Hi,
1. I want to know if someone can tell me if it is common to use SSE or the newer VXA to do matrix and vector calculations for games?
2. Does D3DX make use of SSE/VXA when calling functions like D3DXMatrixMultiply or when doing any arithmetic on D3DXMATRIX or D3DXVECTOR? If not, I'm assuming one has to explicitely implement SSE on your matrix and vector arithmetic.
3. If D3DX does not implement SSE/VXA implecitely, where can I start to look into SSE? Apparently Intel has an SSE/VXA lib?
Im currently using D3DXMATRIX and D3DVECTOR to store and manipulate my matrices and vertices and Im just contemplating if it is the fastest structure to use.
Using VXA/SSE in game programming
1) YES VERY! But there are a LOT of caveats. See, it isn't so quick to actually be shuffling things back and forth from vector registers to cpu registers. So if you don't have enough work to do, it costs more time to setup the vector operations than to just work in the fpu.
There is a LOT of gameplay code where you end up doing things like:
and so any gains you'd get from using SSE/VMX in the Add are instantly lost when you try to retrieve a.y.
So, you'll see the vectorized code in places where it counts: Particle Systems, Skeletal Animation systems, Physics Systems. All of those systems can manage to really stream though a lot of data in the vector unit without ever having to pull it back into the fpu.
2) iirc, yes it can and does.
3) Sony's Math Lib was released through the Bullet Physics SDK. It is a very well done and complete library if you wanted to look into it.
There is a LOT of gameplay code where you end up doing things like:
AddVec( a, b );if( a.y <= groundHeight ) do splat!and so any gains you'd get from using SSE/VMX in the Add are instantly lost when you try to retrieve a.y.
So, you'll see the vectorized code in places where it counts: Particle Systems, Skeletal Animation systems, Physics Systems. All of those systems can manage to really stream though a lot of data in the vector unit without ever having to pull it back into the fpu.
2) iirc, yes it can and does.
3) Sony's Math Lib was released through the Bullet Physics SDK. It is a very well done and complete library if you wanted to look into it.
Author
KulSeran, so if I understand you correctly, D3DXMATRIX and D3DXVECTOR uses/implements SSE? So then I dont need to go and explicitely implement it if I use those 2 data structures.
Those structs are defined in the header files, so you can see most of the code there if you want. There are also aligned variants (eg D3DXMATRIXA16) Some of the big operations like matrix multiply and inverse use SIMD instructions, but most of the simple vector ops are simple inline C code, because vectorized code doesn't help there.
Quote:Original post by G-Man9566
1. I want to know if someone can tell me if it is common to use SSE or the newer VXA to do matrix and vector calculations for games?
SSE is extremely common. AVX (not VXA) is brand new, and is likely to gain some ground given time, but for the time being I doubt many people have done a huge amount with the instruction set. SSE4.1 Still has it's place though, and for most people would be where the dev effort still goes (since pretty much every PC will support SSE, not so many support AVX as yet).
Quote:Original post by G-Man9566
2. Does D3DX make use of SSE/VXA when calling functions like D3DXMatrixMultiply or when doing any arithmetic on D3DXMATRIX or D3DXVECTOR? If not, I'm assuming one has to explicitely implement SSE on your matrix and vector arithmetic.
It certainly used to (postfixed with 'A'), but not sure of the current version. I imagine you should be able to beat the performance in most cases with your own SSE code though (since it doesn't seem to care about alignment, so I'm guessing it loads/stores using the un-aligned versions? It's been a while since I used D3DX)
Quote:Original post by G-Man9566
3. If D3DX does not implement SSE/VXA implecitely, where can I start to look into SSE? Apparently Intel has an SSE/VXA lib?
As does AMD.... (lacking quats IIRC)
As for looking into SSE
Agner's blog
msdn
Intel optimization manual
Quote:Original post by RobTheBloke
It certainly used to (postfixed with 'A'), but not sure of the current version. I imagine you should be able to beat the performance in most cases with your own SSE code though (since it doesn't seem to care about alignment, so I'm guessing it loads/stores using the un-aligned versions? It's been a while since I used D3DX)
The newer XNA Math library for DX has 'aligned' versions of the various structs etc which I would guess get dumped via the aligned instructions, however if using those as members of your own classes you have to be careful as you need to make sure the aligned versions ARE aligned or you'll get weird and wonderfull crashes and access errors.
(This is something I ran into recently as you might have guessed, took a few minutes to figure out why a simple vector assignment was blowing up in my face then I realised wtf was going on [grin])
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement
