The R-LRPD test: speculative parallelization of partially parallel loops

Set-Congruence Dynamic Analysis for Thread-Level Speculation (TLS)

“…Serial-commit tls models [5,6,17] isolate the speculative state from the global state: each thread buffers its write-accesses, and commits them when it becomes master. It follows that war and waw dependencies are implicitly satisfied.…”

“…For the modulo relation we have: {(0, 1)}<8> ∪ {(0, 1)}<18> ⊆ {(0, 1)}<2>, which implies that the program cannot be parallelized (as iterations 0 and 1 modulo 2 are executed on the same thread). However, with the step relation we get: {(0, 1)}|8> ∪ {(0, 1)}|18> ⊆ S|18>, where S = {(0, 1), (8,9), (16,17), (6,7), (14,15), (4,5), (12,13), (2,3), (10,11)}. In this case we can run 9 concurrent threads while satisfying the observed dependencies: …”

“…We denote by a and e the start address of arrays a and e. We observe that accesses of a form three disjoint partitions: at e, a decreasing one starting at e + N and a buffer-interval in the middle. While the first two partitions may use aggressive fine-grained partitioning, the middle-one should use a precise partitioning 6 .…”

“…We keep into account that serial commit models [17,5,6] implicitly satisfy waw and war dependencies, while in-place models [16] do not. Also, [17] is inefficient in the presence of many iteration-independent raw.…”

“…Current software-tls approaches [5,6,18] exhibit heavy-transactional support, and can yield good speed-up when (i) the static compiler disambiguates 1 enough accesses to amortize the speculation overhead, and (ii) the iteration granularity is high enough to amortize the transactional overhead related to starting a new iteration. Their design assumes little about patterns of accesses to memory, merely that these accesses generate relatively few dependencies.…”