AMD ATI CTM, Техническое справочное руководство

Страница: 47 / 56

Texture Floating Point 43
© 2006 Advanced Micro Devices, Inc. ATI CTM Guide v. 1.01
3.6.4 Texture Floating Point
Projected texture coordinates are processed in the X1K FP block before being sent to the texture unit. The texture
unit does not accept NaN, so NaN coordinates are converted to +infinity before being sent to the texture unit. As with
the ALU, denormal inputs and denormal results are converted to pure zero, preserving sign.
The multiplier used for projection and cubemapping does not follow IEEE rules when handling special values. This
will become apparent only when you attempt to project or cubemap a coordinate that contains an infinite or NaN
component.
You should use caution when generating very large values for use as coordinates in a texture lookup. These values
may generate infinite values when scaled by the texture dimensions or projected.
3.7 Errata
There is a problem with texture semaphores: if an instruction that does a wait on a texture semaphore is preceeded by
a flow-control instruction, that second instruction may execute before the texture semaphore has returned.
As a workaround, always do a tex sem wait on a flow-control instruction if any possible next instruction contains a
tex sem wait (including branch and call/return). You should also remove the tex sem wait on the other instructions
(unless you are in full-flow control mode) or you risk waiting on a younger thread, which shouldn't cause a deadlock
but could be bad for performance. In full-flow control mode, you can leave the tex sem wait bit set for those other
instructions.