are you using popsift-demo? If so can u show the log?

@simogasp yes, i using popsift-demo, but But I set the config parameter to：

and crash at get_descriptor function(sift_pyramid.cu file 317 line).
Before that, I printed out the number of feature points： ext_total: 19558, ori_total: 44599.
The error log:

Strange. That error was typical for ext_total==0, but not when ext_total>0. It is possible that the error is actually somewhere else.

Could you enable PopSift_ERRCHK_AFTER_KERNEL in CMake and recompile? That should activate error checking in lines that say POP_SYNC_CHK. Your code will run slower and still crash, but closer to the true origin of the crash.

What is your GPU? Do your CPU and GPU share memory?

@griwodz ,Thank you for your reply. I've found out the reason. I set _max_extrema to 10000, and when initializing pyramid ，it set dbuf_shadow.desc = malloc(2 * _max_extrema). I extracted 44599 descriptors in this image

@lizhaohu Thanks for finding the reason for the crash.

That is still very strange. In the figure of the Config that you show above, you are setting _max_extrema = 10000, but you have found 19558 (ext_total).

That should not be happening. There is a line:
if( indicator && write_index < d_consts.max_extrema )
in s_extrema.cu:537 that should prevent this from happening.
d_consts.max_extrema is filled by PopSift::configure with the MaxExtrema value from your Config object.

Do you perhaps change the max_extrema to 20000 somewhere else?
That would explain the bug: I did not expect that you could find 125% more descriptors than extrema. Lowe wrote in the original SIFT paper that we could expect about 25% more descriptors than extrema. There is no safety check for the 125% case.

PS: a hotfix to prevent a crash over 100% additional descriptor seems to be simple.
The kernel ori_prefix_sum() in s_orientation.cu (lines 344-345) could compare with d_consts.max_extrema and (2* d_consts.max_extrema) and ignore all extrema and descriptors above the limit.

It would be better, of course, to reallocate memory after that kernel.

@lizhaohu Thanks for finding the reason for the crash.

That is still very strange. In the figure of the Config that you show above, you are setting _max_extrema = 10000, but you have found 19558 (ext_total).

That should not be happening. There is a line:
if( indicator && write_index < d_consts.max_extrema )
in s_extrema.cu:537 that should prevent this from happening.
d_consts.max_extrema is filled by PopSift::configure with the MaxExtrema value from your Config object.

Do you perhaps change the max_extrema to 20000 somewhere else?
That would explain the bug: I did not expect that you could find 125% more descriptors than extrema. Lowe wrote in the original SIFT paper that we could expect about 25% more descriptors than extrema. There is no safety check for the 125% case.

@griwodz I didn't set max_extrema = 20000 anywhere else. In my understanding, max_extrema means the maximum number of features that can be extracted from each layer of pyramid. So I got that 19558 was the right result

@griwodz In sift_ pyramid.cu file the reallocaxtreme function will be re-malloc the memory

PS: a hotfix to prevent a crash over 100% additional descriptor seems to be simple.
The kernel ori_prefix_sum() in s_orientation.cu (lines 344-345) could compare with d_consts.max_extrema and (2* d_consts.max_extrema) and ignore all extrema and descriptors above the limit.

It would be better, of course, to reallocate memory after that kernel.

@griwodz Great. Thank you very much