[Bf-cycles] UV Pass and Self Overlap

[Brecht Van Lommel]

>From the kernel point of view, the clear point to write deep EXR would
be kernel_write_data_passes
and kernel_write_light_passes The tricky part though is how to store them
though, we currently assume pixel buffers of fixed size in the Blender
render API, compositor, image APIs, Cycles host and device side buffers,
etc. So it's a big undertaking to implement that throughout the entire
pipeline. The easier solution is to write to a fixed number of passes, like
Cryptomatte (see https://developer.blender.org/D2106 for a proof of concept
implementation). That avoids a lot of the trickiness of each pixel having
an arbitrary number of samples.

On Sat, Jul 29, 2017 at 7:01 PM, Dylan McDiarmid <dylanmcd at gmail.com> wrote:

> Thank for the info, Brecht. I’ll keep the downsampling solution as a last
> resort, and try implementing a pass that gives me enough information to do
> the compositing operations in post. I don’t need to support blur or DOF,
> but overlapping transparency is a possibility I didn't think of.
>
>
>
> Now that I know I’ll need to hack this together, I’ll work on getting a
> good development environment setup and familiarizing myself with the code.
> If Blender did support deep EXR, what would your high-level strategy be for
> implementing this feature for a solo object render? Is there a single place
> where the UV Pass is written, or is this something that happens from
> multiple places?
>
> *From: *Brecht Van Lommel <brechtvanlommel at pandora.be>
> *Sent: *Friday, July 28, 2017 8:03 PM
> *To: *Discussion list to assist Cycles render engine developers
> <bf-cycles at blender.org>
> *Subject: *Re: [Bf-cycles] UV Pass and Self Overlap
>
>
>
> There is currently no good way to output such information. Rendering in a
> high resolution and then scaling down may be the best way to try to recover
> this information. Modifying the code to write only the first sample might
> help with that, like we do for the Z pass.
>
>
>
> In general it's not just "top" and "bottom", but rather an arbitrary
> number of objects that might contribute to a pixel, either due to
> antialiasing, transparency, depth of field or motion blur. A better
> solution to this is deep EXR, which we do not support currently.
>
>
>
> On Thu, Jul 27, 2017 at 9:25 AM, Dylan McDiarmid <dylanmcd at gmail.com>
> wrote:
>
> I'm doing some work that involves mapping textures post-render, and I'm
> trying to figure out how to get anti-aliasing information out of the UV
> pass. From the UV pass EXR, it looks like for edges that overlap the
> background, we get a nice B value that indicates opacity. So, to find the
> actual UV position we would do U = R / B, V = G / B, and then do an alpha
> blend with what we want as the background using the B value.
>
>
>
> However, when an object overlaps itself, the B value stays at 1, and
> instead we get the actual alpha blend of the top R,G with the bottom R,G.
> So when we're doing our mapping, this puts the UV at a completely wrong
> place.
>
>
>
> To properly represent the antialiasing at places that self-overlap, we
> need 5 values, bottom U, bottom V, top U, top V, top opacity. I don't know
> HDR formats very well, are these values already being captured somewhere in
> some kind of metadata field? If not, is there currently a way to do a
> separate pass to capture this information?
>
>
>
> I've done a bit of poking around to the source, but I'm afraid I've been
> hindered by a lack of domain knowledge. It seems that the pass information
> is being read from
>
>
>
> kernel_passes.h : kernel_write_data_passes -> geom_primitive.h :
> primitive_uv -> geom_attribute.h : find_attribute -> kernel_compat_cpu.h:
> kernel_tex_fetch -> KernelGlobals->__attributes_map.fetch(index)
>
>
>
> but I'm having trouble figuring out where and when __attributes_map gets
> filled (and with what, and why :).
>
>
>
> As a newcomer to graphics programming, it's a real pleasure to find such a
> great renderer that I can actual read the source to. So thanks to everyone
> involved.
>
>
>
>
> _______________________________________________
> Bf-cycles mailing list
> Bf-cycles at blender.org
> https://lists.blender.org/mailman/listinfo/bf-cycles
>
>
>
>
>
> _______________________________________________
> Bf-cycles mailing list
> Bf-cycles at blender.org
> https://lists.blender.org/mailman/listinfo/bf-cycles
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://lists.blender.org/pipermail/bf-cycles/attachments/20170801/2cad50cb/attachment.htm