|
There seems to be a general view that if you want to do some serious compute bound work C++ will always win over C# because of its unmanagedness. This may be the case, but then there's an alternative argument that goes if you sidestep garbage collection, and use unsafe constructs you get close, and because your code gets JITed, there may be a chance to hone the code for the actual CPU it will run on which could actually make .NET faster over a generalised native binary. There seems to be better scope for this if you use the SIMD enabled System.Numerics.
I've been mucking around with sound synthesis in .NET recently, and by using unmanaged memory (Marshal.AllocHGlobal) and unsafe pointers the performance is good enough. In fact I've got it so there are virtually no garbage collections at all (beware of Linq - it creates enumerators all over the place). GCs are terrible news for audio, because a delay can mean a buffer not being ready in time and you get nasty clicks. It means a different approach to coding, but it's still miles preferable to header files and linking libs and all that 32/64 bit nastiness you get with C++.
Then I stumbled on this:
https://www.bepuentertainment.com/
This is a stunning masterclass in what C# can do. I recommend having a look at the video, and reading the bits about GPU vs. CPU and 'a different kind of c#' entries. I'm fairly hard to impress these days, but I've downloaded the code built it and played with it and just... wow.
Moreover, I'd argue its the last nail in the coffin of the argument that C# is not a viable choice for high performance compute bound work.
Regards,
Rob Philpott.
|
|
|
|
|
(I haven't followed the link yet - it's Saturday, and I'm feeling lazy.)
It's probably rather like the old C / Assembler debate. Yes, a skilled assembler programmer can produce faster, more compact code than a skilled C programmer purely because he can tell the machine exactly what he wants it to do rather than adding a layer of "interpretation" via a compiler. But ... it'll take a lot longer to code, and an unskilled assembler programmer can still make a serious dogs dinner of the same job!
I suspect that an average C# coder will produce code that is less efficient than a skilled C++ coder to do the same job - but I also suspect that he'll produce it in less time, and it'll be more easily maintainable by an average developer. And with the performance of modern machines that's a critical factor in most cases. Additionally, I suspect that a skilled C# developer will produce better, faster code than an average C++ dev, and get it out the door quicker as well.
Don't get me wrong, C++ is a good language, I used it for many years - but C# produces good code as well which is often a lot more readable and less prone to silly and avoidable bugs.
Sent from my Amstrad PC 1640
Never throw anything away, Griff
Bad command or file name. Bad, bad command! Sit! Stay! Staaaay...
AntiTwitter: @DalekDave is now a follower!
|
|
|
|
|
I like the simple readability of structured C 
|
|
|
|
|
Yes, and I wrote a lot of embedded C, often with assembler for the time critical bits.
But you have to pay attention, and it's far too easy to create nasty bugs if you don't know what you are doing:
int* GetMemory(int n)
{
int arr[n];
return arr;
}Things get nasty real quick with that ... 
Even if you do it right:
int* GetMemory(int n)
{
int* arr = (int*) malloc(n * sizeof(int));
return arr;
}It's still far to easy to leave a memory leak because the caller is responsible for freeing the memory and unless he looks at the function he may not be aware of that.
Sent from my Amstrad PC 1640
Never throw anything away, Griff
Bad command or file name. Bad, bad command! Sit! Stay! Staaaay...
AntiTwitter: @DalekDave is now a follower!
|
|
|
|
|
This is the trouble. Twenty years of living in the .NET nanny state with the parental controls maxed out means you forget (well I have) the myriad of ways you can kill yourself in the unmanaged world.
I'm not going out there!
Regards,
Rob Philpott.
|
|
|
|
|
Returning a stack variable? Tut tut!
Re that and memory leaks though, you just dont write code that bad. And that is it really, you write 100% tight, very simple code (in fact the simpler it is the easier it is to get 100% tight).
|
|
|
|
|
No, I don't. And you don't. But ... we've both seen it done, and sometimes in production code. 
