// CHECK-DAG: %[[EPS_BCAST:.+]] = mhlo.constant dense<1.001000e-05> : tensor<256xf32>
  // CHECK-DAG: %[[VARIANCE_EPS:.+]] = mhlo.add %[[VARIANCE]], %[[EPS_BCAST]] : tensor<256xf32>
  // CHECK-DAG: %[[STDDEV:.+]] = mhlo.sqrt %[[VARIANCE_EPS]] : tensor<256xf32>
  // CHECK-DAG: %[[STDDEV_BCAST:.+]] = "mhlo.broadcast_in_dim"(%[[STDDEV]]) <{broadcast_dimensions = dense<1> : tensor<1xi64>}> : (tensor<256xf32>) -> tensor<4x256xf32>

	}

	// Expect a uniform distribution of byte values, which lie in [0, 255].
	var (
		mean       = 255.0 / 2
		stddev     = 256.0 / math.Sqrt(12.0)
		errorScale = stddev / math.Sqrt(float64(n))
	)

	expected := &statsResults{mean, stddev, 0.10 * errorScale, 0.08 * errorScale}

	// Cast bytes as floats to use the common distribution-validity checks.

 private:
  AbstractTensorHandlePtr exp_;
};

class SqrtGradientFunction : public GradientFunction {
 public:
  explicit SqrtGradientFunction(AbstractTensorHandle* sqrt) : sqrt_(sqrt) {
    sqrt->Ref();
  }
  Status Compute(AbstractContext* ctx,
                 absl::Span<AbstractTensorHandle* const> grad_outputs,
                 absl::Span<AbstractTensorHandle*> grad_inputs) override {

        non_feature_dims.push_back(i);
      }
      auto reduce_dims = GetI64ElementsAttr(non_feature_dims, &rewriter);
      auto scalar_broadcast_dims = rewriter.getDenseI64ArrayAttr({});

      // scratch1 = rsqrt(var + epsilon)
      RankedTensorType scalar_float =
          tensorflow::GetTypeFromTFTensorShape({}, kernel_type);
      auto epsilon = rewriter.create<ConstantOp>(

            sumSquares = sumSquares.add(diff);
        }
        // This isn't quite right, as we may lose precision when converting to a double
        BigDecimal result = BigDecimal.valueOf(Math.sqrt(sumSquares.divide(BigDecimal.valueOf(size()), RoundingMode.HALF_UP).doubleValue())).setScale(2, RoundingMode.HALF_UP);

        standardError = Amount.valueOf(result, baseUnits);
    }

    public Amount<Q> getAverage() {

	var c float64
	// find c such that Prob(mean-c*stddev < x < mean+c*stddev)^N > .9999
	for c = 0.0; math.Pow(math.Erf(c/math.Sqrt(2)), float64(N)) < .9999; c += .1 {
	}
	c *= 11.0 // allowed slack: 40% to 60% - we don't need to be perfectly random
	mean := .5 * REP
	stddev := .5 * math.Sqrt(REP)
	low := int(mean - c*stddev)
	high := int(mean + c*stddev)
	for i := 0; i < n; i++ {
		for j := 0; j < hashSize; j++ {

	"FCPSGNCC",
	"FRES",
	"FRESCC",
	"FRIM",
	"FRIMCC",
	"FRIP",
	"FRIPCC",
	"FRIZ",
	"FRIZCC",
	"FRIN",
	"FRINCC",
	"FRSQRTE",
	"FRSQRTECC",
	"FSEL",
	"FSELCC",
	"FSQRT",
	"FSQRTCC",
	"FSQRTS",
	"FSQRTSCC",
	"CNTLZD",
	"CNTLZDCC",
	"CMPW",
	"CMPWU",
	"CMPB",
	"FTDIV",
	"FTSQRT",
	"DIVD",
	"DIVDCC",
	"DIVDE",
	"DIVDECC",
	"DIVDEU",

		tables[tableIndex].Select(t, windowValue)
		p.Add(p, t)
		tableIndex--
	}

	return p, nil
}

// {{.p}}Sqrt sets e to a square root of x. If x is not a square, {{.p}}Sqrt returns
// false and e is unchanged. e and x can overlap.
func {{.p}}Sqrt(e, x *{{ .Element }}) (isSquare bool) {
	candidate := new({{ .Element }})
	{{.p}}SqrtCandidate(candidate, x)
	square := new({{ .Element }}).Square(candidate)

	res := new(statsResults)
	var sum, squaresum float64
	for _, s := range samples {
		sum += s
		squaresum += s * s
	}
	res.mean = sum / float64(len(samples))
	res.stddev = math.Sqrt(squaresum/float64(len(samples)) - res.mean*res.mean)
	return res
}

func checkSampleDistribution(t *testing.T, samples []float64, expected *statsResults) {
	t.Helper()
	actual := getStatsResults(samples)

		{name: "FSQRT", argLength: 1, reg: fp11, asm: "FSQRT"},                                          // sqrt(arg0) (floating point)
		{name: "FSQRTS", argLength: 1, reg: fp11, asm: "FSQRTS"},                                        // sqrt(arg0) (floating point, single precision)