v1.5.0

Added

Native C/C++ performance extensions (homrec_core, hr_encoder_helpers,
hr_preview, hr_ringbuf, hr_framequeue, hr_stopwatch, hr_display_info,
hr_dxgi_capture, hr_pipeline) — compiled as .dll/.so, loaded at runtime
via ctypes. Auto-discovered next to homrec.py — no config needed.
Graceful fallback to pure Python if any module is missing.

hr_pipeline: unified C++ capture thread — DXGI Desktop Duplication →
BGRA→YUV420p → named pipe (FFmpeg reads) + preview thumbnail.
Eliminates Python-side mss.grab() loop entirely on Windows.

hr_stopwatch: sub-millisecond frame-pacing timer using
QueryPerformanceCounter (Windows) / clock_gettime CLOCK_MONOTONIC (POSIX).
Fixes the 15 ms sleep granularity jitter on Windows that caused frame drops.

hr_dxgi_capture: GPU-accelerated screen capture via DXGI Desktop
Duplication API. Captures frames the moment the compositor presents them
(~1 display cycle latency vs 1-2 frame lag with GDI/BitBlt).

hr_display_info: fast monitor enumeration with per-monitor DPI via
GetDpiForMonitor (Windows) / xrandr (Linux).

homrec_native.py: Python ctypes wrapper for all native libraries with
full Python fallbacks for every function.

"Disable Preview" option in Settings → Advanced → Performance.
Shows a placeholder screen instead of live preview — saves CPU during
recording on low-end hardware.

"Update" button appears in corner when new version is detected
Welcome window shown on first launch

Changed

Capture loop: preview thread now skips mss.grab() entirely during
recording — no GDI bandwidth competition with FFmpeg.

Codec pipeline: QP range clamped (min qp=23) to prevent near-lossless
encoding (~150 Mbit/s) that saturated weak GPU encoders.

QSV encoder: removed -async_depth 1 which blocked the capture thread
waiting for GPU acknowledgement on each frame, halving FPS on i3/i5.

QSV encoder on UHD Graphics: added -low_power 1 (VDENC fixed-function
path, 2-3× faster than shader PAK) and -look_ahead 0 (removes CPU-side
lookahead overhead).

CPU encoder (libx264): thread count now set to 1 on systems with ≤4
logical cores, avoiding competition with FFmpeg for CPU time.

GIL switch interval raised from 5 ms to 20 ms — reduces GIL contention
between capture, audio and UI threads on weak CPUs. One GIL switch per
frame at 60 FPS instead of three.

OpenCV thread count set to 1 on systems with ≤4 logical cores — preview
resize does not need multiple threads.

gdigrab input: thread_queue_size reduced from 512 to 128, added
-rtbufsize 16M to limit memory pressure on systems with shared CPU+GPU RAM.

Preview update interval during recording raised from 1000 ms to 2000 ms —
capture thread sleeps 500 ms anyway so polling faster was pointless.

Audio recording stop: non-blocking teardown — streams closed before
joining threads so read() unblocks immediately. Eliminates the previous
30-second hang on stop.

Fixed

hr_audio_rms: int32_t accumulator replaced with int64_t — prevented
integer overflow on buffers larger than ~2M samples.

hr_blend_rgba: alpha compositing rounding bias corrected from +127 to
+128, fixing a systematic 1-LSB error in badge/watermark blending.

hr_framequeue: push() race condition — slot is now written before head
is advanced so the consumer can never observe a null pointer.

hr_framequeue: size() used mismatched memory orders (acquire/relaxed),
now both acquire for correct cross-thread visibility.

hr_ringbuf: wrap-around memcpy off-by-one at exact power-of-2 boundaries
fixed by computing first from mask(h) rather than raw cursor.

hr_ringbuf: reset() now zeroes buffer contents — prevents stale PCM data
from a previous recording being read after reset.

hr_display_info: missing #include caused std::stable_sort
to be invisible under MinGW, producing a build error.

hr_encoder_helpers: hr_memcpy_nt now uses _mm_stream_si128 (real NT
stores) instead of _mm_storeu_si128 (cached stores) — the function
previously claimed to be non-temporal but was not.

hr_encoder_helpers: BGRA→YUV420p and RGB→YUV420p chroma now averages
all four luma pixels in each 2×2 block instead of sampling only the
top-left pixel — visibly reduces chroma blur on fine detail.

hr_preview: restrict keyword replaced with HR_RESTRICT macro — restrict
is C99-only and caused a cascade of "not declared in this scope" errors
when compiled as C++ under MinGW g++.

homrec_native.py: duplicate _DxgiCaptureAPI singleton at module level
removed — dxcap was created twice, second instance silently overwrote
the first.

homrec_native.py: preview thumbnail now uses ctypes data_as() instead
of .tobytes() — eliminates a full buffer copy on every preview frame.

Quality loaded from old settings.json is now clamped to 50–100% to
prevent a saved value of 95–100 (quality=100 → qp=18 → ~150 Mbit/s)
from saturating the encoder on first launch after upgrade.

Application freeze after stopping recording
General optimization improvements
Tray icon logo visibility improved on dark/light taskbars

Performance

BGRX→RGB (1080p): ~6 ms → ~1.2 ms (5× faster)
Bilinear preview resize: ~4 ms → ~0.8 ms (5× faster)
Audio RMS per chunk: ~0.15 ms → ~0.04 ms (3.5× faster)
Badge blend (RGBA): ~1.2 ms → ~0.25 ms (5× faster)
Capture loop latency: ~30 ms → ~12 ms (2.5× faster)
UI thread block time: ~20 ms → ~3 ms (7× faster)
Preview thumbnail: no intermediate RGB copy (BGRA→thumb direct)