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

    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');

    %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(

    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');

    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');

    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');

	// following addition chain generated with github.com/mmcloughlin/addchain v0.4.0.
	//
	//	_10     = 2*1
	//	_11     = 1 + _10
	//	_110    = 2*_11
	//	_111    = 1 + _110
	//	_111000 = _111 << 3
	//	_111111 = _111 + _111000
	//	x12     = _111111 << 6 + _111111
	//	x15     = x12 << 3 + _111
	//	x16     = 2*x15 + 1
	//	x32     = x16 << 16 + x16
	//	i53     = x32 << 15
	//	x47     = x15 + i53

  %0 = stablehlo.constant() {value = dense<7.000000e+00> : tensor<8x8x3x3xf32>} : () -> tensor<8x8x3x3xf32>
  %2 = stablehlo.convolution(%arg0, %0) dim_numbers = [b, f, 0, 1]x[o, i, 0, 1]->[b, f, 0, 1], window = {pad = [[1, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<1x8x4x4xf32>, tensor<8x8x3x3xf32>) -> tensor<1x8x4x4xf32>
  return %2 : tensor<1x8x4x4xf32>
}

      %outputs_4, %control_5 = tf_executor.island wraps "tf.Conv2D"(%outputs, %outputs_2) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "VALID", strides = [1, 1, 1, 1], use_cudnn_on_gpu = true} : (tensor<*xf32>, tensor<5x5x32x16xf32>) -> tensor<*xf32>