x4 := new(Rat).SetInt(new(Int).Binomial(int64(i+j), int64(i)))

			x1.Mul(x1, x2)
			x1.Mul(x1, x3)
			x1.Mul(x1, x4)
			x1.Mul(x1, x4)

			if (i+j)&1 != 0 {
				x1.Neg(x1)
			}

			a.set(i, j, x1)
		}
	}
	return a
}

func (a *matrix) mul(b *matrix) *matrix {
	if a.m != b.n {
		panic("illegal matrix multiply")
	}
	c := newMatrix(a.n, b.m)

(Select0 m:(LoweredMuluhilo x y)) && m.Uses == 1 => (MULHU x y)
(Select1 m:(LoweredMuluhilo x y)) && m.Uses == 1 => (MUL x y)

(FADD(S|D) a (FMUL(S|D) x y)) && a.Block.Func.useFMA(v) => (FMADD(S|D) x y a)
(FSUB(S|D) a (FMUL(S|D) x y)) && a.Block.Func.useFMA(v) => (FNMSUB(S|D) x y a)
(FSUB(S|D) (FMUL(S|D) x y) a) && a.Block.Func.useFMA(v) => (FMSUB(S|D) x y a)

// Merge negation into fused multiply-add and multiply-subtract.
//

  %1 = "tfl.squared_difference"(%arg0, %0) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("squared_difference")
  %2 = "tfl.mul"(%arg0, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("mul")
  %3 = "tfl.div"(%2, %1) {fused_activation_function = "NONE"} : (tensor<4xf32>, tensor<4xf32>) -> tensor<4xf32> loc("div")

  %1 = "tfl.mul"(%arg0, %0) {fused_activation_function = "NONE"} : (tensor<3x!quant.uniform<i8:f32, 0.1>>, tensor<3x!quant.uniform<i8:f32, 0.1>>) -> tensor<3x!quant.uniform<i8:f32, 0.1>> loc("mul")
  func.return %1 : tensor<3x!quant.uniform<i8:f32, 0.1>>

  %1 = "tfl.mul"(%arg0, %0) {fused_activation_function = "NONE"} : (tensor<3x!quant.uniform<i8:f32, 1.0>>, tensor<3x!quant.uniform<i8:f32, 1.0>>) -> tensor<3x!quant.uniform<i8:f32, 1.0>> loc("mul")
  func.return %1 : tensor<3x!quant.uniform<i8:f32, 1.0>>

// conditions are also hoisted (i.e., `tf.Add` and `tf.Mul`).

// CHECK-LABEL: readvariableop_is_hoisted_if_readonly
// CHECK:       [[CST_0:%.*]] = "tf.Const"
// CHECK:       [[VAR:%.*]] = "tf.VarHandleOp"
// CHECK:       [[CST_1:%.*]] = "tf.Const"
// CHECK:       [[VAR_VAL:%.*]] = "tf.ReadVariableOp"
// CHECK:       [[RES_1:%.*]] = "tf.Add"([[VAR_VAL]], [[CST_0]])
// CHECK:       [[RES_2:%.*]] = "tf.Mul"([[RES_1]], [[CST_1]])

	WFMADB	V1, V4, V5, V4
	FMOVD	16(R5), F5
	WFMADB	V6, V3, V7, V3
	FMOVD	8(R5), F7
	WFMADB	V2, V7, V5, V7
	FMOVD	0(R5), F5
	WFMADB	V2, V5, V16, V5
	WFMADB	V1, V4, V3, V4
	WFMADB	V6, V7, V5, V6
	FMUL	F0, F2
	FMADD	F4, F1, F6
	FMADD	F6, F2, F0
	MOVW	R2, R6
	MOVW	R1, R7
	CMPUBLE	R6, R7, L1
	MOVD	$·atanxpim<>+0(SB), R1
	WORD	$0xED801000	//madb	%f0,%f8,0(%r1)
	BYTE	$0x00
	BYTE	$0x1E
L1:

	// X=Z1; Y=Z1; MUL; T-   // T1 = Z1*Z1
	// X-  ; Y=T ; MUL; R=T  // R  = Z1*T1
	// X=X2; Y-  ; MUL; H=T  // H  = X2*T1
	// X=Z2; Y=Z2; MUL; T-   // T2 = Z2*Z2
	// X-  ; Y=T ; MUL; S1=T // S1 = Z2*T2
	// X=X1; Y-  ; MUL; U1=T // U1 = X1*T2
	// SUB(H<H-T)            // H  = H-U1
	// X=Z1; Y=Z2; MUL; T-   // Z3 = Z1*Z2

  op: "Mul"
  input: "loss_1/res4c_branch2c/kernel/Regularizer/mul/x"
  input: "loss_1/res2a_branch1/kernel/Regularizer/Sum"
  device: "/job:localhost/replica:0/task:0/device:GPU:0"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
}
node {
  name: "loss_1/conv1/kernel/Regularizer/mul"
  op: "Mul"
  input: "loss_1/res4c_branch2c/kernel/Regularizer/mul/x"

    long mul = K3;
    int topBit = 0x7;

    int lengthAligned = length & ~topBit;
    int lengthRemainder = length & topBit;
    long hash = seed ^ (length * mul);

    for (int i = 0; i < lengthAligned; i += 8) {
      long loaded = load64(bytes, offset + i);
      long data = shiftMix(loaded * mul) * mul;
      hash ^= data;
      hash *= mul;
    }