// CHECK-NEXT: [[CONV:%.*]] = "tf.Conv2D"([[INPUT]], [[FILTER]]) <{dilations = [1, 2, 2, 1], padding = "SAME", strides = [1, 1, 1, 1]}> : (tensor<1x128x128x3xf32>, tensor<5x5x3x8xf32>) -> tensor<1x128x128x8xf32>
  // CHECK-NEXT: [[RESULT:%.*]] = "tf.BiasAdd"([[CONV]], [[BIAS]]) : (tensor<1x128x128x8xf32>, tensor<8xf32>) -> tensor<1x128x128x8xf32>

def UndoBroadcastFullyConnectedBiasAdd : Pat<
  (TFL_AddOp $lhs, (Arith_ConstantOp:$const_value $bias), TFL_AF_None),
  (TFL_AddOp $lhs, (Arith_ConstantOp (FlattenTo1D $bias)), TFL_AF_None),
  [(AnyStaticShapeTensor $lhs),
   (IsLastDimEqualToNumElements $bias, $bias),
   (HasRankAtMost<4> $bias),
   (HasRankAtLeast<2> $bias),
   (IsDefinedByFullyConnectedOp $lhs),

    'NewFullyConnected',
    inputs=['input_: T', 'filter_: T', 'bias: T'],
    attrs=['act: {"", "RELU", "RELU6", "TANH"} = ""'],
    derived_attrs=['T: {float, int8}'],
    outputs=['o: T'])
def _composite_fully_connected(input_, filter_, bias, act):
  res = tf.raw_ops.MatMul(
      a=input_, b=filter_, transpose_a=False, transpose_b=True)
  res = tf.raw_ops.Add(x=res, y=bias)
  if act == 'RELU':
    return tf.raw_ops.Relu(features=res)

  func.return %1 : tensor<256x8x7x3xf32>

  // CHECK: %[[weight:.*]] = arith.constant dense<3.000000e+00> : tensor<3x3x3x3xf32>
  // CHECK: %[[bias:.*]] = arith.constant dense<[1.500000e+00, 3.000000e+00, 4.500000e+00]>
  // CHECK: %[[conv:.*]] = "tfl.conv_2d"(%arg0, %[[weight]], %[[bias]])
  // CHECK: return %[[conv]] : tensor<256x8x7x3xf32>

  }
}

// Creates a new `tfl.qconst` op for the bias. The bias values are 0s, because
// this bias a dummy bias (note that bias fusion is not considered for this
// transformation). The quantization scale for the bias is input scale *
// filter scale. `filter_const_op` is used to retrieve the filter scales and
// the size of the bias constant.
TFL::QConstOp CreateTflConstOpForDummyBias(

  func.func @serving_default(%arg0: tensor<!tf_type.resource<tensor<100x50xf32>>> {tf.resource_name = "dense/kernel"}, %arg1: tensor<!tf_type.resource<tensor<50xf32>>> {tf.resource_name = "dense/bias"}) -> (tensor<100x50xf32> {tf_saved_model.index_path = ["dense_2"]})
  attributes {tf.entry_function = {control_outputs = "", inputs = "", outputs = "dense_2/Add:0"}, tf_saved_model.exported_names = ["serving_default"]} {

		std.offset = -int(i.Bias) * 60
		l.cacheStart = alpha
		l.cacheEnd = omega
		l.cacheZone = std
		l.tx = make([]zoneTrans, 1)
		l.tx[0].when = l.cacheStart
		l.tx[0].index = 0
		return
	}

	// StandardBias must be ignored if StandardDate is not set,
	// so this computation is delayed until after the nzone==1
	// return above.
	std.offset = -int(i.Bias+i.StandardBias) * 60

	dst := &l.zone[1]

    }
    const int64_t cost_per_col = 2 * weight_type.getNumElements();

    *count = cost_per_col * cols;

    auto bias = op->getOperand(2);
    if (bias) {
      auto bias_type =
          mlir::dyn_cast_or_null<mlir::RankedTensorType>(bias.getType());
      if (bias_type && bias_type.hasStaticShape()) {
        *count += output_type.getNumElements();
      }
    }

    return true;
  }

	exp += e
	x := Float64bits(frac)
	exp += int(x>>shift)&mask - bias
	if exp < -1075 {
		return Copysign(0, frac) // underflow
	}
	if exp > 1023 { // overflow
		if frac < 0 {
			return Inf(-1)
		}
		return Inf(1)
	}
	var m float64 = 1
	if exp < -1022 { // denormal
		exp += 53
		m = 1.0 / (1 << 53) // 2**-53
	}
	x &^= mask << shift
	x |= uint64(exp+bias) << shift
	return m * Float64frombits(x)

  def optimizeConsecutiveConv#OpsTuple[0]#OpsTuple[1] : Pat<
    (OpsTuple[1]
      (OpsTuple[0] $input, $zp_offset, $broadcast_dims_1),
      $bias, $broadcast_dims_2),
    (OpsTuple[0]
      $input,
      (OpsTuple[2] $zp_offset, $bias, $broadcast_dims_2), $broadcast_dims_1),
    [
      (IsNull $broadcast_dims_1),
      (IsNull $broadcast_dims_2),
      (TensorOf<[AnyInteger]> $input),