*/
      sqrt0 = sqrtApproxWithDoubles(x.shiftRight(shift)).shiftLeft(shift >> 1);
    }
    BigInteger sqrt1 = sqrt0.add(x.divide(sqrt0)).shiftRight(1);
    if (sqrt0.equals(sqrt1)) {
      return sqrt0;
    }
    do {
      sqrt0 = sqrt1;
      sqrt1 = sqrt0.add(x.divide(sqrt0)).shiftRight(1);
    } while (sqrt1.compareTo(sqrt0) < 0);
    return sqrt0;
  }

  @GwtIncompatible // TODO

       */
      sqrt0 = sqrtApproxWithDoubles(x.shiftRight(shift)).shiftLeft(shift >> 1);
    }
    BigInteger sqrt1 = sqrt0.add(x.divide(sqrt0)).shiftRight(1);
    if (sqrt0.equals(sqrt1)) {
      return sqrt0;
    }
    do {
      sqrt0 = sqrt1;
      sqrt1 = sqrt0.add(x.divide(sqrt0)).shiftRight(1);
    } while (sqrt1.compareTo(sqrt0) < 0);
    return sqrt0;
  }

  @GwtIncompatible // TODO

	sink64[4] = math.RoundToEven(x)
}

func sqrt(x float64) float64 {
	// amd64:"SQRTSD"
	// 386/sse2:"SQRTSD" 386/softfloat:-"SQRTD"
	// arm64:"FSQRTD"
	// arm/7:"SQRTD"
	// mips/hardfloat:"SQRTD" mips/softfloat:-"SQRTD"
	// mips64/hardfloat:"SQRTD" mips64/softfloat:-"SQRTD"
	// wasm:"F64Sqrt"
	// ppc64x:"FSQRT"
	// riscv64: "FSQRTD"
	return math.Sqrt(x)
}

func sqrt32(x float32) float32 {
	// amd64:"SQRTSS"

	"MULU",
	"MULH",
	"MULHU",
	"MULW",
	"NEGD",
	"NEGF",
	"NEGW",
	"NEGV",
	"NOOP",
	"NOR",
	"OR",
	"REM",
	"REMU",
	"RFE",
	"SC",
	"SCV",
	"SGT",
	"SGTU",
	"SLL",
	"SQRTD",
	"SQRTF",
	"SRA",
	"SRL",
	"ROTR",
	"SUB",
	"SUBD",
	"SUBF",
	"SUBU",
	"SUBW",
	"DBAR",
	"SYSCALL",
	"TEQ",
	"TNE",
	"WORD",
	"XOR",
	"MASKEQZ",
	"MASKNEZ",
	"MOVV",

 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 {
    std::string name = "Sqrt_Grad";
    TF_RETURN_IF_ERROR(SqrtGrad(ctx, sqrt_.get(), grad_outputs[0],

      BigInteger plusHalfSquared = result.pow(2).add(result).shiftLeft(2).add(ONE);
      BigInteger x4 = x.shiftLeft(2);
      // sqrt(x) <= result + 0.5, so 4 * x <= (result + 0.5)^2 * 4
      // (result + 0.5)^2 * 4 = (result^2 + result)*4 + 1
      assertTrue(x4.compareTo(plusHalfSquared) <= 0);
      BigInteger minusHalfSquared = result.pow(2).subtract(result).shiftLeft(2).add(ONE);

        assertEquals(BigIntegerMath.sqrt(valueOf(x), mode), valueOf(IntMath.sqrt(x, mode)));
      }
    }
  }

  /* Relies on the correctness of sqrt(int, FLOOR). */
  @GwtIncompatible // sqrt
  public void testSqrtExactMatchesFloorOrThrows() {
    for (int x : POSITIVE_INTEGER_CANDIDATES) {
      int floor = IntMath.sqrt(x, FLOOR);

	MOVDW	F4, F5			// 85081b01
	NEGF	F4, F5			// 85141401
	NEGD	F4, F5			// 85181401
	ABSD	F4, F5			// 85081401
	TRUNCDW	F4, F5			// 85881a01
	TRUNCFW	F4, F5			// 85841a01
	SQRTF	F4, F5			// 85441401
	SQRTD	F4, F5			// 85481401

	DBAR	 			// 00007238
	NOOP	 			// 00004003

	MOVWR	R4, result+16(FP) 	// 6460402f
	MOVWR	R4, 1(R5) 		// a404402f
	MOVWR	y+8(FP), R4 		// 6440402e

// This decomposition is only correct inside XLA as it ignores use_locking
// attribute.
// alpha <- learning_rate * sqrt(1 - beta2^t) / (1 - beta1^t)
// m_t <- beta1 * m_{t-1} + (1 - beta1) * g_t
// v_t <- beta2 * v_{t-1} + (1 - beta2) * g_t * g_t
// variable <- variable - alpha * m_t / (sqrt(v_t) + epsilon)
def DecomposeResourceApplyAdamNonNesterov :
  Pattern<
    (TF_ResourceApplyAdamOp:$src_op

        assertEquals(BigIntegerMath.sqrt(valueOf(x), mode), valueOf(LongMath.sqrt(x, mode)));
      }
    }
  }

  /* Relies on the correctness of sqrt(long, FLOOR). */
  @GwtIncompatible // TODO
  public void testSqrtExactMatchesFloorOrThrows() {
    for (long x : POSITIVE_LONG_CANDIDATES) {
      long sqrtFloor = LongMath.sqrt(x, FLOOR);