nx_upsample.nx

nx_upsample.nx -- 2x upsample for 4D feature maps. Closes the missing-dep gap for VAE decoder + UNet up-blocks. Modern image-gen architectures use 2x nearest-neighbor or 2x bilinear upsample inside the decoder stages: VAE decoder: 4 stages of 2x upsample (8x total = 64x64 latent -> 512x512 image) UNet up-path: 2x upsample at each scale level Super-resolution networks: 2x or 4x upsample L3 canonical primitive composing: NxTensor (L1, 4D NCHW input/output) nx_loop.LoopVerdict (control) -- no other deps; pure pixel-replication / linear-interp math ===== Tensor layout ============================================= Input: [N, C, H, W] NCHW Q10 Output: [N, C, 2H, 2W] 2x upscaled in spatial dims; channels unchanged ===== Filter modes ============================================== NX_UP_NEAREST -- each output pixel = nearest input pixel. Blocky but cheap. Common in VAE decoder when followed by a 3x3 conv that smooths the artifacts. NX_UP_BILINEAR -- 4-tap bilinear from the 4 nearest input pixels. Smoother; slightly more expensive. ===== Pixel-center convention =================================== Output pixel (oy, ox) maps to input pixel at: iy = (oy + 0.5) / 2 - 0.5 ix = (ox + 0.5) / 2 - 0.5 This is the "align_corners=False" convention from PyTorch -- the standard for VAE-class upsampling. In Q10 fixed-point: iy_q10 = (oy + 1) * Q10 / 2 - Q10 / 2 ; if (... ) ... simplifies