locked
HLSL Branching Performance

 > 

    Question

  • Question
    Sign in to vote
    0
    Sign in to vote
    I'm aware there are efficiency issues surrounding the use of branching instructions in HLSL, but I'm having trouble finding coding guidelines covering the use of branching. So, I would like to know which of the following three HLSL code snippets is the most efficient:

    // 1:
//
result = (a >= b) ? c : d;

// 2:
//
if (a >= b)
   result = c;
else
   result = d;

// 3:
//
x = step(a, b);
result = (x * c) + ((1 - x) * d);

    Assume the underlying hardware supports dynamic flow control. Also, if you happen to know of a resource that provides information on this subject that would be great too.
    Thursday, November 8, 2012 5:54 AM

Answers

  • Question
    Sign in to vote
    0
    Sign in to vote

    The first 2 are basically equivalent.

    Most GPU performance questions are very specific to the hardware design, so vendor and Feature Level matter a great deal here. For feature level 10.0 and up, there's a number of recent performance presentations on DirectX 11 that can help explain the trade-offs. f you are mostly interested in Feature Level 9.x implications, you should see older presentations on Direct3D 9 and Shader Model 2.0 / 3.0 generation hardware.

    Feature Level 9.1/9.2 probably should use option #3. Feature level 9.3 and up are probably better done as branching (#1 or #2). The real question is if all the pixels in a neighbhorhood tend to all pick the branch the same way or not. That has the most impact on performance.

    Tuesday, November 13, 2012 6:48 PM