I think we have to accept that developers are not what they used to be (and "Hoorah!" for that in some cases), projects generally are a lot larger and more complex than they were, and that we have to change to languages which facilitate that.
Sent from my Amstrad PC 1640
Never throw anything away, Griff
Bad command or file name. Bad, bad command! Sit! Stay! Staaaay...
AntiTwitter: @DalekDave is now a follower!
|
|
|
|
|
OriginalGriff wrote: But ... we've both seen it done, and sometimes in production code.
I have seen it many times in production code. And would have been more except that I started reviewing the code of others and catching it. I worked at a place with formal code reviews and I was still the only one finding them.
Not to mention that these sorts of problems impact the server far more.
OriginalGriff wrote: I think we have to accept that developers are not what they used to be
But applications now are not the same either. When I started I wrote the UI and the back end. Now I can't even effectively review the UI code and the developers cannot write the back end code.
|
|
|
|
|
Oh, I had a programmer returning stack pointers.
The third time he claimed it worked fine when he tested it (inside the debugger, duh),
I fired him. The fact that he could not comprehend that a stack is temporary on invocation of the function...
And that is part of the problem. People understand less and less. Coding used to be understand first, analysis, and development. Now it feels more like "crank out code"... And nobody realizes the how deep the problems are.
Then the often hidden fact: Development is only 20% of the budget. Maintenance is 80% over the long haul. So SIMPLE, CLEAN Code that is easier to maintain wins. When you need performance. Do that. Do it right with the right tools, and keep it clean and simple. And if it is unmanaged code, then you must run a memory analyzer/leak finder. Or your code is not properly tested!
Having worked on a project with 250K lines of code. They said it had a memory leak. A quick review of the code and I labelled it a Memory Sieve! The design guaranteed a leak! Nothing would be freed as children referenced their parents!
|
|
|
|
|
I did a little experiment a while ago - calculating prime numbers in C++ and a handcrafted assembly version, same algorithm obviously. The C++ one ran quicker, never found out why. CPUs are so damn complicated these days. In the glory days you could look at your instruction and know how many clock cycles it took. I think the on chip caches change all that.
In my assembly days you'd get everything into registers and keep it there as long as possible. Perhaps these days the cache is good enough. Reluctantly I concede it's best left to the compiler.
The main thing in the C#/C++ for me though is that the C++ you see is so damn cryptic. I don't think it needs to be, but usually just is. Maybe its the mindset.
Regards,
Rob Philpott.
|
|
|
|
|
Indeed, many of today's optimization techniques revolve around hitting the caches as much as possible. The amount of cycles lost when performing a main memory read are mind boggling. 
|
|
|
|
|
OriginalGriff wrote: I suspect that an average C# coder will produce code that is less efficient than a skilled C++ coder to do the same job - but I also suspect that he'll produce it in less time, and it'll be more easily maintainable by an average developer. And with the performance of modern machines that's a critical factor in most cases. Additionally, I suspect that a skilled C# developer will produce better, faster code than an average C++ dev, and get it out the door quicker as well.
by the same token with newer machines having so much more memory than before you could argue garbage collection / free-ing unused memory can also be ignored in short running programs or where only alloc-ing space for relatively small buffers/structures. [which could also have dramatic results for performance.]
OTOH, using the excuse: "todays machines are faster / bigger memory... so optimisation and/or garbage collection matters little" is when people copy or extend that code into small/short running programs into big / long running ones.
OP's "sound" app may work fine in optimised C# [and/or without cleaning up unused memory] but when added to a video editing suite would that still be true?
On my bicycle I can match (if not beat) the bus on a 10 mile commute and get away without refueling on the way, but let's make that 100 miles.
right tools for the job.
Message Signature
(Click to edit ->)
|
|
|
|
|
Lopatir wrote: but let's make that 100 miles.
Nice analogy.
