Re:Increasing the D Compiler Speed by Over 75%
by WalterBright on Thursday July 25, 2013 @04:14PM
(#6104669)
As I mentioned elsewhere, I like gprof because it reports on fan-in and fan-out. This gives a picture of how the tree is executing, rather than just the functions.
Re:Increasing the D Compiler Speed by Over 75%
by nkurz on Thursday July 25, 2013 @04:39PM
(#6104824)
Maybe I misunderstand what you need, but generating a call graph and showing what percentage of time is spent in a function by caller is possible with all of these. Example for perf: http://lwn.net/Articles/340010/
Re:Increasing the D Compiler Speed by Over 75%
by WalterBright on Thursday July 25, 2013 @04:41PM
(#6104833)
I didn't know it could do that. Thanks for the info.
Re:Increasing the D Compiler Speed by Over 75%
by mrich on Thursday July 25, 2013 @05:20PM
(#6105078)
Vtune and Zoom need to be mentioned, although they are not free. Vtune is even able to do power consumption analysis these days.
Re:Increasing the D Compiler Speed by Over 75%
by nkurz on Thursday July 25, 2013 @07:17PM
(#6105579)
If appropriate for your use, VTune for Linux (and the rest of the Intel Compiler Suite) is available for free under a "Non-Commercial Software Development" license.
http://software.intel.com/en-us/non-commercial-software-deve...
Re:Increasing the D Compiler Speed by Over 75%
by gngeal on Thursday July 25, 2013 @05:22PM
(#6105090)
Vtune is even able to do power consumption analysis these days.
...on AMD machines, too? (Not to mention mobile development on ARM.)
Re:Increasing the D Compiler Speed by Over 75%
by mcpherrinm on Thursday July 25, 2013 @06:51PM
(#6105472)
As far as I understand, having not used VTune, some basic features work on both, but the interesting stuff relies on Intel-specific counters in the CPU.
This doesn't mean it's worthless, though: Finding your performance bottlenecks on Intel will surely be at least a first-order approximation of the performance on AMD, too.
ARM is a little bit of a different story since the memory model is different, and you might be getting killed by something like unaligned accesses, but that doesn't mean the information is worthless; it just means you should probably use more than one tool. After all, not everything is a nail, but hammers are still good tools.
It's been a long time since I've used gprof. I switched to Valgrind and OProfile about 10 years ago, and more recently to 'perf' and 'likwid'. If the goal is finding hot-spots, these last might be more convenient since they run with minimal overhead --- a couple percent rather than 100x.
Are there benefits to gprof that I've forgotten?
Are there newer and better profiling tools I don't know about?