| orphan: |
|---|
| vendor: | xross |
|---|---|
| version: | 0.1.0 |
| scope: | Experimental |
| description: | MPEG video decoding for XCORE with selectable backends |
| category: | Vision |
| keywords: | MPEG-1, MPEG-2, video, decoder |
| devices: | xcore.ai |
lib_mpeg is an experimental MPEG video and MP2 audio decoder for XCORE devices.
The default backend is PL_MPEG, decoding MPEG-1 video elementary streams up to 320 x 240, and MP2 audio, through small push-input APIs that use memory supplied by the application. An optional libmpeg2 backend can be selected instead to evaluate MPEG-2 decoding. Only one backend is compiled into a build.
Both backends run on one thread. On XS3 the default backend's inverse DCT, colour converter and optional integer audio synthesis window use the vector unit; the rest is scalar C.
- MPEG-1 video decoding (I, P and B pictures) to progressive YUV420, up to 320 x 240.
- Push-input API: the application owns storage I/O and feeds compressed bytes.
- Storage-independent MPEG-1 program-stream reader for video E0 and MP2 audio C0, using caller-supplied I/O callbacks and a configurable fixed-size cache.
- One automatic arena layout for benchmarks and applications: after the initial
sequence header, reserve decoder state, three pictures and allocation guards,
then use the remaining memory for compressed input.
mpeg_input_capacity()reports the capacity; feeding and pacing remain the caller's responsibility. - MP2 (MPEG-1 Audio Layer II) decoding to interleaved stereo
float(default) or saturating full-scale signed 32-bit PCM, independent of the video decoder and usable on another task or tile. - No heap use: the application passes each decoder its memory, 384 KiB for 320 x 240 video (less for smaller pictures) and 28 KiB for audio.
- YUV 4:2:0 to RGB565 colour conversion for 16-bit displays.
- Bit-exact XS3 VPU colour conversion, with a scalar build option and no extra frame buffer.
- XS3 VPU inverse DCT, about five times faster per block than the original in simulation, with a bit-identical C model for host tests. The original PL_MPEG transform remains selectable.
- Options set in an application
mpeg_conf.hheader: inverse DCT, colour converter, and the MP2 decoder's fused synthesis window and reciprocal output scaling, or integer PCM with scalar/XS3 synthesis selection, and the PS cache size. - Optional libmpeg2 backend for MPEG-2 evaluation, with a 400 KiB guarded arena.
- Standalone XU316 benchmark examples for each backend with decode timing.
- Native reference tests checked against unmodified upstream and FFmpeg output.
- PL_MPEG fork with speed optimisations; output is identical to upstream when the original transform is selected.
- The default video inverse DCT is not bit-exact with upstream PL_MPEG. Select
MPEG_VIDEO_IDCT_PLMPEGwhen bit-exact output is required. - Decoder inputs are elementary streams (
.m1vvideo, MP2 audio). Use the optionalmpeg_ps_reader.hhelper to demultiplex MPEG-1.mpgfiles; MPEG-2 program streams are not supported by that helper. - No display, storage or audio output, and no audio/video synchronisation, is included; those are the application's responsibility.
- Trusted input only: this is not a hardened validator for malformed or truncated streams.
- One video and one audio decoder per tile. Resolution changes, concatenated streams and seeking are unsupported; close and reopen to replay.
- A compressed picture plus decoder lookahead must fit in the automatically sized input buffer (at least 32 KiB). Not every encode/mux layout fits a given arena; increase supplied memory or reduce quality/bitrate if necessary. There is no heap fallback or midstream growth.
- The example stack reservation (16 KiB) is provisional, not a proved worst-case bound.
mpeg.his an experimental API. The optional libmpeg2 backend uses its ownmpeg2.hinterface; the two are not interchangeable.- The portable S32 synthesis path models exact XS3 rounding/symmetric saturation using 64-bit arithmetic; it is not faster than float on XS3. Float remains the library default. S32 clips at full scale whereas float is not clipped.
- Several XS3 kernels here are general enough to belong in
lib_xcore_math, which has no equivalent today. MPEG-specific tables and history management would stay inlib_mpeg:plm_idct_pass_xs3(): one pass of a 16-bit 8 x 8 inverse DCT as a matrix product with a rounding shift.lib_xcore_math's 8 x 8 DCT is 8-bit only.mpeg_rgb565_vpu_block(): YUV 4:2:0 to RGB565 colour conversion.plm_xs3_copy_block(): a masked vector copy of 8 or 16 bytes per row.plm_audio_window_s32_xs3(): a windowed dot product giving 32 outputs from 16 strided taps each; a polyphase FIR.plm_audio_idct36_s32(): the 32-point MP2 transform.dct32_forward()inlib_xcore_mathcomputes the same transform about twice as fast, but its fixed output scaling loses four bits of precision here; it needs a variant with selectable scaling before it can replace this.
- XMOS XTC Tools: 15.3.1
- None.
- None
This is experimental software provided without formal support. Issues can be raised using GitHub issues.