction 
I can get this going even if only on LDC it would already be a 
great help.

Cheers!
May 08 2020
next sibling parent reply Simen =?UTF-8?B?S2rDpnLDpXM=?= <simen.kjaras gmail.com> writes:
On Friday, 8 May 2020 at 12:38:51 UTC, Marcio Martins wrote:

 Hi,

 I am building a CRC32C implementation using SSE for D, because 
 I couldn't find any readily available :[

 However, I am unable to find any documentation regarding which 
 SSE instructions are available and how I could use them in D. I 
 can see core.simd can't seem to make any use of it.

 How would I go about calling _mm_* functions in D in a way that 
 is portable between D compilers?


You would use core.simd:

https://dlang.org/library/core/simd.html


Sadly, the supported instructions are not documented on that 
page, and you'll need to look at the source to get at them:

https://github.com/dlang/druntime/blob/master/src/core/simd.d#L85-L384

(I'll file a bug on this)


There's a bit more info on D SIMD instructions here:

https://dlang.org/spec/simd.html

--
   Siomen
May 08 2020
parent reply kinke <noone nowhere.com> writes:
On Friday, 8 May 2020 at 12:49:00 UTC, Simen Kjærås wrote:

 How would I go about calling _mm_* functions in D in a way 
 that is portable between D compilers?


 You would use core.simd:


Nope one wouldn't, because that horrible interface isn't 
supported by LDC, and I guess GDC neither.

The intel-intrinsics dub package aims to provide a 
compiler-independent layer: 
https://code.dlang.org/packages/intel-intrinsics
May 08 2020
parent reply Simen =?UTF-8?B?S2rDpnLDpXM=?= <simen.kjaras gmail.com> writes:
On Friday, 8 May 2020 at 13:09:49 UTC, kinke wrote:

 On Friday, 8 May 2020 at 12:49:00 UTC, Simen Kjærås wrote:

 How would I go about calling _mm_* functions in D in a way 
 that is portable between D compilers?


 You would use core.simd:


 Nope one wouldn't, because that horrible interface isn't 
 supported by LDC, and I guess GDC neither.

 The intel-intrinsics dub package aims to provide a 
 compiler-independent layer: 
 https://code.dlang.org/packages/intel-intrinsics


TIL, thanks! :)

--
   Simen
May 08 2020
parent Mike Parker <aldacron gmail.com> writes:
On Friday, 8 May 2020 at 20:14:05 UTC, Simen Kjærås wrote:


 The intel-intrinsics dub package aims to provide a 
 compiler-independent layer: 
 https://code.dlang.org/packages/intel-intrinsics


 TIL, thanks! :)

 --
   Simen


DConf 2019: Not intrinsically about intrinsics -- Guillaume Piolat

https://youtu.be/cmswsx1_BUQ
May 08 2020
prev sibling next sibling parent reply Guillaume Piolat <first.last gmail.com> writes:
On Friday, 8 May 2020 at 12:38:51 UTC, Marcio Martins wrote:

 How would I go about calling _mm_* functions in D in a way that 
 is portable between D compilers?


Hello,

I've made this library for that exact purpose: 
https://github.com/AuburnSounds/intel-intrinsics


Supports every intrinsic listed under MMX/SSE/SSE2/SSE3 in 
https://software.intel.com/sites/landingpage/IntrinsicsGuide/
May 08 2020
parent reply Marcio Martins <marcioapm gmail.com> writes:
On Friday, 8 May 2020 at 13:11:02 UTC, Guillaume Piolat wrote:

 On Friday, 8 May 2020 at 12:38:51 UTC, Marcio Martins wrote:

 How would I go about calling _mm_* functions in D in a way 
 that is portable between D compilers?


 Hello,

 I've made this library for that exact purpose: 
 https://github.com/AuburnSounds/intel-intrinsics


 Supports every intrinsic listed under MMX/SSE/SSE2/SSE3 in 
 https://software.intel.com/sites/landingpage/IntrinsicsGuide/


I saw the intel-intrinsics package, but unfortunately it stops at 
SEE3 and I need SSE4.2 for this.

How is this library working? Will LDC/LLVM detect the name and 
replace it with the right instructions? If so, could I just 
provide an empty _mm_crc32_u8 and it'd pick it up correctly?


Thanks guys!
May 08 2020
parent reply kinke <noone nowhere.com> writes:
On Friday, 8 May 2020 at 13:30:42 UTC, Marcio Martins wrote:

 I saw the intel-intrinsics package, but unfortunately it stops 
 at SEE3 and I need SSE4.2 for this.

 How is this library working?


It's open source. :)


 Will LDC/LLVM detect the name and replace it with the right 
 instructions? If so, could I just provide an empty _mm_crc32_u8 
 and it'd pick it up correctly?


Nope. But take a look at ldc.gccbuiltins_x86, it offers these 4 
builtins:

int __builtin_ia32_crc32hi(int, short);
int __builtin_ia32_crc32si(int, int);
int __builtin_ia32_crc32qi(int, byte);
long __builtin_ia32_crc32di(long, long);

Make sure to compile with `-mattr=+sse4.2` to enable these 
instructions.
May 08 2020
parent Marcio Martins <marcioapm gmail.com> writes:
On Friday, 8 May 2020 at 13:56:18 UTC, kinke wrote:

 On Friday, 8 May 2020 at 13:30:42 UTC, Marcio Martins wrote:

 I saw the intel-intrinsics package, but unfortunately it stops 
 at SEE3 and I need SSE4.2 for this.

 How is this library working?


 It's open source. :)


I know, but I don't have time to implement it and make it pretty, 
I am prototyping  with very little time.


 Will LDC/LLVM detect the name and replace it with the right 
 instructions? If so, could I just provide an empty 
 _mm_crc32_u8 and it'd pick it up correctly?


 Nope. But take a look at ldc.gccbuiltins_x86, it offers these 4 
 builtins:

 int __builtin_ia32_crc32hi(int, short);
 int __builtin_ia32_crc32si(int, int);
 int __builtin_ia32_crc32qi(int, byte);
 long __builtin_ia32_crc32di(long, long);

 Make sure to compile with `-mattr=+sse4.2` to enable these 
 instructions.


YES! Thank you, this is exactly what I was looking for.
I suppose I will have to implement a scalar version for DMD 
later... I don't even know why I bother with DMD...
May 08 2020
prev sibling parent Olivier Pisano <olivier.pisano laposte.net> writes:
On Friday, 8 May 2020 at 12:38:51 UTC, Marcio Martins wrote:

 Hi,

 I am building a CRC32C implementation using SSE for D, because 
 I couldn't find any readily available :[


Here is mine:

https://github.com/opisano/crc32c/blob/master/crc32c.d
May 09 2020