|
All the .NET languages are slow. Despite the dubious claims by some people that they can be as fast as unmanaged code, .NET programs are usually sluggish. Managed code does JITing, boxing/unboxing, and run-time checking (e.g. bounds checking on array accesses) for example.
However, .NET programs are more reliable. Recently the same subsystem was implemented at my company in C# and unmanaged C++. The C++ code was faster but would regularly crash mysteriously. The C# code was solid.
The best approach is to write the bulk of your code in C#, but do the image processing in unmanaged C++. This gives you the reliability of managed code, and the speed of unmanaged code for the time-consuming repetitive tasks.
The fastest processing is in small C++ loops that fit entirely into the cache, with no contained branches. This allows out-of-order execution which speeds up processing.
Since a loop is a branch, which makes out-of-order execution difficult, you can process two or more cases inside your loop (instead of one) to do more processing before the branch. This is called "loop unrolling", and can speed up processing.
Processing the image from low addresses to high addresses minimizes memory accesses and makes use of the cache, which fills a 32 (or 64)-byte buffer (called a "cache line") with a single memory access (even one pixel). Subsequent memory accesses at addresses just above this DON'T access memory because the contents are already in the cache. This saves time by avoiding the need for the processor to wait for a memory access.
The Intel (and AMD) SSE and MMX extensions to the instruction set (http://en.wikipedia.org/wiki/Streaming_SIMD_Extensions[^]) allow the use of 128-bit registers which may allow you to do image-processing operations in parallel for more speed.
|
|
|
|
|
Alan Balkany wrote: The C++ code was faster but would regularly crash mysteriously. The C# code was solid.
I would suggest the C++ code was written by an incompetent or hadn't been debugged properly. I've written a ton of C++ code over the years and my code does not typically "just crash mysteriously". Or it was some Microsoft benchmark slanted to show how their proprietary garbage is superior.
You measure democracy by the freedom it gives its dissidents, not the freedom it gives its assimilated conformists.
|
|
|
|
|
"I would suggest the C++ code was written by an incompetent or hadn't been debugged properly."
Probably. But my point is that C++ lets you get away with that. C# quickly detects that type of problem.
In this case, a few months previously the C++ programmer had been deriding the need for C# memory protections, saying "My code doesn't have memory leaks!". People aren't perfect, and a language that protects you against some mistakes will produce more reliable programs.
|
|
|
|
|
Alan Balkany wrote: C# quickly detects that type of problem.
Don't count on it.
|
|
|
|
|
The whole memory leak problem that .NET is supposed to cure is a big Microsoft FUD. Since at least the early 90s, Microsoft as provided a debug heap that's instrumented to detect memory leaks. All you have to do is enable its use and test your debug version. When you exit it will not only tell you if you have memory leaks but where the unfreed memory was allocated. This works both for C and C++. The bigger problem is resource leaks which Microsoft didn't do much to address. Forget to call Dispose and/or not implement it properly and it's no better than forgetting to call free or not handling your destructor properly.
You measure democracy by the freedom it gives its dissidents, not the freedom it gives its assimilated conformists.
|
|
|
|
|
.NET detects memory problems that the unmanaged debug heap won't, such as array indexes out of bounds.
.NET code is more reliable. Of course if you write perfect code, you can have a reliable unmanaged program, but who writes perfect code? At the current level of technology, we can't prove that ANY program is correct.
|
|
|
|
|
I guess we'll just have to agree to disagree about the benefits of locking yourself into a proprietary language for dubious benefits.
You measure democracy by the freedom it gives its dissidents, not the freedom it gives its assimilated conformists.
|
|
|
|
|
I had implemented an image processing program in Delphi 7, and it showed real-time performance. However, the support for Delphi is not very strong. Let me hasten to add that I found absolutely no problems with the Delphi 7 executable - no crashing, etc. for big images. Also, Delphi has a function to assign all bits of a single scan-line - so you don't need nested 'for' loops (outer loop over height, and inner loop over width); such a function is missing in C#. I am amazed at the way Delphi achieves high performance.
Java is yet another option, which I have not yet explored.
modified on Tuesday, September 22, 2009 2:35 AM
|
|
|
|
|
I don't think Java is a good approach if you're looking for maximum performance. It's an interpreted language, with bytecodes being executed by a software virtual machine, so it's slower than native code.
|
|
|
|
|
Alan Balkany wrote: It's an interpreted language, with bytecodes being executed by a software virtual machine
That of course is highly debatable. Both Java and C# compile to an intermediate language, which is then compiled to and stored and executed as native code (at "run-time", which actually means just before it runs, so not really different from "at build-time" except that it adds to your app's start-u[ time). An interpreter would never generate native code.
Whether the end result is worse, equal or better performance-wise is mainly determined by the amount of effort they have chosen to spend in the compiler and virtual machine. After all, the intermediate code, containing a lot of meta information, is a perfect representation of the original source code.
BTW: most/all regular compilers also have a front-end dealing with the source language, and a back-end generating the final instructions. With the two parts communicating through a rather language-agnostic internal representation of the source; that basically is what bytecode and IL also are.
