// PerTensor: %[[out_2:.*]] = "tf.PartitionedCall"(%arg0, %[[q_w]], %[[scale]], %[[zp]]) <{config = "", config_proto = "", executor_type = "",
// PerTensor-SAME: f = @quantized_conv2d_fn_0}> : (tensor<1x2x2x3xf32>, tensor<2x3x3x2xi8>, tensor<f32>, tensor<i32>) -> tensor<*xf32>
// PerTensor: return %[[out_1]], %[[out_2]]

// PerChannel-LABEL: func @conv

				fallthrough
			case ssStarter:
				if outp > 0 {
					copy(i.buf[outp:], d)
					return i.buf[:p]
				}
				return d
			}
			copy(i.buf[outp:], d)
			outp = p
			inCopyStart, outCopyStart = i.p, outp
			if i.info.ccc < prevCC {
				goto doNorm
			}
			continue
		} else if r := i.rb.src.hangul(i.p); r != 0 {
			outp = decomposeHangul(i.buf[:], r)
			i.p += hangulUTF8Size

    private static class OptionComparator implements Comparator<Option> {
        public int compare(Option opt1, Option opt2) {
            String s1 = opt1.getOpt() != null ? opt1.getOpt() : opt1.getLongOpt();
            String s2 = opt2.getOpt() != null ? opt2.getOpt() : opt2.getLongOpt();
            return s1.compareToIgnoreCase(s2);
        }
    }

    private static class CLIManagerExtension extends CLIManager {

}

func (c *CodeDirectory) put(out []byte) []byte {
	out = put32be(out, c.magic)
	out = put32be(out, c.length)
	out = put32be(out, c.version)
	out = put32be(out, c.flags)
	out = put32be(out, c.hashOffset)
	out = put32be(out, c.identOffset)
	out = put32be(out, c.nSpecialSlots)
	out = put32be(out, c.nCodeSlots)
	out = put32be(out, c.codeLimit)
	out = put8(out, c.hashSize)
	out = put8(out, c.hashType)

  );

  let results = (outs
    TF_NumberTensor:$out
  );

  TF_DerivedOperandTypeAttr T = TF_DerivedOperandTypeAttr<0>;
  TF_DerivedOperandTypeAttr Tidx = TF_DerivedOperandTypeAttr<1>;

  let hasVerifier = 1;
}

def TF_CumsumOp : TF_Op<"Cumsum", [Pure, TF_AllTypesMatch<["x", "out"]>]> {

  let description = [{
    Transfer a CPU tensor to device.

    Example:
      %device_tensor = gpurt.transfer_to_device %cpu_tensor
  }];

  let arguments = (ins TFTensorType);
  let results = (outs TFTensorType);
  let assemblyFormat = "operands attr-dict";
}

// TODO(b/260267885): We may add a device argument when we want to support
// GPU MIG.
def TransferFromDeviceOp: Gpu_Op<"transfer_from_device"> {

	SUBVU	R9, R5, R7
	AND	$7, R7
	BNE	R7, out1

	// if less than 8 bytes, do byte copying
	SGTU	$8, R6, R7
	BNE	R7, out1

	// do one byte at a time until 8-aligned
	AND	$7, R9, R8
	BEQ	R8, words1
	ADDV	$-1, R5
	MOVB	(R5), R7
	ADDV	$-1, R9
	MOVB	R7, (R9)
	JMP	-6(PC)

words1:
	// do 8 bytes at a time if there is room
	ADDV	$7, R4, R6 // R6 is start pointer+7

	SUBVU	R6, R2, R4
	AND	$7, R4
	BNE	R4, out1

	// if less than 8 bytes, do byte copying
	SGTU	$8, R3, R4
	BNE	R4, out1

	// do one byte at a time until 8-aligned
	AND	$7, R6, R5
	BEQ	R5, words1
	ADDV	$-1, R2
	MOVB	(R2), R4
	ADDV	$-1, R6
	MOVB	R4, (R6)
	JMP	-6(PC)

words1:
	// do 8 bytes at a time if there is room
	ADDV	$7, R1, R3 // R3 is start pointer+7

	SGTU	R6, R3, R5

		w.Close()
		os.Stdout, os.Stderr = stdout, stderr
		out := <-outC

		// report any errors
		tstr := fmt.Sprintf("(%.2f seconds)", dt.Seconds())
		if g, e := strings.TrimSpace(out), strings.TrimSpace(eg.Output); g != e {
			fmt.Printf("--- FAIL: %s %s\ngot:\n%s\nwant:\n%s\n",
				eg.Name, tstr, g, e)
			ok = false
		} else if *chatty {
			fmt.Printf("--- PASS: %s %s\n", eg.Name, tstr)
		}
	}

    $future: A value of type !mlrt.future. The underlying value must be a tensorflow tensor.

    $result: a tensorflow tensor.
  }];

  let arguments = (ins
    MlrtFutureType:$future
  );

  let results = (outs
    TF_Tensor:$result
  );
}

// tf_mlrt.tf_promise takes a tensorflow Tensor. It is a fake op that is only
// used during parallelization and has no runtime implementation.