Was messing around with SIMD, SSE stuff..  didn't know how much faster it could be!  Its been a few years since I did any assembly.
Though Id just share this, any word of adding this to the lib or compiler optimizations for this?


my results:
Parallel	 Single

SQRTPS:0.000120  FSQRT:0.001021
SQRTPS:0.000114  FSQRT:0.001026
SQRTPS:0.000114  FSQRT:0.001021
SQRTPS:0.000114  FSQRT:0.001026


codepad if you wish to play with it..
http://codepad.org/oqq5jsbJ

...times from codepad
SQRTPS:0.000291  FSQRT:0.000634
SQRTPS:0.000289  FSQRT:0.000632
SQRTPS:0.000300  FSQRT:0.000642
SQRTPS:0.000291  FSQRT:0.000632


//used these sites as a resource..
http://www.neilkemp.us/v3/tutorials/SSE_Tutorial_1.html
http://www.tommesani.com/SSEArithmetic.html



import std.stdio : putr = writefln;
import tango.time.StopWatch;


void fastsqrt( float[] a, float[] b )
{
  if(a.length != b.length || a.length % 4 != 0)
    throw new Exception("fsqrt bad params!");

  float* pa = a.ptr, pb = b.ptr;
  uint times = a.length>>2;

  asm
  {
    mov ECX, times;
    mov EAX, [pa];
    mov EBX, [pb];

  REP:
    movups XMM0, [EAX];
    sqrtps XMM0, XMM0;
    movups [EBX], XMM0;
    add EAX, 16;
    add EBX, 16;
    loop REP;
  }
}

void sqrt( float[] a, float[] b )
{
  if(a.length != b.length)
    throw new Exception("fsqrt bad params!");

  float* pa = a.ptr, pb = b.ptr;
  uint times = a.length;

  asm
  {
    mov EAX, [pa];
    mov EBX, [pb];
    mov ECX, times; //error on a.length

  REP2:
    fldpi float ptr[EAX];
    fsqrt;
    fstp float ptr[EBX];
    add EAX, 4;
    add EBX, 4;
    loop REP2;
  }
}

void main()
{
  float[40_000] a, b, c;

  foreach( k, ref i; a )
    i = cast(float)k;

  double A, B;
  StopWatch timer;

  timer.start;
   fastsqrt( a, b );
  A = timer.stop;

  timer.start;
   sqrt( a, c );
  B = timer.stop;

 //putr(a, "\n", b, "\n", c);

 foreach(k, i; b)
    assert( b[k] == c[k] );

 putr("SQRTPS:%.6f  FSQRT:%.6f", A, B);
}

Wyverex wrote:
> Was messing around with SIMD, SSE stuff..  didn't know how much faster it could be!  Its been a few years since I did any assembly.
> Though Id just share this, any word of adding this to the lib or compiler optimizations for this?
> 
> 
> my results:
> Parallel     Single
> 
> SQRTPS:0.000120  FSQRT:0.001021
> SQRTPS:0.000114  FSQRT:0.001026
> SQRTPS:0.000114  FSQRT:0.001021
> SQRTPS:0.000114  FSQRT:0.001026
> 
> 
> codepad if you wish to play with it..
> http://codepad.org/oqq5jsbJ
> 
> ...times from codepad
> SQRTPS:0.000291  FSQRT:0.000634
> SQRTPS:0.000289  FSQRT:0.000632

>   asm
>   {
>     mov EAX, [pa];
>     mov EBX, [pb];
>     mov ECX, times; //error on a.length
> 
>   REP2:
>     fldpi float ptr[EAX];

shouldn't that be fld ?
fldpi loads 3.1415.... !
Doesn't make any difference to the time, though.

This is exactly why D just got array operations.

Don wrote:
> Wyverex wrote:
>> Was messing around with SIMD, SSE stuff..  didn't know how much faster it could be!  Its been a few years since I did any assembly.
>> Though Id just share this, any word of adding this to the lib or compiler optimizations for this?
>>
>>
>> my results:
>> Parallel     Single
>>
>> SQRTPS:0.000120  FSQRT:0.001021
>> SQRTPS:0.000114  FSQRT:0.001026
>> SQRTPS:0.000114  FSQRT:0.001021
>> SQRTPS:0.000114  FSQRT:0.001026
>>
>>
>> codepad if you wish to play with it..
>> http://codepad.org/oqq5jsbJ
>>
>> ...times from codepad
>> SQRTPS:0.000291  FSQRT:0.000634
>> SQRTPS:0.000289  FSQRT:0.000632
> 
>>   asm
>>   {
>>     mov EAX, [pa];
>>     mov EBX, [pb];
>>     mov ECX, times; //error on a.length
>>
>>   REP2:
>>     fldpi float ptr[EAX];
> 
> shouldn't that be fld ?
> fldpi loads 3.1415.... !
> Doesn't make any difference to the time, though.
> 
> This is exactly why D just got array operations.

i had  fld  first but sqrt of 5 or higher came back as -nan....