SqrtFloat ASM replacement !?

Helle · Post by **Helle** » Wed Dec 20, 2006 11:29 am

@ Trond: Sorry!
For CPU´s with SSE, but not SSE2:

#Tries = 100000000 

z.f 
time = GetTickCount_() 
For I = 0 To #Tries 
  ;z = I 
  !cvtsi2ss xmm1,[v_I] 
  !sqrtss xmm0,xmm1 
Next 
!movss [v_z],xmm0 
MessageRequester("SSE", Str(GetTickCount_()-time)+#CRLF$+StrF(z))

Test of SIMD:

Code: Select all

;- CPU-SIMD-Test, "Helle" Klaus Helbing, 20.12.2006, PB4.02

Global mmx.c   = $2d         ;"-"
Global dnow.c  = $2d
Global ednow.c = $2d
Global cmov.c  = $2d
Global sse.c   = $2d
Global sse2.c  = $2d
Global sse3.c  = $2d
Global ssse3.c = $2d

Global Bit0.l  = $1          ;for SSE3
Global Bit9.l  = $200        ;for SSSE3
Global Bit15.l = $8000       ;for CMOVcc
Global Bit21.l = $200000     ;for EFlag
Global Bit23.l = $800000     ;for MMX 
Global Bit25.l = $2000000    ;for SSE
Global Bit26.l = $4000000    ;for SSE2
Global Bit30.l = $40000000   ;for extended 3DNow!
Global Bit31.l = $80000000   ;for 3DNow!

Global Name$   = "The tested CPU supported :"
Global MMX$    = "MMX :            "
Global DNOW$   = "3DNow! :       "
Global EDNOW$  = "ext3DNow! :  "
Global CMOV$   = "CMOVcc :      "
Global SSE$    = "SSE :             "
Global SSE2$   = "SSE2 :           "
Global SSE3$   = "SSE3 :           "
Global SSSE3$  = "SSSE3 :         "

;-------- Test, if CPUID is possible
    !pushfd                  ;EFlag-Register (32-Bit) on the Stack
    !pop eax                 ;copy to EAX
    !mov edx,eax             ;save EAX
    !xor edx,[v_Bit21]       ;toggle Bit21 
    !push edx                          
    !popfd                   ;write in EFlag
    !pushfd                  ;back to Stack
    !pop edx
    !push eax                ;EAX back
    !popfd   
    !cmp eax,edx
    !jne l_iscpuid           ;not equal -> CPUID is supported

 MessageRequester("Status", "The tested CPU give no support for CPUID; no MMX or SSE !")
End  

IsCPUID:
;------------------------------------------------------------------------------

;-------- Test of MMX, SSE, SSE2, SSE3 and SSSE3
    !mov eax,1h
    !cpuid
    !test edx,[v_Bit23]      ;MMX
    !jz l_nommx
    !mov [v_mmx],2bh         ;"+" 
NOMMX:
    !test edx,[v_Bit25]      ;SSE  
    !jz l_nosse
    !mov [v_sse],2bh
NOSSE:
    !test edx,[v_Bit26]      ;SSE2  
    !jz l_nosse2
    !mov [v_sse2],2bh
NOSSE2:
    !test ecx,[v_Bit0]       ;SSE3
    !jz l_nosse3
    !mov [v_sse3],2bh    
NOSSE3:
    !test ecx,[v_Bit9]       ;SSSE3 ("old" SSE4)
    !jz l_nossse3
    !mov [v_ssse3],2bh 
NOSSSE3:
 ;-------- Test of CMOVcc (conditional move)
    !test edx,[v_Bit15]    
    !jz l_nocmov
    !mov [v_cmov],2bh
NOCMOV:       
;------------------------------------------------------------------------------    

;-------- value of extended levels (for 3DNow!-Test)
;-------- back-value in EAX (-80000000h) = value of extended level
    !mov eax,80000000h
    !cpuid
    !cmp eax,80000000h
    !jbe l_noext             ;no extended levels, no 3DNow!
;------------------------------------------------------------------------------
    
;-------- Test of 3DNow! and extended 3DNow!
    !mov eax,80000001h       ;80000001h is the first extended level, no supported from Intel-CPU´s!
    !cpuid                   ;Intel-CPU´s gives EAX=0 return
    !or eax,eax
    !je l_noext              ;is Intel-Prozessor
    !test edx,[v_Bit31]      ;3DNow! 
    !jz l_noext
    !mov [v_dnow],2bh
    !test edx,[v_Bit30]      ;extended 3DNow!  
    !jz l_noext
    !mov [v_ednow],2bh
NOEXT:    
;------------------------------------------------------------------------------
   
 MessageRequester(Name$,MMX$+Chr(mmx)+Chr(10)+DNOW$+Chr(dnow)+Chr(10)+EDNOW$+Chr(ednow)+Chr(10)+CMOV$+Chr(cmov)+Chr(10)+SSE$+Chr(sse)+Chr(10)+SSE2$+Chr(sse2)+Chr(10)+SSE3$+Chr(sse3)+Chr(10)+SSSE3$+Chr(ssse3))
End

