Pat<(TF_MatrixSetDiagV2Op $input, $diagonal,
            (Arith_ConstantOp ConstantAttr<I32ElementsAttr, "{0}">)),
          (TF_MatrixSetDiagOp $input, $diagonal)>;

// `align` attribute can be ignored because we only support converting
// `MatrixSetDiagV3` to `MatrixSetDiag` with default `k` inputs.
def ConvertMatrixSetDiagV3ToMatrixSetDiag :
      Pat<(TF_MatrixSetDiagV3Op $input, $diagonal,

		var buf bytes.Buffer
		if err := Encode(&buf, src, nil); err != nil {
			t.Errorf("gray-diagonal: Encode: %v", err)
			return
		}
		m, err := Decode(&buf)
		if err != nil {
			t.Errorf("gray-diagonal: Decode: %v", err)
			return
		}
		if got, want := m.Bounds(), image.Rect(0, 0, 6, 6); got != want {
			t.Errorf("gray-diagonal: got %v, want %v", got, want)
			return
		}

const blockSize = 64

// quarterRound is the core of ChaCha20. It shuffles the bits of 4 state words.
// It's executed 4 times for each of the 20 ChaCha20 rounds, operating on all 16
// words each round, in columnar or diagonal groups of 4 at a time.
func quarterRound(a, b, c, d uint32) (uint32, uint32, uint32, uint32) {
	a += b
	d ^= a
	d = bits.RotateLeft32(d, 16)
	c += d
	b ^= c
	b = bits.RotateLeft32(b, 12)
	a += b

  let summary = "Returns a diagonal tensor with a given diagonal values.";

  let description = [{
Given a `diagonal`, this operation returns a tensor with the `diagonal` and
everything else padded with zeros. The diagonal is computed as follows:

Assume `diagonal` has dimensions [D1,..., Dk], then the output is a tensor of

  (TFL_NonMaxSuppressionV5Op $boxes, $scores, $max_output_size, $iou_threshold,
    $score_threshold, $soft_nms_sigma)>;

def LegalizeMatrixDiag : Pat<(TF_MatrixDiagOp $diagonal),
                             (TFL_MatrixDiagOp $diagonal)>;

def LegalizeConv2DBackpropInput : Pat<
  (TF_Conv2DBackpropInputOp $input_sizes, $filter, $out_backprop,
     IsIntList1XY1:$strides,
     BoolAttr:$use_cudnn_on_gpu,

  // LSTM `func::FuncOps` with indy behavior always have the `tf.api_implements`
  // function attribute prefixed with `"indy_lstm_"`.
  // IndyLSTMs have diagonal recurrent weight matrices and can benefit from
  // more efficent operations in TFLite with the correct conversion (i.e. when
  // the diagonal recurrent weight matrices are provided as vectors).
  if (attr.getValue().starts_with("indy_lstm_")) {

  //   VJP are einsums with the equations "ji->ij" and "i->ii" respectively,
  //   where the latter represents 'un-tracing', or filling the diagonal with
  //   the input axis and non-diagonal entries are zeros.
  //        Furthermore, recall that matrix multiplication, which is
  //   represented by the equation "ab,bc->ac", has its VJPs given by the

  }

  /**
   * Asserts that {@code transformation} is diagonal (i.e. neither horizontal nor vertical) and
   * passes through both {@code (x1, y1)} and {@code (x1 + xDelta, y1 + yDelta)}. Includes
   * assertions about all the public instance methods of {@link LinearTransformation} (on both
   * {@code transformation} and its inverse). Since the transformation is expected to be diagonal,
   * neither {@code xDelta} nor {@code yDelta} may be zero.

    TFL_TCresVTEtIsSameAsOp<0, 0>>]> {
  let summary = [{
    Returns a tensor with the provided diagonal and everything else padded with zeros.
  }];

  let description = [{
    Given a diagonal, returns a tensor with the diagonal and everything else padded with zeros.
    Assume diagonal has k dimensions `[I, J, K, ..., N]`, then the output is a tensor of rank `k+1`
    with dimensions `[I, J, K, ..., N, N]` where:

    // We extract the diagonals from k[0] up to and including k[1].
    // Addressing is 0 for the main diagonal. (So k = [0, 0] would just extract
    // the main diagonal). It's negative for subdiagonals (under and to the left
    // of the main diagonal) and positive for superdiagonals (above and to the
    // right of the main diagonal).
    int64_t k[2];