Mimir

Realtime cameras, microphones, chirplets, and field evidence organized into one coherent stream machine.

Low-Level Implementation Notes

3 min read

Low-Level Implementation Notes

Purpose

This file collects the implementation details that are easy to lose when the conversation stays at the architecture level. None of these notes are promises; they are candidates to benchmark.

Audio Memory Layout

Preferred Analysis Layout

For decoder kernels:

channel0 samples: [s0 s1 s2 ...]
channel1 samples: [s0 s1 s2 ...]
...

Separate channel spans make per-channel scoring straightforward and avoid striding through interleaved audio during analysis. Interleaving is useful for device I/O and some DSP kernels; it is not automatically the right analysis shape.

Preferred Scoring Layout

candidateStartSamples[]
candidateEnergy[]
symbolReliability[]
binFrequencies[]
binSin[]
binCos[]

Dense arrays let SIMD/native/GPU kernels process candidates without object chasing. The codebook trellis can stay managed/CPU because it is small and branchy.

SIMD Chirp-Bin Scoring

Likely CPU kernel shape:

  1. load candidate window samples in 8-float AVX2 chunks;
  2. multiply by dechirp oscillator/window;
  3. accumulate bin real/imag sums;
  4. compute magnitude or normalized energy;
  5. write top K or full dense bin vector.

Important:

  • keep oscillator tables aligned;
  • keep candidate windows contiguous;
  • batch multiple candidates per interop call;
  • avoid denormals;
  • normalize outside the deepest loop where possible.

FFT / Goertzel Decision

Use Goertzel/fixed-bin DFT when:

  • symbol count is small;
  • bins are known;
  • candidate count is modest;
  • native dispatch overhead matters.

Use FFT/CZT when:

  • many bins are needed;
  • frequency offset interpolation dominates;
  • candidate batches are large enough;
  • GPU/native FFT plans can be reused.

Do not choose FFT because it sounds more canonical. Canonical is not a profiler.

GPU Chirp Scoring

The GPU scoring pass only makes sense if data is already resident or batches are large enough to pay dispatch overhead.

Good GPU candidates:

  • many mic streams;
  • many remote receiver windows;
  • wide candidate sets during acquisition;
  • offline/artifact calibration sweeps.

Bad GPU candidates:

  • one or two channels;
  • tiny candidate counts;
  • paths that require CPU readback before codebook solving every event.

ASIO Capture Hot Path

Current observed risks:

  • per-block std::vector<float> resize;
  • std::deque queue;
  • native-to-managed polling copies into float[];
  • managed byte array allocation for sample payload;
  • ConcurrentQueue.Count pressure.

Preferred cut:

  • native SPSC ring;
  • fixed block pool;
  • monotonic frame counters;
  • block descriptors with handle/index;
  • C# metadata only until DSP needs samples.

Camera Capture Hot Path

Current probe strengths:

  • KS async queueing is already close to the driver;
  • PS3 Eye raw path proves high fps;
  • JSON frame events make cadence visible.

Current probe risks:

  • JSON must not become production payload;
  • CPU frame buffers are diagnostic until Fensalir import/upload path exists;
  • one-off command-line knobs must become configuration/stateful driver policy.

Preferred cut:

  • one native capture worker per camera family;
  • fixed frame ring;
  • descriptor ABI;
  • optional shared GPU texture handle;
  • Fensalir owns texture import and GPU staging.

Fensalir GPU Fusion

Fensalir contracts already distinguish:

  • GpuSensorFrame for calibrated camera inputs and external textures;
  • AcousticFieldFrame for timing/room constraints;
  • CalibrationEventFrame for high-confidence event anchors;
  • GpuFusionField for point buffers/seeds;
  • TemporalGaussianField for stable gaussian observations.

Mimir should produce these rows. It should not pack shader structs directly.

Gaussian Splat Rendering Levers

From modern 3DGS implementations:

  • tile size matters;
  • sorting strategy matters;
  • cull before sort;
  • packed buffers matter;
  • anti-aliasing and opacity/radius thresholds matter;
  • view-consistency fixes cost real time;
  • depth/feature render modes are useful for diagnostics.

Mimir-specific cut:

  • first update live claims;
  • then render simple claims;
  • only later optimize dynamic radiance fields.

Benchmark Counters To Add

Decoder:

  • samples consumed;
  • proposals created;
  • candidates scored;
  • anchors accepted/rejected;
  • elapsed scoring time;
  • allocations per analysis tick.

ASIO:

  • callback count;
  • block count;
  • queue/ring high-water mark;
  • dropped blocks;
  • conversion time if measurable.

Camera:

  • frame interval p50/p95/p99;
  • async queue depth;
  • dropped/late frames;
  • payload copy time;
  • upload/import time.

Fensalir:

  • contract lowering time;
  • upload time;
  • GPU pass timings;
  • temporal field count;
  • acoustic constraint count.

Backlinks