In migrating custom Core Image filter kernels to Metal Shading Language, I encountered an error with building the default Metal library (default.metallib):
metallib: error: exactly one input file required
I was under the impression these could be in separate .metal files. Attempting to merge them into one file leads to this error:
Metal library creation failed: Error Domain=MTLLibraryErrorDomain Code=3 "Filters module must contain no vertex/fragment/kernel functions but contains 1 kernel function"
Namespacing to metal and to coreimage prevent the compute kernel from showing up as an available function in the default library.
Found this SO answer, which recommends building separate libraries:
Metal: vertexFunction defined in .metal file becomes nil once setting Compiler and Linker Options for MSL cikernel
You may create multiple Foo.metal and Bar.metal files.
Just don't add them as linker targets.
Instead #include "Foo.metal" and #include "Bar.metal" in a Main.metal file.
And only add the Main.metal file as linker target.
That way there is only one .metal file, which includes all the other .metal files. Simple.
Therefore the content of the Main.metal file may very simply look like:
#include "Foo.metal"
#include "Bar.metal"
You can’t use the default Metal build pipeline for compiling multiple .metal files containing Core Image kernels into one library right now. The linker doesn’t allow merging multiple .air files into one .metallib when setting the -cikernel flag.
You either have to put all your kernels into one .metal file or use the solution I posted in the answer you linked to above.
Related
I've recently made a foray into building Linux-based embedded systems, a far cry from my usual embedded stuff where I have total control over everything.
As part of that, I'm looking into the Yocto/bitbake/OpenEmbedded build system.
There's one thing I'm grappling with and that's the layering concept, so I'm trying to both figure out the way in which layers use/affect other layers.
From my understanding to date, a .bb recipe file uses require to simply include another file, similar to C's #include "myheader.h" which generally looks locally.
A .bbappend file in an "upper" layer will auto-magically include the base file then make changes to it, sort of an inherent require.
In contrast, the inherit keyword looks for a .bbclass class file in much the same way as it locates the .bb files, and inherits all the detials from them (sort of like #include <stdio.h> which, again generally, looks in the system area(a)).
So the first part of my question is: is my understanding correct? Or am I being too simplistic?
The second part of my question then involves the use of BBEXTENDS in the light of my current understanding. If we already have the ability to extend a recipe by using require, what is the purpose of listing said recipes in a BBEXTENDS variable?
(a) Yes, I'm aware they're both totally implementation dependent in terms of where the headers come from, I'm simply talking about their common use.
The learning curve for Yocto is different than other building systems, that's why I understand your confusion. But trust me, this is worth it. Your questions are related to BitBake so I recommend the BitBake User Manual. Just ensure that you're reading the same version as your poky revision.
require and include.
require is similar to include and can be compared to #include from C and C++ just like you have written.
Although generally both of them should be used to add some extensions to a recipe (*.bb) which are common to some amount of recipes (simply - can be reused).
For instance: definitions of paths, custom tasks used by couple recipes. The common purpose is to make recipe cleaner and separate some constants for re-usage.
The very important thing -> difference between include and require (from BitBake manual):
The include directive does not produce an error when the file cannot be found. Consequently, it is recommended that if the file you are including is expected to exist, you should use require instead of include. Doing so makes sure that an error is produced if the file cannot be found.
As a result: when you include a file to *.bb and it hasn't been found, the BitBake will not raise an error during parsing this recipe.
If you would use require, the error will be raised. You should use require when the pointed file must exist because it contains important variables/tasks that are mandatory to process.
*.bbappend mechanism.
In the case of *.bbappend - it's very powerful. The typical usage is whey you are adding some custom modifications to the recipe from other layer (located above layer where original recipe is) by *.bbappend because (e.g): you are not the maintainer of original recipe or the modifications are only used in your project (then it should be located in your meta-layer). But you can also bbappend the recipe on the same layer. BitBake parses all layers and then 'creates' an output and executes it. More in chapter Execution from BitBake man.
inherit.
