weight_new_shape, &rewriter);

  // Replace the fc with conv.
  // The output would be [1, 1, width, output].
  auto conv_output_type = RankedTensorType::get({1, 1, width, output_size},
                                                output_type.getElementType());
  auto conv = rewriter.create<TFL::Conv2DOp>(
      fc_op.getLoc(), conv_output_type, reshaped_input, reshaped_weight,

      mlir::mhlo::createLegalizeDotToDotGeneralPass());
  // Unfuse mhlo BatchNorm to primitive ops.
  pm.addNestedPass<mlir::func::FuncOp>(mlir::odml::createUnfuseBatchNormPass());
  // Fuse Conv + Mul to Conv.
  pm.addNestedPass<mlir::func::FuncOp>(mlir::odml::createFuseConvolutionPass());
  // Fold broadcast_in_dim + Mul.
  pm.addNestedPass<mlir::func::FuncOp>(mlir::odml::createFoldBroadcastPass());

// CHECK: %[[conv:.*]] = "tf.Conv2D"(%[[dq_input]], %[[dq_weight]])
// CHECK: %[[bias:.*]] = "tf.BiasAdd"(%[[conv]], %[[cst_0]]) <{data_format = "NHWC"}>
// CHECK: %[[relu6:.*]] = "tf.Relu6"(%[[bias]])

// -----

// CHECK: %[[cst1:.*]] = "tfl.pseudo_qconst"() <{qtype = tensor<32x3x3x3x!quant.uniform<u8<1:255>:f32, 1.000000e-01>>, value = dense<1> : tensor<32x3x3x3xi8>}>
// CHECK: %[[conv:.*]] = "tfl.conv_2d"(%arg0, %[[cst1]], %[[cst0]])
// CHECK: return %[[conv]] : tensor<1x112x112x32x!quant.uniform<u8:f32, 0.023528476789885875>>
}

// CHECK-LABEL: QuantizeConv2D4Bit

# MLIR:         %[[conv:.*]] = "tfl.conv_2d"(%[[ARG_0]], %[[weight]], %[[bias]]) <{dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32}
# MLIR:         %[[reshape:.*]] = "tfl.reshape"(%[[conv]], %[[shape]]) : (tensor<1x1x1x186x!quant.uniform<i8:f32, 0.09363494573854933:22>>, tensor<3xi32>)

# MLIR:         %[[conv:.*]] = "tfl.conv_2d"(%[[ARG_0]], %[[weight]], %[[bias]]) <{dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32}
# MLIR:         %[[reshape:.*]] = "tfl.reshape"(%[[conv]], %[[shape]]) : (tensor<1x1x1x186x!quant.uniform<i8:f32, 0.09363494573854933:22>>, tensor<3xi32>)

        "MLIR dump file name.">,
    Option<"merge_fusion_with_dequantize_",
        "merge-fusion-with-dequantize",
        "bool", /*default=*/"false",
        "Whether to merge quantized conv/dot_general fusion with subsequent dequantize.">,
  ];
  let dependentDialects = [
    "mlir::arith::ArithDialect",
    "mlir::stablehlo::StablehloDialect",
    "mlir::quant::QuantizationDialect",

	if types.IsComplex[et] && n.Op() == ir.ODIV {
		t := n.Type()
		call := mkcall("complex128div", types.Types[types.TCOMPLEX128], init, typecheck.Conv(n.X, types.Types[types.TCOMPLEX128]), typecheck.Conv(n.Y, types.Types[types.TCOMPLEX128]))
		return typecheck.Conv(call, t)
	}

	// Nothing to do for float divisions.
	if types.IsFloat[et] {
		return n
	}

	// rewrite 64-bit div and mod on 32-bit architectures.

# MLIR:         %[[conv:.*]] = "tfl.conv_2d"(%[[ARG_0]], %[[weight]], %[[bias]]) <{dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32}
# MLIR:         %[[reshape:.*]] = "tfl.reshape"(%[[conv]], %[[shape]]) : (tensor<1x1x1x186x!quant.uniform<i8:f32, 0.09363494573854933:22>>, tensor<3xi32>)

# MLIR:         %[[conv:.*]] = "tfl.conv_2d"(%[[ARG_0]], %[[weight]], %[[bias]]) <{dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32}
# MLIR:         %[[reshape:.*]] = "tfl.reshape"(%[[conv]], %[[shape]]) : (tensor<1x1x1x186x!quant.uniform<i8:f32, 0.09363494573854933:22>>, tensor<3xi32>)