Optimal Newton 3×3 solver step-kernel (in: DSP and Plugin Devel…)

this seems to be 3 times faster and more accurate

<code>
/* 3×3 Newton-step kernel, Cramer / adjugate version.
*
* Drop-in for newton3x3_solve(): same signature, same return convention.
* Solves J·dx = -F for dx from the augmented matrix A[3][4] (columns 0..2
* hold J, column 3 the RHS). A is only read, not destroyed.
*
* dx = adj(J) · rhs / det(J): no pivoting, no branches apart from the
* singular check, one division. jmax is caller-computed over the 3×3
* Jacobian, as before…

Go to this post

Go to first unread post in this topic