%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>

	// Compute new scale matrix in current mouse position
	var k = root.createSVGMatrix().translate(p.x, p.y).scale(z).translate(-p.x, -p.y);

	setCTM(g, g.getCTM().multiply(k));

	if(typeof(stateTf) == "undefined")
		stateTf = g.getCTM().inverse();

	stateTf = stateTf.multiply(k.inverse());
}

/**
 * Handle mouse move event.
 */
function handleMouseMove(evt) {
	if(evt.preventDefault)
		evt.preventDefault();

        }

        @Override
        public BigDecimal scaleTo(BigDecimal value, Units<Q> units) {
            if (units == this) {
                return value;
            }
            if (units.equals(baseUnits)) {
                return value.multiply(factor);
            }

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")

//   ⎡x / c⎤ = ⎣x * m / 2^(n+s)⎦ + 1
// (TODO: derivation?)
//
// The multiply is a bit odd, as it is a signed n-bit value
// times an unsigned n-bit value.  For n smaller than the
// word size, we can extend x and m appropriately and use the
// signed multiply instruction.  For n == word size,
// we must use the signed multiply high and correct
// the result by adding x*2^n.
//

	// 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
	)

    %16 = stablehlo.subtract %10, %15 : tensor<1x3x3x4xf32>
    %17 = stablehlo.broadcast_in_dim %4, dims = [0, 1, 2, 3] : (tensor<1x1x1x4xf32>) -> tensor<1x3x3x4xf32>
    %18 = stablehlo.multiply %16, %17 : tensor<1x3x3x4xf32>
    %19 = call @uniform_quantize_1(%18, %5, %6) : (tensor<1x3x3x4xf32>, tensor<1x1x1x1xf32>, tensor<1x1x1x1xi8>) -> tensor<1x3x3x4xi8>

	MOVD 16(R3), R10 // h2
	MOVD 24(R3), R11 // r0
	MOVD 32(R3), R12 // r1

	CMP R5, $16
	BLT bytes_between_0_and_15

loop:
	POLY1305_ADD(R4, R8, R9, R10, R20, R21, R22)

	PCALIGN $16
multiply:
	POLY1305_MUL(R8, R9, R10, R11, R12, R16, R17, R18, R14, R20, R21)
	ADD $-16, R5
	CMP R5, $16
	BGE loop

bytes_between_0_and_15:
	CMP  R5, $0
	BEQ  done
	MOVD $0, R16 // h0
	MOVD $0, R17 // h1

	HasAVX512IFMA       bool // Advanced vector extension 512 Integer Fused Multiply Add
	HasAVX512VBMI       bool // Advanced vector extension 512 Vector Byte Manipulation Instructions
	HasAVX5124VNNIW     bool // Advanced vector extension 512 Vector Neural Network Instructions Word variable precision
	HasAVX5124FMAPS     bool // Advanced vector extension 512 Fused Multiply Accumulation Packed Single precision