The inherit mechanism can be used to inherit *.bbclass where common tasks for some specific purpose are defined so you don't need to write them on your own, e.g: you use inherit cmake or inherit autotools to your recipe when it needs to provide output for sources that are built correspondingly by CMake (and you have CMakeLists.txt defined) or autotools (Makefile.am etc.).
The definitions of classes provided by OpenEmbedded are located under /meta/classes/ if you are using Yocto release with poky.
You can check them and you will see that for example autotools.bbclass has defined (among others) task: autotools_do_configure() so you don't need to write it from the scratch.
However, you can redefine it in your recipe (by just providing your own definition of this function). If the recipe can't be changed, then you can simply create a *.bbappend file and write your own function do_configure() which will override the function from *.bbclass. Just like in OO languages such as C++ or Java.
I have just one question related to Linux shared library files.
I saw lot of links related to dynamically shared library for the Linux O.S
http://www.yolinux.com/TUTORIALS/LibraryArchives-StaticAndDynamic.html
Here in above link it is mentioned ---
include file for the library: ctest.h
Now in LINUX to use the libdl in build function --- dlopen, dlsym, dlclose.
Do we really need to include the prototype file -- ctest.h -- for dynamic lybrary ?
Please give some suggestion related to above post.
You don't really need to include the header or prototype file for the dynamic library, you do however need to at least the specific type information for the value returned by dlsym.
See here and here for examples that don't contain the include file for the dynamic library.
In the example you posted, they started off with their library functions not having header files / function prototypes, which along with providing instruction on how to avoid C++ name mangling is why they included the header file in this case.
If you define your own libraries without function prototypes, either in the source file or the header file, then you will need to include the header file when using dlsym, otherwise the inclusion of the header file of the dynamic library is unnecessary as its function prototypes were already included in the generated shared object.
The function prototypes included in header file are so that functions implemented can be resolved by name by linker. Where as the shared object file regardless of how it is linked, contains the implementation of the library which the linker links to.
The short explanation is that header files that are included with the #include are processed by the preprocessor, which means the resulting source file / files passed to the linker has knowledge of who each function call is because it looks up the function call prototype that was in the include file and has been included in the modified source. Include files tell the linker about who the function call is.
Object files, Shared Object files and other libraries files tell either the linker about what the implementation of the function call prototype does.
To answer your question in the comment, you will only have to add the libdl.so path to LD_LIBRARY_PATH or to /etc/ld.so.conf and run ldconfig, if that library or its relevant symlink hierarchy isn't in a standard location such as /usr/lib/ or /lib/.
See the following relevant StackOverflow answer answer for more information.
Further information can be found at
Program Library Howto
C Libraries Howto
Compilers, Assemblers, Linkers, Loaders: A Short Course
Linkers and Loaders
Beginner's Guide to Linkers
This is probably a totally noob question but I have missing links in my mind when thinking about linking libraries in iOS. I usually just add a new library that's been cross compiled and set the build and linker paths without really know what I'm doing. I'm hoping someone can help me fill in some gaps.
Let's take the OpenCV library for instance. I have this totally working btw because of a really well written tutorial( http://niw.at/articles/2009/03/14/using-opencv-on-iphone/en ), but I'm just wanting to know what is exactly going on.
What I'm thinking is happening is that when I build OpenCV for iOS is that your creating object code that gets placed in the .a files. This object code is just the implementation files( .m ) compiled. One reason you would want to do this is to make it hard to see the source code and so that you don't have to compile that source code every time.
The .h files won't be put in the library ( .a ). You include the .h in your source files and these header files communicate with the object code library ( .a ) in some way.
You also have to include the header files for your library in the Build Path and the Library itself in the Linker Path.
So, is the way I view linking libraries correct? If , not can someone correct me on this ?
Basically, you are correct.
Compiling the source code of a library produces one object file for each of the source files (in more than one, if compiled multiply times against different architectures). Then all the object files are archived (or packaged) into one .a file (or .lib on Windows). The code is not yet linked at this stage.
The .h files provide an interface for the functionality exposed by the library. They contain constants, function prototypes, possibly global declarations (e.g. extern int bad_global;), etc. -- basically, everything that is required to compile the code which is using the library.
