Canvas на битмапе кадра больше не используется вообще (кроме редкого
ADC-overload пути с восстановлением альфы) — убраны все raw<->canvas
переходы, каждый из которых синкал весь битмап на Qt6 (~2мс):
- RawBegin/RawEnd: прямой доступ к пикселям через RawImage.Data +
BytesPerLine (bottom-up DIB win32 учтён отрицательным шагом)
- RawVLine/RawHLine (с пунктиром), RawFillRect, RawTriangleDown,
RawCurve (кривая = связные вертикальные сегменты по колонкам)
- кэш текстовых масок (TRawTextMask, LRU 48): строка рендерится canvas'ом
один раз белым-на-чёрном в свой мини-битмап, блитится raw с цветом;
все тексты кадра теперь Courier New
- конвертированы: маркер (метка без случайной цветной плашки), бикон-
маркеры, буквы слайсов, слайс-линии, AGC (пунктир 4/2 и 1/2, подпись
с фоном SpecFilter как раньше), края/VFO/TX-курсоры с треугольниками
- пофреймовый альфа-OR проход удалён: альфа чинится один раз в кэше
СЕТКИ после ребилда (canvas-текст там), кадр наследует её через memcpy
Дворд-блендеры (два 16-бит лейна на умножение, /256 вместо /255,
расхождение <0.4%) вместо побайтовых с div 255:
- BandPlanOverlay.CompositeStrip (~1мс -> ~0.2мс)
- SampleRateOverlay.CompositeCache (per-px альфа)
- VfoOverlay.BlendBitmapKey (keyed-магента дворд-сравнением)
- SpectrumView.DrawSpectrumGradient (source-лейны 1 раз на строку,
~1.6мс -> ~0.8мс)
Порог перестройки бэндплана: полпикселя вместо 1 Гц (на QO-100 с
decoder-lock центр подтюнивается непрерывно — был canvas-ребилд полоски
каждый кадр). Композиты внутри raw-лока (их BeginUpdate вложенные).
PerfLog: зона ui.composites отделена от ui.spec_alpha (=RawEnd).
Замер на эфире (Pluto, 576k, 52fps): кадр 8.3мс -> 1.67мс (5x),
со слайсом 2.1-2.3мс; UI-поток ~45% -> ~26% ядра, из них 15% — блиты
paint_spectrum/paint_waterfall (потолок Qt6-вывода). Визуально проверено.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Fix: dragging/wheeling the spectrum in the web UI clamped frequency to a
hardcoded HF ceiling (vfoMax=60 MHz), so on Pluto VHF/UHF it snapped down
to 60 MHz and couldn't pan up (only the VFO wheel worked, via a different
limit). Push the backend's tuning range (BackendCaps.MinFreqHz/MaxFreqHz)
to the web alongside the band plan; JS vfoMin/vfoMax use freq_min/freq_max
from state. The transverter branch (XVTR_CUR>=0) is unchanged — it keeps
the wide-open range since the displayed RF freq is translated to IF and
clamped server-side.
Dead-code cleanup (audit of this session's work):
- SampleRateMode/TSampleRateMode: orphaned when the sample-rate
unification replaced the srmContinuous branch in ApplyBackendCapsToUI;
no readers left. Removed (+ now-unused PLUTO_MAX_SR).
- Pre-existing write-only/unused: TBackendCaps.MaxSampleRate,
SampleRateOverlay.SPAN_NAMES, MainForm.BAND_FREQ.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add ADALM-Pluto / AD9361 support as a second hardware backend alongside
openHPSDR, sharing the WDSP DSP pipeline and the existing controller API
(UI stays decoupled from logic).
- RadioBackend.pas: abstract TRadioBackend + TBackendCaps + TRadioDevice
(Kind/URI/Serial). THPSDRNetwork now derives from it (state via virtual
getters); TRadioController.FNetwork is the base type.
- IIOBindings.pas: dynamic libiio loader (runs without libiio present).
- PlutoBackend.pas: scan/probe-by-URI, connect, LO/rate/bandwidth/gain
control, RX streaming thread (int16->24bit BE -> OnDDCIQ), Q conjugated
to match WDSP IQ convention. Verified on LibreSDR (AD9361) over network.
- Unified discovery: TDiscoverThread scans both backends; network Plutos
found via direct ProbeURI (no mDNS needed). ConnectDevice dispatches by
Dev.Kind (EnsureBackend swaps backend, preserving callbacks).
- DeviceStore/DeviceForm: persist Kind/URI/Serial; save Pluto via
AddSavedPluto so saved/autostart devices reconnect across restarts.
- VHF/UHF band plan (BoardUtils, kind-aware): 6m..ADS-B for Pluto; fixes
HF clamps (band detect, mouse-wheel 60 MHz cap, freq-display max).
- SampleRateOverlay: configurable presets (Pluto 576k..5760k, >520 ksps),
auto-width to fit.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The span-selector panel was rendered directly onto the spectrum canvas
every frame via DrawSelf, doing ~30 TextWidth + ~11 TextOut Canvas calls.
On Cocoa these text ops are very expensive (~10.6 ms/frame, ~30% of a
core) and dominated DrawSpectrum.
Render the panel once into a per-pixel-alpha cache bitmap and only
composite it over the spectrum each frame; rebuild only when state
changes (rate, hover, hidden, bounds). The cache is built with the
two-background method (render over black and white, derive straight
alpha + color) so layered semi-transparent button fills and antialiased
text composite over the live spectrum exactly as before.
Cuts the overlay phase from ~10.6 ms to ~0.2 ms per frame; DrawSpectrum
drops ~13.8 ms -> ~2.7 ms. Mirrors the caching VfoOverlay already uses.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
On macOS Cocoa, TBitmap.pf32bit stores pixels in ARGB order
(byte0=A, byte1=R, byte2=G, byte3=B) rather than BGRA used on
Linux/Windows. All raw ScanLine pixel blending and channel mapping
is now guarded with { DARWIN} to use the correct offsets.
Also fix PortAudio library discovery on macOS: probe .dylib names
(including Homebrew paths) before falling through to .so/.dll.
- Add SampleRateOverlay.pas: TCustomControl overlay (top-left of spectrum)
with hide-panel button and 6 span buttons (48k–1536k), same hit-test
pattern as VfoOverlay
- Consolidate toolbar: move PanelTopVfoGroup and PanelSMeterRight into
PanelToolbar (height 72) instead of floating on the form
- Remove PanelSpanButtons (alTop spacer) that caused the ~1cm black strip
at the top of the spectrum
- Add separator line inside PanelToolbar anchored to its bottom edge,
with 16px left/right margins
- Fix left panel blocks starting Y offset (was 38, now 0)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
ew8bakandClaude Fable 5