nona-gpu result width is padded to a multiple of 8

[Yuv]

initial tests by Thomas Modes show that the result of nona -g sometimes differ from the expected result by a few pixels on the horizontal. This is likely caused by the padding to a multiple of 8 for faster GPU transfer. We need to identify the padding pixels and remove them from the end result. It would be a bug if an equirect of 398x199 would return as 400x199.

this is critical for the upcoming 2009.2 release.

[tmodes (tmodes)]

This bug could also result in a crash of nona.
For a test project it crashes when output non-cropped tiffs with a cropped area. When output cropped tiffs with all other settings unchanged it works.

It crashs in saveRemapped with a heap corruption.

[Yuv (yuv)]

seems not to affect nVidia video cards, unless I misunderstand
http://groups.google.com/group/hugin-ptx/msg/40b305e318f3aa72

?

@Thomas: to avoid misunderstandings, can you add to this tracker the necessary files to run the test case to this tracker? then we can ask Terry to run them on his machine and we'll know for sure.

[tmodes (tmodes)]

The file test.zip was added: Test case

[tmodes (tmodes)]

I added a test project.

When output cropped tiffs, the resulting image size are

no crop area (project gpu_cnc.pto)
nona: 293x197
nona -g: 296x197

with crop area (gpu_cc.pto)
nona: image 1: 260x169, image 2: 269x169
nona -g: image 1: 264x169, image 2: 272x169

When normal tiffs are selected as output (gpu_ncnc.pto and gpu_ncc.pto), "nona -g" outputs the same image sizes as "nona".
But with gpu_ncc.pto there is an artefact from padding to multiply of 8: The first image has on the left border a column of 4 pixel which belongs to the right border (They should be cropped from the right border to get a correct output, but not add to the left border) With bigger output images I'm getting in this case an crash in nona.

[pmllc (pmllc-yahoo)]

I confirm this bug on ATI video card (HD5400 series)

[Yuv (yuv)]

I set importance to 'High' because I would like to use 'Critical' only for things that we'd like to see in the next release (the next tarball of the current release cycle, not the next release cycle)

[tmodes (tmodes)]

Tried to fix in changeset e741dddbdba1