komrad36 / kfast Goto Github PK

Hi, firstly thank you very much for your excellent work on KFAST!

I have a question about the 32 byte loads performed in processCols. I noticed that when I run valgrind on KFAST I get some complaints like this:

==21867== Invalid read of size 32
==21867== at 0x5A2D57: _mm256_cmpgt_epi8 (avx2intrin.h:261)
==21867== by 0x5A2D57: processCols<false, false> (KFastAVX2.h:207)
==21867== by 0x5A2D57: void _KFAST<false, true, true>(unsigned char const*, int, int, int, int, std::vector<Keypoint, std::allocator >&, unsigned char) (KFastAVX2.h:401)
==21867== by 0x5A1D5A: KFAST<false, false> (KFastAVX2.h:493)
==21867== Address 0x630ef43 is 307,187 bytes inside a block of size 307,200 alloc'd

From reading the code what I can understand is that, on each row when processing the final set of columns (i.e. < 32 columns left), you load in chunks of 32 bytes and mask out the data beyond the end of the row. This means that for the final row you will load data beyond the end of the buffer, and valgrind thinks this is bad.

I couldn't really find any info on the safety of doing this with AVX, so I wondering if you could comment on it, or maybe point me to some documentation?