── SAM3 derived helpers ────────────────────────────────────────────

Save a tensor as raw float32 binary + shape file.

Save tensor as .bin + .shape files.

Generates sinusoidal PE matching PositionEmbeddingSine from Python. num_pos_feats = d_model / 2 = 128, temperature = 10000, normalize = true, scale =

Save tensor as .bin + .shape (PyTorch native layout).

Save [S, B, D] (seq-first) as ggml [D, S, B].

Fuse a standard Conv + BN pair. Returns (weight, bias) as numpy arrays.

Bilinear interpolation of a flat mask [H_in * W_in] to [H_out * W_out]. Uses double for coordinate math to match PyTorch F.interpolate precision.

Test-only: encode from pre-preprocessed float data (bypasses C++ resize/normalize). chw_data is in [C, H, W] layout, already normalized, with C=3, H=W

Read tensor data from backend into a float buffer, handling F32, F16, and quantized types. n = number of float elements to produce.

fast 0..255 * 0..255 => 0..255 rounded multiplication

Conv2d bias addition helper: adds bias [ch] to conv output [W, H, ch, B]

── EdgeTAM derived helpers ────────────────────────────────────────

── SAM2 derived helpers ────────────────────────────────────────────

if there's a pending marker from the entropy stream, return that otherwise, fetch from the stream and get a marker. if there's no marker, return 0xff,

Compare raw flat data (same byte layout between Python and C++)

Run win_part only (no roundtrip) — compare intermediate output.

Run a graph: build win_part -> win_unpart on the given backend. Returns the final output as a flat float vector.

Apply RoPE to Q and K tensors using complex multiplication. x shape: [head_dim, N, num_heads*B] in ggml layout freqs_cis shape: [2, 32, N] in ggml lay

take a -128..127 value and stbi__clamp it and convert to 0..255

returns 1 if "a*b*c + add" has no negative terms/factors and doesn't overflow

Compare two float arrays with tolerance. Returns max absolute difference.

Feature size for the active backbone

Extract object pointer from SAM output token via 3-layer MLP (CPU-side).

initialize a callback-based context

initialize a memory-decode context

1×1 convolution as mul_mat: avoids im2col+cont overhead. Input x: [W, H, Cin, 1], weight w: [1, 1, Cin, Cout], bias b: [Cout] Returns: [W, H, Cout, 1]