if err != nil {
			t.Fatal(err)
		}
		aliceSecret, err := aliceKey.ECDH(bobKey.PublicKey())
		if err != nil {
			t.Fatal(err)
		}

		if !bytes.Equal(bobSecret, aliceSecret) {
			t.Error("two ECDH computations came out different")
		}
	})
}

type countingReader struct {
	r io.Reader
	n int
}

func (r *countingReader) Read(p []byte) (int, error) {
	n, err := r.r.Read(p)
	r.n += n
	return n, err

void AppendMarkForCompilationPassFlagsInternal(std::vector<Flag>* flag_list) {
  std::vector<Flag> new_flags = {
      Flag("tf_xla_auto_jit", SetterForXlaAutoJitFlag, "0",
           "Control compilation of operators into XLA computations on CPU and "
           "GPU devices.  0 = use ConfigProto setting; -1 = off; 1 = on for "
           "things very likely to be improved; 2 = on for everything; "

	// 	x1:x0 := r.Uint64()
	// 	0:hi, lo1:lo0 := bits.Mul64(x1:x0, 0:n)
	// Writing out the multiplication in terms of bits.Mul32 allows
	// using direct hardware instructions and avoiding
	// the computations involving these zeros.
	x := r.Uint64()
	lo1a, lo0 := bits.Mul32(uint32(x), n)
	hi, lo1b := bits.Mul32(uint32(x>>32), n)
	lo1, c := bits.Add32(lo1a, lo1b, 0)
	hi += c
	if lo1 == 0 && lo0 < uint32(n) {

	PAND POLY, T0
	PSRLL $31, T1
	PSLLDQ $4, T1
	PSLLL $1, B0
	PXOR T0, B0
	PXOR T1, B0
	// Karatsuba pre-computations
	MOVOU B0, (16*14)(dst)
	PSHUFD $78, B0, B1
	PXOR B0, B1
	MOVOU B1, (16*15)(dst)

	MOVOU B0, B2
	MOVOU B1, B3
	// Now prepare powers of H and pre-computations for them
	MOVQ $7, AX

initLoop:
		MOVOU B2, T0
		MOVOU B2, T1
		MOVOU B3, T2
		PCLMULQDQ $0x00, B0, T0

    // done().
    // The latter may be too conservative, but given the host is involved in
    // waiting for the transfer to complete anyway there is probably little
    // downside. If we were to add the ability for computations to wait directly
    // on transfers, then we might want to rethink this property.
    // Also ideally this host callback should be on source stream rather than

        //   ApplyWeightUpdates_2 -> "iteration++"
        //   ApplyWeightUpdates_0 -> Computation_A
        //   ApplyWeightUpdates_1 -> Computation_B
        //   ApplyWeightUpdates_2 -> Computation_C
        //   Computation_A -> NoOp
        //   Computation_B -> NoOp
        //   Computation_C -> NoOp
        //   "iteration++" -> NoOp
        // }
        //

	fieldCache.RLock()
	f := fieldCache.m[t]
	fieldCache.RUnlock()
	if f != nil {
		return f
	}

	// Compute fields without lock.
	// Might duplicate effort but won't hold other computations back.
	f = typeFields(t)
	if f == nil {
		f = []field{}
	}

	fieldCache.Lock()
	if fieldCache.m == nil {
		fieldCache.m = map[reflect.Type][]field{}
	}
	fieldCache.m[t] = f

			ok = true
		}
	}
	if !ok {
		return bigFtoa(dst, prec, fmt, neg, mant, exp, flt)
	}
	return formatDigits(dst, shortest, neg, digs, prec, fmt)
}

// bigFtoa uses multiprecision computations to format a float.
func bigFtoa(dst []byte, prec int, fmt byte, neg bool, mant uint64, exp int, flt *floatInfo) []byte {
	d := new(decimal)
	d.Assign(mant)
	d.Shift(exp - int(flt.mantbits))
	var digs decimalSlice

// between theoretical and numerical jacobians:
//   fabs(jac_t - jac_n) / max(fabs(jac_t), fabs(jac_n))

// TODO(andydavis) Vectorize and/or multi-thread Jacobian computations if
// performance becomes an issue.

// BaseUnitsForType provides a list of typed unit values for each basis in the
// requested type.
// When T is real,

  // `forward_function` is null, a GradientTape is used on the backward function
  // to compute the jvp, which will waste computation when executing eagerly.
  //
  // Unlike GradientTape::RecordOperation, Accumulate runs gradient computation
  // immediately. It stores the results, which feed into Accumulate for future
  // operations and may be fetched by calling FetchJVP. ForwardAccumulator