Printable View
Is there any good books for this? I really don't have any idea on how to code it in C++. Can anyone give me a good starting learning code for it?
I read in wiki that it is widely use in 3D graphics programming so I went here and ask it. Do I need math here? I'm good in Calculus and Linear Algebra.
I found this site. Is it a good introduction? http://www.codeproject.com/KB/recipes/sseintro.aspx
Sarah22
And you are already thinking about SIMD?Quote:
I really don't have any idea on how to code it in C++
Learn basic C++ first. This site has a tutorial, for example.
Yes, because I read somewhere on the net that it enhance performance.Quote:
Originally Posted by cyberfish
Basic C++? Like what?Quote:
Originally Posted by cyberfish
IIRC, several C++ compilers offer support for automatically generating SSE(1/2/3/4) where possible. MSVC, GCC, and the Intel compiler can all do it for you. IIRC, MSVC and GCC can do SSE and SSE2, but the Intel compiler can do 1, 2, 3, and 4.
It does, but its not for the faint of heart, it basically involves either using inline assembly, or the macros Microsoft so 'graciously' replaced inline assembly with.Quote:
Originally Posted by sarah22
Well, for starters have you written Hello World in C/C++ yet?Quote:
Basic C++? Like what?
As of MSVC 2005 all MSVS compilers support SIMD instructions and will use them if you set it in the options. MSVC 2005 by default has this turned on which is why the new games would not run on my old AMD system which did not support SIMD.
Intel and AMD's tech refs explain SIMD in great detail but you had better have a good understanding of the x86 architecture. I would not recommend hand coding SIMD into an application unless you really knew what you were doing. SIMD for 3D is sort of a moot point now since the video card is just as fast or faster than the equivalent SIMD. For accelerated 2D multimedia I could see some usage for it but I'm not sure it's worth the pain.
I stand corrected, originally 2005 did not support inline assembly SIMD instructions, which are not the same as intrinsics. 2008 appears to do so again though. I guess I can stop using VS 6.0 now, although I still need to keep using VS 4.0.Quote:
Originally Posted by Bubba
OK, I thought I was correct in my original assertion so I looked back at some notes. VS does NOT support inline assembly for 64 bit mode.
and to the OP, this demonstrates how much faster SIMD is than C++, at least on Intel processors, AMD SIMD performance is less than spectacular last time I checked.
Code:#include <windows.h>
#include <stdio.h>
int main(){
double* pData = (double*)_aligned_malloc(sizeof(double) * 2 * 67108864, 16);
double* pOutput = (double*)_aligned_malloc(sizeof(double) * 2, 16);
for(int x = 0;x<67108864 * 2;x++) pData[x] = 1.0;
DWORD SIMF_Start = GetTickCount();
_asm {
pushad;
mov eax , pData;
prefetcht0 [eax];
mov edx , pOutput;
mov edi , edx;
mov edx , 0x00000010;
mov ecx , 0x04000000;
xorpd xmm0 , xmm0;
the_loop:
prefetcht0 [eax+16];
addpd xmm0 , [eax];
add eax , edx;
loop the_loop;
mov edx , edi;
movapd [edx] , xmm0;
popad;
}
DWORD SIMD_Stop = GetTickCount();
DWORD CPP_Start = GetTickCount();
for(int x = 0;x<2 * 67108864;x++) pOutput[0] += pData[x];
DWORD CPP_Stop = GetTickCount();
printf("SIMD took %d\nCPP took %d\n" , SIMD_Stop-SIMF_Start , CPP_Stop-CPP_Start);
return 0;
}
Provided your compiler would not have already used SIMD in place of your original C++ code. Without looking at the output asm from the compiler I would say it's a bit dangerous to try to out-optimize the compiler without first doing some research.