uv3 := new(Element).Multiply(u, t0.Multiply(v2, v))
	uv7 := new(Element).Multiply(uv3, t0.Square(v2))
	rr := new(Element).Multiply(uv3, t0.Pow22523(uv7))

	check := new(Element).Multiply(v, t0.Square(rr)) // check = v * r^2

	uNeg := new(Element).Negate(u)
	correctSignSqrt := check.Equal(u)
	flippedSignSqrt := check.Equal(uNeg)
	flippedSignSqrtI := check.Equal(t0.Multiply(uNeg, sqrtM1))

  // CHECK-DAG: %[[VARIANCE_EPS_RSQRT:.+]] = mhlo.rsqrt %[[VARIANCE_EPS]] : tensor<256xf32>
  // CHECK-DAG: %[[MULTIPLIER:.+]] = mhlo.multiply %[[VARIANCE_EPS_RSQRT]], %[[SCALE]] : tensor<256xf32>
  // CHECK-DAG: %[[MUL_MEAN:.+]] = mhlo.multiply %[[MULTIPLIER]], %[[MEAN]] : tensor<256xf32>
  // CHECK-DAG: %[[RHS:.+]] = mhlo.subtract %[[OFFSET]], %[[MUL_MEAN]] : tensor<256xf32>

  %0 = "mhlo.broadcast_in_dim"(%cst0) <{broadcast_dimensions = dense<3> : tensor<1xi64>}> : (tensor<4xi32>) -> tensor<1x1x2x4xi32>
  // CHECK: %[[MUL:.*]] = mhlo.multiply %[[BROADCAST]], %[[ARG]] : tensor<1x1x2x4xi32>
  %1 = mhlo.multiply %0, %arg0 : tensor<1x1x2x4xi32>
  // CHECK:      return %[[MUL]] : tensor<1x1x2x4xi32>
  func.return %1 : tensor<1x1x2x4xi32>

  // CHECK-DAG: %[[CST_BCAST:.+]] = "mhlo.broadcast_in_dim"(%[[CST]]) <{broadcast_dimensions = dense<3> : tensor<1xi64>}> : (tensor<2xf32>) -> tensor<1x1x3x2xf32>
  // CHECK-DAG: %[[NEW_FILTER:.+]] =  mhlo.multiply %[[CST_BCAST]], %[[FILTER]] : tensor<1x1x3x2xf32>

	op_VME    uint32 = 0xE7A6 // 	VRR-c	VECTOR MULTIPLY EVEN
	op_VMH    uint32 = 0xE7A3 // 	VRR-c	VECTOR MULTIPLY HIGH
	op_VMLE   uint32 = 0xE7A4 // 	VRR-c	VECTOR MULTIPLY EVEN LOGICAL
	op_VMLH   uint32 = 0xE7A1 // 	VRR-c	VECTOR MULTIPLY HIGH LOGICAL
	op_VMLO   uint32 = 0xE7A5 // 	VRR-c	VECTOR MULTIPLY ODD LOGICAL
	op_VML    uint32 = 0xE7A2 // 	VRR-c	VECTOR MULTIPLY LOW
	op_VMO    uint32 = 0xE7A7 // 	VRR-c	VECTOR MULTIPLY ODD

func NewScalar() *Scalar {
	return &Scalar{}
}

// MultiplyAdd sets s = x * y + z mod l, and returns s. It is equivalent to
// using Multiply and then Add.
func (s *Scalar) MultiplyAdd(x, y, z *Scalar) *Scalar {
	// Make a copy of z in case it aliases s.
	zCopy := new(Scalar).Set(z)
	return s.Multiply(x, y).Add(s, zCopy)
}

// Add sets s = x + y mod l, and returns s.
func (s *Scalar) Add(x, y *Scalar) *Scalar {
	// s = 1 * x + y mod l

  %2 = stablehlo.divide %arg0, %0 : tensor<2xf32>
  return %2 : tensor<2xf32>
}

// CHECK-LABEL: divideToMulReciprocalSplat
// CHECK: stablehlo.constant dense<5.000000e-01> : tensor<2xf32>
// CHECK: stablehlo.multiply

// -----

func.func @divideToMulReciprocal(%arg0: tensor<2xf32>) -> tensor<2xf32> {
  %0 = stablehlo.constant dense<[2.0, 3.0]> : tensor<2xf32>
  %2 = stablehlo.divide %arg0, %0 : tensor<2xf32>

def TFL_ArithmeticCount : OpInterface<"TflArithmeticCountOpInterface"> {
  let description = [{
    Interface for TFLite ops to calculate arithmetic count (Multiply-Add Count).
  }];

  let methods = [
    StaticInterfaceMethod<
      [{Returns an integer representing the op's arithmetic count (Multiply-Add
      Count), return -1 if the arithmetic count cannot be determined.}],
       "int64_t", "GetArithmeticCount", (ins "Operation*":$op)
    >,

					}
					atomic.AddInt64(&ops, int64(n))
					atomic.AddInt64(&lat, latency)
				})
				spent := time.Since(t)
				if spent > 0 && n > 0 {
					// Since we are benchmarking n parallel servers we need to multiply by n.
					// This will give an estimate of the total ops/s.
					latency := float64(atomic.LoadInt64(&lat)) / float64(time.Millisecond)
					b.ReportMetric(float64(n)*float64(ops)/spent.Seconds(), "vops/s")

	// This does not seem worthwhile, at least for Go: not having any high
	// bits in the multiplier reduces the effect of low bits on the highest bits,
	// and it only saves 1 multiply out of 3.
	// (On 32-bit systems, it saves 1 out of 6, since Mul64 is doing 4.)
	const (
		mulHi = 2549297995355413924
		mulLo = 4865540595714422341
		incHi = 6364136223846793005
		incLo = 1442695040888963407
	)