VideoReader:
- Guard get_source_info/allocate_conversion_buffers behind have_video
check — prevents crash on audio-only files
- Remove unused audiobuf parameter from get_next_frame
- Remove dead return statements after throw
- Remove unused SourceInfo::stride field and AudioInfo struct
- Pass const std::string& instead of by value in constructor/open_file
- Remove redundant struct keyword on SourceInfo source_info{}
- Fix video_stream_index never guarded against -1 in open_file
- Check av_image_alloc and avcodec_parameters_to_context return values
- Remove extra av_packet_unref after seek (was harmless but confusing)
- Use SWS_BILINEAR for sws_getContext flags instead of 0
- Replace NULL with nullptr in sws_getContext
- Remove unused #include <libavutil/pixdesc.h>
videoin main.cpp:
- Early return if have_video is false after open
- Inline make_video_flow_def call (remove intermediate variable)
- Add grain count log line matching other nodes
- Change continue to break when get_next_frame returns false
- Make video_rate const
- Remove double blank line before main()
- Remove trailing spaces
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Pass MXL grain stride (from sliceSizes[0]) into get_next_frame so
YUV422P10toV210 writes with the correct line width instead of a
self-computed value that may not match the MXL buffer.
Fix audio stream detection: condition was inverted (!= -1 → == -1),
so the first audio stream was never picked up.
Add return false at end of get_next_frame to fix UB when g_running
goes false and the loop exits without returning.
Replace av_frame_unref/av_packet_unref with av_frame_free/av_packet_free
in destructor — unref only releases data, not the struct itself.
Add av_freep(&p10_data[0]) in destructor to free av_image_alloc memory.
Remove unused p10_buffer and v210_buffer allocations.
Read filename from config("file") instead of hardcoded path.
Add early return if mxlCreateFlowWriter fails.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Edges that share the same source node and port now get the same MXL flow
UUID. One writer, multiple independent readers — each tracks its own
position in the ring buffer. Enables 1-to-N routing in graph.json without
any splitter node.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
studio-manager now accepts an optional path argument:
dmf-studio-manager graph.json
load_graph() parses nodes and edges from JSON, generating fresh UUIDs
for each MXL flow on every launch. Falls back to hardcoded build_graph()
when no argument is given.
graph.json at repo root defines the current testpattern→ndiout pipeline.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Readers (ndiout, fakesink) now handle MXL_ERR_FLOW_INVALID by releasing
and recreating the flow reader, then realigning the index to current time.
This lets consumer nodes survive a producer restart without exiting.
All mxlCreateInstance/FlowWriter/FlowReader options args changed from
nullptr to "" to match MXL reference implementation style.
mxlReleaseFlowReader calls guarded with null checks so cleanup is safe
when a mid-run reconnect attempt fails.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
studio-manager: call mxlGarbageCollectFlows before launching nodes to
clean up stale flow directories left by previous crashed runs.
shared/NodeBase.hpp: add mxl_status_str(mxlStatus) — converts error
codes to readable names (e.g. MXL_ERR_OUT_OF_RANGE_TOO_LATE). All
nodes now log these names instead of raw integers.
shared/FlowDef.hpp: use "audio/float32" as media_type for 32-bit audio
flows, matching the MXL SDK examples. audio/L{n} kept for other depths.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
flow_id is now optional (config().contains). When absent: skip the
MXL reader, active-wait, and NDI video path entirely. Video buffers
and NDI video frames are heap-allocated only when has_video.
Loop pacing: video branch sleeps 1ms on TOO_EARLY as before; audio-only
path sleeps 1ms when no chunk was available (audio_advanced == false)
to avoid busy-spinning.
NDI sender name changed from the video flow UUID to node_id(), which
is stable and human-readable for both video and audio-only modes.
To use audio-only: wire only audio_flow_id in build_graph, omit flow_id.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
sample_rate / fps_num gave 48000/30000 = 1 for 29.97 fps, producing
silence. Correct formula is sample_rate * fps_den / fps_num, matching
the calculation already used in testpattern.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
ndiout: p216_frame.FourCC was zero-initialized — NDI needs it explicitly
set to NDIlib_FourCC_video_type_P216 to send the correct format.
All three nodes: audio_rate = {sample_rate, 1} was always declared even
when has_audio is false (sample_rate = 0 in ndiout). Scoped into the
has_audio block so the bad rate can never be passed to mxlGetCurrentIndex.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Fix: c * 1920 → c * samples_per_frame (broke non-25fps or non-48kHz).
Fix: audio channel loop now iterates `channels` not `audio_slices.count`.
Rename: mxl_stride → video_stride, no_samples → samples_per_frame,
ndi_frame_10bit/16bit → v210_frame/p216_frame, ndi_audio_frame → ndi_audio.
Scope `ast` locally to its use block.
