Thanks to for clarifying mod 1 concept , for constant $\mathcal{A}$ I got the proper explanation in CLRS itself it says :
Although this method works with any value of the constant $\mathcal{A}$, it works better
with some values than with others. The optimal choice depends on the characteristics of the data being hashed. Knuth suggests that
$$
\mathcal{A} \approx (\sqrt(5) - 1)/div 2 = 0.6180339887 ...
$$
is likely to work reasonably well.
As an example, suppose we have $ \mathcal{k} = 123456, \mathcal{p} = 14, \mathcal{m} = 2^{14} = 16384,$ and $ \mathcal{w} = 32.$ Adapting Knuth’s suggestion, we choose $\mathcal{A}$ to be the fraction of the form $$\mathcal{s}/2^{32}$$ that is closest to $$(\sqrt(5) - 1)/ 2,$$ so that $$ \mathcal{A} = 2654435769/2^{32}.$$
Then $$\mathcal{k} . \mathcal{s} = 327706022297664 =(76300 . 2^{32} )+ 17612864,$$ and so $\mathcal{r_{1}} = 76300$ and $\mathcal{r_{0}} = 17612864.$ The 14 most significant bits of $ \mathcal{r_{0}} $ yield the value $\mathcal{h}(k) = 67$.