[Bf-committers] MB/DOF Further optimizations without current render architecture

[Magnus Löfgren]

Let's change this subject a bit :)

Some other ideas, possibly easier to implement without changes to the
architecture:

Is it possible to render all the sub-passes "silently" per render tile?
What I mean by that is you calculate all the scene transforms over the
specified number of sub-passes before the scene is rendered, and then you
render all of the sub-passes in each render tile/bucket and only display the
result when it's finished.

So, if you have specified 32 full scene motion blur passes, you do:
1: Calculate all transforms in the scene + raytree + shadow maps etc 32
times
2:    Silently render all 32 passes in the FIRST render tile
3:       Display the result in the FIRST render tile
4:    Silently render all 32 passes in the SECOND render tile
5:       Display the result in the SECOND render tile
6:    Etc

Possible further opimizations on this method:
* Render objects that has no animated deformations as instances (i.e. object
tranformation only)
* Render sub-passes with dithering or random noise, omitting pixels
completely from the rendering process to save render time (this will be
filled when the subpasses are merged after completion of the render
tile/bucket. Besides faster rendertimes this will reduce strobing effects
somewhat
* Deep Shadow maps can be combined intelligently in step 1 (calculate
transforms, raytree, shadow maps)


And this could be extended to depth of field, for example with 32
sub-passes:
1: Attach camera to a disc with a user specified radius
2:    Randomly place camera over surface of the disc 32 times
3:        Silently render all 32 passes in the FIRST render tile
4:             Silently render all 32 passes in the SECOND render tile
5:                  Etc


For stills only, the DOF method would greatly benefint from rendering
instances.

Combined with motion blur, this process could be:

1: Attach camera to disc with a user specified radius
2:    Calculate all transforms in the scene (including placing camera
randomly on the disc) + raytree + shadow maps etc 32 times
3:        Silently render all 32 passes in the FIRST render tile/bucket
4:            Display result in the FIRST render tile
5:        Silently render all 32 passes in the SECOND render tile/bucket
6:             Display the result in the SECOND render tile


Pros/Cons of this method:

+ Will require minimal change to the render architecture (hopefully)
+ Increased render time by optimizing the process, reducing alot of steps in
the motion blur process
+ Increased quality, especially if dithering/and/or/noise method is added.
Approaching raytraced or REYES like motion blur results
+ Blender actually gets a real depth of field feature that's not a 2.5d
image based filter

- Will possibly require quite alot of render tiles, probably more than the
default 8x8
- Initial scene transform step might be quite memory consuming, in
particular with heavy displacements, might have to move transform step per
tile instead?

Any feedback appreciated

//Magnus


---------- Forwarded message ----------
From: Magnus Löfgren <lofgrenmeister at gmail.com>
Date: 2010/4/21
Subject: Blender render architecture - Interleaved sampling
To: bf-committers at blender.org


Hi,

Being a non-programmer and lacking understanding of blenders render
architecture, I'm not sure if this is a stupid question.
But how much work would it take to implement interleaved sampling in
blenders internal renderer for Motion Blur and Depth of field?
http://www.cs.ubc.ca/labs/imager/tr/2001/keller2001a/

Blender has raytraced area light shadows and glossy reflections after all,
and a similar approach must have been made in this area.

If this is indeed a mammoth task and there isn't enough time to implement
this, how about another approach to optimize the multi-sample motion blur
and multi-sample DOF already present in blender 2.4x?

For example to in Project Messiah, their internal renderer (it shares the
same roots as Arnold from Sony), when you render motion blur, the raytrace
samples are divided by the amount of passes, so that each "sub-frame"
rendered is extremely noisy, but merged together produces noise free
results.
It's still a multi-sample "render-the-whole-scene x amount of times" but
alot smarter than blenders approach, when you can keep the amount of samples
in each subframe to a minimum to reduce total rendertime without sacrificing
quality. So you can increase the amount of passes to reduce the common
strobing effects of multipass motion blur.

In blender, the noise pattern remains identical for each subframe, so
reducing the amount of samples in ambient occlusion or raytraced shadows to
take advantage of multi-pass rendering didn't really give an advantage like
in Project Messiah.

Of coarse interleaved sampling would still be ideal since strobing is
completely eliminated and the only artifact you have to take care of is
noise, so you only need to increase the pixel samples.

Thanks