Drop unused #include "V210.hpp" and #include "FlowDef.hpp".
ndi_audio struct zero-initialized then filled only when has_audio.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
testpattern now optionally writes an audio flow when audio_flow_id is
wired in the graph. Each channel generates a sine wave at (c+1)*1000 Hz
(ch0=1kHz, ch1=2kHz, ...) at -18 dBFS broadcast reference level.
Phase is derived from the absolute audio_index so there are no clicks
at frame boundaries. One audio chunk per video frame (1920 samples at
48kHz/25fps), written via mxlFlowWriterOpenSamples with ring-buffer
fragment and channel stride handling matching ndiin.
build_graph: separate UUIDs for testpattern and ndiin flows; wire
testpattern audio_flow_id (no sink yet).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
V210.hpp: add fill_colorbars, fill_ire_ramp, fill_black, fill_white.
Shared fill_solid helper stamps the first line across all rows via
memcpy. Generic write_palette_line template handles both SMPTE and
IRE bar palettes. fill_frame kept as a backward-compat alias.
testpattern/main.cpp: read "pattern" from node config and dispatch
to the appropriate fill function (bars/ire/black/white).
build_graph: pass {"pattern","bars"} to testpattern node params.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
NDIlib_audio_frame_v3_t carries a FourCC. Only FLTP (float32 planar)
is handled — skip and free any other format rather than misinterpreting
compressed or integer audio bytes as floats.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
NDIlib_audio_frame_v3_t.p_data is uint8_t*, not float*. Calling
vector<float>::assign(uint8_t*, uint8_t*+N) converted each individual
byte (0-255) into a separate float instead of reinterpreting 4 bytes
as one float sample — producing complete noise on playback.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Audio flows in MXL use mxlFlowWriterOpenSamples/CommitSamples, not
the grain API (mxlFlowWriterOpenGrain is for discrete video flows).
Using the grain API silently did nothing — hence "last published index: 0".
MXL audio is float32 planar: each channel occupies its own ring buffer
region accessed at base.fragments[*].pointer + c * slice.stride.
Two fragments handle ring buffer wraparound. No int32 conversion needed.
Also log continuous.channelCount/bufferLength instead of the discrete
sliceSizes field (wrong union member for audio).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
MXL reads channel_count (plain integer) to determine grain buffer
geometry, defaulting to 1 if absent. The NMOS IS-04 channels array
is ignored by MXL. Without channel_count, stereo audio grains were
half the required size.
Also: MXL only accepts bit_depth 32 or 64 for audio flows.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
MXL FlowParser requires urn:x-nmos:tag:grouphint/v1.0 in the tags
object of every flow definition. make_audio_flow_def was missing tags
entirely, causing FlowParser.cpp:192 "Invalid group hint tag".
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- FlowDef.hpp: add make_audio_flow_def (NMOS IS-04 audio/Lnn flow)
- NDIReceiver: replace capture_v210 with capture() — single recv call
dispatches to FrameKind::{Video,Audio,None}; AudioInfo struct carries
sample_rate/channels/samples/channel_stride
- ndiin: optional audio writer when graph wires audio_flow_id; single
loop with 5ms NDI poll; audio grains written immediately on arrival;
video grains written when mxlGetCurrentIndex reaches video_index
- build_graph: rename ndiin video port flow_id → video_flow_id; add
audio_flow_id edge (no sink yet — wired to ndiin only)
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
On status_change, probe() the new format and compare against the
format used to create the MXL flow. If resolution or fps changed,
throw with a descriptive message so the caller exits cleanly.
The existing catch in NDIInNode::run() logs the reason and breaks
out of the loop, causing the process to exit. Studio-manager detects
the exit and logs it — operator can restart to pick up the new format.
Same-format status changes (e.g. metadata only) still return false
and repeat the last frame as before.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Share a single UUID across both edges so both readers connect to the
same MXL flow. MXL supports multiple readers per flow natively — each
reader has its own read pointer into the ring buffer.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
- NDIContext struct handles NDIlib_initialize/send_create/send_destroy/destroy
so NDIlib_destroy() is guaranteed even if send_create fails (was leaked before)
- malloc/free for 10-bit and 16-bit frame buffers → std::vector<uint8_t>
- NDI frame structs point into vector data, no manual lifetime management
- static_cast for FourCC instead of C-style cast
- Tidy: ndi_frame_count replaces ndi_frame_counter, ++prefix form
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
On Linux, spdlog and fmt are compiled as static libs then linked into
libmxl.so. Without -fPIC the linker rejects TPOFF32 relocations inside
a shared object. Setting PIC globally fixes this for all fetched deps.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
UUID_SYSTEM_GENERATOR=ON requires libuuid on Linux and CoreFoundation
on macOS. Turning it off uses stduuid's built-in mt19937 generator,
which needs no system packages and works identically on all platforms.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Removes the Homebrew dependency so the build works on any platform
with only CMake 3.24+ and a C++20 compiler. fmt and spdlog are now
fetched and compiled from source alongside the other FetchContent deps.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
itten
itten