Architectures for Stereo Vision - Signal Processing Systems - page 489

Fig. 8 Performance of the

3 3 median filter: on

1280

×

960 images as the
parallelization configuration
changes. Block width is fixed
to tdx

×

32. Best performance
is achieved with
tdx

=

× tdy = 32 × 4and n ppt = 4

Fig. 9 3 median filter: comparison of the texture memory kernel and the

proposed shared memory kernel on a Tesla C2050 GPU for the best-performing parallelization
configuration

Performance of the 3 ×

The median filter is always compute bound and performs best with tdx × tdy =

32 × 4 threads and n ppt = 4. The results of the parameter study for tdx = 32 are shown

in Fig. 8 . Configurations with n ppt = 8 perform slightly worse although redundant

memory access is further reduced because of inefficient pipeline utilization. Pro-
cessing times for a 3 3 median filter (i.e. kernel radius K = × 1) are given in Fig. 9

resulting in 0 64 ms for the new shared memory based kernel. For a texture-memory

based kernel, which is the most often suggested way of implementing a 2D non-
separable filter, processing time is which is 2

.

. 77 ms. In comparison, this yields a

speed-up of 4 . 3 when processing a 1280 × 960 image.

9 rank transform (i.e. K = For a 9 × 4) experiments showed that a block size

of tdx × tdy = 32 × 4 with n ppt = 4 yields best performance. A speed up of 4 . 0is

obtained switching from the texture-based kernel (3 . 13 ms) to the shared memory

kernel (0 . 78 ms) for 1280 × 960 images.