Age | Commit message (Collapse) | Author |
|
|
|
|
|
This is needed in preparation for eager evaluation and it also
simplifies things by removing pre execution logic in operations.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
This also accidentally includes refactoring of some of the components
like the scheduler.
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
Bound texture loads by their size to support single results and
non-robust GL implementations.
|
|
|
|
Allocate results on domains directly to avoid having to propogate domain
transformation manually in all nodes.
|
|
Insert memory barriers just before reading from textures as opposed to
after writing to them.
|
|
Declare input domain priorities for existing nodes.
|
|
Inputs now declare a domain priority value instead of declaring
themselves as domain inputs. This allows multiple inputs to assume
domain inference role, and the best option is chosen based on the
priority and other factors.
|
|
Delay mapping the result of the of the last input processor to the
operation input until after adding all input processors.
|