package a

func Conv(v interface{}) string {
	return conv[string](v)
}

func conv[T any](v interface{}) T {
	return v.(T)
}

func Conv2(v interface{}) (string, bool) {
	return conv2[string](v)
}

func conv2[T any](v interface{}) (T, bool) {
	x, ok := v.(T)
	return x, ok
}

func Conv3(v interface{}) string {
	return conv3[string](v)
}

}

// Ensure OCONVNOP is zero cost.
func Conv(v uint64) uint64 { // ERROR "can inline Conv"
	return conv2(conv2(conv2(v))) // ERROR "inlining call to (conv1|conv2)"
}
func conv2(v uint64) uint64 { // ERROR "can inline conv2"
	return conv1(conv1(conv1(conv1(v)))) // ERROR "inlining call to conv1"
}
func conv1(v uint64) uint64 { // ERROR "can inline conv1"

	"./a"
	"fmt"
)

func main() {
	s := "foo"
	x := a.Conv(s)
	if x != s {
		panic(fmt.Sprintf("got %s wanted %s", x, s))
	}
	y, ok := a.Conv2(s)
	if !ok {
		panic("conversion failed")
	}
	if y != s {
		panic(fmt.Sprintf("got %s wanted %s", y, s))
	}
	z := a.Conv3(s)
	if z != s {
		panic(fmt.Sprintf("got %s wanted %s", z, s))
	}
	w := a.Conv4(a.Mystring(s))
	if w != a.Mystring(s) {

}

// Ensure OCONVNOP is zero cost.
func Conv(v uint64) uint64 { // ERROR "can inline Conv"
	return conv2(conv2(conv2(v))) // ERROR "inlining call to (conv1|conv2)"
}
func conv2(v uint64) uint64 { // ERROR "can inline conv2"
	return conv1(conv1(conv1(conv1(v)))) // ERROR "inlining call to conv1"
}
func conv1(v uint64) uint64 { // ERROR "can inline conv1"

    conv2 = gen_mnist_ops.new_conv2d(max_pool1, self.weights['f2'],
                                     self.biases['b2'], 1, 1, 1, 1, 'SAME',
                                     'RELU')

    # output shape: [-1, 7, 7, 64]
    max_pool2 = gen_mnist_ops.new_max_pool(conv2, 2, 2, 2, 2, 'SAME')

    # Reshape the feature map cuboid into a 2D matrix to feed it to the
    # fully connected layers.
    # output shape: [-1, 7*7*64]

  %5 = "tf.Conv2D"(%4, %arg3)
        {
          data_format = "NCHW",
          dilations = [1, 1, 1, 1],
          explicit_paddings = [],
          padding = "VALID",
          strides = [1, 1, 2, 2]
        } : (tensor<?x3x230x230xf32>, tensor<7x7x3x64xf32>) -> tensor<?x64x112x112xf32>

  // CHECK: %[[CONV0:[0-9]*]] = "tf.Conv2D"
  // CHECK-SAME: %[[PAD]]

  %conv2 = "tfl.conv_2d"(%4, %5, %cst) {dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32} : (tensor<1x112x112x32xf32>, tensor<32x3x3x3xf32>, tensor<32xf32>) -> tensor<1x56x56x32xf32>

  %b2 = arith.constant dense<[1.0e-2, 2.1473647e1, -2.1473647e2]> : tensor<3xf32>
  %conv = "tfl.conv_2d"(%0, %w, %b) {
    dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "RELU",
    padding = "SAME", stride_h = 1 : i32, stride_w = 1 : i32
  } : (tensor<1x5x5x2xf32>, tensor<3x1x1x2xf32>, tensor<3xf32>) -> tensor<1x5x5x3xf32>
  %conv2 = "tfl.conv_2d"(%0, %w, %b2) {

  // 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>

    ```python
    conv0 = tf.compat.v1.layers.Conv2D(
     filters=filters,
     kernel_size=kernel_size,
     strides=2,
     padding=('SAME' if strides == 1 else 'VALID'),
     use_bias=False,
     kernel_initializer=tf.variance_scaling_initializer(),
     data_format=data_format)

    # Use the image size without space-to-depth transform as the input of conv0.