Search Unity

HDR Mode clarification required

Discussion in 'General Graphics' started by dotsquid, Mar 21, 2019.

  1. dotsquid

    dotsquid

    Joined:
    Aug 11, 2016
    Posts:
    224
    So I was thinking about optimization of post-processing effects and decided to try lowering the precision of back buffer (HDR mode) from FP16 to R11G11B10. I was sure to see a drastic performance boost since we're kinda saving half of the bandwidth. But there result was not impressing:
    • HDR FP16......~21.1ms
    • HDR R11G11B10.~19.7ms
    • no HDR:.......~15.6ms
    Can anyone explain why my expectations were wrong?
     
  2. aleksandrk

    aleksandrk

    Unity Technologies

    Joined:
    Jul 3, 2017
    Posts:
    3,014
    What hardware did you try this on?
    Which Unity version? Which platform?
     
  3. drcrck

    drcrck

    Joined:
    May 23, 2017
    Posts:
    328
    well it seems you're not bandwidth-bound.
     
    BrainSlugs83, Tartiflette and bgolus like this.
  4. dotsquid

    dotsquid

    Joined:
    Aug 11, 2016
    Posts:
    224
    It was tested on Unity 2017.1.4f1 / Windows 10 / DX11 / Geforce GT 440 (this old GPU is used on purpose to test the game on low-end specs).
    Just to clarify: I actually turned off all post-process effects for this test, so those timings are the result of switching between HDR Modes.
     
  5. aleksandrk

    aleksandrk

    Unity Technologies

    Joined:
    Jul 3, 2017
    Posts:
    3,014
    Well, switching from RGBA16F to RG11B10 saved you 1.4ms. Even with post-processing disabled, you probably still render stuff, and your bottleneck is probably somewhere else now :)
     
  6. dotsquid

    dotsquid

    Joined:
    Aug 11, 2016
    Posts:
    224
    I render that stuff even when HDR is off. My question was why switching from ARGB32 to RG11B10 gives me extra ~4ms.
     
    Last edited: Mar 22, 2019
  7. aleksandrk

    aleksandrk

    Unity Technologies

    Joined:
    Jul 3, 2017
    Posts:
    3,014
    It might be due to conversion to RG11B10. It might be something else.
    A profiler would tell much more accurately :)
     
    BrainSlugs83 likes this.