Adrian Cochrane
How did we hardware-parallelise 3D (depth-buffered) graphics before we could stuff that hardware full of compute units?
How we assembled the different steps from multiply-adders is fairly straightforward, I'll leave that to your imagination. Though some early chips like in the PlayStation didn't include all the steps, as a cost-cutting measure.
The image data is split into tiles (who's pixels are ordered a space-filling curve) to simplify its processing.
1/?
Adrian Cochrane
That tiling helps being allowing output to be parallelized per-tile, & by making improving data-locality when fetching neighbouring pixels. Also there was often hardwired decompression to reduce memory needs.
---
Again the interesting bits, the transitions between the steps, is not public knowledge. So I'm echoing some educated guesses.
The GPU's input comes via a queue dispatching to the appropriate destination. Though Nintendo tasked this queue with more than outputting to the GPU.
2/?
Adrian Cochrane
Once the vertices are pipelined through the vertex-shader (which may be preceded by a stage to programatically generate them), that data needs to be routed to the appropriate triangle.
This could be done by storing the vertex-indices for each triangle in some dedicated memory, including circuitry to populate it with an incrementing sequence. Once each vertex finishes it'd be tested each of those indices to set a flag, indicating when to emit the triangle to the next stage.
3/4?
０７６萌ＳＮＳ is a social network, courtesy of ０７６. It runs on GNU social, version 2.0.2-beta0, available under the GNU Affero General Public License.
All ０７６萌ＳＮＳ content and data are available under the Creative Commons Attribution 3.0 license.