Luc Pattyn
Have a look at my entry for the lean-and-mean competition; please provide comments, feedback, discussion, and don’t forget to vote for it! Thank you.
Local announcement (Antwerp region): Lange Wapper? Neen!
|
|
|
|
|
Luc Pattyn wrote: That of course is highly debatable
However, experience shows that java programs are (usually) deadly slow.
If the Lord God Almighty had consulted me before embarking upon the Creation, I would have recommended something simpler.
-- Alfonso the Wise, 13th Century King of Castile.
This is going on my arrogant assumptions. You may have a superb reason why I'm completely wrong.
-- Iain Clarke
[My articles]
|
|
|
|
|
Experience also show that VB code is sh!t, this is often due to the monkeys rather than the tree.
Panic, Chaos, Destruction.
My work here is done.
|
|
|
|
|
That's true. Anyway I wouldn't suggest anyone to use java for programming 'damned-fast' applications. While java has many many many qualities, alas speed is not one of (of course this is going on my...).
If the Lord God Almighty had consulted me before embarking upon the Creation, I would have recommended something simpler.
-- Alfonso the Wise, 13th Century King of Castile.
This is going on my arrogant assumptions. You may have a superb reason why I'm completely wrong.
-- Iain Clarke
[My articles]
|
|
|
|
|
I have been using a very fast and proprietary Java system for many years, well before .NET came to be. The average Java implementation being slow is due to the fact that everyone can create a JVM (just like everyone can write a compiler), it takes a professional and performance oriented approach to create a good one. "It works, lets ship it" isn't good enough. Not here, not anywhere if performance matters.
Luc Pattyn
Local announcement (Antwerp region): Lange Wapper? Neen!
|
|
|
|
|
No: nothing can even approach plain C language performance (well assembly can be better).
I'm talking about well written java applications vs well written C ones.
Believe me there's a reason why light speed is c  .
On the other hand java has many many many good features, but, you know it is the "compile once, slow down everywhere" language (well, everywhere but the proprietary implementation you experienced...) 
If the Lord God Almighty had consulted me before embarking upon the Creation, I would have recommended something simpler.
-- Alfonso the Wise, 13th Century King of Castile.
This is going on my arrogant assumptions. You may have a superb reason why I'm completely wrong.
-- Iain Clarke
[My articles]
modified on Friday, September 25, 2009 3:45 PM
|
|
|
|
|
unmanaged code kills managed code for this type of thing.
we do all of our stuff in C++, with a bit of assembly for the MMX/SSE optimizations. sometimes we use the MMX/SSE intrinsics (which are essentially macros), but that's just being lazy. hand-coded assembly can beat the intrinsics.
nothing will beat an unmanaged pointer zipping across the image data for sheer speed.
of course, a decent algorithm can do wonders, too.
|
|
|
|
|
why isn't talking about matlab?
|
|
|
|
|
Hello Good people,am not so sure if this should go here,i am student who is interested in developing a search engine that indexes pages from my country.I have been doing my research on Algorithm to use for sometime now and i have found HITS and PageRank as the best out there.I have choosing to go with PageRank since it is more stable than the HITS Algorithm(so i read).
I have found countless articles and university researches about PageRank but my problem is that i do not understand most of the mathematical symbols that form the algorithm in this papers.Currently,i cannot understand how the Google Matrix(the irreducible,stochastic matrix) was calculated with the algorithm,i do not seem to understand the Algorithm used.
I did my reading from the articles below:
PDF 1
PDF 2
Please i need you to help me go through it,i need a basic explanation(examples will be nice) with less mathematical symbols.
Thanks in advance.
|
|
|
|
|
kentipsy wrote: Please i need you to help me go through it,i need a basic explanation(examples will be nice) with less mathematical symbols.
I don't think you will get a tutorial on statisitics and matrix arithmetic. I suspect that you will need to study the mathematics quite closely if you want to have any possibility of implementing this sort of product. However if you want a ready made solution you could always try one of these[^].
|
|
|
|
|
Hi i have a problem.
Thing like this:
This is multi-thread program, 1 master thread processing the msgs generated by child threads(typical consumer-producer model). I create a large share buf, all the children will put their msgs into this buf if they can get the buf lock. When a child put its msg into the buf completely, it will notify the master there is a msg waiting for process, so master aquire the buf lock and hold it and begin to deal with the msgs.
The problem is when there are a lot of child threads startup, there will be a lot of msgs waiting for process, then master thread will hold the the buf lock for a while which is not acceptable  , as no child thread can put msg into it. This problem apparently causes poor performance of this program.
So what i try to do is to separete the large share buf to a array of small buf, and children can put their msgs into the available buf, and master can process msgs in each small share buf one by one.
But i don't know whether there is a good and mature algorithm or way to manage the share buf array.
Any good idea is welcome! Thanks.