.h files do not 'communicate' with object code in any way. They simply provide clues for the compiler. Consider this header file:
// library.h
extern int bad_global;
int public_func(int, const void*);
By including this file in your own code, you're simply telling the compiler to copy and paste these declarations into your source file. You could have written declarations for OpenCV library and not use the headers provided with it. In other words, you're asking the compiler to not issue errors about undefined symbols, saying "I have those symbols elsewhere, ok? Here are their declarations, now leave me alone!".
The header files need to be included in the search path in order for compiler to find them. You could simply include them via the full path, e.g. #include "path/to/file.h", or supply an -I option for your compiler, telling him where to look for additional headers, and use #include <file.h> instead.
When your code is compiled, the declarations in header files serve as an indication that symbols your code is using are defined somewhere. Note the difference between the words declaration and definition. Header files contain only declarations most of the time.
Now, when your code is compiled, it must be linked in order to produce the final executable. This is where the actual object code stored in the library comes into play. The linker will look at each symbol, function call, etc. in your object code and then try to find the corresponding definition for each such symbol. If it doesn't find one in the object code of your program, it will look the standard library and any other library you've provided it with.
Thus, it is important to understand that compilation and linkage are two separate stages. You could write any function prototypes at all and use them in your code, it will compile cleanly. However, when it comes to the linking stage, you have to provide implementation for symbols used in your code, or you won't get your executable.
Hope that makes sense!
The .a is the compiled version of the code.
The header files provided with a library are its public interface. They show what classes, methods, properties are available. They do not "communicate" with the binary code.
The compiler needs the headers to know that a symbol (a method name for example) is defined somewhere else. They are associated with the right "piece of code" in the library binary later during the "link" step.
Can you guide me how to properly link static library to iPhone project. I use static library project added to app project as direct dependency (target -> general -> direct dependencies) and all works OK, but categories. A category defined in static library is not working in app.
So my question is how to add static library with some categories into other project?
And in general, what is best practice to use in app project code from other projects?
Solution: As of Xcode 4.2, you only need to go to the application that is linking against the library (not the library itself) and click the project in the Project Navigator, click your app's target, then build settings, then search for "Other Linker Flags", click the + button, and add '-ObjC'. '-all_load' and '-force_load' are no longer needed.
Details:
I found some answers on various forums, blogs and apple docs. Now I try make short summary of my searches and experiments.
Problem was caused by (citation from apple Technical Q&A QA1490 https://developer.apple.com/library/content/qa/qa1490/_index.html):
Objective-C does not define linker
symbols for each function (or method,
in Objective-C) - instead, linker
symbols are only generated for each
class. If you extend a pre-existing
class with categories, the linker does
not know to associate the object code
of the core class implementation and
the category implementation. This
prevents objects created in the
resulting application from responding
to a selector that is defined in the
category.
And their solution:
To resolve this issue, the static
library should pass the -ObjC option
to the linker. This flag causes the
linker to load every object file in
the library that defines an
Objective-C class or category. While
this option will typically result in a
larger executable (due to additional
object code loaded into the
application), it will allow the
successful creation of effective
Objective-C static libraries that
contain categories on existing
classes.
and there is also recommendation in iPhone Development FAQ:
How do I link all the Objective-C
classes in a static library? Set the
Other Linker Flags build setting to
-ObjC.
and flags descriptions:
-all_load Loads all members of static archive libraries.
-ObjC Loads all members of static archive libraries that implement an
Objective-C class or category.
-force_load (path_to_archive) Loads all members of the specified static
archive library. Note: -all_load
forces all members of all archives to
be loaded. This option allows you to
target a specific archive.
*we can use force_load to reduce app binary size and to avoid conflicts which all_load can cause in some cases.
Yes, it works with *.a files added to the project.
Yet I had troubles with lib project added as direct dependency. But later I found that it was my fault - direct dependency project possibly was not added properly. When I remove it and add again with steps:
Drag&drop lib project file in app project (or add it with Project->Add to project…).
Click on arrow at lib project icon - mylib.a file name shown, drag this mylib.a file and drop it into Target -> Link Binary With Library group.
