Performances Tricks

[jdelauney]

After our discussion here #27 i've made some tests for improving performances we can use

XMMWORD PTR

examples :

• in instruction like andps xmm0, [RIP+cSSE_MASK_ONLY_W] we can do andps xmm0, XMMWORD PTR [RIP+cSSE_MASK_ONLY_W]

• In AngleBetween instead :

  mov rax, [A]            // A is a pointer on the stack
  movups xmm1, [RAX] 

  we can do

  mov rax, [A]            // A is a pointer on the stack
  movaps xmm1, XMMWORD PTR [RAX] 

• in operators or in function like min, max, clamp, negate, abs, divideby2, MulAdd, MulDiv.... for operator +, instead of

  movaps xmm1,[B]
  addps  xmm0, xmm1 

  we can do

  //movaps xmm1,[B]
  addps  xmm0, XMMWORD PTR [B] //xmm1  

  Using XMMWORD PTR increase performances a little or a lot (depend of case)

[dicepd]

Thinking about this one using word pointer is fine for the first two cases. As for using it on parameters which are fetched from global memory I am not so sure as there is no guarantee that the parameter is in memory which is addressable with just a word. I have 20 gig of ram and I do not think a word pointer can address all that.

[jdelauney]

XMMWORD is a 128bit pointer no ? I'm following you, it will be wiser not to use it on the parameters. But our paramters are normaly aligned and exist so......need more tests at end. But by using this trick for the first two is sure it improves performance a bit