Is it proportional to the samples count?
Whats the performance difference between binding a multi-sampler and a regular sampler?
2
-
2$\begingroup$ "How does it affect the performance when resolving it immediately or later?" How does what affect the performance? $\endgroup$– Nicol BolasNov 30, 2021 at 16:40
-
$\begingroup$ Sorry for my poor English, I 've re-edit it. $\endgroup$– DanielhuDec 1, 2021 at 15:18
Add a comment
|
1 Answer
$\begingroup$
$\endgroup$
There is a lot of implication in binding a multi-sampler image.
Basically, multi-sampling will run multiple fragments for the same pixel. AFAIK if the same triangle is rasterized on multiple samples of the same pixel, it will be shaded only once (This is implementation-specific!). This is the advantage of supersampling.
More sample count you have more chance you get to get more shaded fragments, therefore, bandwidth, instructions count, memory, everything is impacted.
Also, the multi-sampled image takes more space and might(will) have different internal alignment.
To conclude, here is a summary of the differences between the techniques:
- normal sampler: 1 fragment per pixel, good coherency, fast, might have aliasing
- ms sampler: 2/4/8/16 fragments per pixel when a lot of triangles are partially covering the triangle, less good coherency, quite fast, need a resolve pass, less aliasing.
- supersampling: 2/4/8 fragments run for each resolved pixel. very expensive, need a beefy gc good coherency, not fast, best mathematical antialiasing.
Basically, multi sampling is a "cheap supersampling".
Is it proportional to the samples count?
Yes but not linear.
Whats the performance difference between binding a multi-sampler and a regular sampler?
Benchmarking is the only way to know for your renderer on one card model with one driver. But it will affect speed significantly.
