slow_conv2d_update_output_frame Class — pytorch Architecture

Source Code

aten/src/ATen/native/ConvolutionMM2d.cpp lines 222–284

template <typename scalar_t>
void slow_conv2d_update_output_frame(
    TensorAccessor<const scalar_t, 3> input,
    TensorAccessor<scalar_t, 3> output,
    TensorAccessor<const scalar_t, 2> weight,
    bool has_bias,
    TensorAccessor<scalar_t, 2> finput,
    int64_t kernel_height,
    int64_t kernel_width,
    int64_t stride_height,
    int64_t stride_width,
    int64_t pad_height,
    int64_t pad_width,
    int64_t n_input_plane,
    int64_t input_height,
    int64_t input_width,
    int64_t n_output_plane,
    int64_t output_height,
    int64_t output_width,
    bool is_channels_last) {
  const int beta = has_bias ? 1 : 0;

  // Compute out = weight * input
  // Note gemm expects fortran order, so all 3 matrices are transposed.
  // Swapping argument order cancels this, since C == AB <=> T(C) == T(B)T(A)
  if (is_channels_last) {
    const int64_t m = n_output_plane;
    const int64_t n = output_height * output_width;
    const int64_t k = n_input_plane * kernel_height * kernel_width;

    const int64_t lda = k;
    const int64_t ldb = k;
    const int64_t ldc = m;

    at::native::cpublas::gemm(
        TransposeType::Transpose,
        TransposeType::NoTranspose,
        m, n, k,
        static_cast<scalar_t>(1),
        weight.data(), lda,
        finput.data(), ldb,
        static_cast<scalar_t>(beta),
        output.data(), ldc);
  } else {
    const int64_t m = output_height * output_width;
    const int64_t n = n_output_plane;
    const int64_t k = n_input_plane * kernel_height * kernel_width;

    const int64_t lda = m;
    const int64_t ldb = k;
    const int64_t ldc = m;

    at::native::cpublas::gemm(
        TransposeType::NoTranspose,
        TransposeType::NoTranspose,
        m, n, k,
        static_cast<scalar_t>(1),
        finput.data(), lda,
        weight.data(), ldb,
        static_cast<scalar_t>(beta),
        output.data(), ldc);
  }
}