Depthwise Separable Convolution

Why aren’t depthwise separable convolutions used more than regular convolution when it has significantly less parameters and performs comparably?

Because they don’t perform comparably, they perform worse.