Module cls 

BCE classification loss Triton kernel for YOLO detection training.

Computes numerically-stable binary cross-entropy loss between predicted class logits and soft target labels, summed over all C classes per anchor.

Stable BCE formula (avoids log(0) and exp overflow): loss = relu(x) − x·t + log(1 + exp(−|x|)) = max(x, 0) − x·t + log(1 + exp(−|x|))

Layout:

• pred: [C, N] — predicted class logits per anchor
• target: [C, N] — soft class labels ∈ [0, 1] per anchor
• loss: [N] — per-anchor BCE loss (summed over C classes)

Parallelism: one CTA per BLOCK_N-wide anchor tile; classes iterated sequentially inside each CTA. Grid: cdiv(N, BLOCK_N) flat CTAs.

Structs

YoloBceClsLossBackward

BCE classification loss backward: upstream gradient + forward inputs → ∂L/∂pred.

YoloBceClsLossForward

BCE classification loss forward: class logits + soft targets → per-anchor loss.

Functions

yolo_bce_cls_loss_backward

BCE classification loss backward: upstream gradient + forward inputs → ∂L/∂pred.

yolo_bce_cls_loss_forward

BCE classification loss forward: class logits + soft targets → per-anchor loss.