Here's a sample program output. In this program I have calculated the features' range of coordinates in the horizontal axis, as well as on the vertical axis. Notice how on the wide image ("Two squares horizontal"), the CPU's max X (horizontal) value is significantly less than CUDA's max X value. Similarly, on the tall image, the CPU's max Y (vertical) value is significantly less than CUDA's max Y value. Also notice how the CPU detects less features (actually, half) than CUDA does:

(This is using arrayfire at 8ecbe61)

*** Two Squares Horizontal ***
Rows: 400
Cols: 800
CPU features: 16
CUDA features: 32
CPU range: 
X: 100,299
Y: 100,299
CUDA range: 
X: 100,699
Y: 100,299
*** Two Squares Vertical ***
Rows: 800
Cols: 400
CPU features: 16
CUDA features: 32
CPU range: 
X: 100,299
Y: 100,299
CUDA range: 
X: 100,299
Y: 100,699