Accelerating intersections with SIMD instructions(lemire.me)
lemire.me
Accelerating intersections with SIMD instructions
http://lemire.me/blog/archives/2015/03/25/accelerating-intersections-with-simd-instructions/
1 comments
For interested readers who want follow up reading: learn about loop unrolling. Commonly used with SIMD vectorization to get rid of loop overhead. Then we can also start doing cache based optimization, cache blocking etc
Actually, I'd probably go the other way on that: loop unrolling is much less useful than it used to be. As long as your loops are predictable I'd suggest not worrying about them. Correctly predicted branches are very close to free, and fitting the entire loop in the decoded µop cache is more important. There are still cases where it helps, but there are also cases where where the instruction mix works out so that there is zero overhead. Not just low, but zero. So until you've measured, keep your loops short and simple.
Instead, I'd suggest that if you want to write really high performance code, you need to be thinking in assembly regardless of the language you are writing the actual code in. Look at the code that's actually executing, either by disassembling the compiled object or profiling with something like 'perf' on Linux.[1] Then, you need to be aware of how the processor handles that assembly, in terms of latencies and execution ports. Agner Fog's manuals in combination with those from Intel and AMD are incredibly valuable.
Instead, I'd suggest that if you want to write really high performance code, you need to be thinking in assembly regardless of the language you are writing the actual code in. Look at the code that's actually executing, either by disassembling the compiled object or profiling with something like 'perf' on Linux.[1] Then, you need to be aware of how the processor handles that assembly, in terms of latencies and execution ports. Agner Fog's manuals in combination with those from Intel and AMD are incredibly valuable.
Wouldn't there be cases where loop unrolling, is what enables SIMD instructions to be used? Is that the exception to your "rule"
Yeah, I agree with nkurz. Unrolling does better in micro benchmarks, but in a context of a larger program seems to rarely help nowadays. Maybe in some very tight loops, that don't do much.
Aggressive unrolling consumes more instruction cache space and hurts performance elsewhere.
Aggressive unrolling consumes more instruction cache space and hurts performance elsewhere.