Open target info in fist page (General) and add my lib to dependencies list
after that all works OK. "-ObjC" flag was enough in my case.
I also was interested with idea from http://iphonedevelopmentexperiences.blogspot.com/2010/03/categories-in-static-library.html blog. Author say he can use category from lib without setting -all_load or -ObjC flag. He just add to category h/m files empty dummy class interface/implementation to force linker use this file. And yes, this trick do the job.
But author also said he even not instantiated dummy object. Mm… As I've found we should explicitly call some "real" code from category file. So at least class function should be called.
And we even need not dummy class. Single c function do the same.
So if we write lib files as:
// mylib.h
void useMyLib();
#interface NSObject (Logger)
-(void)logSelf;
#end
// mylib.m
void useMyLib(){
NSLog(#"do nothing, just for make mylib linked");
}
#implementation NSObject (Logger)
-(void)logSelf{
NSLog(#"self is:%#", [self description]);
}
#end
and if we call useMyLib(); anywhere in App project
then in any class we can use logSelf category method;
[self logSelf];
And more blogs on theme:
http://t-machine.org/index.php/2009/10/13/how-to-make-an-iphone-static-library-part-1/
http://blog.costan.us/2009/12/fat-iphone-static-libraries-device-and.html
The answer from Vladimir is actually pretty good, however, I'd like to give some more background knowledge here. Maybe one day somebody finds my reply and may find it helpful.
The compiler transforms source files (.c, .cc, .cpp, .m) into object files (.o). There is one object file per source file. Object files contain symbols, code and data. Object files are not directly usable by the operating system.
Now when building a dynamic library (.dylib), a framework, a loadable bundle (.bundle) or an executable binary, these object files are linked together by the linker to produce something the operating system considers "usable", e.g. something it can directly load to a specific memory address.
However when building a static library, all these object files are simply added to a big archive file, hence the extension of static libraries (.a for archive). So an .a file is nothing than an archive of object (.o) files. Think of a TAR archive or a ZIP archive without compression. It's just easier to copy a single .a file around than a whole bunch of .o files (similar to Java, where you pack .class files into a .jar archive for easy distribution).
When linking a binary to a static library (= archive), the linker will get a table of all symbols in the archive and check which of these symbols are referenced by the binaries. Only the object files containing referenced symbols are actually loaded by the linker and are considered by the linking process. E.g. if your archive has 50 object files, but only 20 contain symbols used by the binary, only those 20 are loaded by the linker, the other 30 are entirely ignored in the linking process.
This works quite well for C and C++ code, as these languages try to do as much as possible at compile time (though C++ also has some runtime-only features). Obj-C, however, is a different kind of language. Obj-C heavily depends on runtime features and many Obj-C features are actually runtime-only features. Obj-C classes actually have symbols comparable to C functions or global C variables (at least in current Obj-C runtime). A linker can see if a class is referenced or not, so it can determine a class being in use or not. If you use a class from an object file in a static library, this object file will be loaded by the linker because the linker sees a symbol being in use. Categories are a runtime-only feature, categories aren't symbols like classes or functions and that also means a linker cannot determine if a category is in use or not.
If the linker loads an object file containing Obj-C code, all Obj-C parts of it are always part of the linking stage. So if an object file containing categories is loaded because any symbol from it is considered "in use" (be it a class, be it a function, be it a global variable), the categories are loaded as well and will be available at runtime. Yet if the object file itself is not loaded, the categories in it will not be available at runtime. An object file containing only categories is never loaded because it contains no symbols the linker would ever consider "in use". And this is the whole problem here.
Several solutions have been proposed and now that you know how all this plays together, let's have another look on the proposed solution:
One solution is to add -all_load to the linker call. What will that linker flag actually do? Actually it tells the linker the following "Load all object files of all archives regardless if you see any symbol in use or not'. Of course, that will work; but it may also produce rather big binaries.
Another solution is to add -force_load to the linker call including the path to the archive. This flag works exactly like -all_load, but only for the specified archive. Of course this will work as well.
