Customising Open_CL max group size settings

[Coalisgood]

morning guys.

we have setup a workstation, along the lines of my previous post with the following:

Xeon E5-2687w v2 (3.4ghz) 8 cores
64gb ram
12gb Titan X
4gb Quadro K5000

Our current workflow is processing orthophotos and 3d models from an RTK UAV with a GSD of around 4cm.

We currently process with:

Photo Alignment: High
Dense point cloud: Medium
Mesh: Medium

Orthophoto: 50cm

The above settings are more than adequate for our output purpose and provide a very fast turnaround on the data.

One issue we are seeing with this upgraded workstation is the 'max group size' setting (which i believe to be the amount of data sent to the GPU's at once) appears to be too low.

The Titan X processes each piece in less than 1 second and doesnt get to max performance, and only reaches 45% utilisation. (reported performance 800 million with 0/16 CPU enabled, or 625 million with 12/16 CPU enabled)
The K5000 takes a little bit longer and hits 92% utilisation. (330 million)

I understand this is due to the resampled photo size (set by dense point cloud on medium) because when it is set to ultra, the Titan X hits full performance. (1200 million)

My question is: is it possible to edit the max_group size settings in Open_CL to send more data to the GPU's (it seemingly only sends 1 picture at a time.  my thoughts are if this could happen, the GPU would spend longer processing and there hit full performance?

is this the case? or are we 'doomed' to a reduced performance score on the Titan  X due to the processing settings?

thanks!

[SamT]

Bump!

We are in the same situation.

Ta,

Sam

[Coalisgood]

is this a valid concern? i see no other way of maximising the GPU usage with the resampled picture size

[Alexey Pasumansky]

Hello Coalisgood,

In the version 1.3 usage of GPU was optimized, also for NVIDIA cards now CUDA is used instead of OpenCL. Additionally, some optimizations and fixes related to GPU processing would be included in 1.3.1.