%graph = tf_executor.graph {
            %read0, %read0_control = tf_executor.island wraps "tf.ReadVariableOp"(%arg0) : (tensor<*x!tf_type.resource<tensor<32xf32>>>) -> tensor<32xf32>
            %assign0_control = tf_executor.island wraps "tf.AssignVariableOp"(%arg0, %arg1) : (tensor<*x!tf_type.resource<tensor<32xf32>>>, tensor<32xf32>) -> ()

		return fallback_nanotime()
	}
	return int64((1e9 * uint64(bt.sec)) + ((1e9 * uint64(bt.frac>>32)) >> 32))
}

func walltime() (sec int64, nsec int32) {
	bt := vdsoClockGettime(_CLOCK_REALTIME)
	if bt == zeroBintime {
		return fallback_walltime()
	}
	return int64(bt.sec), int32((1e9 * uint64(bt.frac>>32)) >> 32)

      %0 = tf_executor.graph {
        %read0, %read0_control = tf_executor.island wraps "tf.ReadVariableOp"(%arg0) : (tensor<*x!tf_type.resource<tensor<32xf32>>>) -> tensor<32xf32>
        %assign0_control = tf_executor.island(%read0_control) wraps "tf.AssignVariableOp"(%arg0, %arg1) : (tensor<*x!tf_type.resource<tensor<32xf32>>>, tensor<32xf32>) -> ()

  func.return %0 : tensor<256x30x30x16xf32>
}

// -----

func.func @testDepthwiseConv(%arg0: tensor<1x112x112x32xf32>, %arg1: tensor<1x3x3x32xf32>, %arg2: tensor<32xf32>) -> tensor<1x112x112x32xf32> {
  // CHECK: tac.device = "GPU", tac.inference_type = "FLOAT"

func.func @bias_add(%arg0: tensor<1x10x10x32xf32>, %arg1: tensor<32xf32>) -> tensor<1x10x10x32xf32> {
  %0 = "tf.BiasAdd"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", data_format = "NHWC"} : (tensor<1x10x10x32xf32>, tensor<32xf32>) -> tensor<1x10x10x32xf32>
  func.return %0 : tensor<1x10x10x32xf32>
// CHECK: %[[BIASADD_0:.*]] = "tf.BiasAdd"(%arg0, %arg1) <{data_format = "NHWC"}> {T = "tfdtype$DT_FLOAT"} : (tensor<1x10x10x32xf32>, tensor<32xf32>) -> tensor<1x10x10x32xf32>

    -> (tensor<?x2xi64>, tensor<?x2xi64>, tensor<?x!tf_type.string>, tensor<?xi64>, tensor<2xi64>, tensor<2xi64>, tensor<32xf32>, tensor<32xf32>)

  // CHECK: tfrt.return {{.*}}, [[outputs]]#0
  func.return %outputs#0 : tensor<?x2xi64>
}

// CHECK-LABEL: func @no_native

  %0 = "tfl.transpose_conv"(%cst_1, %cst_2, %arg0, %cst_3) {padding = "SAME", stride_h = 2 : i32, stride_w = 2 : i32, fused_activation_function = "NONE"} : (tensor<4xi32>, tensor<32x4x4x128xf32>, tensor<1x32x42x128xf32>, tensor<32xf32>) -> tensor<1x64x84x32xf32>
  %1 = "tfl.add"(%0, %cst) {fused_activation_function = "NONE"} : (tensor<1x64x84x32xf32>, tensor<32xf32>) -> tensor<1x64x84x32xf32>

// CHECK:           %[[VAL_2:.*]] = "tfl.pseudo_const"{{.*}}dense<0.000000e+00> : tensor<32xf32>
// CHECK:           %[[VAL_3:.*]] = "tfl.conv_2d"(%[[VAL_0]], %[[VAL_1]], %[[VAL_2]]) <{dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "VALID", stride_h = 1 : i32, stride_w = 1 : i32}> : (tensor<128x32x32x3xf32>, tensor<32x1x1x3xf32>, tensor<32xf32>) -> tensor<128x32x32x32xf32>

  // CHECK-DAG: %[[VAL1:.+]] = "tfl.pseudo_qconst"() <{qtype = tensor<32x!quant.uniform<{{.+}}>>, value = dense<0> : tensor<32xi32>}>
  // CHECK-DAG: %[[VAL2:.+]] = "tfl.dequantize"(%[[VAL0]]) : (tensor<32x3x3x3x!quant.uniform<{{.+}}>>)
  // CHECK-DAG: %[[VAL3:.+]] = "tfl.dequantize"(%[[VAL1]]) : (tensor<32x!quant.uniform<{{.+}})

		new := uint64(seed+i)<<32 | uint64(seed+i)<<32>>32
		old := uint64(addr.Swap(int64(new)))
		if old>>32 != old<<32>>32 {
			panic(fmt.Sprintf("SwapInt64 is not atomic: %v", old))
		}
	}
}

func hammerSwapUint64(addr *uint64, count int) {
	seed := int(uintptr(unsafe.Pointer(&count)))
	for i := 0; i < count; i++ {
		new := uint64(seed+i)<<32 | uint64(seed+i)<<32>>32
		old := SwapUint64(addr, new)