Could not create texture object

[natowi]

When testing Meshroom with popsift in FeatureExtraction using huge images (12000x9000px)
popsift fails with:

popsift/src/popsift/s_image.cu

Line 292 in 3e624d2

POP_CUDA_FATAL_TEST( err, "Could not create texture object: " );

I assume this is due to the huge image size that my exceeds the limit. If that is the case, a more informative error message would be useful.

[Xel233]

I'v got
Could not destroy texture object: driver shutting down

[natowi]

@Xel233 only with huge images or with normal sized images, too?

[simogasp]

Should ask @griwodz but it is very likely that there is an upper bound on the size of the images. And a hw limitation too, ie it depends on the size of the RAM of your GPU.

[griwodz]

It is a limitation that comes from CUDA because we use Textures.
Image loading uses linear memory, and the max size for newer cards is 65000x65000.
SIFT octave computation uses CudaArray, and the max size for newer cards is 65536x65535.

These numbers are just from the CUDA page on Wikipedia, but they fit with my experience. The exact number for each card can be found in the cudaDeviceProp structure (for example used in device_prop_t::print in popsift/common/device_prop.cu), but we're not checking it yet.

It is not very difficult to add a size check to enqueue() and reject images that are too big, I've just not thought about it.

[fabiencastan]

Yes, if we have a way to get this information in the popsift API, we could switch to the CPU version on the AliceVision side.

[simogasp]

in AV we have the same check/print in gpu/gpu.cpp
https://github.com/alicevision/AliceVision/blob/develop/src/aliceVision/gpu/gpu.cpp#L68

[griwodz]

I propose PR #89 to give a better and earlier warning. The function can also be used to find out if you can use PopSift for a particular width and height or if you should use a CPU alternative.
Please note that the check depends on the current Config of that PopSift object. Cards with different CCs may have different limits, it's important to cudaSetDevice to the desired one before testing.

[simogasp]

closed by #89