Skip to content

Fixed various issues found in development #4

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
kudanowen wants to merge 3 commits into
base: master
Choose a base branch
from
Open

Fixed various issues found in development #4

Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
dbeccfa
Fix performance issues close to image boundary.
kudanowen Jul 8, 2019
5b975a3
Fix undefined behaviour when cols == 35.
kudanowen Jul 8, 2019
6e7fd63
Fixed reading into the next row which could generate false keypoints.
kudanowen Jul 8, 2019
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
68 changes: 50 additions & 18 deletions KFAST.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,6 +62,7 @@

#include <algorithm>
#include <cstdint>
#include <cstring>
#include <future>
#include <immintrin.h>
#include <vector>
Expand All @@ -88,7 +89,38 @@ void processCols(int32_t& num_corners, const uint8_t* __restrict & ptr, int32_t&
const __m256i& consec, int32_t* const __restrict corners, uint8_t* const __restrict cur,
std::vector<Keypoint>& keypoints, const int32_t i, const int32_t start_row) {
// ppt is an integer vector that now holds 32 of point p
__m256i ppt = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr));
__m256i ppt, p9, p5, p1, p13;
if (full) {
ppt = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr));

// Rosten's point 9 for all 32 pixels in consideration
p9 = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr + *offsets));

// Rosten's point 5
p5 = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr + offsets[4]));

// Rosten's point 1
p1 = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr + offsets[8]));

// Rosten's point 13
p13 = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr + offsets[12]));
}
else {
// to avoid reading into the next row and generating false features.
ppt = _mm256_setzero_si256();
p9 = p5 = p1 = p13 = ppt;

memcpy(&ppt, ptr, cols - j - 3);
memcpy(&p9, ptr + offsets[0], cols - j - 3);
memcpy(&p5, ptr + offsets[4], cols - j - 3);
memcpy(&p1, ptr + offsets[8], cols - j - 3);
memcpy(&p13, ptr + offsets[12], cols - j - 3);
}

p9 = _mm256_xor_si256(p9, ushft);
p5 = _mm256_xor_si256(p5, ushft);
p1 = _mm256_xor_si256(p1, ushft);
p13 = _mm256_xor_si256(p13, ushft);

// we subtract (and clamp) the threshold value from all 32 pixels
// pmt represents p - t
Expand All @@ -98,18 +130,6 @@ void processCols(int32_t& num_corners, const uint8_t* __restrict & ptr, int32_t&
// ppt represents p + t
ppt = _mm256_xor_si256(_mm256_adds_epu8(ppt, t), ushft);

// Rosten's point 9 for all 32 pixels in consideration
__m256i p9 = _mm256_xor_si256(_mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr + *offsets)), ushft);

// Rosten's point 5
__m256i p5 = _mm256_xor_si256(_mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr + offsets[4])), ushft);

// Rosten's point 1
__m256i p1 = _mm256_xor_si256(_mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr + offsets[8])), ushft);

// Rosten's point 13
__m256i p13 = _mm256_xor_si256(_mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr + offsets[12])), ushft);

// compare p's against p9's
// compare p's against p5's
// AND the result into ppt_accum
Expand Down Expand Up @@ -156,7 +176,7 @@ void processCols(int32_t& num_corners, const uint8_t* __restrict & ptr, int32_t&
// the normal full 32 columns, or special handling for the last few columns if they
// don't divide up evenly into 32
uint32_t last_cols_mask;
if (!full) last_cols_mask = (1 << (cols - j - 3)) - 1;
if (!full) last_cols_mask = static_cast<uint32_t>((1ull << (cols - j - 3)) - 1);

// 'mask' now contains one bit for each element
// which is SET if that element COULD be a corner based on the 2 consective cardinal point test
Expand Down Expand Up @@ -187,7 +207,17 @@ void processCols(int32_t& num_corners, const uint8_t* __restrict & ptr, int32_t&
// for each of the 24 pixels in the circle (wrapping around extra 8 at the end)
for (int32_t k = 0; k < 24; ++k) {
// x is a vector of the kth member of the circle
__m256i p = _mm256_xor_si256(_mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr + offsets[k])), ushft);
__m256i p;
if (full) {
p = _mm256_xor_si256(
_mm256_loadu_si256(reinterpret_cast<const __m256i*>(ptr + offsets[k])), ushft);
}
else {
p = _mm256_setzero_si256();

memcpy(&p, ptr + offsets[k], cols - j - 3);
p = _mm256_xor_si256(p, ushft);
}

ppt_accum = _mm256_cmpgt_epi8(p, ppt);
pmt_accum = _mm256_cmpgt_epi8(pmt, p);
Expand Down Expand Up @@ -372,9 +402,11 @@ void _KFAST(const uint8_t* __restrict const data, const int32_t cols, const int3
processCols<true, nonmax_suppression>(num_corners, ptr, j, offsets, ushft, t,
cols, consec, corners, cur, keypoints, i, start_row);
}
// handle last few columns
processCols<false, nonmax_suppression>(num_corners, ptr, j, offsets, ushft, t,
cols, consec, corners, cur, keypoints, i, start_row);
if (cols - j - 3 > 0) {
// handle last few columns
processCols<false, nonmax_suppression>(num_corners, ptr, j, offsets, ushft, t,
cols, consec, corners, cur, keypoints, i, start_row);
}
}

if (nonmax_suppression) {
Expand Down