// workbufAlloc is the number of bytes to allocate at a time
	// for new workbufs. This must be a multiple of pageSize and
	// should be a multiple of _WorkbufSize.
	//
	// Larger values reduce workbuf allocation overhead. Smaller
	// values reduce heap fragmentation.
	workbufAlloc = 32 << 10
)

func init() {
	if workbufAlloc%pageSize != 0 || workbufAlloc%_WorkbufSize != 0 {
		throw("bad workbufAlloc")
	}
}

	n := uint(len(rr.limbs))
	mLen := uint(m.BitLen())
	logR := _W * n

	// We start by computing R = 2^(_W * n) mod m. We can get pretty close, to
	// 2^⌊log₂m⌋, by setting the highest bit we can without having to reduce.
	rr.limbs[n-1] = 1 << ((mLen - 1) % _W)
	// Then we double until we reach 2^(_W * n).
	for i := mLen - 1; i < logR; i++ {
		rr.Add(rr, m)
	}

	}

	f2 := func(fe, r Element) bool {
		r.SetBytes(fe.Bytes())

		// Intentionally not using Equal not to go through Bytes again.
		// Calling reduce because both Generate and SetBytes can produce
		// non-canonical representations.
		fe.reduce()
		r.reduce()
		return fe == r
	}
	if err := quick.Check(f2, nil); err != nil {
		t.Errorf("failed FE->bytes->FE round-trip: %v", err)
	}

	// carry and partially reduce the partial products
	REDUCE(H_0, H_1, H_2, H_3, H_4)

	CMPBNE R3, $0, loop

finish:
	// sum lane 0 and lane 1 and put the result in lane 1
	VZERO  T_0
	VSUMQG H_0, T_0, H_0
	VSUMQG H_3, T_0, H_3
	VSUMQG H_1, T_0, H_1
	VSUMQG H_4, T_0, H_4
	VSUMQG H_2, T_0, H_2

	// reduce again after summation

==== Compile-time dependencies

Maven automatically exposes dependencies using its implicit `compile` scope to the consumers of that project.
This behavior is undesirable, and Gradle takes steps to help library authors reduce their API footprint using the `api` and `implementation` configurations of the `java-library` plugin.

mapreduce.fileoutputcommitter.algorithm.version=2 # Use the latest committer version
mapreduce.job.reduce.slowstart.completedmaps=0.99 # 99% map, then reduce
mapreduce.reduce.shuffle.input.buffer.percent=0.9 # Min % buffer in RAM
mapreduce.reduce.shuffle.merge.percent=0.9 # Minimum % merges in RAM
mapreduce.reduce.speculative=false # Disable speculation for reducing
mapreduce.task.io.sort.factor=999 # Threshold before writing to disk

    private
    fun partialEvaluationOf(
        program: Program,
        programKind: ProgramKind,
        programTarget: ProgramTarget
    ): ResidualProgram = PartialEvaluator(programKind, programTarget).reduce(program)

    private
    fun isResidualProgram(program: ResidualProgram) =
        equalTo(program)

                                        out_dims);
  }

  Value reduced = builder.create<TF::EinsumOp>(
      loc, RankedTensorType::get(output_shape, builder.getIntegerType(32)),
      input_arguments, builder.getStringAttr(einsum_equation));

  reduced.getDefiningOp()->setAttr(
      kTfQuantCreatedEinsum,
      BoolAttr::get(reduced.getDefiningOp()->getContext(), true));

  %1 = mhlo.constant dense<false> : tensor<i1>
  %2 = mhlo.constant dense<0> : tensor<i32>
  %3:2 = mhlo.reduce(%arg0 init: %1), (%0 init: %2) across dimensions = [0] : (tensor<2xi1>, tensor<2xi32>, tensor<i1>, tensor<i32>) -> (tensor<i1>, tensor<i32>)
    reducer(%arg1: tensor<i1>, %arg3: tensor<i1>) (%arg2: tensor<i32>, %arg4: tensor<i32>)  {

  if (!has_repeated_labels &&
      input_subs_without_reduced_labels == output_subs) {
    // Obtain the shape of the output, as if keepdims=True on reduce sum. E.g.
    // for the equation "abcd->ac" with input shape [2,5,3,4], we get the
    // reduced shape [2,1,3,1].
    auto reduced_shape = ReducedShapeHelper(scope, input_shape, reduced_axes);
    // Reshaping the gradient (wrt "ac") to [2,1,3,1] and broadcasting it to