I have finished a prototype of a beast application tool, but feel uncertain if I have the passion to keep improving that.
I have finalized the basic idea of project Xe-Diagram – aiming to show case a project that can visualize public architecture view. I am very satisfied on the current status. With sub pages reworked, reviewed and published, I think my project can sit better than the public Programmer’s Reference Manuals and can serve better as a knowledge base like MCP server for Intel post-si GPUs.
My beast project is done as a prototype locally, referring to xe-diagram and serve as a project aiming to completely change the “tune” process of game enabling on Intel GPU, in this “agent-based” workflow era. I have finished a solid project with codex in the past weeks, and I like it so much.
However, I have found a missing layer in between.
I initialized a version that the application tool behave purely based on Xe-Diagram knowledge and global knowledge base, and have found that the conclusion we got from this combo (xe-diagram, global knowledge, captured data) is not accurate enough, unfortunately.
I reviewed the decision tree wrote by codex, and it looks promising and nice, in theory. But there are gaps between theory results and reality, including but not limited to the real cost of a profile, micro architecture behavior in each module, and potential defect on existing layers. That’s why incrementally improvements to the knowledge base of the beast are necessary. And that is the missing “experience layer” between the projects.
Those tune experiences can be done by just experiments, treating the GPU/MDapi/driver as pure black boxes. I believe if I spend enough time I can eventually get things figure out. I love our GPU product and am willing to contribute to the customer community, but unfortunately I cannot do these things.
Because I belong to Xe-Architecture team already, I have the insider access to these types of knowledge. Which means, if I continue to “tune” the project and make adjustments, I may occasionally expose some confidential micro architecture. The “tune knowledge base” becomes confidential to stop me from continuing this path.
Will consider stop here and find a way to publish the low-value prototype.
I am happy that I can now develop my own tool individually. Enjoy it a lot!
Leave a Reply