fix illegal access when there are more orientations than allocated

[mitjap]

This also resolves #105.

I am also attaching a sample image where this error happens. 1.pgm.txt

Description

This PR resolves an issue with illegal memory access when there are more orientations/descriptors than allocated. This error happens when there at least 2 times more orientations than there are extremas. There must also be at least 100k extremas for this crash to happen. You can reduce this value to get this error easier.

popsift/src/popsift/sift_conf.cu

Line 33 in 4b4b247

, _max_extrema( 100000 )

I'm pasting some link for easier understanding which parts are relevant for memory allocation and its amount.

popsift/src/popsift/sift_constants.cu

Lines 30 to 31 in 4b4b247

	h_consts.max_extrema = max_extrema;
	h_consts.max_orientations = max_extrema + max_extrema/4;

popsift/src/popsift/sift_pyramid.cu

Line 154 in 4b4b247

sz = max( 2 * h_consts.max_extrema, h_consts.max_orientations );

popsift/src/popsift/sift_pyramid.cu

Line 192 in 4b4b247

numExtrema *= 2;

This part here is interesting because there is a constant max_orientations = max_extrema + max_extrema/4; but the amount allocated is then max( 2 * h_consts.max_extrema, h_consts.max_orientations ) which does not make much sense. But that in itself is not a problem. The problem is that excess orientations are never pruned which causes problem later in the process. This makes it mandatory for checking if orientation index is within allocated memory.

Implementation remarks

Instead of pruning excess orientations I've implemented index checking at relevant places. Pruning might be better solution but it is not as easy to implement (maybe I am wrong). If pruning was to happen I think it should be done here:

popsift/src/popsift/s_orientation.cu

Lines 328 to 331 in 4b4b247

	ExtremaWrt w( extremum );
	ExtremaTot t( total_ori );
	ExtremaWrtMap wrtm( feat_to_ext_map, max( d_consts.max_orientations, dbuf.ori_allocated ) );
	ExclusivePrefixSum::Block<ExtremaRead,ExtremaWrt,ExtremaTot,ExtremaWrtMap>( total_ext_ct, r, w, t, wrtm );

[griwodz]

Instead of changing all the descriptor extraction types, I would prefer to fix the out-of-bounds problem in s_orientation.cu as you propose.

But instead of changing ExclusivePrefixSum::Block, wouldn't it be better to fix the loop

        for( int o=1; o<MAX_OCTAVES; o++ ) {
            dct.ori_ps[o] = dct.ori_ps[o-1] + dct.ori_ct[o-1];
        }

by adding a max( ..., dbuf.ori_allocated-1 ) ?