The usual approach is to plot the normalised values, not log (base 2) fold-changes. One first identifies statistically significant genes an then subsets their normalised expression matrix such that it only includes the statistically significant genes. From this subset, hierarchical clustering can be performed (with a heatmap being generated, in addition) for the purposes of determining how well your selection of statistically significant genes can actually segregate your conditions that were compared during differential expression.

I and others have posted a lot of answers with code showing how you can generate a dendrogram with heatmap from any matrix of values:

gplots

• A: Heatmap based with FPKM values
• A: Hierarchical Clustering in single-channel agilent microarray experiment

ComplexHeatmap

• A: Complex Heatmap: Changing order of clusters
• C: Hierarchical Clustering in single-channel agilent microarray experiment
• C: how to cluster genes in heatmap
• A: How to plot a heatmap with two different distance matrices for X and Y

pheatmap:

• A: extract dendrogram cluster from pheatmap

Other:

• A: how to draw circular dendrogram with distance information

Kevin

I have the data from DESeq2 but I don't know if I can use it directly in the heatmap or I need to do some additional normalization before that.

You can, in the sense it will produce a heatmap, but it won't be useful. You certainly need to perform further normalization (but some heatmap functions have normalization built-in) for the heatmap produce meaningful visualization. I would suggest you follow one of the outstanding tutorial out there, from DESeq2 vignette to Ming Tang's Heatmap demystified, not to mention some of Kevin Blighe's posts on the subject.