Because God said so. That and the aim of the system is to create something akin to the Commodore 64 in which the entire system is open to the user. http://www.templeos.org/Wb/Doc/Charter.html#l1
"Graphics operations should be transparent, not hidden in
a GPU."
Is GPU programming opaque, I've not done any low level stuff? I imagine it's a whole bunch of different API calls, but based on triangles, rather than points and lines?
Could one write a simple GPU in code, for example, I wonder?
You'd be surprised just how many layers of abstraction there are between getting something done 'outside' the context of the GPU, and across the CPU/GPU bridge, and getting it done on the GPU in a modern 3D stack these days .. You can do things any one of a number of different ways - pass off a blob of data for rendering, or write shader programs that get compiled for the GPU when the app requires it .. those compilers are not open (shader compilers are an arcane and highly contentious realm of IP-rights-holders in a very competitive and volatile industry) and often-times, the hard work of a 3D developer is spent in moving existing assets (code/resources) from one 3D-pipeline-fashion-runway de-jeur to the next ..
It is pretty arcane.
That said, of course you can write a software renderer and simulate a fair amount of the work that the GPU will usually offload from the CPU - and in many cases this has been successfully applied - e.g. the Emulation world - to the task of maintaining legacy binary assets in lieu of having source code to port. The emu guys have amazing stats in that regard.
Well with how modern GPU's, their driver and API's work you write code to do something then the driver "decides" on the "best" way of doing it.
Display drivers can replace entire shaders and modify pretty much any instruction and they often do so if nothing than for the code to actually work because game developer not only have shifted allot of the "performance" optimization burden to the GPU vendors but also quite often ship completely non complaint (and nonfunctional) graphics code, there have been plenty of gem posts on gamedev.net including multiple AAA studio's that botched their code so poorly on multiple titles that if it there wasn't a generic fix in the driver already it would not display anything at by not calling D3D::Create, D3D::EndFrame/Begin frame properly or at all in many cases.
Overall the majority of the driver codebase today is abstraction and fixes, only about a 3rd of it is actual API implementation.
So this is one reason why GPU driver code is necessarily closed :( to save face.
There are some software-based graphics cores out there, and one or two VHDL/FPGA efforts, but the performance/watt ratio between those and mainstream GPUs is laughable without a second thought.
Here's hoping AMD's vision to try and be more open in the future really works out. Also that Vulkan is at least mildly sane with open-ness.
Because the current state of things reinforces ideas like "the GPU is opaque" - architecturally, graphics processing is not a magic box, and while it would take a long time to fully understand it's not technically insurmountable; but the current status quo with drivers makes it so.
How much performance would a pure and thus leaner driver gain, I wonder.
Also a fair bit of backwards compatibility will be involved, like with x86 cpus. And even then, the same fixup is required to run opengl on directX cards, I believe.
Considering that even with API's that have some what good internal compliance testing like DirectX developers still ship utterly broken and un-optimized code - none.
This is also why Vulcan will probably not succeed (at least now how people think it will).
The last thing that say Nvidia wants is to maintain a code base of 3-4M LOC's which 50-60% of it is intended to allow for games to run anywhere from running at all to running well.
With how the current market works the driver is the "secret" sauce that GPU makers use to compete in the market and is just as important (or even more in some cases) as the hardware it self.
I wonder what the rationale behind this point is, then - especially considering the GPU is quite a bit more parallel than the CPU.