On 08/22, Kenneth Graunke wrote:
Previously, the state upload code listened to a broad set of dirty flags that corresponded to all possible state dependencies for a shader stage. This is somewhat overkill. For example, if a shader has no textures, there is no need to listen to _NEW_TEXTURE. Although these extra dependencies are harmless for correctness, they cause us to recompute the program keys and search the program cache more often than necessary.This patch introduces a new brw_program::nos field, containing the dirty flags which cover a shader's non-orthogonal state dependencies. We look at what the shader actually requires, and record that at link time (or fixed function program generation time). Then, the state upload code simply checks the current dirty flags against those. This should reduce CPU overhead when drawing with the same shaders multiple times, but changing state or resources (such as binding new textures or changing uniforms). Improves performance in GFXBench4's gl_driver2_off on Apollolake at 1280x720 by 3.06834% +/- 0.722141% (n=25).
That all makes sense and seems like an obviously good idea. Reviewed-by: Matt Turner <matts...@gmail.com>
signature.asc
Description: Digital signature
_______________________________________________ mesa-dev mailing list mesa-dev@lists.freedesktop.org https://lists.freedesktop.org/mailman/listinfo/mesa-dev
