Pascal has native FP16 operations and can execute 2 FP16's at once ( https://docs.nvidia.com/cuda/pascal-tuning-guide/index.html#... )
BUT, and this is where things get fucked up, Nvidia then neutered that in the GeForce lineup because market segmentation. In fact, it's slower than FP32 operations: "GTX 1080’s FP16 instruction rate is 1/128th its FP32 instruction rate" https://www.anandtech.com/show/10325/the-nvidia-geforce-gtx-...
You can workaround this by doing most of your math in local coordinate spaces though, so it's fine for games.
The target architecture was a 33MHz 486 PC running in "real" (ie sixteen-bit) mode. While hardware floating point was sometimes available (DX systems) it was quite slow.
Perf? Which hardware does have support for math on fixed-integers? (e.g. sqrt, inverse sqrt, ray triangle intersection accelerations, etc.)
Maybe the vc4-dri.so driver hasen't implemented it properly yet one year later? But does it still save the RAM?
So the texture uvs could be accelerated? But most of the data is in vec3... (vertex and normal)
For vertex arrays, while in theory the driver could load vec2/vec4 F16 attributes into the vertex shader as-is, AFAICT the Mesa driver does not attempt to do this; they get converted to F32 as they are loaded from main memory. There would be downsides to getting rid of this conversion... in particular the hardware does not support F32->F16 conversions when loading attributes, so the driver would have to recompile the shader if you switched the vertex arrays to F32 (or perhaps more likely recompile the shader when it realises you are using F16 vertex arrays rather than F32).
In any case, the hardware does support zero overhead F16->F32 conversions when loading attributes, so you should not see a performance drop when switching vertex arrays to F16! The performance should go up slightly as less data needs to be loaded from main memory. If you're seeing an 80% performance drop something has gone terribly wrong!
Desktop Nvidia simply converts them to 32-bit floats up front so the performance is kept but memory is wasted, but raspberry does the 16 -> 32 bit calculations every time resulting in horrible performance.
I still have to test the engine on Jetson Nano with half floats, but I'm pretty sure I will be dissapointed again, and since raspberry doesn't support it I need to backtrack the code anyway!
After some further research I heard snapdragon has 16-bit support in the hardware and hopefully this is where we are heading! 32-bit is completely overkill for model data, it wastes memory and cycles! 16-bit for model data: vertex, normal, texture and index and 8-bit for bone-index and weights! Back to the 80-90s!
This is the last memory & performance increase we can grab now at "5"nm without adding too much complexity!
You can try the engine without half float here: http://talk.binarytask.com/task?id=5959519327505901449