[Sidefx-houdini-list] Film at 16 bit

Sean Lewkiw seanl at rainmaker.com
Wed Mar 12 13:14:08 EDT 2008


Interestingly, we were looking at this a while back.

One of our senior compers reports that shake converts exrs to 16 bit on 
load, so 32 bit exrs are unnecessary.  However, we also found that 16 
bit float EXRs were not entirely accurate.  (For example, rendering UV 
maps out of Houdini and loading into Shake results in some very subtle 
banding.)

We have not done any tests on loading times/image quality and image 
sizes of 32 bit VS 16 bit exr floats.  I'm wondering if Nuke deals with 
this differently and/or better.

So, our policy here is 16 bit float with embedded DODs for most CG 
renders, except where great precision is necessary, then 32 bit TIFFs.

Sean




Simon Kapeniak wrote:
> Hello Mark,
> This is extremely interesting. Are you saying that any application working
> on any current platform has to deal with a 16bit float data by converting it
> up to a 32bit float internally? Or it does have to do it if wants to use
> MMX/SSE? How this happens in Nuke working with EXR for example
> (theoretically)? Assuming that it doesn't use GPU which supports half floats
> natively, which isn't the case I bet. Surely Shake converts EXRs to 32bit on
> load, which illustrate that issue.
> 
> Hmm, interesting... I've never though this way although I'm complete freak
> about compositing performance (working with Shake's iffs mostly for that
> very reason). I was silently assuming that half float means means half space
> and half work for CPU (roughly).
> 
> 
> 2008/3/7, Mark Alexander <malexender at sidefx.com>:
>>
>> Both 16b Int and 32b FP are XMMX and SSE optimized. 16b FP is the
>> slowest of the bunch, being a format that is not native to the CPU (it
>> is pretty much always up-converted to FP32). 16b Int is probably the
>> most work, since you have to worry about the white point (setting it
>> somewhere around 4000 for really HDR stuff).
>>
>> However, you can save all your files in 16b FP and force the compositor
>> to load them as FP32 files. Edit->Composite Project Settings, "Pixel
>> Format" = "32 bit FP" and "File Operators use the Project Pixel Depth by
>> Default" = "On". That save disk space and network bandwidth, while
>> optimizing performance in COPs by comp'ing everything in 32b FP. You can
>> still save out 16b FP files at the end without any problems.
>>
>> Cheers,
>>
>> M.
>>
>> _______________________________________________
>> Sidefx-houdini-list mailing list
>> Sidefx-houdini-list at sidefx.com
>> https://lists.sidefx.com:443/mailman/listinfo/sidefx-houdini-list
>>
> _______________________________________________
> Sidefx-houdini-list mailing list
> Sidefx-houdini-list at sidefx.com
> https://lists.sidefx.com:443/mailman/listinfo/sidefx-houdini-list

-- 
Sean Lewkiw
--
CIS London
130 Shaftesbury Avenue
London W1D 5EU
--
+44 (0) 207.031.1138 (P) (Direct)
+44 (0) 207.031.1136 (F)
+44 (0) 795.606.7245 (M)
--
Composite Image Systems London Ltd
Registered Office: Denham Media Park, North Orbital Road, Denham, 
Uxbridge, Middlesex UB9 5HQ
Company Registration No. 5577229 England




More information about the Sidefx-houdini-list mailing list