Cherry-picked src/ changes from edge branch commit 17450d122.
Adds ParsePercentagePolygon() to convert percentage-based zone coordinates
to pixel values using monitor dimensions. Zone::Load() now checks the Units
column to dispatch between legacy pixel parsing and percentage-based parsing.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Instead of copying the entire delta_image into the per-zone mask buffer
every frame (4MB memcpy for 4MP), allocate a persistent grayscale mask
and only zero the polygon bounding-box rows. alarmedpixels_row now
reads from delta_image (const, read-only) and writes threshold results
to the separate mask buffer.
Key changes:
- alarmedpixels_row takes separate pdelta (read) and pmask (write) pointers
- std_alarmedpixels accepts both delta_image (const) and mask_image
- CheckAlarms allocates mask with explicit linesize == width to match
filter/blob pointer arithmetic (avoids FFALIGN padding mismatch)
- Full buffer zero on allocation; bbox-only zero on reuse
- No functional change to filter, blob, or HighlightEdges stages
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Restructure std_alarmedpixels inner loop so GCC/Clang can auto-vectorize
it at -O2/-O3. The compiler now emits 16-byte SIMD (SSE2/NEON) processing
16 pixels per iteration instead of 1.
Three changes enable this:
- Extract inner loop into static alarmedpixels_row() with __restrict__
on function parameters, giving the compiler a strong no-alias guarantee
- Use branchless bitwise AND instead of short-circuit && to avoid
branches that block vectorization
- Remove per-row Debug(7) call that clobbered memory from the compiler's
perspective, invalidating pointer analysis
The original hand-written SSE2 ASM (removed in 2011, commit 8e9ccfe1e)
had alignment restrictions and didn't use per-row polygon ranges. This
approach is portable, maintainable, and achieves equivalent throughput.
GCC confirms: "loop vectorized using 16 byte vectors"
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Replace delete/new Image cycle with lazy-alloc + Assign(). When the
buffer already exists (every frame after the first), Assign() detects
matching dimensions and does a plain memcpy into the existing
allocation, eliminating an aligned malloc+free of ~2 MB per zone per
analyzed frame.
With 4 zones at 15 fps this removes 60 alloc/free cycles per second
from the analysis hot path. The HighlightEdges code path (analysis
images) still allocates a new Image and deletes the old diff buffer,
which is correct — the next Assign() will reallocate once to restore
the single-channel format, then resume reuse.
Behaviorally equivalent: Zone dimensions are constant during zone
lifetime, the destructor already handles cleanup via delete image,
and the only external consumer (Monitor::Analyse → AlarmImage →
Overlay) reads the image without storing pointers.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- Apply credentials to secondary stream URL in FFmpegCamera (was causing 401 Unauthorized)
- Add empty check for rtsp_second_path in RTSP2WebManager before applying credentials
- Replace unsafe sprintf pattern in Monitor::DumpSettings with std::string + stringtf
- Refactor Zone::DumpSettings to return std::string instead of writing to char buffer
- Add decimal precision to event duration debug output
Co-Authored-By: Claude Opus 4.5 <noreply@anthropic.com>
Move the definitions to zm_define.h and include the header at the appropriate locations.
These macros have not been adopted by the C++11 standard.
However glibc 2.17 (CentOS 7) still depends on them to provide the macros which are guarded by these defines.
With this commit a unified structure for includes is introduced.
The general rules:
* Only include what you need
* Include wherever possible in the cpp and forward-declare in the header
The includes are sorted in a local to global fashion. This means for the include order:
0. If cpp file: The corresponding h file and an empty line
1. Includes from the project sorted alphabetically
2. System/library includes sorted alphabetically
3. Conditional includes
Cleanup Analysis a bit. We can't skip packets just because they are audio. Clean up the state machine transitions a bit to make them a little more readable.
Change logic of PrimeCapture, success MUST return 1. 0 means no error but also no success.
Debugging and braces improvements in local_camera.
