It's still binning per the patent because the hardware is still exercised and the additional latency is still there.
The binning is ineffective. It's also potentially injecting a specific order to where ordering isn't required.
In that case it's just a longer pipeline which is effectively not binning. Geometry is not collected, accumulated and then processed in batches but triangle after triangle. That's how I would define it that it's not binning. Binning implies at least some accumulation within a bin. Otherwise it's sequential processing of triangles as known from traditional IMRs.
Even if the binning were in play, it doesn't seem like skipping to the end of a shader that is supposed to be keeping a global count incremented by every pixel is valid.
I still don't get what you are referring to. The Vega processes geometry sequentially and not binned in tiles in that test. That is obvious. Otherwise it would complete tiles while others have not even started.
Depending on how the hardware works, all the hardware locally evaluating a tile may know is that bin 1 has some unspecified serial dependence on bin 0. The binning process is at least locally sequential. It creates a first bin, then iteratively sends primitives that intersect with other tiles to those tiles for binning.
But the bins/tiles should be independent of each other. And as the access order to an unordered access view in a pixel shader (what the triangle bin test uses for pixel counting) isn't guaranteed at all that means there are no actual dependencies between tiles (only false ones maybe).
This first bin is reducing everything it is doing to one triangle, and the hardware is flagged as needing to pause any issue for the next tile until this tile is done.
Again, if it is processing triangles one by one, is basically equivalent to saying it disables binning. And by the way, the binning rasterizer should process the geometry tile by tile anyway (with a small numbers of tiles potentially overlapping execution). But it should process all geometry for a given tile instead of one triangle after triangle (for all bins it covers).
If the hardware/driver are conservative, they may say a screen tile cannot issue primitives in bin N if N-1 is still in progress and there was a dependence flagged for the initial bin.
A bin is a screen tile. One bins into screen tiles. I'm afraid, that somehow we are not using the same terminology.
And bins are supposed to be processed somewhat sequentially (save for some overlap if cache size and shader resources allow it). Even in absence of any potential dependencies.
In this case there's just one primitive, and there all tiles have the exact same sequences of bins due to identical coverage. The patent also indicates something like double-buffering the binning process, so it may also only queue up a few bins before stalling.
Again, for me the screen tiles are the bins the geometry is binned to. Therefore, the first sentence doesn't make too much sense to me.
And as said, bins are supposed to be processed more or less sequentially with a binning rasterizer. That's the whole difference: process bins/screen tiles (with accumulated geometry) sequentially instead of processing the geometry directly in a sequential manner.
If binning is inactive (for whatever reason, maybe because of a false dependency), the rasterizer(s) work effectively like in an IMR. Primitives are processed sequentially (pixel waves are started over all screentiles the triangle covers before the next triangle is rastered [as mentioned before, some small overlap between different triangles in different screen tiles is possible]).
