Блог
|
Степанов об STL| By complexity I do not mean just the asymptotic complexity but the machine cycle count. In order to learn about it, it is necessary to acquire a habit of writing benchmarks. Time and time again I discovered that my beautiful designs were totally wrong after writing a little benchmark. The most embarrassing case was when after claiming publicly on multiple occasions that STL had the performance of hand-written assembly code, I published my Abstraction Penalty Benchmark that showed that my claims were only true if you were using a specialized preprocessor from KAI. It was particularly embarrassing because it showed that the compiler produced by my employer – Silicon Graphics – was the worst in terms of abstraction penalty and compiling STL. The SGI compiler was eventually fixed, but the performance of STL on the major platforms keeps getting worse precisely because customers as well as vendors do not do benchmarking and seem to be totally unconcerned about performance degradation. Occasionally there will be assignments that require benchmarking. Please do them. |
|
Будь в курсе того, что думает Nappy!
Подпишись на
Nappy