6, 5, 4, 5, 5,
		4, 3, 3, 5, 9,
		6, 2, 5, 6, 5,
		4, 5, 5, 3, 10,
		2, 5, 5, 3, 7, 6}
	muls := []*p256OrdElement{
		_101111, _111, _11, _1111, _10101,
		_101, _101, _101, _111, _101111,
		_1111, _1, _1, _1111, _111,
		_111, _111, _101, _11, _101111,
		_11, _11, _11, _1, _10101, _1111}

	for i, s := range sqrs {
		p256OrdSqr(x, x, s)
		p256OrdMul(x, x, muls[i])
	}

	// following addition chain generated with github.com/mmcloughlin/addchain v0.4.0.
	//
	//	_10       = 2*1
	//	_11       = 1 + _10
	//	_1100     = _11 << 2
	//	_1111     = _11 + _1100
	//	_11110000 = _1111 << 4
	//	_11111111 = _1111 + _11110000
	//	x16       = _11111111 << 8 + _11111111
	//	x32       = x16 << 16 + x16
	//	x64       = x32 << 32 + x32
	//	x65       = 2*x64 + 1
	//	x129      = x65 << 64 + x64

	// following addition chain generated with github.com/mmcloughlin/addchain v0.4.0.
	//
	//	_10       = 2*1
	//	_11       = 1 + _10
	//	_1100     = _11 << 2
	//	_1111     = _11 + _1100
	//	_11110000 = _1111 << 4
	//	_11111111 = _1111 + _11110000
	//	x16       = _11111111 << 8 + _11111111
	//	x32       = x16 << 16 + x16
	//	return      ((x32 << 32 + 1) << 96 + 1) << 94
	//
	var t0 = new(fiat.P256Element)

	// following addition chain generated with github.com/mmcloughlin/addchain v0.4.0.
	//
	//	_10       = 2*1
	//	_11       = 1 + _10
	//	_1100     = _11 << 2
	//	_1111     = _11 + _1100
	//	_11110000 = _1111 << 4
	//	_11111111 = _1111 + _11110000
	//	x16       = _11111111 << 8 + _11111111
	//	x32       = x16 << 16 + x16
	//	return      ((x32 << 32 + 1) << 96 + 1) << 94
	//
	p256Sqr(t0, x, 1)

    data_format = "NDHWC", device = "", dilations = [1, 1, 1, 1, 1], padding = "SAME", strides = [1, 1, 2, 1, 1]
  } : (tensor<1x3x4x3x3xf32>, tensor<2x3x3x3x2xf32>) -> tensor<1x3x2x3x2xf32>
  %1 = "tf.Relu"(%0) {device = ""} : (tensor<1x3x2x3x2xf32>) -> tensor<1x3x2x3x2xf32>

  %2 = "tf.Conv3D"(%arg0, %cst) {
    data_format = "NDHWC", device = "", dilations = [1, 1, 1, 1, 1], padding = "SAME", strides = [1, 1, 2, 1, 1]

  Tensor filter = test::AsTensor<float>({0.5f}, {1, 1, 1, 1});
  const std::vector<int> strides{1, 1, 1, 1};
  auto y = Conv2D(scope_, x, filter, strides, "SAME");
  RunTest(x, shape, y, shape);
}

TEST_F(NNGradTest, MaxPoolGradHelper) {
  TensorShape x_shape({1, 2, 2, 1});
  TensorShape y_shape({1, 1, 1, 1});
  auto x = Placeholder(scope_, DT_FLOAT, Placeholder::Shape(x_shape));

  %cst = "tf.Const"() {device = "", value = dense<1.0> : tensor<2x3x3x3x2xf32>} : () -> tensor<2x3x3x3x2xf32>
  %0 = "tf.Conv3D"(%arg0, %cst) {
    data_format = "NDHWC", device = "", dilations = [1, 1, 1, 1, 1], padding = "SAME", strides = [1, 1, 2, 1, 1]
  } : (tensor<1x3x4x3x3xf32>, tensor<2x3x3x3x2xf32>) -> tensor<1x3x2x3x2xf32>
  %1 = "tf.Relu"(%0) {device = ""} : (tensor<1x3x2x3x2xf32>) -> tensor<1x3x2x3x2xf32>

    assertEscaping(e, "%7F", '\u007f'); // del
    assertEscaping(e, "%C2%80", '\u0080'); // xx-00010,x-000000
    assertEscaping(e, "%DF%BF", '\u07ff'); // xx-11111,x-111111
    assertEscaping(e, "%E0%A0%80", '\u0800'); // xxx-0000,x-100000,x-00,0000
    assertEscaping(e, "%EF%BF%BF", '\uffff'); // xxx-1111,x-111111,x-11,1111
    assertUnicodeEscaping(e, "%F0%90%80%80", '\uD800', '\uDC00');
    assertUnicodeEscaping(e, "%F4%8F%BF%BF", '\uDBFF', '\uDFFF');

                    return -1;

                case STRING_ITEM:
                    return 1;
                case COMBINATION_ITEM:
                    return 1; // 1.1 > 1-sp

                case LIST_ITEM:
                    return 1; // 1.1 > 1-1

                default:
                    throw new IllegalStateException("invalid item: " + item.getClass());
            }
        }

        @Override

    %identity = "tf.Identity"(%filter) : (tensor<*xi8>) -> tensor<*xi8>
    %2 = "tf.Cast"(%identity) {Truncate = false} : (tensor<*xi8>) -> tensor<*xf32>
    %3 = "tf.Conv3D"(%input, %2) {
      padding = "VALID", strides = [1, 1, 1, 1, 1],
      attr_map = "strides:0,padding:1,dilations:2"
    } : (tensor<*xf32>, tensor<*xf32>) -> tensor<*xf32>
    func.return %3 : tensor<*xf32>
  }

  func.func private @internal_batch_matmul_fn(