slow_conv2d_backward_weight_frame Class — pytorch Architecture

Source Code

aten/src/ATen/native/ConvolutionMM2d.cpp lines 435–481

template <typename scalar_t>
void slow_conv2d_backward_weight_frame(
    TensorAccessor<scalar_t, 2> grad_weight,
    TensorAccessor<const scalar_t, 3> grad_output,
    TensorAccessor<const scalar_t, 2> finput,
    bool is_channels_last) {
  // Compute grad_weight += grad_output.reshape({grad_output.shape(0), -1}) * finput.T
  // 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 = finput.size(1);
    const int64_t n = grad_output.size(0);
    const int64_t k = grad_output.size(1) * grad_output.size(2);

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

    at::native::cpublas::gemm(
        TransposeType::NoTranspose,
        TransposeType::Transpose,
        m, n, k,
        static_cast<scalar_t>(1),
        finput.data(), lda,
        grad_output.data(), ldb,
        static_cast<scalar_t>(1),
        grad_weight.data(), ldc);
  } else {
    const int64_t m = finput.size(0);
    const int64_t n = grad_output.size(0);
    const int64_t k = grad_output.size(1) * grad_output.size(2);

    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),
        finput.data(), lda,
        grad_output.data(), ldb,
        static_cast<scalar_t>(1),
        grad_weight.data(), ldc);
  }
}