Non-separable 2D, 3D, and 4D Filtering with CUDA - GPU Pro: Advanced Rendering Techniques - page 480

if (( x < DATA_W )&&( y < DATA_H ))
Filter_Response [ Get2DIndex ( x , y , DATA_W )] =
Conv2D ( s_Image , threadIdx . y + HALO , threadIdx . x + HALO );

if ((( x + 32) < DATA_W )&&( y < DATA_H ))
Filter_Response [ Get2DIndex ( x +32, y , DATA_W )] =
Conv2D ( s_Image , threadIdx . y + HALO , threadIdx . x +32+ HALO );

− HALO ￿ 2))
{ if ((( x + 64) < DATA_W )&&( y < DATA_H ))
Filter_Response [ Get2DIndex ( x +64, y , DATA_W )] =
Conv2D ( s_Image , threadIdx . y + HALO , threadIdx . x +64+ HALO );

if ( threadIdx . x < (32

}

if ( threadIdx . y < (32 − HALO ￿ 2))
{ if (( x < DATA_W )&&(( y + 32) < DATA_H ))
Filter_Response [ Get2DIndex ( x , y +32, DATA_W )] =
Conv2D ( s_Image , threadIdx . y +32+ HALO , threadIdx . x + HALO );

}

if ( threadIdx . y < (32 − HALO ￿ 2))
{ if ((( x + 32) < DATA_W )&&(( y + 32) < DATA_H ))
Filter_Response [ Get2DIndex ( x +32, y +32, DATA_W )] =
Conv2D ( s_Image , threadIdx . y +32+ HALO , threadIdx . x +32+ HALO );

}

if (( threadIdx . x < (32 − HALO ￿ 2)) &&
( threadIdx . y < (32 − HALO ￿ 2)) )
{ if ((( x + 64) < DATA_W )&&(( y + 32) < DATA_H ))
Filter_Response [ Get2DIndex ( x +64, y +32, DATA_W )] =
Conv2D ( s_Image , threadIdx . y +32+ HALO , threadIdx . x +64+ HALO );

}

}

Listing 5.3. Non-separable 2D convolution using shared memory. This listing represents
the second part of the kernel where the convolutions are performed, by calling a device
function for each block of filter responses. For a filter size of 17 × 17, HALO is 8 and the
first two blocks yield filter responses for 32 × 32 pixels, the third block yields 16 × 32
filter responses, the fourth and fifth blocks yield 32

16 filter responses and the sixth
block yields the last 16 × 16 filter responses. The parameter HALO can easily be changed
to optimize the code for different filter sizes. The code for the device function Conv2D
is given in the repository.

×

5.6 Non-separable 3D Convolution

Three-dimensional convolution between a signal s and a filter f for position
[ x, y, z ] is defined as

f x = N/ 2 f y = N/ 2 f z = N/ 2

( s

∗ f )[ x, y, z ]= s [ x − f x ,y − f y ,z − f z ] · f [ f x ,f y ,f z ] .

f x = −N/ 2 f y = −N/ 2 f z = −N/ 2