There are a number of functions for creating XlaOps from native C++ values. I'm trying to figure out how to use each to construct a graph. I've gone through xla_builder.h and picked out some candidates, omitting overloads and convenience wrappers. The two most likely candidates seem to be
// Enqueues a "retrieve parameter value" instruction for a parameter that was
// passed to the computation.
XlaOp Parameter(XlaBuilder* builder, int64 parameter_number, const Shape& shape,
const string& name);
// Enqueues a constant with the value of the given literal onto the
// computation.
XlaOp ConstantLiteral(XlaBuilder* builder, const LiteralSlice& literal);
Am I right in thinking Parameter is for "symbols", while ConstantLiteral is for constant values? For example, in f(x) = x + 1, we'd encode 1 as a ConstantLiteral, and then for x we could either
write f(x) as a C++ function, and at application site use another ConstantLiteral for our value of x, or
encode x using Parameter and build an XlaComputation from the corresponding XlaBuilder. That said, I'm not clear on how to actually call the XlaComputation with a Literal, other than with LocalClient which doesn't work with to multiple XlaComputations afaict.
What's the difference between these two approaches? Is one better than the other? I notice the former doesn't appear possible for higher-order functions: those which accept XlaComputations.
Next there's
Infeed, which I'd guess is a streaming version of Parameter.
Recv which looks like a way to pass data between computations, but doesn't actually create a completely new XlaOp itself.
ReplicaId, Iota, and XlaOp CreateToken(XlaBuilder* builder); appear largely irrelevant for this discussion.
Have I got this right? Are there any other important functions I've missed?
Related
Suppose we have a set of 2 functions with multiple common arguments (x,y,z), let f_i(x,y,z) be one of those functions. When these arguments are evaluated by specific real numbers, Mathematica provides a solution which contains a real part and a very small irreal number (which I think is a calculation error), for the two functions.
I would like to create a new function with the same arguments that, when evaluated, chooses only the real part of the result of the function whose real result satisfies a certain criteria (e.g. be between -1 and 0).
This final function should allow me to plot the real part that matches the criteria in terms of any of its variables and to create other new functions.
I have tried the Chop and If functions in multiple orders without any success. My problem is that the chop function operates directly on the unevaluated functions and thus, does not allow me to attain the mentioned objective.
f[x_, y_, z_] =If[-1 <= Chop[f_1[x, y, z]] <= 0,Chop[f_1[x, y, z]],Chop[f_2[x, y, z]]]
Thank you very much.
Before saying anything, please let me make it clear that this question is NOT language-specific but part of my work on an interpreter.
Let's say we have an enum of Value types. So a value can be:
SV, // stringValue
IV, // integerValue
AV, // arrayValue
etc, etc
then let's say we have a function F which takes one of the following combinations of arguments:
[
[SV],
[SV,IV],
[AV]
]
Now, the function is called, we calculate the values passed, and get their types. Let's say we get [XV,YV].
The question is:
What is the most efficient way to check if the passed values are allowed?
(The original interpreter is written in Nim, so one could say we could lookup the value array in the array of accepted value arrays like: accepted.contains(passed) - but this is not efficient)
P.S. ^ That's how I'm currently doing it, although I've explored the option of using bitmasks too. However, I cannot seem how it would help, since order plays an important part too.
I've got a function which looks like this:
bool generate_script (bool net, bool tv, bool phone,
std::string clientsID,
std::string password,
int index, std::string number,
std::string Iport, std::string sernoID,
std::string VoiP_number, std::string VoiP_pass,
std::string target, int slot, int port,
int onu, int extra, std::string IP, std::string MAC);
In my opinion it looks ugly. What is the proper way of handling this problem? Should I create few vectors with different data types (int, string and bool) and pass them as arguments to this function?
If all these parameters are meaningfully related, pack them in a structure.
Put them in a struct
Create a structure
struct GenerateScriptParams { /* ... */ };
and put all the parameters in there. You can actually provide default values for the initialization of the struct as well by implementing a default constructor or, in C++11, by providing default initialization of individual members. You can then change the values that are not supposed to be defaulted. This selective picking of non-default parameters is not possible for a function call with lots of parameters in C++.
Making the interface nice for the caller
Yet, the usage is a little ugly, since you have to create a temporary name object, then change the values that should not be default and then pass the object to the function:
GenerateScriptParams gsp;
gsp.net = true;
gsp.phone = false;
gps.extra = 10;
generate_script( gsp );
If you call that function in several different places, it makes sense to avoid this uglyness by providing mutating member functions that can be chained:
GenerateScriptParams & GenerateScriptParams::setNet ( bool val );
GenerateScriptParams & GenerateScriptParams::setTV ( bool val );
GenerateScriptParams & GenerateScriptParams::setPhone( bool val );
// ... //
Then calling code can write
generate_script( GenerateScriptParams()
.setNet(true),
.setPhone(false),
.setExtra(10) );
without the above uglyness. This avoids the named object that is only used once.
I personally do not believe that moving all the arguments in one struct will make the code much better. You just move dirt under the carpet. When you are going to deal with the creation of the struct you have the same problem.
