Sign in to follow this  
Followers 0
Ryan_001

Fast exp2() function in shader

3 posts in this topic

I'm writing a fairly extensive pixel shader, using D3D11 and targetting shader model 4.  As it stands I currently mix float's and int's a fair amount.  I'm under the assumption, but haven't been able to find any definitive information, that casting between int and float in a pixel shader would incur some performance overhead.  I could simplify a lot of the code and change it all to float's, but in a few places I need to do a fast 2^x multiply or divide, which with int's can be performed with a bit-shift.

 

Is there anyway to do a fast 2^x multiply with floats?  Or, I was under the impression that with older hlsl targets that ints were just emulated with the mantissa portion of the float and perhaps that's still the case and casting between and float and int is simply a signal to the compiler and actually doesn't require an op?

 

I'd like to write both versions and give it a try, but its fairly extensive and at the moment I don't have the time.  Given the circumstances what would you choose?

0

Share this post


Link to post
Share on other sites

exp2
http://msdn.microsoft.com/en-us/library/windows/desktop/bb509596%28v=vs.85%29.aspx

The only way to find out the answer to your question is to test performance.  You can also get an idea by looking at the token assembly of your shader.

 

Also, integers operations were emulated with floats in DirectX 9, but 10 requires full integer support.  I don't see how they could simulate it with a float, and if they did you would notice a huge performance impact.

1

Share this post


Link to post
Share on other sites

exp2
http://msdn.microsoft.com/en-us/library/windows/desktop/bb509596%28v=vs.85%29.aspx

The only way to find out the answer to your question is to test performance.  You can also get an idea by looking at the token assembly of your shader.

 

Also, integers operations were emulated with floats in DirectX 9, but 10 requires full integer support.  I don't see how they could simulate it with a float, and if they did you would notice a huge performance impact.

 

Seconded.  Just use exp2 for now and worry about performance later.  Unless you've got profiling information to back up your "I need a faster exp2" desire, you're indulging in pre-emptive optimization.

 

Most modern hardware (even quite low-end stuff) will crunch through ALU operations and come back looking for more; you may find it very difficult to slow things down.  ALU can even be free if the compiler determines that it can overlap ALU with texture lookups (and sometimes even just a simple code re-ordering can be enough).  I doubt if any home-grown replacement can outperform the built-in instruction.

0

Share this post


Link to post
Share on other sites
Almost all GPU's have a native exp2 instruction in their ALU's, so you're not going to make a faster version on your own. Converting to integer does often have a performance cost, and on most GPU's integer instructions run at 1/2 or 1/4 rate which means its unlikely you'll get better performance with bit shifts. You'll have to check the available docs on various architectures to find out the specifics.
1

Share this post


Link to post
Share on other sites

Create an account or sign in to comment

You need to be a member in order to leave a comment

Create an account

Sign up for a new account in our community. It's easy!


Register a new account

Sign in

Already have an account? Sign in here.


Sign In Now
Sign in to follow this  
Followers 0