Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Possible performance improvements to half float conversion #76

Open
elasota opened this issue Aug 18, 2018 · 0 comments
Open

Possible performance improvements to half float conversion #76

elasota opened this issue Aug 18, 2018 · 0 comments

Comments

@elasota
Copy link

elasota commented Aug 18, 2018

XMConvertHalfToFloat and XMConvertFloatToHalf both use a large number of integer ops when F16 intrinsics aren't available. It may be faster to do it with floating point operations. XMConvertHalfToFloat has a while loop for denormals, which is particularly slow.

Float-to-half conversion can use a trick: For positive numbers, (f + max(f, 2^-24)) will produce a float with an exponent at a fixed bias from the half float, and handle denormals and zero, and only needs 2 ops. (Bit-exactness in this case is sensitive to handling of the dropped mantissa bits in the denormal case though.)

Half-to-float can handle denormals (and zero) by converting the mantissa to float and multiplying it by 2^-24, which should be faster than the loop.

@walbourn walbourn added this to TODO in Version316 May 14, 2020
@walbourn walbourn removed this from Consider in Version316 Jul 7, 2020
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Projects
None yet
Development

No branches or pull requests

2 participants