call this function in network */

general avg pooling int8 kernel */

* Convert an uint8 BGR / BGRA image to nchw float blob */

* Convert a nchw float blob to bgr uint8 image * input blob must have 3 channel */

* Convert a nchw float blob to BGRA * input blob must have 3 or 4 channels */

get cmatrix stride */

* Output[i] = max(input0[i], input1[i], input..n[i]) * Broadcast is supported on each dimension of NCHW */

* Output[i] = min(input0[i], input1[i], input..n[i]) * Broadcast is supported on n, c, h, w dims */

* Output[i] = input0[i] - input1[i] - ... - input..n[i] * CPU_SQUARED_DIFFERENCE supports broadcast on all dimensions */

@brief matmul op to MatMul matrix_a_dims and matrix_b_dims see https://pytorch.org/docs/stable/generated/torch.matmul.html?highlight=matmul#torch.matm

ic_mapped is alligned to 4 when pad_c_b % 4 == 0

Channel pad in channel, height and width directions

ic_mapped is not alligned to 4 when pad_c_b % 4 != 0

Common pad in height and width directions