The question is how much reusable this struct will be? If you end up with a 18 parameters for one function call something it is not quite right in your design. After further analysis you may discover that those parameters can be group in several different classes and those classes could be aggregated to one single object that will be the input of your function. You may want also prefer classes to struct in order to protect your data.
EDIT
I will give you a small example to describe why several classes are better than one monolithic struct. Let's start counting the tests that you need to write to cover the function above. There are 18 parameters as input (3 boolean). So we are going to need at least 15 tests only to validate the input (assuming the values are not interconnected).
The overall number of tests is impossible to be calculated without the implementation, but we can have an idea of the magnitude. Let take the lower bound all the input can be treat as boolean the number of possible combination are 2^18 so around 262000 tests.
Now, what happen if we split the input in several objects?
First of all, the code to validate the input is moved away from the function to the body of every single object (and it can be reused).
But more importantly the number of tests will collapse, let say in group of four (4,4,4 and 4 params per object) the total number of tests is only:
2^4 + 2^4 + 2^4 + 2^4 + 2^4 = 80
The fifth attributes is due to the permutation of the objects with themselves.
So, what is more cost demanding? Write thousand of tests or few more classes?
Obviously, this is a crude simplification, however, it will underlying the core of the problem. A clutter interface is not just matter of style or an inconvenient for the developer it is a true impediment to produce quality code.
This is the most important lesson I ever learnt in my career as a professional developer: "Big classes and fat interfaces are evil". That's just my heuristic version of the single responsibility principle (I have notice that the SRP can be tricky to get it right, what it seems reasonable to be single responsibility it can be not quite the same after a hour coding, so I used some heuristic rule to help me to revaulate my initial choices).
Or you could use a fluent interface. It would look like this:
script my_script(mandatory, parameters);
my_script.net(true).tv(false).phone(true);
This is applicable if you have default values for your specified parameters or it is allowed to have a partially constructed script.
Ignoring the possibility or desirability of changing the function or program in some way as to reduce the number of parameters...
I have seen coding standards that specify how long parameter lists should be formatted, for cases where refactoring is not possible. One such example is using double indentations and one parameter per line (Not for all functions - only for those that have multiple-lines of parameters).
E.g.
bool generate_script (
bool net,
bool tv,
bool phone,
std::string clientsID,
std::string password,
int index,
std::string number,
std::string Iport,
std::string sernoID,
std::string VoiP_number,
std::string VoiP_pass,
std::string target,
int slot,
int port,
int onu,
int extra,
std::string IP,
std::string MAC);
The point here is to create a consistent layout and look for all functions with a large number of parameters.
A bit late here, but since nobody has done it yet, I'd like to point out an obvious aspect of the issue: to me, a function which takes so many arguments is likely to do a lot of computation, so consider the possibility of decomposing it in smaller functions as a first step.
This should help you structuring your data.
So I have a function that takes in 2 different inputs.
I've ran into the situation, however, where I very occasionally need a third input. Most of the time I don't though.
The solution I currently have is that the actual function I want to use is only called by 2 other functions. These two functions have the same name, but 1 takes 3 input and the other 2 (with this one just setting a null value to the third input before calling the original function).
This works quite well, but it feels like there might be a much better way of handling this type of problem. The only other solution I have is to declare a null value of the third input every time I go to call the first function, but that seems even messier.
Is there a better way to do this? Is it bad form the way I've written it?
Default arguments:
void foo (int x, int y, int z = 0);
Unless you pass a third value, z will be 0 by default inside the function.
I'm wondering if there is a library like Boost Format, but which supports named parameters rather than positional ones. This is a common idiom in e.g. Python, where you have a context to format strings with that may or may not use all available arguments, e.g.
mouse_state = {}
mouse_state['button'] = 0
mouse_state['x'] = 50
mouse_state['y'] = 30
#...
"You clicked %(button)s at %(x)d,%(y)d." % mouse_state
"Targeting %(x)d, %(y)d." % mouse_state
Are there any libraries that offer the functionality of those last two lines? I would expect it to offer a API something like:
PrintFMap(string format, map<string, string> args);
In Googling I have found many libraries offering variations of positional parameters, but none that support named ones. Ideally the library has few dependencies so I can drop it easily into my code. C++ won't be quite as idiomatic for collecting named arguments, but probably someone out there has thought more about it than me.
Performance is important, in particular I'd like to keep memory allocations down (always tricky in C++), since this may be run on devices without virtual memory. But having even a slow one to start from will probably be faster than writing it from scratch myself.
The fmt library supports named arguments:
print("You clicked {button} at {x},{y}.",
arg("button", "b1"), arg("x", 50), arg("y", 30));
And as a syntactic sugar you can even (ab)use user-defined literals to pass arguments:
print("You clicked {button} at {x},{y}.",
"button"_a="b1", "x"_a=50, "y"_a=30);
For brevity the namespace fmt is omitted in the above examples.
Disclaimer: I'm the author of this library.
