* Pavel Machek <pa...@ucw.cz> wrote: > > > > Yeah, so generic memcpy() replacement is only feasible I think if the > > > > most > > > > optimistic implementation is actually correct: > > > > > > > > - if no preempt disable()/enable() is required > > > > > > > > - if direct access to the AVX[2] registers does not disturb legacy FPU > > > > state in > > > > any fashion > > > > > > > > - if direct access to the AVX[2] registers cannot raise weird > > > > exceptions or have > > > > weird behavior if the FPU control word is modified to non-standard > > > > values by > > > > untrusted user-space > > > > > > > > If we have to touch the FPU tag or control words then it's probably > > > > only good for > > > > a specialized API. > > > > > > I did not mean to have a general memcpy replacement. Rather something like > > > magic_memcpy() which falls back to memcpy when AVX is not usable or the > > > length does not justify the AVX stuff at all. > > > > OK, fair enough. > > > > Note that a generic version might still be worth trying out, if and only if > > it's > > safe to access those vector registers directly: modern x86 CPUs will do > > their > > non-constant memcpy()s via the common memcpy_erms() function - which could > > in > > theory be an easy common point to be (cpufeatures-) patched to an AVX2 > > variant, if > > size (and alignment, perhaps) is a multiple of 32 bytes or so. > > How is AVX2 supposed to help the memcpy speed? > > If the copy is small, constant overhead will dominate, and I don't > think AVX2 is going to be win there.
There are several advantages: 1) "REP; MOVS" (also called ERMS) has a significant constant "setup cost". In the scheme I suggested (and if it's possible) then single-register AVX2 access on the other hand has a setup cost on the "few cycles" order of magnitude. 2) AVX2 have various non-temporary load and store behavioral variants - while "REP; MOVS" doesn't (or rather, any such caching optimizations, to the extent they exist, are hidden in the microcode). > If the copy is big, well, the copy loop will likely run out of L1 and maybe > even > out of L2, and at that point speed of the loop does not matter because memory > is > slow...? In many cases "memory" will be something very fast, such as another level of cache. Also, on NUMA "memory" can also be something locally wired to the CPU - again accessible at ridiculous bandwidths. Nevertheless ERMS is probably wins for the regular bulk memcpy by a few percentage points, so I don't think AVX2 is a win in the generic large-memcpy case, as long as continued caching of both the loads and the stores is beneficial. Thanks, Ingo