The most popular solution is to add -ObjC to the linker call. What will that linker flag actually do? This flag tells the linker "Load all object files from all archives if you see that they contain any Obj-C code". And "any Obj-C code" includes categories. This will work as well and it will not force loading of object files containing no Obj-C code (these are still only loaded on demand).
Another solution is the rather new Xcode build setting Perform Single-Object Prelink. What will this setting do? If enabled, all the object files (remember, there is one per source file) are merged together into a single object file (that is not real linking, hence the name PreLink) and this single object file (sometimes also called a "master object file") is then added to the archive. If now any symbol of the master object file is considered in use, the whole master object file is considered in use and thus all Objective-C parts of it are always loaded. And since classes are normal symbols, it's enough to use a single class from such a static library to also get all the categories.
The final solution is the trick Vladimir added at the very end of his answer. Place a "fake symbol" into any source file declaring only categories. If you want to use any of the categories at runtime, make sure you somehow reference the fake symbol at compile time, as this causes the object file to be loaded by the linker and thus also all Obj-C code in it. E.g. it could be a function with an empty function body (which will do nothing when being called) or it could be a global variable accessed (e.g. a global int once read or once written, this is sufficient). Unlike all other solutions above, this solution shifts control about which categories are available at runtime to the compiled code (if it wants them to be linked and available, it accesses the symbol, otherwise it doesn't access the symbol and the linker will ignore it).
That's all folks.
Oh, wait, there's one more thing:
The linker has an option named -dead_strip. What does this option do? If the linker decided to load an object file, all symbols of the object file become part of the linked binary, whether they are used or not. E.g. an object file contains 100 functions, but only one of them is used by the binary, all 100 functions are still added to the binary because object files are either added as a whole or they are not added at all. Adding an object file partially is usually not supported by linkers.
However, if you tell the linker to "dead strip", the linker will first add all the object files to the binary, resolve all the references and finally scan the binary for symbols not in use (or only in use by other symbols not in use). All the symbols found to be not in use are then removed as part of the optimization stage. In the example above, the 99 unused functions are removed again. This is very useful if you use options like -load_all, -force_load or Perform Single-Object Prelink because these options can easily blow up binary sizes dramatically in some cases and the dead stripping will remove unused code and data again.
Dead stripping works very well for C code (e.g. unused functions, variables and constants are removed as expected) and it also works quite good for C++ (e.g. unused classes are removed). It is not perfect, in some cases some symbols are not removed even though it would be okay to remove them, but in most cases it works quite well for these languages.
What about Obj-C? Forget about it! There is no dead stripping for Obj-C. As Obj-C is a runtime-feature language, the compiler cannot say at compile time whether a symbol is really in use or not. E.g. an Obj-C class is not in use if there is no code directly referencing it, correct? Wrong! You can dynamically build a string containing a class name, request a class pointer for that name and dynamically allocate the class. E.g. instead of
MyCoolClass * mcc = [[MyCoolClass alloc] init];
I could also write
NSString * cname = #"CoolClass";
NSString * cnameFull = [NSString stringWithFormat:#"My%#", cname];
Class mmcClass = NSClassFromString(cnameFull);
id mmc = [[mmcClass alloc] init];
In both cases mmc is a reference to an object of the class "MyCoolClass", but there is no direct reference to this class in the second code sample (not even the class name as a static string). Everything happens only at runtime. And that's even though classes are actually real symbols. It's even worse for categories, as they are not even real symbols.
So if you have a static library with hundreds of objects, yet most of your binaries only need a few of them, you may prefer not to use the solutions (1) to (4) above. Otherwise you end up with very big binaries containing all these classes, even though most of them are never used. For classes you usually don't need any special solution at all since classes have real symbols and as long as you reference them directly (not as in the second code sample), the linker will identify their usage pretty well on its own. For categories, though, consider solution (5), as it makes it possible to only include the categories you really need.
E.g. if you want a category for NSData, e.g. adding a compression/decompression method to it, you'd create a header file:
// NSData+Compress.h
#interface NSData (Compression)
- (NSData *)compressedData;
- (NSData *)decompressedData;
#end
void import_NSData_Compression ( );
and an implementation file
// NSData+Compress
#implementation NSData (Compression)
- (NSData *)compressedData
{
// ... magic ...
