QConv1dInt8 Class — pytorch Architecture

Source Code

aten/src/ATen/native/quantized/cudnn/Conv.cpp lines 345–365

class QConv1dInt8 final {
 public:
  static Tensor run(
      Tensor act,
      const c10::intrusive_ptr<ConvPackedParamsBase<2>>& packed_weight,
      double output_scale,
      int64_t output_zero_point) {
    at::Tensor output;
    // we currently use conv2d kernel for conv1d by making the input and weight tensors
    // 4D rather than 3D. we add a dummy width dimension of size 1
    // N, C, L -> N, C, 1, L
    act = act.unsqueeze(-2);
    if (kReluFused) {
      output = packed_weight->apply_relu(act, output_scale, output_zero_point);
    } else {
      output = packed_weight->apply(act, output_scale, output_zero_point);
    }
    // N, C, 1, L -> N, C, L
    return output.squeeze_(-2);
  }
};