}, {
  "tf.Yield"() : () -> ()
  }) { is_stateless = true } : (tensor<i1>) -> ()
```
### `-tf-move-transposes`

_Move transposes pass._


#### Options
```
-fold-transpose-in-ops : Whether to fold transposes in ops which can support folding.
-direction             : Move transposes to the beginning or the end of the block where they are defined.
```
### `-tf-name-anonymous-iterators`

  // CHECK: %[[TRANSPOSE:.*]] = "tf.Transpose"(%[[DEQUANT]], %[[CST]]) : (tensor<3x4xf32>, tensor<?xi32>) -> tensor<*xf32>
  // CHECK: %[[MATMUL:.*]] = "tf.MatMul"(%arg0, %[[TRANSPOSE]]) <{grad_a = false, grad_b = false, transpose_a = false, transpose_b = true}> : (tensor<2x3xf32>, tensor<*xf32>) -> tensor<2x4xf32>
  // CHECK: return %[[MATMUL]] : tensor<2x4xf32>
}

  // CHECK: %[[MATMUL_4:.*]] = "tf.MatMul"(%[[LHS_4]], %[[RHS_4]]) <{grad_a = false, grad_b = false, transpose_a = false, transpose_b = false}> : (tensor<4x5xf32>, tensor<5x6xf32>) -> tensor<4x6xf32>

    return %0 : tensor<2x2xf32>
  }
  func.func private @composite_matmul_fn_2_0(%arg0: tensor<2x2xf32>, %arg1: tensor<2x2xf32>) -> tensor<2x2xf32> attributes {tf_quant.composite_function} {

                        std::vector<Output>* grad_outputs) {
  if (is_batch == false) {
    auto dx =
        MatMul(scope, x0, x1, MatMul::TransposeA(adj_x0).TransposeB(adj_x1));
    grad_outputs->push_back(dx);
    auto dy =
        MatMul(scope, y0, y1, MatMul::TransposeA(adj_y0).TransposeB(adj_y1));
    grad_outputs->push_back(dy);
  } else {
    auto dx = BatchMatMulV3(scope, x0, x1, x_data_type,

  [(HasRank<2> $input),
   (AreLastTwoDimsTransposed $perm_value),
   (IsBoolAttrEqual<"false"> $adj_y)]>;

// Replace conv-->transpose-->add with conv-->add-->transpose
// The bias needs only reshape (i.e. ReshapeNCHWBiasToNHWC) and not transpose
// because the bias's shape simply changes from NxCx1x1 to Nx1x1xC.
def ReorderNCHWTransposeAdd : Pat <
  (TFL_AddOp:$add_op

    rewriter.replaceOpWithNewOp<ConcatV2Op>(op, op.getType(), expanded_inputs,
                                            axis_value);
    return success();
  }
};

// Lowers SpaceToBatchND by reducing to reshape(transpose(reshape(pad(input)))).
//
// Before rewrite:
//   output = SpaceToBatchND(input, block_shape, paddings)
// Let:
//   [batch] + spatial_shape + remaining_shape = input.shape
//   M = spatial_shape.rank

    %8 = "tf.Cast"(%7) {Truncate = false, device = ""} : (tensor<1024x3xi8>) -> tensor<1024x3xi32>
    %9 = "tf.MatMul"(%6, %8) {device = "", transpose_a = false, transpose_b = false} : (tensor<1x1024xi32>, tensor<1024x3xi32>) -> tensor<1x3xi32>
    %10 = "tf.Cast"(%9) {Truncate = false, device = ""} : (tensor<1x3xi32>) -> tensor<1x3xf32>

                       TF_Operation* r, const char* name,
                       bool transpose_a = false, bool transpose_b = false) {
    TF_OperationDescription* desc = TF_NewOperation(graph, "MatMul", name);
    if (transpose_a) {
      TF_SetAttrBool(desc, "transpose_a", 1);
    }
    if (transpose_b) {
      TF_SetAttrBool(desc, "transpose_b", 1);
    }
    TF_AddInput(desc, {l, 0});
    TF_AddInput(desc, {r, 0});

  TFL::populateWithGenerated(patterns);
  // TODO(fengliuai): Implement similar rule in the QuantizePass if the constant
  // folding hook of tfl.transpose and tfl.reshape are implemented.
  patterns.add<ReorderFakeQuantPattern<TF::ReshapeOp>,
               ReorderFakeQuantPattern<TF::TransposeOp>>(ctx);
  // Remove redundant reshape ops.
  TF::ReshapeOp::getCanonicalizationPatterns(patterns, ctx);