Architectures for Stereo Vision - Signal Processing Systems - page 492

Fig. 12 Impact of the parallelization configuration on the performance of the concurrent path cost
calculation for eight paths of the SGM for 1280 × 960 images and 128 disparity levels. Block width

and tile width are both fixed to tdx = 32. Best performance is achieved with tdx × tdy = 32 × 4(i.e.

each inner loop processes 32 disparity levels) and t y = 16

of keeping the processing implementation unchanged but rearranging the data in
the memory creates an inherently contradictory situation: if the GPU is used to
rearrange the data, the re-sorting causes additional memory access with is not even
coalesced.
Again, parameters adjustment allows to navigate between the performance
optimization principles. The first parameter ( tdy ) trades thread parallelism against
sequential computation in the inner loop for all kernels. The second parameter
( t y ) trades the number of parallelly processable blocks versus launch overhead and
memory overhead for the four diagonal paths. Figure 12 shows the result of the
parameter study. Choosing tdy = 4and t y = 16 results in best performance (39 . 8ms

and 39 960 image. If the concurrent kernel execution is not

used, performance is approximately halved (75

. 7GB / s) for a 1280 ×

. 7 ms and 20 . 9GB / s). Both kernel

sets, concurrent and sequential, are latency bound.
Summation of the eight path cost spaces ( 7 ) and winner-takes-all disparity
selection ( 9 ) can be performed independently for each pixel allowing for the same
parallelization scheme as for the MC calculation. This kernel ( sum wta ) requires
15

. 1 ms and is memory bound with 117 . 4GB / s.

3.6.4 Performance

The processing time for the complete disparity estimation including rank transform,
semi-global matching for eight paths, disparity map generation (without left/right
check ( 3 ) ) and median filtering on a Tesla C2050 Fermi architecture GPU is
summarized in Table 2 . Overall, a 1280 × 960 image with 128 disparity levels

requires 56 2 ms. The processing times do not include data transfer between host

and GPU because it can be effectively hidden using concurrent data transfer
when processing image streams. When processing 1280

.

× 960 image sets ca. 5 ms

additional transfer time is required.