4 ms·

Not in my experience! Though it was a few years ago, I'll have to re-evaluate. In the past, when doing numeric/scientific programming -- lots of dataflow, it

by loxias 4y ago

Not in my experience! Though it was a few years ago, I'll have to re-evaluate. In the past, when doing numeric/scientific programming -- lots of dataflow, it was much easier to "coax" clang to generate the right (performant) assembly than GCC. I guess my use is less typical -- the internet agrees with you. :) Hurrah for competition! (of course, neither GCC or clang could _ever_ beat ICC at some of my tests.. grumble)

DannyBee 4y ago
"(of course, neither GCC or clang could _ever_ beat ICC at some of my tests.. grumble)" They never tried, honestly - ICC had plenty of defaults that were targeted at performance above correctness. That's fine - but it wasn't what either clang or gcc was going to go for. Even when trying to compare apples to apples, folks often compared them by trying to get GCC/Clang to emulate the correctness level of ICC (IE give GCC/Clang more freedom), rather than the other way around :) Which, again, similarly understandable, but also a thing that you'd have to spend a while on in GCC/Clang to get to a reasonable place) Small-number-of-target compilers like ICC are also fundamentally easier. Lots of techniques (IE optimal register allocation[1], etc) that add up to performance gains are a lot more tractable to really good applied engineering when they don't have to be so general. Similarly small-target-market compilers are also easier. Over time, high performance was literally ICC's only remaining market. So they can spend their days working on that. GCC/Clang had to care about a lot more. [1] This happens to now be a bad example because advances finally made this particular thing tractable. But it took years more, and feel free to replace it with something like "optimal integrated register allocation and scheduling" or whatever is still intractable to generalize.
- loxias 4y ago
  You sound like you know what you're talking about. > They never tried, honestly - ICC had plenty of defaults that were targeted at performance above correctness. Is it possible to get ICC level performance out of open source tools? Much of my CPU bound work relates to array signal processing, which, if you can code it right :) lends itself heavily to SIMD branchless pipelines. Plus some scatter gathers on a group of other cores to calculate a sparse crosscorrelation tensor. I would love to be able to get same or better performing code out of clang or gcc, even it it takes 2-4x more work than with ICC...

by loxias 4y ago

DannyBee 4y ago

"(of course, neither GCC or clang could _ever_ beat ICC at some of my tests.. grumble)" They never tried, honestly - ICC had plenty of defaults that were targeted at performance above correctness. That's fine - but it wasn't what either clang or gcc was going to go for. Even when trying to compare apples to apples, folks often compared them by trying to get GCC/Clang to emulate the correctness level of ICC (IE give GCC/Clang more freedom), rather than the other way around :) Which, again, similarly understandable, but also a thing that you'd have to spend a while on in GCC/Clang to get to a reasonable place) Small-number-of-target compilers like ICC are also fundamentally easier. Lots of techniques (IE optimal register allocation[1], etc) that add up to performance gains are a lot more tractable to really good applied engineering when they don't have to be so general. Similarly small-target-market compilers are also easier. Over time, high performance was literally ICC's only remaining market. So they can spend their days working on that. GCC/Clang had to care about a lot more. [1] This happens to now be a bad example because advances finally made this particular thing tractable. But it took years more, and feel free to replace it with something like "optimal integrated register allocation and scheduling" or whatever is still intractable to generalize.

loxias 4y ago
You sound like you know what you're talking about. > They never tried, honestly - ICC had plenty of defaults that were targeted at performance above correctness. Is it possible to get ICC level performance out of open source tools? Much of my CPU bound work relates to array signal processing, which, if you can code it right :) lends itself heavily to SIMD branchless pipelines. Plus some scatter gathers on a group of other cores to calculate a sparse crosscorrelation tensor. I would love to be able to get same or better performing code out of clang or gcc, even it it takes 2-4x more work than with ICC...