}];

  let description = [{
That is for rows we have grad for, we update var, accum and linear as follows:
accum_new = accum + grad * grad
linear += grad - (accum_new^(-lr_power) - accum^(-lr_power)) / lr * var
quadratic = 1.0 / (accum_new^(lr_power) * lr) + 2 * l2
var = (sign(linear) * l1 - linear) / quadratic if |linear| > l1 else 0.0
accum = accum_new
  }];

  let arguments = (ins

1003B         ; disallowed                             # NA   <reserved-1003B>
1003C..1003D  ; valid                                  # 4.0  LINEAR B SYLLABLE B017 ZA..LINEAR B SYLLABLE B074 ZE
1003E         ; disallowed                             # NA   <reserved-1003E>
1003F..1004D  ; valid                                  # 4.0  LINEAR B SYLLABLE B020 ZO..LINEAR B SYLLABLE B091 TWO

    It significantly reduces the peak memory required and latency incurred
    during restore.
*   Added a new library for library of matrix-free (iterative) solvers for
    linear equations, linear least-squares, eigenvalues and singular values in
    tensorflow/contrib/solvers. Initial version has lanczos bidiagonalization,
    conjugate gradients and CGLS.