But most business programming these days is more akin to a trip to china or even the moon.
|
|
|
|
|
Agreed wholeheartedly.
If performance can be improved just by throwing more hardware at it, it probably makes more sense to have an average developer write good, maintainable code than having a rock star developer write code that only he has the ability to read and maintain, even if that code manages to squeeze every little bit of performance out of the hardware.
Simply because developer time is a lot more expensive than hardware.
|
|
|
|
|
How do you find the Audio performance on Windows though, do you notice the lag through the kernel?
(For example to run recording studio type stuff you need to bin it all and put in ASIO)
|
|
|
|
|
No not really. I think Microsoft likes to occasionally upload your entire browsing history so it can 'improve your windows experience' and if that were to happen the same moment as you're filling your buffers you may come unstuck. Ultimately, we're all slaves to the thread scheduler.
But, that's just windows. I've mucked about with thread and process priorities, but it doesn't seem to make much difference.
In terms of the driver, I am using ASIO. And so far it's all single threaded too.
Regards,
Rob Philpott.
|
|
|
|
|
Rob Philpott wrote: I am using ASIO
Good, you have fixed 99% of your problems right there. 
I did some audio work, it was a service that had to open a pin on two sound cards and send the data between them, so it used the KS API.
It was OK, ish, but much better when I put it in the kernel (same KS API, so pretty easy to move it over).
(Since it used KS it did the same thing as ASIO, it kicks the windows audio mixer off the hardware and give you pretty much real time IO)
|
|
|
|
|
Quote: "See, it's easy," you begin, but are interrupted by the sound of rapid British-sounding footsteps approaching the doorway.
I have to ask, British footsteps sound different from, say, German footsteps? 
|
|
|
|
|
British footsteps sound like an old man shuffling along a rough track.
German footsteps sound like an army goosestepping along a wide autobahn (to the sound of an oompah band).
|
|
|
|
|
And French footsteps just get quieter and quieter...
Sent from my Amstrad PC 1640
Never throw anything away, Griff
Bad command or file name. Bad, bad command! Sit! Stay! Staaaay...
AntiTwitter: @DalekDave is now a follower!
|
|
|
|
|
And the Dutch, the Dutch with their woody clutter !
|
|
|
|
|
The click of wet thongs (flip flops to you ethnics) for the Aussies
Never underestimate the power of human stupidity -
RAH
I'm old. I know stuff - JSOP
|
|
|
|
|
Or footsteps of a person looking down at their phone?
|
|
|
|
|
I used unmanaged methods with C# to do some image processing.
For fun I was writing something to detect hand-drawn rectangles and circles - so that you can turn a hand-drawn diagram into vectors.
I found that the unmanaged methods took around three times less time than using managed code - so it does make a bit of a difference.
“That which can be asserted without evidence, can be dismissed without evidence.”
― Christopher Hitchens
|
|
|
|
|
System.Numerics is better than nothing (I recommend it over "doing nothing" at least), but it's really difficult. By which I mean more difficult than using SIMD intrinsics in C++. The API is full of holes (no right shift?? no shuffle? no special operations?) and landmines (you'd think that eg multiplying a vector by a scalar is equivalent to broadcasting that scalar and then doing a vector multiply, but no, you have to do that manually and ban the scalar*vec from your code).
Of course programming is an exercise in "poking the code until the asm looks good" either way, but the System.Numerics API makes it harder to get it right. Sometimes impossible.
The newer SIMD API in .NET Core 3 (System.Runtime.Intrinsics) is more complete and probably will be better. Still tricky to get good codegen in all cases. For example it's not so easy to force it to "broadcast from memory" and avoid the "load-then-broadcast" anti-pattern (this costs a shuffle µop on p5 in addition to the load, broadcast from memory only costs a load).
|
|
|
|
General 
News 
Suggestion 
Question 
Bug 
Answer 
Joke 
Praise 
Rant 
Admin 
Submit an article or tip
