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)
}

	"./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) {

                             %arg2: tensor<256xf32>,          // batch_norm args
                             %arg3: tensor<7x7x3x64xf32>,    // conv filter #0
                             %arg4: tensor<1x1x64x256xf32>   // conv filter #1
                            ) -> tensor<?x256xf32> {

  // This is a simplified ResNet layer that gets input in NHWC format, converts

}

// 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"

}

// 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"

    # output shape: [-1, 28, 28, 32]
    conv1 = gen_mnist_ops.new_conv2d(x, self.weights['f1'], self.biases['b1'],
                                     1, 1, 1, 1, 'SAME', 'RELU')

    # Max pooling. The kernel size spec {ksize} also follows the layout of
    # the data. Here we have a pooling window of 2, and a stride of 2.
    # output shape: [-1, 14, 14, 32]
    max_pool1 = gen_mnist_ops.new_max_pool(conv1, 2, 2, 2, 2, 'SAME')

  // 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.

// CHECK: %[[conv3d:.*]] = "tfl.conv_3d"(%arg0, %[[w]], %[[const]]) <{dilation_d_factor = 1 : i32, dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "VALID", stride_d = 1 : i32, stride_h = 1 : i32, stride_w = 1 : i32}> : (tensor<?x28x28x28x8xf32>, tensor<3x3x3x8x16xf32>, none) -> tensor<?x26x26x26x16xf32>
// CHECK: %2 = "tfl.shape"(%[[conv3d]]) : (tensor<?x26x26x26x16xf32>) -> tensor<5xi64>

  [(HasRankOf<1> $add_rhs_value),
   (HasEqualElementSize<[-1], [0]> $conv_out, $add_rhs)], [], (addBenefit -1)>;

// Convert conv+sub+mul pattern to conv+mul+add.
// (conv - sub) * mul -> conv * mul + (-sub) * mul
//
// This is needed to support Conv+BatchNorm pattern from Jax models converted
// using jax2tf w/o native serialization. Note that Jax2tf patterns always