Hello,
I am developing a visualization application which at the moment has a very simple scene content: a static but quite large rectangular triangle mesh defined as a VBO, and which extends well beyond the viewport.
When I run the Grapher, I see significant variations in such metrics as Reused vertices/Second, % Vertex Fetch Stall, GPU % Busy, etc.
I wonder if these variations in the graphs are caused by the Profiler accessing the Adreno registeres where the performance data is stored, or the Profiler accesses GMEM directly via adb, and calculates this data at time intervals defined by frequency in the Profiler's GUI?
I also noticed that the number of clipped primitives is always zero. What is the size of the guardband? Can I control it in my app?
Thank you
.
Hi,
In grapher mode, Profiler does not significantly impact performance, since the process of reading counters is lightweight. There are a number of factors that can cause the variation you are seeing right now; the only thing that I can really suggest that might reduce the variation is to increase the sampling frequency, and pay more attention to the trends. The next version of Profiler (due out in March) will change the way we sample counters and should improve this significantly.
I don't remember offhand how big our guardband is, but it seems like it's 2x. It is not possible to control it.