BTW, i'm not sure whether i make the problem clear.(Because english is not my mother tongue, forgive me if it is not clear enough.) if you are interested in this problem, please feel free to contact me and ask me more details by email or
MSN: donniehan@live.cn
Thanks again!
|
|
|
|
|
Hi,
if you have multiple producers and a single consumer, then the following seem logical:
1.
the normal way of communication would be through a single queue; all producers can enqueue messages, the consumer will dequeue the messages. Other means of communication could be used if there is a need for them, e.g. when message priorities have to be established.
2.
the queue (or other comm means) should be locked as short as possible; it is only while dequeueing that the consumer needs to lock the queue, the processing of a dequeued messages should happen with the queue unlocked (otherwise the queue would locked all the time).
3.
on average the consumer should be faster than each of the consumers, as it has to handle all the messages, whereas each producer only needs to handle its share of them. This can be implemented by giving the consumer a higher thread priority (tricky on Windows), or by limiting the size of the queue (or other comm means), so when the queue gets full, the producers get blocked.
Luc Pattyn
Have a look at my entry for the lean-and-mean competition; please provide comments, feedback, discussion, and don’t forget to vote for it! Thank you.
Local announcement (Antwerp region): Lange Wapper? Neen!
|
|
|
|
|
Thanks for your reply and the idea.
But it seems not solve my problem. Firstly, i can't higher this master thread priority, actually the child threads are more important for me. Because each thread serves a client during its lifecycle. So as you said when the master hold the queue lock, it will lock the buf for a while, and it makes client thread waiting for a long time, which is my problem
Anyway thanks your answer! I will vote for you when i know how to vote!
|
|
|
|
|
donniehan wrote: the child threads are more important for me
that may well be, but then you MUST solve the locked problem.
donniehan wrote: Because each thread serves a client
that doesn't tell me much. If it were say serial ports where new data is coming in all the time, and risking data loss, then yes.
Whatever, if your consumer can't cope, the overall system will fail in the end, unless the lifespan of the app is limited, so you can just buffer sufficient data and have the consumer catch up after all the producers are done.
Could you show some code, I've got a hunch something is not the way it should be?
Luc Pattyn
Have a look at my entry for the lean-and-mean competition; please provide comments, feedback, discussion, and don’t forget to vote for it! Thank you.
Local announcement (Antwerp region): Lange Wapper? Neen!
|
|
|
|
|
My code is based on the pg src(Postgres Sql).Do you know well about the pg source?
The child thread serves the client,
it means it has a lot of other works to do and has to response to the client. it is not acceptable when the child thread halt and no response for a long time. And what i try to do is to allocate serval share small buf for all children use. When any child needs to use the buf, it can find one available from buf array and just put the data into it, release the buf quickly and do its other jobs. Master thread can process all msgs in the buf array one by one. It takes very short time for each msg to be handled by master.And buf is large enough for their using. That's the situation.
|
|
|
|
|
Hi,
no I have no experience with Postgres SQL, but that won't matter much I think.
donniehan wrote: It takes very short time for each msg to be handled by master.
That is very good. It will be key to your system working well.
My simple design would consist of:
- a single Message class that will be instantiated a number of times; the instances will be recycled without loading the garbage collector.
- an "emptyCollection" holding a number of empty message objects; think of it as a queue, although other kinds of collections could work too.
- an "emptyLock" object used to guard operations on emptyCollection;
- a "filledCollection" holding a number of filled message objects; again think of it as a queue.
- a "filledLock" object used to guard operations on filledCollection.
- emptyCollection+emptyLock and filledCollection+filledLock could be two instances of a single class MySafeCollection.
The initialization code would create a sufficient number of message instances and put them in the emptyCollection; the filledCollection is still empty.
Then the system really starts and performs these steps:
1. the producers obtain a message object quickly (getting one from emptyCollection, this must be protected by locking on emptyLock, which never will take long);
2. let the producer fill the message object;
3. let the producer put its filled message into the "filledCollection", using the filledLock, which again will not take long.
4. let the consumer obtain a filled message object from the filledCollection(again shortly using the filledLock).
5. let the consumer process the message; as this is a short operation and should never become a bottleneck I would consider doing this on a thread with slightly higher priority.
6. let the consumer return the message to the "emptyCollection" using the emptyLock.
Possible refinements:
A. You still need a signaling mechanism: each time a producer adds something to the filledCollection, it should send a signal to the single consumer, so it goes and fetches one (or more) messages from the filledCollection. This avoids the need for a polling loop in which the consumer would waste CPU cycles when no messages got filles.
B. A producer needing a message object and not finding one in the "emptyCollection" could create a new one; that would make the initialization unnecessary, however when things go terribly wrong, you may end up with a huge number of messages in the filledCollection.
I have implemented the above scheme numerous times inside embedded systems (i.e.using a proprietary operating system, not .NET or Windows). I know it works very well and is very performant.
Luc Pattyn
Have a look at my entry for the lean-and-mean competition; please provide comments, feedback, discussion, and don’t forget to vote for it! Thank you.
Local announcement (Antwerp region): Lange Wapper? Neen!
|
|
|
|
|
General 
News 
Suggestion 
Question 
Bug 
Answer 
Joke 
Praise 
Rant 
Admin 
Submit an article or tip
