About SMS algorithm for L1D Load prefetching

[eeeeewwz]

Hi, I've got a question about SMS prefetching algorithm attached to L1D:
SMS in Xiangshan only uses hashed PC, instead of PC+Offset, to tag and index PHT:

XiangShan/src/main/scala/xiangshan/mem/prefetch/SMSPrefetcher.scala

Lines 96 to 104 in 31fae68

	def pht_index(pc: UInt): UInt = {
	val low_bits = pc(PHT_INDEX_BITS, 2)
	val hi_bit = pc(1) ^ pc(PHT_INDEX_BITS+1)
	Cat(hi_bit, low_bits)
	}
	def pht_tag(pc: UInt): UInt = {
	pc(PHT_INDEX_BITS + 2 + PHT_TAG_BITS - 1, PHT_INDEX_BITS + 2)
	}

but this could cause pattern aliasing (i.e., the same PC is correlated to different trigger offsets), thus is less accurate because it cannot distinguish among distinct access patterns to different data structures by the same code (i.e., PC), as the original paper (Spatial Memory Streaming, Somogyi, et.al., 2006) said. Is there any special considerations or additional performance evaluations?

[shinezyy]

Because we employ a very small table for SMS and a PC-only index requires less storage than PC+index.
If use PC + offset, then one PC will occupy more table entries, which is possible to collide with others.[1]

[1] Jiang, S., Yang, Q., & Ci, Y. (2022). Merging Similar Patterns for Hardware Prefetching. Proceedings of the Annual International Symposium on Microarchitecture, MICRO, 2022-Octob, 1012–1026. https://doi.org/10.1109/MICRO56248.2022.00071

[eeeeewwz]

Thanks for replying! I've noticed this paper before, but here comes another question: it turns out that pattern collision and duplicate are contradictory (see the figure below), while the PMP proposal adopt the method to merge patterns with the same trigger offset, which is more aggressive than that with the same (hashed) PC, so does the PC-indexed method performs better than the offset-indexed one, just as PMP does?

[shinezyy]

It depends on workloads. Yes for SPECCPU 2006 on our configuration

[eeeeewwz]

Is these SPEC results from simulator (like XS-GEM5), or cycle-accurate RTL emulation? And how about the detailed performance gain statistics (sorry but if convenient)?