benchmarks

From memory, Benchmark.NET will run benchmarks for all frameworks with some internal wizardry. So instead of using the existing preprocessor symbols it's probably better to split your tests across two classes with different RuntimeMoniker attributes. For example:

You can do it e.g. like this:

Yes, data misalignment could explain your 2x slowdown for small arrays that fit in L1d. You'd hope that with every other load/store being a cache-line split, it might only slow down by a factor of 1.5x, not 2, if a split load or store cost 2 accesses to L1d instead of 1.

But it has extra effects like replays of uops dependent on the load result that apparently account for the rest of the problem, either making out-of-order exec less able to overlap work and hide latency, or directly running into bottlenecks like "split registers".

ld_blocks.no_sr counts number of times cache-line split loads are temporarily blocked because all resources for handling the split accesses are in use.

When a load execution unit detects that the load splits across a cache line, it has to save the first part somewhere (apparently in a "split register") and then access the 2nd cache line. On Intel SnB-family CPUs like yours, this 2nd access doesn't require the RS to dispatch the load uop to the port again; the load execution unit just does it a few cycles later. (But presumably can't accept another load in the same cycle as that 2nd access.)

• https://chat.stackoverflow.com/transcript/message/48426108#48426108 - uops waiting for the result of a cache-split load will get replayed.
• Are load ops deallocated from the RS when they dispatch, complete or some other time? But the load itself can leave the RS earlier.
• How can I accurately benchmark unaligned access speed on x86_64? general stuff on split load penalties.

The extra latency of split loads, and also the potential replays of uops waiting for those loads results, is another factor, but those are also fairly direct consequences of misaligned loads. Lots of counts for ld_blocks.no_sr tells you that the CPU actually ran out of split registers and could otherwise be doing more work, but had to stall because of the unaligned load itself, not just other effects.

You could also look for the front-end stalling due to the ROB or RS being full, if you want to investigate the details, but not being able to execute split loads will make that happen more. So probably all the back-end stalling is a consequence of the unaligned loads (and maybe stores if commit from store buffer to L1d is also a bottleneck.)

On a 100KB I reproduce the issue: 1075ns vs 1412ns. On 1 MB I don't think I see it.

Data alignment doesn't normally make that much difference for large arrays (except with 512-bit vectors). With a cache line (2x YMM vectors) arriving less frequently, the back-end has time to work through the extra overhead of unaligned loads / stores and still keep up. HW prefetch does a good enough job that it can still max out the per-core L3 bandwidth. Seeing a smaller effect for a size that fits in L2 but not L1d (like 100kiB) is expected.

Of course, most kinds of execution bottlenecks would show similar effects, even something as simple as un-optimized code that does some extra store/reloads for each vector of array data. So this alone doesn't prove that it was misalignment causing the slowdowns for small sizes that do fit in L1d, like your 10 KiB. But that's clearly the most sensible conclusion.

Code alignment or other front-end bottlenecks seem not to be the problem; most of your uops are coming from the DSB, according to idq.dsb_uops. (A significant number aren't, but not a big percentage difference between slow vs. fast.)

How can I mitigate the impact of the Intel jcc erratum on gcc? can be important on Skylake-derived microarchitectures like yours; it's even possible that's why your idq.dsb_uops isn't closer to your uops_issued.any.

Update: Version 1.6.9 has been released and should fix this issue! 🎉

This is actually a known bug, which is now open for quite a while: OSSRH-66257. There are two known workarounds:

As a workaround, use --add-opens to give the library causing the problem access to the required classes:

You want:

TL;DR: You are observing what happens when JIT compiler cannot trust that values are not changing inside the loop. Additionally, in the tiny benchmark like this, Blackhole.consume costs dominate, obscuring the results.

Simplifying the test:

It seems that all your threads use the same database connection and statement objects. Therefore you have a race-condition (even in SERIALIZED threading model), as multiple threads are binding, stepping, and resetting the same statement. Asking 'why is it slow' becomes irrelevant until you fix this problem.

Instead you should wrap your sql_insert with a mutex to guarantee that at most one thread is accessing the database connection:

Zen2 has 3 cycle latency for vaddpd, 5 cycle latency for vfma...pd. (https://uops.info/).

Your code with 8 accumulators has enough ILP that you'd expect close to two FMA per clock, about 8 per 5 clocks (if there aren't other bottlenecks) which is a bit less than the 10/5 theoretical max.

vaddpd and vmulpd actually run on different ports on Zen2 (unlike Intel), port FP2/3 and FP0/1 respectively, so it can in theory sustain 2/clock vaddpd and vmulpd. Since the latency of the loop-carried dependency is shorter, 8 accumulators are enough to hide the vaddpd latency if scheduling doesn't let one dep chain get behind. (But at least multiplies aren't stealing cycles from it.)

Zen2's front-end is 5 instructions wide (or 6 uops if there are any multi-uop instructions), and it can decode memory-source instructions as a single uop. So it might well be doing 2/clock each multiply and add with the non-FMA version.

If you can unroll by 10 or 12, that might hide enough FMA latency and make it equal to the non-FMA version, but with less power consumption and more SMT-friendly to code running on the other logical core. (10 = 5 x 2 would be just barely enough, which means any scheduling imperfections lose progress on a dep chain which is on the critical path. See Why does mulss take only 3 cycles on Haswell, different from Agner's instruction tables? (Unrolling FP loops with multiple accumulators) for some testing on Intel.)

(By comparison, Intel Skylake runs vaddpd/vmulpd on the same ports with the same latency as vfma...pd, all with 4c latency, 0.5c throughput.)

I didn't look at your code super carefully, but 10 YMM vectors might be a tradeoff between touching two pairs of cache lines vs. touching 5 total lines, which might be worse if a spatial prefetcher tries to complete an aligned pair. Or might be fine. 12 YMM vectors would be three pairs, which should be fine.

Depending on matrix size, out-of-order exec may be able to overlap inner loop dep chains between separate iterations of the outer loop, especially if the loop exit condition can execute sooner and resolve the mispredict (if there is one) while FP work is still in flight. That's an advantage to having fewer total uops for the same work, favouring FMA.

You can just use List.RemoveAll and pass your predicate - https://docs.microsoft.com/en-us/dotnet/api/system.collections.generic.list-1.removeall?view=net-6.0#System_Collections_Generic_List_1_RemoveAll_System_Predicate__0__ . This guaranteed to be linear complexity O(list.Count)