Conservative Rasterization EarlyZ Optimization via Inner Coverage
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conservative rasterization in GPUs conflicts with EarlyZ optimization, leading to performance issues due to erroneous Z test and Z-write operations when generating extrapolated depth values, prompting the disabling of EarlyZ, which limits the usage of conservative rasterization.
Innovation Solution
Exposing inner coverage information to components of the conservative rasterization pipeline allows for the enabling of EarlyZ by employing three approaches: enabling EarlyZ for fully covered pixels, using LateZ optimization for partially covered pixels, or conditionally processing pixels based on inner coverage data to clamp extrapolated depth values.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Power
If EarlyZ optimization is enabled for conservative rasterization, then GPU processing power is improved by avoiding shading of occluded pixels, but the Z test and Z-write operations become erroneous due to extrapolated depth values extending beyond the original primitive Z-range
Solution Approach 1:
The patent applies local quality by differentiating the treatment of depth values based on pixel coverage status. Fully covered pixels receive extrapolated depth values for EarlyZ optimization, while partially covered pixels use clamped depth values within the primitive Z-range. This localized differentiation resolves the contradiction by applying the appropriate depth value strategy to each pixel type, maintaining both processing efficiency and operational accuracy.
Solution Approach 2:
The patent changes the depth value parameter based on coverage information. For partially covered pixels, it clamps the depth value to the minimum and maximum Z-values of the primitive, transforming the erroneous extrapolated depth into a reliable bounded depth value. This parameter transformation resolves the contradiction by ensuring depth values remain within valid ranges for Z test and Z-write operations while preserving EarlyZ benefits for fully covered pixels.
2Reliability
If EarlyZ is disabled for conservative rasterization, then the accuracy of depth operations is maintained, but GPU processing power is wasted by shading pixels that would be occluded
Solution Approach 1:
The patent applies local quality by differentiating the treatment of depth values based on pixel coverage status. Fully covered pixels receive extrapolated depth values for EarlyZ optimization, while partially covered pixels use clamped depth values within the primitive Z-range. This localized differentiation resolves the contradiction by applying the appropriate depth value strategy to each pixel type, maintaining both processing efficiency and operational accuracy.
Solution Approach 2:
The patent segments the pixel processing into two distinct paths: fully covered pixels processed with EarlyZ optimization using extrapolated depth values, and partially covered pixels processed with clamped depth values for accurate Z test operations. This segmentation allows the system to simultaneously achieve both goals by routing different pixel types through appropriate processing pipelines.
3Manufacturing precision
If conservative rasterization is used to generate fragments for every pixel touched by a primitive, then image quality is improved, but the complexity of depth value generation increases due to extrapolated depth values for partially covered pixels
Solution Approach 1:
The patent applies local quality by differentiating the treatment of depth values based on pixel coverage status. Fully covered pixels receive extrapolated depth values for EarlyZ optimization, while partially covered pixels use clamped depth values within the primitive Z-range. This localized differentiation resolves the contradiction by applying the appropriate depth value strategy to each pixel type, maintaining both processing efficiency and operational accuracy.
Solution Approach 2:
The patent introduces coverage information as an intermediary that mediates between conservative rasterization output and depth value generation. The coverage information indicates whether a pixel is fully or partially covered, enabling the system to automatically select the appropriate depth value strategy without complex manual intervention. This intermediary simplifies the overall system by providing clear decision criteria for depth value generation.
Data Source
AI summary
Embodiments described herein are generally directed to conservative rasterization pipeline configurations that allow EarlyZ to be enabled for conservative rasterization. An embodiment of a method includes receiving, by a conservative rasterizer, a primitive; creating, by the conservative rasterizer, a pixel location stream based on the primitive and inner coverage data for each pixel within the pixel location stream indicative of whether the corresponding pixel is fully covered or partially covered by the primitive; for each block of pixels of the pixel location stream, launching, by the conservative rasterizer, a thread of a pixel shader, including causing EarlyZ to be performed or not for fully covered pixels and partially covered pixels, respectively; and generating, by the pixel shader, a stream of pixel updates by conditionally processing the pixel location stream to incorporate pixel shading characteristics, including for partially covered pixels computing a depth value and causing LateZ to be performed.