%2 = "tf.Relu"(%1) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
    %3 = "tf.FakeQuantWithMinMaxArgs"(%2) {device = "", max = 4.000000e-01 : f32, min = -3.000000e-01 : f32, narrow_range = false, num_bits = 8 : i64} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>

    version = 5 : i64
  } : (tensor<1x3x2x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x2x2x2xf32>
  return %0 : tensor<1x2x2x2xf32>
}

// CHECK-LABEL: func.func @qdq_for_conv_weight_empty
// CHECK-SAME: (%[[ARG_0:.+]]: tensor<1x3x2x3xf32>) -> tensor<1x2x2x2xf32>
// CHECK: %[[CST:.+]] = "tf.Const"() <{value = dense<3.000000e-01> : tensor<2x3x3x2xf32>}> : () -> tensor<2x3x3x2xf32>

    %1 = "tf.BiasAdd"(%0, %arg2) {data_format = "NHWC", device = ""} : (tensor<1x3x4x2xf32>, tensor<2xf32>) -> tensor<1x3x4x2xf32>
    %2 = "tf.Relu6"(%1) {device = ""} : (tensor<1x3x4x2xf32>) -> tensor<1x3x4x2xf32>
    func.return %2 : tensor<1x3x4x2xf32>
  }

// CHECK-LABEL: func @conv_with_per_channel_and_tensor_weight

      } : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>, tensor<2xf32>) -> tensor<1x3x4x2xf32>
    %2 = "quantfork.stats"(%1) {layerStats = dense<[5.00000000e-6, 7.00000000e-1]> : tensor<2xf32>} : (tensor<1x3x4x2xf32>) -> tensor<1x3x4x2xf32>
    return %2 : tensor<1x3x4x2xf32>
  }
// CHECK: func.func private @quantize_conv_with_bias_1d_fn(%[[ARG_0:.+]]: tensor<1x3x4x3xf32>) -> tensor<1x3x4x2xf32> attributes {tf._original_func_name = "main_0"}

            index_vector_dim = 2>,
          slice_sizes = array<i64: 1, 2, 2>,
          indices_are_sorted = false
        } : (tensor<3x4x2xf32>, tensor<2x3x2xi64>) -> tensor<2x3x2x2xf32>
        func.return %2 : tensor<2x3x2x2xf32>
      }
    }
  )mlir";
  OwningOpRef<ModuleOp> module_op = ParseModuleOpString(kModuleGather);
  ASSERT_TRUE(module_op);

  func::FuncOp main_fn = FindMainFuncOp(*module_op);

    indices_are_sorted = false
  } : (tensor<3x4x2xf32>, tensor<2x3x2xi64>) -> tensor<2x3x2x2xf32>
    %8 = "quantfork.qcast"(%7) {volatile} : (tensor<2x3x2x2xf32>) -> tensor<2x3x2x2x!quant.uniform<i8:f32, 0.13170163023705575:-1>>
    %9 = "quantfork.dcast"(%8) : (tensor<2x3x2x2x!quant.uniform<i8:f32, 0.13170163023705575:-1>>) -> tensor<2x3x2x2xf32>
    return %9 : tensor<2x3x2x2xf32>
  }

  // CHECK: quantized_dot_general_fn_1

}
// CHECK: func.func @main(%[[ARG:.+]]: tensor<1x3x4x3xf32>) -> tensor<1x3x2x2xf32> {
// CHECK-DAG: %[[CONST_0:.*]] = stablehlo.constant dense<[{{.*}}]> : tensor<2xf32>
// CHECK-DAG: %[[CONST_1:.*]] = stablehlo.constant dense<[{{.*}}]> : tensor<2x3x3x2xf32>
// CHECK-DAG: %[[CONV:.*]] = stablehlo.convolution(%[[ARG]], %[[CONST_1]]) {{.*}} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x2x2xf32>

    %0 = "quantfork.stats"(%arg1) {layerStats = dense<[-3.54062747, 0.54742622]> : tensor<2xf32>} : (tensor<2x3x3x2xf32>) -> tensor<2x3x3x2xf32>
    %1 = "quantfork.stats"(%arg0) {layerStats = dense<[1.27501142, 2.824783]> : tensor<2xf32>} : (tensor<1x3x2x3xf32>) -> tensor<1x3x2x3xf32>

      index_vector_dim = 2>,
    slice_sizes = array<i64: 1, 1, 2, 2>,
    indices_are_sorted = false
  } : (tensor<3x4x2x2xf32>, tensor<2x3x2xi64>) -> tensor<2x3x2x2xf32>
  return %0 : tensor<2x3x2x2xf32>
}

// CHECK-LABEL: func @gather_float
// CHECK: stablehlo.gather
// CHECK-NOT: tfl.gather_nd
// CHECK-NOT: tfl.gather

// -----

  %1 = "tf.Sub"(%0, %cst_0) {data_format = "NHWC"} : (tensor<1x3x2x2xf32>, tensor<1x1x1x2xf32>) -> tensor<1x3x2x2xf32>
  %2 = "tf.Mul"(%1, %cst_1) : (tensor<1x3x2x2xf32>, tensor<1x1x1x2xf32>) -> tensor<1x3x2x2xf32>
  %3 = "tf.AddV2"(%2, %cst_2) : (tensor<1x3x2x2xf32>, tensor<1x1x1x2xf32>) -> tensor<1x3x2x2xf32>
  func.return %3 : tensor<1x3x2x2xf32>
}