nx_conv2d.nx

nx_conv2d.nx -- multi-channel 2D convolution forward pass. L3 canonical primitive. Closes the diffusion / vision-model gap: every modern image-gen architecture (VAE encoder/decoder, UNet stages, attention pre/post conv blocks, ResNet / EfficientNet classifiers) is built around Conv2D layers. Substrate had nx_winograd_conv (per-tile math) but no multi-channel composer. L4 consumers queued: nx_unet_block, nx_vae_decode, nx_resnet_block. Each composes nx_conv2d in their per-layer flow. ===== Tensor layout ============================================= Input: [N, C_in, H, W] NHWC? No -- NCHW (PyTorch/Caffe convention; canonical for the Q10 substrate). Weight: [C_out, C_in, KH, KW] KH, KW = kernel height/width. v1 hardcodes KH=KW=3. Bias: [C_out] nullable. Output: [N, C_out, H', W'] stride 1 + padding 1 -> H'=H, W'=W. Padding mode: zero-padding by (KH-1)/2 = 1 around input. Same as PyTorch default for "padding=1, stride=1". All tensors in Q10 (substrate convention). Caller is responsible for ensuring intermediate sums don't overflow i64 -- for typical dimensions (C_in <= 1024, KH*KW = 9, Q10 * Q10 = ~1e6) the accumulator stays under 1e10, plenty of i64 headroom. ===== Algorithm ================================================= v1: direct 3x3 convolution. 6 nested loops: n, c_out, h, w, c_in, k. O(N * C_out * H * W * C_in * 9) ops. v2 (queued): Winograd F(2x2, 3x3) per-tile via nx_winograd_conv (already shipped 2026-05-15). 2.25x reduction in multiplies. Plumbed by adding a `winograd` flag dispatching to a different inner loop. Per the bits-up cardinal + bounded-loop discipline.