VertexArrayState: dirty() forces every active array to be dispatched again - #51
VertexArrayState: dirty() forces every active array to be dispatched again#51patricktodd wants to merge 1 commit into
Conversation
|
There are plenty of other things that have been dirtied along the way before we get to the point where the offset's changed and the dispatch needs redoing as a result, and I expect that at least one of them will always need a re-dispatch when it's changed, so there should be a simpler way to fix this problem by cleverly reusing something that already exists rather than picking out the offset as a new thing that needs special tracking. |
…again A vertex array object records each attribute pointer (buffer and offset) at dispatch time. When one of a drawable's arrays is replaced, Drawable:: dirtyGLObjects() already reaches the per-context VertexArrayState through dirty(), but that only set _requiresSetArrays, and setArray() still skipped arrays whose pointer and modified count were unchanged. Those arrays share a VertexBufferObject with the replaced one and move inside it when the buffer is recompiled, so the vertex array object kept their old offsets. Have dirty() invalidate the recorded modified count of every active array so the next setArrays() dispatches them all.
ca21a1e to
03aac0d
Compare
|
OK yeah, I rebuilt around dirty() since that's what Dropping the other half (giving buffer-less arrays a VBO) turned up OpenMW's NifOsg::ParticleSystem sending a one-element overall-bound normal through the vertex array path under a VAO, so it errored constantly. Fixed that on the OpenMW side instead, so my test stations render clouds correctly again. Let me know if that fits, thanks! |
patricktodd
changed the title
2 participants
A vertex array object records the attribute pointer (buffer and offset) at dispatch time. When another array sharing the same VertexBufferObject is replaced or resized, the buffer is re-laid out and unchanged arrays move, but setArray() only re-dispatched on a new array pointer or modified count, so the vertex array object kept a stale offset. Track the buffer object and offset each array was last dispatched with and re-dispatch when they differ; reset them in dirty(). With a vertex array object in use, arrays that were never assigned a buffer object get one, as client-side pointers are not allowed there.
Observed while working on OpenMW on a GL 4.1 core profile context on macOS, alongside PR #50: OpenMW's cloud layer read its colour array from the wrong offset after the sky geometry's other arrays were regenerated, so the clouds drew with the wrong data.
Two things for review: the array gets its buffer object through a const_cast (osg::Geometry does the same internally), and the last-dispatched buffer object is compared by address. The check is gated on a VAO being in use, so the client-array path is untouched.
Mainline OpenSceneGraph master still has the offset-only check.