Clarification on Throughput

[cod3monk]

Since I read it in #1 and nowhere else, I want to clarify this: the throughput numbers in benchmark/throughput/*.csv, are they the computed average as the timing-harness prints (i.e., (measuremnt(200it) - measurement(100it)) / 100) or are they something else?

For example:

# hex encoded block: 81fdffff0000
.section	__TEXT,__text,regular,pure_instructions
	cmpl	$65535, %ebp            ## imm = 0xFFFF
# bhive measured throughput: 33.000000

does this say that cmpl immediate, gpr64 has been measured with a derived (reciprocal) throughput of 0.33cy (i.e. 3 instruction per cycle) or 33cy?

Maybe this could be mentioned in the README.md :)

[ychen306]

It's throughput times 100 (so 0.33). Note that this number is unfortunately wrong should be 50.

[andreas-abel]

50 does not seem to be correct.

[ychen306]

My mistake. I double-checked, and it should have been 37 (not 50 or 33).

[andreas-abel]

For Haswell and Skylake, it's 25 if the uops are delivered from the DSB. ~37 is only correct if they go through the legacy decoder.

Since you are comparing your measurements to IACA, which assumes "ideal front-end conditions", 25 is the only meaningful value for such a comparison. And it's also close to the value that IACA reports for this benchmark.