Can it be guaranteed that if i load a fixed point GL_FIXED matrix into opengles, instruxtions such as rotation using glScalex and glRotatex will be performed using fixed-point calculations?

If the opengl renderer casts (or converts) the values to GLfloat, how does any speed increase happen?

Can i count on the implementation to performs matrix calculations in fixed point when specifying a fixed-point matrix and using the 'x' equivalent functions to manipulate the contents?

Basically, do i have to implement my own matrix class (as proposed by Phil Atkin at NVIDIA) in order to secure fixed point math?