Refine Mesh error 1.7.3 12743

[eltama]

Hello Alexey
Every project will end after loading 61 MB TIFF 8bit compressed images, even with low quality,  any suggestion?
Best regards
Elmar

2021-07-22 17:10:04 Found 2 GPUs in 0 sec (CUDA: 0 sec, OpenCL: 0 sec)
2021-07-22 17:10:04 RefineMesh: quality = High, iterations = 5, smoothness = 1
2021-07-22 17:10:04 Initializing...
2021-07-22 17:10:04 Found 2 GPUs in 0.001 sec (CUDA: 0 sec, OpenCL: 0.001 sec)
2021-07-22 17:10:04 Using device: NVIDIA GeForce RTX 3090, 82 compute units, free memory: 23324/24576 MB, compute capability 8.6
2021-07-22 17:10:04   driver/runtime CUDA: 11040/8000
2021-07-22 17:10:04   max work group size 1024
2021-07-22 17:10:04   max work item sizes [1024, 1024, 64]
2021-07-22 17:10:04 Using device: NVIDIA GeForce RTX 3090, 82 compute units, free memory: 23034/24576 MB, compute capability 8.6
2021-07-22 17:10:04   driver/runtime CUDA: 11040/8000
2021-07-22 17:10:04   max work group size 1024
2021-07-22 17:10:04   max work item sizes [1024, 1024, 64]
2021-07-22 17:10:04 Device 'NVIDIA GeForce RTX 3090' has 23069 Mb of free memory
2021-07-22 17:10:05 Device 'NVIDIA GeForce RTX 3090' has 22779 Mb of free memory
2021-07-22 17:10:05 Analyzing mesh...
2021-07-22 17:10:06 Faces: 500000, Vertices: 250258
2021-07-22 17:10:06 Memory required on each device: 690 Mb + 104 Mb = 795 Mb
2021-07-22 17:10:06 Using device 'NVIDIA GeForce RTX 3090' in concurrent. (2 times)
2021-07-22 17:10:06 Using device 'NVIDIA GeForce RTX 3090' in concurrent. (2 times)
2021-07-22 17:10:26 Stage #1 out of 3
2021-07-22 17:10:27 Faces: 500000, Vertices: 250258
2021-07-22 17:10:27 Memory required on each device: 43 Mb + 104 Mb = 148 Mb
2021-07-22 17:10:27 Subdividing mesh...
2021-07-22 17:10:33 Faces: 1215002, Vertices: 607759
2021-07-22 17:10:33 Memory required on each device: 194 Mb + 254 Mb = 449 Mb
2021-07-22 17:10:34 Loading photos...
2021-07-22 17:12:06 loaded photos in 91.78 seconds
2021-07-22 17:12:06 Refining model...
2021-07-22 17:12:17 Iteration #1 out of 5
2021-07-22 17:12:37 Iteration #2 out of 5
2021-07-22 17:12:56 Iteration #3 out of 5
2021-07-22 17:13:15 Iteration #4 out of 5
2021-07-22 17:13:35 Iteration #5 out of 5
2021-07-22 17:13:54 Stage #2 out of 3
2021-07-22 17:13:54 Faces: 1215002, Vertices: 607759
2021-07-22 17:13:54 Memory required on each device: 172 Mb + 254 Mb = 427 Mb
2021-07-22 17:13:55 Subdividing mesh...
2021-07-22 17:14:10 Faces: 4104200, Vertices: 2052358
2021-07-22 17:14:10 Memory required on each device: 779 Mb + 861 Mb = 1640 Mb
2021-07-22 17:14:11 Loading photos...
2021-07-22 17:14:14 Finished processing in 250.786 sec (exit code 0)
2021-07-22 17:14:14 Error: Kernel failed: invalid device pointer (17) at line 99

[Alexey Pasumansky]

Hello Elmar,

Which NVIDIA driver version is installed on your machine?

[eltama]

Driver studio 471.41, I tried without success to compute with only one GPU

[Alexey Pasumansky]

Hello Elmar,

Can you please check, if the procedure can be completed when only single GPU is selected at a time in Metashape Preferences?

If the same behavior is observed, I would ask you to send the project with alignment results and mesh model to support@agisoft.com, so that you could check additional details.

[Alexey Pasumansky]

Hello Elmar,

Please try to switch to OpenCL implementation by creating main/gpu_enable_cuda parameter via Advanced Preferences -> Tweaks dialog and setting its value to false.

Alternatively, you may consider building mesh directly from the depth maps data using Ultra High quality (skipping dense cloud and refine mesh operations).

[eltama]

Now it's working, thank you so much 

[kb2zuz]

I was curious if this issue with dual matching GPUs had been resolved in any of the Bug fixes since and if so can I remove the CUDA tweak?

[Alexey Pasumansky]

Hello kb2zuz,

At the moment there were no improvements on this matter.

However, in most cases building mesh from the depth maps source should give a result of the expected quality and therefore Refine Mesh operation shouldn't be needed in the most recent versions of Metashape.