The innovation is noted by Huawei and responds to a problem that the company has already solved with OpenGL.
Khronos Group has announced the latest Vulkan API version 1.3.239, which has added a very interesting extension. THE VK_HUAWEI_cluster_culling_shader designed for the related SPIR-V SPV_HUAWEI_cluster_culling_shader developed by Huawei, and is very similar to its counterpart also available on OpenGL.
The innovation enables the more efficient use of two-step GPU cropping solutions, which are common at the PC level and are performed via compute shader. There is no problem with these in themselves, but their operation requires a resource limitation between the compute and the graphics assembly line, which is a factor that can be easily assumed on robust PC GPUs, but it is not so favorable to use such a system at the ultra-mobile level.
The CCS, i.e. the cluster culling shader, provides a solution for transferring data to the IA, i.e. input assembler, with a callable intrinsic function and a number of built-in variables, with which the aforementioned conveyor step provides the possibility to query the vertex points of the clusters from the vertex shader. With this method, there is no need to use a compute shader, which also avoids resource limitations.
This can be of great use at the level of ultra-mobile GPUs, because the architectures intended for this are quite narrow in terms of compute capabilities, which is why the GPU cutting procedures common at the PC level do not even apply to them very much. It’s simply in vain that they have more and more computing power, but the designs are not robust enough to effectively cope with more complex compute shaders. However, with Huawei’s new extension, they don’t even need this.
The big question is whether this development is so useful that the consortium developing the Vulkan API is considering its standardization, or whether they see that by the time they are done with it, the ultra-mobile GPUs will have developed so much that compute shaders will be less of a problem for them. .
