2D convolution over a (C, H, W) chunk.

Torch-layout convolution for in-graph model inference: x is a channels-first (C_in, H, W) array, w a (C_out, C_in/groups, kH, kW) kernel, bias an optional length-C_out vector. stride, padding (symmetric), and dilation are scalars or length-2 (y, x). groups gives grouped/depthwise convolution. Weights and bias are plain R arrays; traced they enter the kernel as constants uploaded once at compile.

Usage

g_conv2d(
  x,
  w,
  bias = NULL,
  stride = 1L,
  padding = 0L,
  dilation = 1L,
  groups = 1L
)

Arguments

x

(C_in, H, W) array (traced or plain).

w

(C_out, C_in/groups, kH, kW) numeric array.

bias

Optional length-C_out numeric.

stride, padding, dilation

Scalar or length-2 integers.

groups

Feature group count.

Value

(C_out, H_out, W_out) array.

Details

The plain-R branch is an im2col matmul: correct at any size, meant for tests, not throughput.