[Paraview] Paraview v5.0.0-RC1 with OpenGL2 backend not running using OSMESA v10.5.5

[Albina, Frank]

Dear Utkarsh!

Thank you very much for your swift reply and for sending the patch to the MESA libraries. I will apply it asap. 

Also, thank you for liaising with the OpenSWR team  in order to check if they support or plan supporting OSMESA. It is of interest for us, since any gain in reducing rendering times during post-processing of our simulations has a substantial impact in the longer term on the number of simulations we can perform (i.e. less time spent in post-processing means that there are more resources left for running simulations).

Concerning the surfaceLIC, we are using it primarily on GPUs but it would have been great to have the possibility to switch to software rendering under ParaView 5.0.0. This was also motivated by the fact that the software has made a good step in reducing rendering times. 
To  underline this, below is a summary of the rendering times in seconds for generating 60 pressure-contours images on the same compute cluster:

	Paraview 3.8.1		Paraview 4.3.1		Paraview 5.0.0-rc2
	Intel MPI 5.0.1.035	Intel MPI 5.0.1.035	Intel MPI 5.0.1.035
	MESA v10.5.5		MESA v10.5.5		llvm-MESA-11.0.7
Cores 	python v2.7.8		python v2.7.8		python v2.7.10
   1	     339.32		    502.49		      320.29
   2    	     177.62	    	    276.19	                      86.43
   4	      95.91		    142.94		      50.64
   8 	      57.17		     87.47		      33.71
  16	      34.91		     50.49		      26.12
  32 	      22.51		     32.44		      17.36

Basically, I can't wait to get my hands on the new ParaView release. 

On my side I will continue today to perform additional tests which are part of the way we use ParaView on our compute cluster. I'll keep you posted on that .

Talk to you soon.

Frank.

-----Original Message-----
From: Utkarsh Ayachit [mailto:utkarsh.ayachit at kitware.com] 
Sent: Donnerstag, 17. Dezember 2015 19:05
To: Albina, Frank
Cc: paraview at paraview.org
Subject: Re: [Paraview] Paraview v5.0.0-RC1 with OpenGL2 backend not running using OSMESA v10.5.5

Frank,

First of all, thanks for all your exhausting testing, it's greatly appreciated!

* OSMesa and MESA_GL_VERSION_OVERRIDE

Communicating with the Mesa team, we learnt that MESA_GL_VERSION_OVERRIDE is a bad idea since it doesnt' really create the specified OpenGL context. Thus if anything tries to use features from the specific OpenGL version, the rendering would crash and burn.
Working with Brian Paul from the Mesa team, Ken recently pushed a fix to VTK (after the RC2) that uses new OSMesa API to create the correct OpenGL context. This new OSMesa API will appear in Mesa 11.2, however.
Earlier version can be patched using the "alpha" patch that Brian provided us (attached).

* Classic v/s llvm

Your observation is indeed correct. llvmpipe is the way to go for software rendering. On my builds, classic mesa was painstakingly slow for rendering even small geometries.

* Issues with LIC

Another correct observation. There are a few regressions in the new implementation for SurfaceLIC. Realistically, we  may not be able to address these before the 5.0 release. If you're using SurfaceLIC heavily, I'd suggest building ParaView 5.0 with the older OpenGL backend (or sticking with 4.4) for the interim.

* OSMesa + SWR

While currently I am not sure openswr+osmesa builds are possible, I wouldn't be surprised if that on the roadmap. I am going to try to contact the OpenSWR team to see if they have any thoughts on that.

Utkarsh