}
  }
}
node {
  name: "loss_1/conv1/kernel/Regularizer/Square/ReadVariableOp"
  op: "ReadVariableOp"
  input: "conv1/kernel"
  device: "/job:localhost/replica:0/task:0/device:GPU:0"
  attr {
    key: "dtype"
    value {
      type: DT_FLOAT
    }
  }
}
node {
  name: "conv1_1/Conv2D/ReadVariableOp"
  op: "ReadVariableOp"
  input: "conv1/kernel"

    value {
      b: false
    }
  }
}
node {
  name: "training/LossScaleOptimizer/gradients/loss_1/conv1/kernel/Regularizer/Square_grad/Mul_1"
  op: "Mul"
  input: "loss_1/conv1/kernel/Regularizer/Square/ReadVariableOp"
  input: "training/LossScaleOptimizer/gradients/loss_1/conv1/kernel/Regularizer/Square_grad/Mul"
  device: "/job:localhost/replica:0/task:0/device:GPU:0"
  attr {
    key: "T"

		// pick up next sudog
		var cond2 sync.Cond
		var mu2 sync.Mutex
		cond2.L = &mu2
		go func() {
			mu2.Lock()
			cond2.Wait()
			mu2.Unlock()
		}()
		time.Sleep(1 * time.Millisecond)

		// put top sudog back
		cond1.Broadcast()
		time.Sleep(1 * time.Millisecond)

		// drop cache on floor
		runtime.GC()

		// release cond2 after select has gotten to run
		release = func() {

    }
  }

  // Handle Conv2D input, stride and filter.
  HandleConv2DInput(conv2d, block_size);
  HandleConv2DStride(conv2d);
  HandleConv2DFilter(conv2d, block_size);

  // Book keeping new filter shape for backprop filter rewrite.
  // Filter shape is defined in HandleConv2DFilter, thus it is RankedTensorType.
  filter_shape =
      mlir::cast<RankedTensorType>(conv2d.getFilter().getType()).getShape();

  // Check that Conv2D computed in NCHW format, and all redundant transpose
  // operations removed from the function.

  // CHECK: %[[CONV:[0-9]*]] = "tf.Conv2D"(%arg0, %arg1)
  // CHECK-SAME: data_format = "NCHW"
  // CHECK-SAME: -> tensor<1x8x32x32xf32>

  // CHECK: return %[[CONV]]

  func.return %4 : tensor<1x8x32x32xf32>

        scale, offset = [1.0] * 2, [0.5] * 2
        mean, variance = scale, offset
        out = nn_ops.conv2d(
            q_input,
            q_filter,
            strides=[1, 1, 2, 1],
            dilations=[1, 1, 1, 1],
            padding='SAME',
            data_format='NHWC',
            name='sample/conv2d',
        )
        if has_bias:
          out = nn_ops.bias_add(out, bias, data_format='NHWC')

    return %3 : tensor<?x?x?x2xf32>
  }

// CHECK-LABEL: func @conv_with_dynamic_shape
// The Conv2D should not be quantized since it has dynamic channel.
// CHECK: "tf.Conv2D"
// CHECK-SAME: (tensor<?x?x?x?xf32>, tensor<2x3x3x2xf32>) -> tensor<?x?x?x2xf32>

  // CHECK: %[[ARG_PERM:.*]] = "tf.Const"() <{value = dense<[0, 3, 1, 2]> : tensor<4xi64>}>
  // CHECK: %[[ARG_TRANSPOSE:[0-9]*]] = "tf.Transpose"(%arg0, %[[ARG_PERM]])

  // CHECK: %[[CONV2D:[0-9]*]] = "tf.Conv2D"(%[[ARG_TRANSPOSE]], %arg1)
  // CHECK-SAME: data_format = "NCHW"
  // CHECK-SAME: dilations = [1, 4, 2, 3]
  // CHECK-SAME: explicit_paddings = [1, 2, 7, 8, 3, 4, 5, 6]
  // CHECK-SAME: padding = "EXPLICIT"

        def conf2 = new ConfigurationDetails('config2', null, false, null, null)

        when:
        renderer.prepareVisit()
        renderer.startConfiguration(conf1);
        renderer.completeConfiguration(conf1);
        renderer.startConfiguration(conf2);
        renderer.completeConfiguration(conf2);

        then:
        textOutput.value.readLines() == [

  // cuDNN prefers NCHW data format for spatial convolutions.
  // CHECK: "tf.Conv2D"(%[[INPUT_TRANSPOSE:[0-9]*]], %arg1)
  // CHECK-SAME: data_format = "NCHW"
  %0 = "tf.Conv2D"(%input, %filter)
       {
         data_format = "NHWC",
         padding = "VALID",
         strides = [1, 1, 1, 1]