memory_model

Runtime model of render-arena peak storage per frame chunk.

The model observes ManualMemory.max_pointer and fits a conservative affine estimate, peak(n) = a + b*n, separately for each render signature. Recent observations, safety margins and growth limits guide the next chunk size. When an affine fit is unsuitable it falls back to a per-frame bound.

This measures allocations made through the arena, including new rendering and post-processing code that uses it. It does not measure every PyTorch allocation, compiler workspace or driver reservation. The surrounding render loop handles those constraints and retries chunks that exhaust the arena. Later, denser frames can exceed a fit derived from earlier frames; the estimate is not a proof that an entire animation will fit.

Classes

AffineFrameCost	What one prepared batch costs, split into its actor and frame parts.
ChunkMemoryModel	Affine fit of arena peak against chunk frame count.
PeakRatioModel	Measured bound on an out-of-arena transient build's peak.

Functions

chunk_signature(*, width, height, channels, dtype, samples_per_pixel, num_triangles, num_circuits)

Key identifying batches whose peak lies on the same line.

Resolution and buffer dtype change the per-frame cost; the primitive counts change it too, and a batch with different geometry is a different line. Geometry counts are bucketed logarithmically so ordinary scene variation does not discard a usable fit, while an order-of-magnitude change starts a fresh one.