if (isStart && !isDirect) { // start of new frag, do trans
            off = in.read(buf, 0, 1024);
        } else {
            off = in.readDirect(buf, 0, buf.length);
        }

        if (buf[0] != 5 && buf[1] != 0)
            throw new IOException("Unexpected DCERPC PDU header");

        flags = buf[3] & 0xFF;
        // next read is start of new frag
        isStart = (flags & DCERPC_LAST_FRAG) == DCERPC_LAST_FRAG;

            if (first) {
                first = false
            } else {
                val tokens = line.replace("\"", "").split(",")

                val (benchmark, _, _, _, score) = tokens.subList(0, 5)
                val (_, _, accessor) = tokens.subList(5, tokens.size)
                val name = benchToScenarioName(benchmark)
                val benchmarksValue = benchmarks.getValue(name)

    tf_device.return %4 : tensor<i32>
  }) {cluster_attr = "cluster_attr"} : () -> tensor<i32>
  %5 = "tf.Const"() {value = dense<5> : tensor<i32>} : () -> tensor<i32>
  %6 = "tf.Add"(%0, %5) : (tensor<i32>, tensor<i32>) -> tensor<i32>
  func.return %6 : tensor<i32>
}

// -----

// CHECK-LABEL: func @no_stateful_partitioned_call_in_cluster_op
// CHECK-NOT: "tf_device.cluster"

        if ( buf.length < getMaxRecv() ) {
            throw new IllegalArgumentException("buffer too small");
        }

        int off = this.handle.recv(buf, 0, buf.length);
        if ( buf[ 0 ] != 5 || buf[ 1 ] != 0 ) {
            throw new IOException("Unexpected DCERPC PDU header");
        }

        int length = Encdec.dec_uint16le(buf, 8);
        if ( length > getMaxRecv() ) {

	}
	f := f1 - 1
	k := float64(ki)

	// compute
	s := f / (2 + f)
	s2 := s * s
	s4 := s2 * s2
	t1 := s2 * (L1 + s4*(L3+s4*(L5+s4*L7)))
	t2 := s4 * (L2 + s4*(L4+s4*L6))
	R := t1 + t2
	hfsq := 0.5 * f * f
	return k*Ln2Hi - ((hfsq - (s*(hfsq+R) + k*Ln2Lo)) - f)

000002b0  d1 4b 19 49 91 b5 70 44  2d 62 a1 04 8f 98 89 ab  |.K.I..pD-b......|
000002c0  75 1f 8a 97 d5 28 dc 88  20 92 7f f2 90 16 1b de  |u....(.. .......|
000002d0  54 7c 5f 05 8f c9 70 a1  f5 8c 99 09 43 ab 09 73  |T|_...p.....C..s|
000002e0  dd 5e 62 c3 9f c8 e1 fa  9e 98 65 18 b5 5e 4e f9  |.^b.......e..^N.|
000002f0  85 a4 cc dc b8 12 1b eb  0a bc bc bd 1c 2f 51 0e  |............./Q.|

000000e0  10 01 58 cc 03 c3 53 04  03 69 80 f4 ad 4d ce 72  |..X...S..i...M.r|
000000f0  26 4e 6c c7 2c 31 69 2b  fd 97 67 5e 7d e0 05 b3  |&Nl.,1i+..g^}...|
00000100  f4 40 64 a1 bd a3 fd a8  f9 7b 18 82 89 8f 25 f9  |.@d......{....%.|
00000110  ca ca c4 8f e4 90 7b 26  7a d5 b2 1e fa 05 db ad  |......{&z.......|
00000120  8a 9f 93 e9 13 5b 28 cc  cb 8b 30 f2 4c 1d 73 09  |.....[(...0.L.s.|

	f(Const)            // OK
	f(Float32)          // ERROR "convert|wrong type|cannot"
	f(Float64)          // ERROR "convert|wrong type|cannot"
	f(ConstFloat)       // ERROR "truncate"
	f(ConstFloat - 0.5) // OK
	f(Big)              // ERROR "convert|wrong type|cannot"
	f(String)           // ERROR "convert|wrong type|cannot|incompatible"
	f(Bool)             // ERROR "convert|wrong type|cannot|incompatible"
}

// RUN: tac-opt-all-backends -tfl-compute-cost %s -split-input-file -verify-diagnostics | FileCheck %s

// CHECK: tac.cost = 7.864320e+05
func.func @func_0_CPU(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<256x32x32x3xf32>) -> tensor<256x32x32x3xf32> attributes {tac.device = "CPU", tac.interface_name = "func_0"} {
  %0 = "tfl.add"(%arg0, %arg1) {fused_activation_function = "RELU", tac.device = "CPU"} : (tensor<256x32x32x3xf32>, tensor<256x32x32x3xf32>) -> tensor<256x32x32x3xf32>

  }

  private Element[] createQueries(Set<Element> elementsInSet, int numQueries) {
    List<Element> queryList = Lists.newArrayListWithCapacity(numQueries);

    int numGoodQueries = (int) (numQueries * hitRate + 0.5);

    // add good queries
    int size = elementsInSet.size();
    if (size > 0) {
      int minCopiesOfEachGoodQuery = numGoodQueries / size;
      int extras = numGoodQueries % size;