// `forward_function` is null, a GradientTape is used on the backward function
  // to compute the jvp, which will waste computation when executing eagerly.
  //
  // Unlike GradientTape::RecordOperation, Accumulate runs gradient computation
  // immediately. It stores the results, which feed into Accumulate for future
  // operations and may be fetched by calling FetchJVP. ForwardAccumulator

  // A non-blocking version of `Execute`. After each call, `Join` must be called
  // before `StartExecute` is called again. Using `StartExecute` with `Join`
  // allows the caller to schedule computation on multiple ParallelDevices
  // without sequencing those operations (first call `StartExecute` on each
  // parallel device, then call `Join` on each; even if some of the `Join`s

public final class Quantiles {

  /** Specifies the computation of a median (i.e. the 1st 2-quantile). */
  public static ScaleAndIndex median() {
    return scale(2).index(1);
  }

  /** Specifies the computation of quartiles (i.e. 4-quantiles). */
  public static Scale quartiles() {
    return scale(4);
  }

  /** Specifies the computation of percentiles (i.e. 100-quantiles). */

    // TODO(ycao): Support tensorlist-type output.
    out_desc.is_tensor_list = false;
  }

  // XLA computation always uses Tuple shape.
  *xla_output_shape = xla::ShapeUtil::MakeTupleShape(shapes);
  return absl::OkStatus();
}

// Creates a vector that maps from the parameters of the XLA computation to
// their original argument positions.
// MLIR-based TF-Compiler bridge doesn't have constant analysis yet, thus no

    const XlaCompiler::CompilationResult& result) {
  VLOG(2) << "Compiling to xla::PjRtLoadedExecutable.";

  TF_ASSIGN_OR_RETURN(auto executable,
                      client_->Compile(*result.computation,
                                       GetPjRtCompileOptions(options, result)));

  VLOG(2) << "Compiled PJRT executable " << executable->name()
          << " num_replicas " << executable->num_replicas()

func blockGeneric(dig *digest, p []byte) {
	var w [64]uint32
	h0, h1, h2, h3, h4, h5, h6, h7 := dig.h[0], dig.h[1], dig.h[2], dig.h[3], dig.h[4], dig.h[5], dig.h[6], dig.h[7]
	for len(p) >= chunk {
		// Can interlace the computation of w with the
		// rounds below if needed for speed.
		for i := 0; i < 16; i++ {
			j := i * 4
			w[i] = uint32(p[j])<<24 | uint32(p[j+1])<<16 | uint32(p[j+2])<<8 | uint32(p[j+3])
		}
		for i := 16; i < 64; i++ {

//
// References:
//
// Baillie and Wagstaff, "Lucas Pseudoprimes", Mathematics of Computation 35(152),
// October 1980, pp. 1391-1417, especially page 1401.
// https://www.ams.org/journals/mcom/1980-35-152/S0025-5718-1980-0583518-6/S0025-5718-1980-0583518-6.pdf
//
// Grantham, "Frobenius Pseudoprimes", Mathematics of Computation 70(234),
// March 2000, pp. 873-891.

func blockGeneric(dig *digest, p []byte) {
	var w [64]uint32
	h0, h1, h2, h3, h4, h5, h6, h7 := dig.h[0], dig.h[1], dig.h[2], dig.h[3], dig.h[4], dig.h[5], dig.h[6], dig.h[7]
	for len(p) >= chunk {
		// Can interlace the computation of w with the
		// rounds below if needed for speed.
		for i := 0; i < 16; i++ {
			j := i * 4
			w[i] = uint32(p[j])<<24 | uint32(p[j+1])<<16 | uint32(p[j+2])<<8 | uint32(p[j+3])
		}
		for i := 16; i < 64; i++ {

   *
   * <p>The {@link OutputStream} should not be written to before or after the hand-off.
   */
  // TODO(user): Evaluate whether it makes sense to always piggyback the computation of a
  // HashCode on an existing OutputStream, compared to creating a separate OutputStream that could
  // be (optionally) be combined with another if needed (with something like
  // MultiplexingOutputStream).

    FunctionDef *xla_fdef = fdl.add_function();
    TF_CHECK_OK(GraphToFunctionDef(*g, "f3", xla_fdef));
  }
  FunctionLibraryDefinition fld(OpRegistry::Global(), fdl);

  // Build the XLA computation graph.
  // "arg0" (T=DT_RESOURCE), "arg1" (T=DT_INT32)
  // "arg0", "arg1" -> "if" (If) -> "ret0", "ret1"
  // "arg0", "arg1" -> "while" (While) -> "ret2", "ret3"
  tensorflow::Scope s = tensorflow::Scope::NewRootScope();