nx_winograd_conv.nx

nx_winograd_conv.nx -- 2D convolution via Winograd F(2x2, 3x3). Algorithm-led mult-reduction kernel. For every 3x3 conv kernel producing a 2x2 output tile: Direct convolution: 2 * 2 * 3 * 3 = 36 multiplies Winograd F(2x2, 3x3): 4 * 4 = 16 multiplies Reduction: 36 / 16 = 2.25x (Lavin & Gray 2016) More additions (transform matrices) -- but adds are 1-cycle ALU on every CPU including MCUs; mults take 3-5. Net win on every hardware floor we target. Per the min-hardware-floor + algo-led cardinal: this brick beats direct conv on a $5 microcontroller as much as on a 5090, without any SIMD / GPU dependency. When SIMD lands (compiler I3.x) the 16 mults become 4 SIMD ops, compounding the win. Reference output composes via nx_compute_node as NX_CN_K_CONV2D kernel kind. Winograd transform matrices (F(2x2, 3x3), Lavin & Gray 2016): B^T = [[1, 0, -1, 0], [0, 1, 1, 0], [0, -1, 1, 0], [0, 1, 0, -1]] G = [[1, 0, 0], [1/2, 1/2, 1/2], [1/2, -1/2, 1/2], [0, 0, 1]] A^T = [[1, 1, 1, 0], [0, 1, -1, -1]] In Q-format on i64 substrate: * Input tile (4x4) is just i64 values * G filter transform: G * g * G^T -> 4x4 transformed kernel * B input transform: B^T * d * B -> 4x4 transformed input