Gruss
Helle

Edit 10.01.2007: New detection for SSE3

Helle · Post by **Helle** » Wed Dec 20, 2006 6:31 pm

This is code only for an AMD-CPU:

Code: Select all

Procedure.f Sqrt(N.f) 
  !mov eax, [p.v_N] 
  !sub eax, $3F800000 
  !shr eax, 1 
  !add eax, $3F800000 
  !mov [esp-4], eax 
  !fld dword [esp-4] 
  CompilerIf #PB_Compiler_Debugger 
    ProcedureReturn 
  CompilerElse 
    !ret 4 
  CompilerEndIf 
EndProcedure 

#Tries = 50000000 

;- ONLY for AMD-CPU´s !!! AMD-K6-2 or better
z.f
time = GetTickCount_() 
For I = 0 To #Tries 
!femms
!movd mm0,[v_I]
!pi2fd mm1,mm0
!movq mm0,mm1
!pfrsqrt mm1,mm0
!punpckldq mm0,mm0
!pfmul mm0,mm1
!movd [v_z],mm0
!femms
Next 

MessageRequester("3DNow!", Str(GetTickCount_()-time)+#CRLF$+StrF(z)) 

z.f 
time = GetTickCount_() 
For I = 0 To #Tries 
  z = I 
  Sqrt(z) 
Next 
z=Sqrt(z)
MessageRequester("Procedure", Str(GetTickCount_()-time)+#CRLF$+StrF(z)) 

z.f 
time = GetTickCount_() 
For I = 0 To #Tries 
  z = I 
  Sqr(z) 
Next 
z=Sqr(z)
MessageRequester("PB", Str(GetTickCount_()-time)+#CRLF$+StrF(z))

Gruss
Helle

P.S.: 3DNow! works with a table on the chip and is very quick!

gebe · Post by **gebe** » Sat Apr 07, 2007 5:28 pm

Hi chaps...

Derek,
I got finally to the test you suggested
and found the results similar with the given test program
I did some changes and saw that the normal sqr is same speed
and a lot more accurate
my machine gives: 625 ,,,,,, 625
and 625 R=7147.7 not important ,,,,,, 625 r=7071.06

That is when I use the given Z=I
Sqr(Z)

If I use the different approach
Z=sqr(I)

I get 422 ,,,,,,, 609
and 422 ,,, r=7147.7 ,,,,,,l, 609 ,,, r=7071.06

That seems to prove the point that accuracy is time costly.

and that test must be very carefully conducted .

(Not that I am a specialist .I just have tried a lot of times)

gebe

here is the more fancy way.

Code: Select all

Procedure.f Sqrt(N.f) 
  !mov eax, [p.v_N] 
  !sub eax, $3F800000 
  !shr eax, 1 
  !add eax, $3F800000 
  !mov [esp-4], eax 
  !fld dword [esp-4] 
  CompilerIf #PB_Compiler_Debugger 
    ProcedureReturn 
  CompilerElse 
    !ret 4 
  CompilerEndIf 
EndProcedure 


#Tries = 50000000 

z.f 
time = GetTickCount_() 
For I = 0 To #Tries 
 ; z = I ;more accurate ????????????????????????????
 ;sqrt(z)
 z= Sqrt(i) ;more accurate ?????????????????????????
Next
stopTime.l= GetTickCount_();dont allow any loss of time
;before the count stop
Duration.l=stopTime - time
result.d=sqrt(i);read the result of square root of 50 000 001
;now we have all the time to get a nicely formatted
;answer
MessageRequester(StrD(result),"With Sqrt   "+ Str(GetTickCount_()-time)+ " mSec.                    ") 
;make enough space for showing accuracy in title
 
time = GetTickCount_() 
For I = 0 To #Tries 
 ; z = I 
 z= Sqr(i) 
Next
stopTime = GetTickCount_() 
Duration =stopTime  - time
result = Sqr(i)
MessageRequester(StrD(result),"With Sqr   "+ Str(GetTickCount_()-time)+ " mSec.                   ") 
;we could also put the results o 20 full runs in an array And then inspect

gebe

am I right ????

gebe · Post by **gebe** » Sat Apr 07, 2007 5:35 pm

Sorry,Idid not see the more fancy stuff ,i started at the begining.
i am going to save and check the other codes

Gebe

gebe · Post by **gebe** » Sat Apr 07, 2007 6:10 pm

Helle ,
That 3D thing SCREAMS !!!! That is SPEED with capital SSSSS !!

Nice to see speeding .. Carefull for the cops

gebe

Derek · Post by **Derek** » Sat Apr 07, 2007 6:59 pm

@gebe, as you can see some people are interested in speed over size.

gebe · Post by **gebe** » Sun Apr 08, 2007 4:36 am

Yah,the speed in <square root> is very important even more than accuracy...
I also checked without inlining the machine code and it barely ,if, changes the time (proc of MC called at every for next loop)

That is interesting stuff.

.
Thanks again for the tip.

gebe