GPU Pro 4

3. Decoupled Deferred Shading on the GPU 91

are “interesting” for the fragments. We therefore did not attempt to simulate an

LRU cache, as suggested by Ragan-Kelley et al. On a current GPU there can

be several thousand fragments rasterized in parallel, thus the global cache should

also be able to hold a similar magnitude of samples to achieve a good hit rate.

In a na¨ıve implementation, a thread could query the buffer tail, and check the

last N items in the CG-buffer. Of course the latency of the iterative memory

accesses would be prohibitively high. We now show a cache implementation that

performs cache lookups with only one buffer load and an atomic ...