Hi Segher, Thanks for the review!
on 2021/9/18 上午6:01, Segher Boessenkool wrote: > Hi! > > On Thu, Sep 16, 2021 at 09:14:15AM +0800, Kewen.Lin wrote: >> The way with nunits * stmt_cost can get one much exaggerated >> penalized cost, such as: for V16QI on P8, it's 16 * 20 = 320, >> that's why we need one bound. To make it scale, this patch >> doesn't use nunits * stmt_cost any more, but it still keeps >> nunits since there are actually nunits scalar loads there. So >> it uses one cost adjusted from stmt_cost, since the current >> stmt_cost sort of considers nunits, we can stablize the cost >> for big nunits and retain the cost for small nunits. After >> some tries, this patch gets the adjusted cost as: >> >> stmt_cost / (log2(nunits) * log2(nunits)) > > So for V16QI it gives *16/(4*4) so *1 > V8HI it gives *8/(3*3) so *8/9 > V4SI it gives *4/(2*2) so *1 > V2DI it gives *2/(1*1) so *2 > and for V1TI it gives *1/(0*0) which is UB (no, does not crash for us, > just gives wildly wrong answers; the div returns 0 on recent systems). > I don't expected we will have V1TI for strided/elementwise load, if it's one unit vector, it's the whole vector itself. Besides, the below assertion should exclude it already. >> For V16QI, the adjusted cost would be 1 and total penalized >> cost is 16, it isn't exaggerated. For V2DI, the adjusted >> cost would be 2 and total penalized cost is 4, which is the >> same as before. btw, I tried to use one single log2(nunits), >> but the penalized cost is still big enough and can't fix the >> degraded bmk blender_r. > > Does it make sense to treat V2DI (and V2DF) as twice more expensive than > other vectors, which are all pretty much equal cost (except those that > end up with cost 0)? If so, there are simpler ways to do that. > Yeah, from the SPEC2017 evaluation, it's good with this. The costing framework of vectorization doesn't consider the dependent insn chain and available #unit etc. like local scheduling (it can't either), so we have to use some heuristics to handle some special cases. For more units vector construction, the used instructions are more. It has more chances to schedule them better (even run in parallelly when enough available units at the time), so we don't need to penalize more for them. For V2DI, the load result is fed into construction directly, the current stmt_cost is to consider merging and only 2, penalizing it with one is not enough from the bwaves experiment. >> + int nunits_log2 = exact_log2 (nunits); >> + gcc_assert (nunits_log2 > 0); >> + unsigned int nunits_sq = nunits_log2 * nunits_log2; > >> = 0 > > This of course is assuming nunits will always be a power of 2, but I'm > sure that we have many other places in the compiler assuming that > already, so that is fine. And if one day this stops being true we will > get a nice ICE, pretty much the best we could hope for. > Yeah, exact_log2 returns -1 for non power of 2 input, for example: input output 0 -> -1 1 -> 0 2 -> 1 3 -> -1 >> + unsigned int adjusted_cost = stmt_cost / nunits_sq; > > But this can divide by 0. Or are we somehow guaranteed that nunits > will never be 1? Yes the log2 check above, sure, but that ICEs if this > is violated; is there anything that actually guarantees it is true? > As I mentioned above, I don't expect we can have nunits 1 strided/ew load, and the ICE should check this and ensure dividing by zero never happens. :) >> + gcc_assert (adjusted_cost > 0); > > I don't see how you guarantee this, either. > It's mainly to prevent that one day we tweak the cost for construction in rs6000_builtin_vectorization_cost then make some unexpected values generated here. But now these expected values are guaranteed as the current costs and the formula. > > A magic crazy formula like this is no good. If you want to make the > cost of everything but V2D* be the same, and that of V2D* be twice that, > that is a weird heuristic, but we can live with that perhaps. But that > beats completely unexplained (and unexplainable) magic! > > Sorry. > That's all right, thanks for the comments! let's improve it. :) How about just assigning 2 for V2DI and 1 for the others for the penalized_cost_per_load with some detailed commentary, it should have the same effect with this "magic crazy formula", but I guess it can be more clear. BR, Kewen