I've always been critic with C++ I/O (especially formatting) because in my opinion is a step backward in respect to C. Formats needs to be dynamic, and makes perfect sense for example to load them from an external resource as a file or a parameter.
I've never tried before however to actually implement an alternative and your question made me making an attempt investing some weekend hours on this idea.
Sure the problem was more complex than I thought (for example just the integer formatting routine is 200+ lines), but I think that this approach (dynamic format strings) is more usable.
You can download my experiment from this link (it's just a .h file) and a test program from this link (test is probably not the correct term, I used it just to see if I was able to compile).
The following is an example
#include "format.h"
#include <iostream>
using format::FormatString;
using format::FormatDict;
int main()
{
std::cout << FormatString("The answer is %{x}") % FormatDict()("x", 42);
return 0;
}
It is different from boost.format approach because uses named parameters and because
the format string and format dictionary are meant to be built separately (and for
example passed around). Also I think that formatting options should be part of the
string (like printf) and not in the code.
FormatDict uses a trick for keeping the syntax reasonable:
FormatDict fd;
fd("x", 12)
("y", 3.141592654)
("z", "A string");
FormatString is instead just parsed from a const std::string& (I decided to preparse format strings but a slower but probably acceptable approach would be just passing the string and reparsing it each time).
The formatting can be extended for user defined types by specializing a conversion function template; for example
struct P2d
{
int x, y;
P2d(int x, int y)
: x(x), y(y)
{
}
};
namespace format {
template<>
std::string toString<P2d>(const P2d& p, const std::string& parms)
{
return FormatString("P2d(%{x}; %{y})") % FormatDict()
("x", p.x)
("y", p.y);
}
}
after that a P2d instance can be simply placed in a formatting dictionary.
Also it's possible to pass parameters to a formatting function by placing them between % and {.
For now I only implemented an integer formatting specialization that supports
Fixed size with left/right/center alignment
Custom filling char
Generic base (2-36), lower or uppercase
Digit separator (with both custom char and count)
Overflow char
Sign display
I've also added some shortcuts for common cases, for example
"%08x{hexdata}"
is an hex number with 8 digits padded with '0's.
"%026/2,8:{bindata}"
is a 24-bit binary number (as required by "/2") with digit separator ":" every 8 bits (as required by ",8:").
Note that the code is just an idea, and for example for now I just prevented copies when probably it's reasonable to allow storing both format strings and dictionaries (for dictionaries it's however important to give the ability to avoid copying an object just because it needs to be added to a FormatDict, and while IMO this is possible it's also something that raises non-trivial problems about lifetimes).
UPDATE
I've made a few changes to the initial approach:
Format strings can now be copied
Formatting for custom types is done using template classes instead of functions (this allows partial specialization)
I've added a formatter for sequences (two iterators). Syntax is still crude.
I've created a github project for it, with boost licensing.
The answer appears to be, no, there is not a C++ library that does this, and C++ programmers apparently do not even see the need for one, based on the comments I have received. I will have to write my own yet again.
Well I'll add my own answer as well, not that I know (or have coded) such a library, but to answer to the "keep the memory allocation down" bit.
As always I can envision some kind of speed / memory trade-off.
On the one hand, you can parse "Just In Time":
class Formater:
def __init__(self, format): self._string = format
def compute(self):
for k,v in context:
while self.__contains(k):
left, variable, right = self.__extract(k)
self._string = left + self.__replace(variable, v) + right
This way you don't keep a "parsed" structure at hand, and hopefully most of the time you'll just insert the new data in place (unlike Python, C++ strings are not immutable).
However it's far from being efficient...
On the other hand, you can build a fully constructed tree representing the parsed format. You will have several classes like: Constant, String, Integer, Real, etc... and probably some subclasses / decorators as well for the formatting itself.
I think however than the most efficient approach would be to have some kind of a mix of the two.
explode the format string into a list of Constant, Variable
index the variables in another structure (a hash table with open-addressing would do nicely, or something akin to Loki::AssocVector).
There you are: you're done with only 2 dynamically allocated arrays (basically). If you want to allow a same key to be repeated multiple times, simply use a std::vector<size_t> as a value of the index: good implementations should not allocate any memory dynamically for small sized vectors (VC++ 2010 doesn't for less than 16 bytes worth of data).
When evaluating the context itself, look up the instances. You then parse the formatter "just in time", check it agaisnt the current type of the value with which to replace it, and process the format.
Pros and cons:
- Just In Time: you scan the string again and again
- One Parse: requires a lot of dedicated classes, possibly many allocations, but the format is validated on input. Like Boost it may be reused.
- Mix: more efficient, especially if you don't replace some values (allow some kind of "null" value), but delaying the parsing of the format delays the reporting of errors.
Personally I would go for the One Parse scheme, trying to keep the allocations down using boost::variant and the Strategy Pattern as much I could.
Given that Python it's self is written in C and that formatting is such a commonly used feature, you might be able (ignoring copy write issues) to rip the relevant code from the python interpreter and port it to use STL maps rather than Pythons native dicts.
I've writen a library for this puporse, check it out on GitHub.
Contributions are wellcome.