}
- (NSData *)decompressedData
{
// ... magic ...
}
#end
void import_NSData_Compression ( ) { }
Now just make sure that anywhere in your code import_NSData_Compression() is called. It doesn't matter where it is called or how often it is called. Actually it doesn't really have to be called at all, it's enough if the linker thinks so. E.g. you could put the following code anywhere in your project:
__attribute__((used)) static void importCategories ()
{
import_NSData_Compression();
// add more import calls here
}
You don't have to ever call importCategories() in your code, the attribute will make the compiler and linker believe that it is called, even in case it is not.
And a final tip:
If you add -whyload to the final link call, the linker will print in the build log which object file from which library it did load because of which symbol in use. It will only print the first symbol considered in use, but that is not necessarily the only symbol in use of that object file.
This issue has been fixed in LLVM. The fix ships as part of LLVM 2.9 The first Xcode version to contain the fix is Xcode 4.2 shipping with LLVM 3.0. The usage of -all_load or -force_load is no longer needed when working with XCode 4.2 -ObjC is still needed.
Here's what you need to do to resolve this problem completely when compiling your static library:
Either go to Xcode Build Settings and set Perform Single-Object Prelink to YES or
GENERATE_MASTER_OBJECT_FILE = YES in your build configuration file.
By default,the linker generates an .o file for each .m file. So categories gets different .o files. When the linker looks at a static library .o files, it doesn't create an index of all symbols per class (Runtime will, doesn't matter what).
This directive will ask the linker to pack all objects together into one big .o file and by this it forces the linker that process the static library to get index all class categories.
Hope that clarifies it.
One factor that is rarely mentioned whenever the static library linking discussion comes up is the fact that you must also include the categories themselves in the build phases->copy files and compile sources of the static library itself.
Apple also doesn't emphasize this fact in their recently published Using Static Libraries in iOS either.
I spent a whole day trying all sorts of variations of -objC and -all_load etc.. but nothing came out of it.. this question brought that issue to my attention. (don't get me wrong.. you still have to do the -objC stuff.. but it's more than just that).
also another action that has always helped me is that I always build the included static library first on its own.. then i build the enclosing application..
You probably need to have the category in you're static library's "public" header: #import "MyStaticLib.h"
I am writing a reusable static library for the iPhone, following the directions provided here.
I want to use minizip in my library internally, but don't want to expose it to the user.
It should be possible for the user to include minizip themselves, possibly a different version, and not cause clashes with my "inner" minizip version.
Is this possible?
Edit:
I've tried adding -fvisibility=hidden to additional compiler flags for minizip files and changing functions to be __private_extern__ and __attribute__((visibility("hidden"))), but it still seems to produce defined external symbols:
00000918 T _unzOpen
0000058e T _unzOpen2
00001d06 T _unzOpenCurrentFile
00001d6b T _unzOpenCurrentFile2
...
Edit #2:
Apparently the symbols marked with these annotations are only made private by the linker, which never happens when Xcode builds the sources, since it adds the -c parameter ("Compile or assemble the source files, but do not link.")
You could rename all exported symbol from minizip with objcopy.
something like
objcopy -redefine-sym=minizip.syms yourstaticlibray.a
and minizip.syms
_unzOpen _yourownprefix_unzOpen
_unzOpen2 _yourownprefix_unzOpen2
... ...
No clash if an executable is linked with an other minizip.a and yourstaticlibray.a, and because you renamed all the symbol in yourstaticlibray.a your call inside yourstaticlibray.a to minizip will use the prefixed symbol, and not the unzOpen one.
Since static library is nothing more than a set of .o files (which are not linked yet, as you have mentioned), the only way to completely hide presence of minizip from the outside world is to somehow compile minizip and your library together as a single compilation unit and make minizip functions/variables static.
You could have a look at how does SQLite do the "amalgamation" process which turns library source code into single .c file for further compilation: The SQLite Amalgamation.
As a bonus you'll get better optimization (really recent GCC and Binutils are able to make link-time optimizations, but this functionality is not released yet).