libvips segfault during mapim with bicubic interpolation

[erdmann]

Hi John,

I'm finding that libvips is segfaulting whenever it performs mapim using bicubic interpolation. Here is a demonstration using a debug build from tag v8.7.2 (though the problem seems to go back a few versions):

$ vips --version
vips-8.7.2-Sun Dec 16 08:18:04 UTC 2018

$ vips black im.v 100 100
memory: high-water mark 16.09 KB

$ vips xyz coords.v 100 100
memory: high-water mark 143.86 KB

$ vips mapim im.v im2_linear.v coords.v
memory: high-water mark 29.54 KB

$ vips mapim im.v im2_cubic.v coords.v --interpolate=bicubic
**
VIPS:ERROR:bicubic.cpp:536:void vips_interpolate_bicubic_interpolate(VipsInterpolate*, void*, VipsRegion*, double, double): assertion failed: (ix - 1 >= in->valid.left)
Aborted

Many thanks in advance for having a look!

[jcupitt]

Oh dear, that's bad. I guess there's an off-by-one problem somewhere.

It seems to work in 8.8, I'll try 8.7. I don't remember touching any of that code, funny.

[jcupitt]

Duh I had DEBUG off. With it on, I see the assert fail on 8.8 as well. Thanks for the great report, I'll investigate.

[erdmann]

As an additional piece of information, it looks like it applies to the other interpolation types, too, suggesting that it might have something to do with boundary conditions when kernel footprint steps outside of the image.

$ vips --version
vips-8.8.0-Sun Dec 16 10:26:51 UTC 2018
memory: high-water mark 0 bytes

$ vips mapim im.v im2_cubic.v coords.v --interpolate=bicubic
**
VIPS:ERROR:bicubic.cpp:536:void vips_interpolate_bicubic_interpolate(VipsInterpolate*, void*, VipsRegion*, double, double): assertion failed: (ix - 1 >= in->valid.left)
Aborted

$ vips mapim im.v im2_nohalo.v coords.v --interpolate=nohalo
**
VIPS:ERROR:nohalo.cpp:1524:void vips_interpolate_nohalo_interpolate(VipsInterpolate*, void*, VipsRegion*, double, double): assertion failed: (ix - 2 >= in->valid.left)
Aborted

$ vips mapim im.v im2_vsqbs.v coords.v --interpolate=vsqbs
**
VIPS:ERROR:vsqbs.cpp:344:void vips_interpolate_vsqbs_interpolate(VipsInterpolate*, void*, VipsRegion*, double, double): assertion failed: (ix - 1 >= in->valid.left)
Aborted

$ vips mapim im.v im2_lbb.v coords.v --interpolate=lbb
**
VIPS:ERROR:lbb.cpp:803:void vips_interpolate_lbb_interpolate(VipsInterpolate*, void*, VipsRegion*, double, double): assertion failed: (ix - 1 >= in->valid.left)
Aborted

Thanks again!

[jcupitt]

OK, fixed in git master plus the head of the 8.7 branch. Thank you very much for finding this awful problem @erdmann !

I'll wait a few days before pushing out 8.7.4, in case anything else comes up.

[kleisauke]

I think this should land in 8.7.3 (instead than 8.7.4) because 8.7.3 isn't released yet.

[jcupitt]

Yes, I goofed. I've moved it back to 8.7.3.

[erdmann]

Hi John,

Thanks for the very fast response. This seems to solve the segfault problem, but I now get artifacts along the right and bottom edges of the image using cubic interpolation. I'm expecting that when I do mapim using an identity mapping, I'll get an identical image to the input, and that's true when I use linear interpolation, but it's not true with bicubic interpolation. With bicubic, the bottom row and right column are difference.

Here is a demonstration of the issue:

$ vips --version
vips-8.8.0-Mon Dec 17 18:06:08 UTC 2018
memory: high-water mark 0 bytes

$ convert logo: logo.tif # make a non-trivial image using imagemagick

$ vipsheader logo.tif 
logo.tif: 640x480 uchar, 3 bands, srgb, tiffload
memory: high-water mark 0 bytes

$ vips xyz coords.v 640 480
memory: high-water mark 3.28 MB

$ vips mapim logo.tif logo_linear.v coords.v 
memory: high-water mark 3.80 MB

$ vips subtract logo.tif logo_linear.v difference_linear.v
memory: high-water mark 2.81 MB

$ vips min difference_linear.v
0.000000
memory: high-water mark 90.38 KB

$ vips max difference_linear.v
0.000000
memory: high-water mark 90.38 KB

$ vips mapim logo.tif logo_cubic.v coords.v --interpolate=bicubic
memory: high-water mark 3.80 MB

$ vips subtract logo.tif logo_cubic.v difference_cubic.v
memory: high-water mark 2.81 MB

$ vips max difference_cubic.v # problem is here
255.000000
memory: high-water mark 90.38 KB

$ vips min difference_cubic.v 
0.000000
memory: high-water mark 90.38 KB

# remove the bottom row and right column
$ vips crop difference_cubic.v difference_cubic_upper_left.v 0 0 639 479 
memory: high-water mark 1.80 MB

$ vips min difference_cubic_upper_left.v 
0.000000
memory: high-water mark 90.23 KB

$ vips max difference_cubic_upper_left.v 
0.000000
memory: high-water mark 90.23 KB

Thanks again!

[erdmann]

Sorry, accidentally closed.

[jcupitt]

You're right, the clip was in the wrong place. I improved the test as well. Could you test again, please?

[erdmann]

Looks great! All of my personal tests pass (tested on master). Thanks again.