-
Notifications
You must be signed in to change notification settings - Fork 91
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Fix #580 by using a fixed-point implementation for unit conversions using integer representations #615
base: master
Are you sure you want to change the base?
Conversation
Th hi; | ||
Tl lo; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Can we make those private
? In case we do, can we name them with the _
postfix to be consistent with all other places in the library?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I don't have strong feelings here. I was thinking it could be relevant for this type to be a literal type, given that we intend to use it mostly at compile-time. Perhaps it becomes useful in some numeric manipulation needed for magnitudes? Perhaps we want to store an offset between two quantity points/origins with 128 bits of precision? However, I'm aware that literal types are only needed if we want instances as NTTP, and neither of these use-cases definitely require a literal type. I'll make them private, we can still revert that later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I think you meant structural type? litearal types are types that can be used in constexpr
functions and those do not need public members.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Ah yes, of course.
}; | ||
|
||
|
||
constexpr double_width_int operator+(Tl rhs) const |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Should it be a non-member two-parameter function to make sure that conversions work for lhs and rhs? Should there be another overload with reversed arguments?
Tl lo; | ||
}; | ||
|
||
#if false && defined(__SIZEOF_INT128__) |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Always false?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
yeah, CI complains about the true-path because __int128_t
is non-standard. Given that this is an implementation-detail, I should probably look for a way to disable that warning instead.
#endif | ||
|
||
template<typename T> | ||
inline constexpr std::size_t integer_rep_width_v = std::numeric_limits<std::make_unsigned_t<T>>::digits; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
inline
not needed from variable templates.
|
||
template<std::integral U> | ||
requires(integer_rep_width_v<U> <= integer_rep_width_v<T>) | ||
auto scale(U v) const |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
constexpr
?
fractional_bits); | ||
} | ||
|
||
repr_t int_repr_is_an_implementation_detail_; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
private
?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Again, literal type, but again, I don't actually need that right now. I'll make it private (... and remove the is_an_implementation_detail_
). We can still change it if a need comes up.
/* using multiplier_type = conditional< | ||
treat_as_floating_point<c_rep_type>, | ||
// ensure that the multiplier is also floating-point | ||
conditional<std::is_arithmetic_v<value_type_t<c_rep_type>>, | ||
// reuse user's type if possible | ||
std::common_type_t<c_mag_type, value_type_t<c_rep_type>>, std::common_type_t<c_mag_type, double>>, | ||
c_mag_type>;*/ |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Why to comment this code instead of removing it?
Most of the review concerns should be fixed now. Also, I "fixed" the CI failure in the trigonometry tests by allowing the affected test-cases to deviate by two instead of one epsilon. I believe this appears because now, for What I did not yet address are the discussions about making the implementation details |
That CI failure for GCC 12 and GCC 13 |
I was not aware of this. This makes sense... But in such a case, I would obfuscate the names much more (similar to what we have in |
I think you did not check if properly. GCC complains about |
You are right of course. It just happens that |
I should probably increase the test coverage for that |
Hi @burnpanck, do you plan to work on this in the near future? I would love to see fixed-point-scaling in action. Also, @JohelEGP is writing API Reference for ISO papers now, and we do not know what to describe. Also, talking about ISO papers, it would be great if you could contribute an add this small chapter about scaling that we talked about some time ago. |
Hey @mpusz , sorry for the long pause, as usual, paid work did interfere. Could you point me to a good location where to start drafting the design rationale behind this change? |
35a5d23
to
8415998
Compare
@burnpanck, it's great to hear from you again 😃 The best way to contribute to the ISO proposal is by doing a PR to a 3045R4 file at https://github.com/mpusz/wg21-papers. Please note that this repo has git submodules. To generate the paper, go to the src directory and do |
I would recommend putting the chapter somewhere inside of https://mpusz.github.io/wg21-papers/papers/3045R4_quantities_and_units_library.html#quantities. |
…nt-multiplication-for-integral-conversion
8415998
to
99d4315
Compare
I just merged latest master, I believe it should pass the tests again now. Given that it does pass all the tests, I think it is potentially mergeable. Perhaps you want to review again, or do you want to wait for a first draft of the rationale doc? Before I remove the "draft" state, there are two areas where I'd like your input:
|
No, I would like to merge it ASAP. Having a rationale discussion in the ISO paper would be great, but it is a separate task. |
This is not exposed in the public API, so it does not affect our users' code. This is why I believe that if it improves performance (compile-time or runtime), we should use such types wherever we can and are sure they exist. A generic and portable implementation should be used only in the remaining cases. |
If those features are available on all the supported compilers, then there is no need for alternatives. We always compile with at least C++20. |
Those tests may be cleaned up and refactored if needed.
Sure, please note that I am also working on refactoring the template<typename T>
concept WeaklyRegular = std::copyable<T> && std::equality_comparable<T>;
template<typename T>
concept Scalar = is_scalar<T>;
template<typename T>
concept ScalarRepresentation = Scalar<T> && WeaklyRegular<T> && requires(T a, T b, std::intmax_t i, long double f) {
// scaling
{ a* i } -> Scalar;
{ i* a } -> Scalar;
{ a / i } -> Scalar;
{ a* f } -> Scalar; // TODO How this affects freestanding?
{ f* a } -> Scalar;
{ a / f } -> Scalar;
// scalar operations
{ a + b } -> Scalar;
{ a - b } -> Scalar;
{ a* b } -> Scalar;
{ a / b } -> Scalar;
};
template<typename T>
concept Representation =
detail::ScalarRepresentation<T> || detail::ComplexRepresentation<T> || detail::VectorRepresentation<T> ||
detail::Tensor2Representation<T> || detail::Tensor4Representation<T>; The scaling part probably needs modifications after your change? |
Probably. I see two options. Option one is to distinguish two types of representations: "floating point" and "fixed point" (integers fall under this). The former ones are easy, because they have a a large enough range so we don't care about over- and underflow, and scaling them usually does not risk loosing precision, because the "epsilon" scales with the number. The later ones require us to carefully reason about range and resolution, because every scaling operation "shifts out" bits either above or below. I should really get a draft in of that rationale, to clarify why these are essential considerations. The more I think of this however, I believe what we really want from a representation is that it can be scaled by a magnitude. Thus, I think we should actually use that as the "defining property"; we define an API such as What would be the preferred mechanism for customisation points? I recall a discussion in the "std::execution" proposal for C++26 where they considered several possible choices for customisation points; they deemed the free-function approach of "std::ranges" not suitable for their use-case, but I believe that was because they essentially need something like "multiple-dispatch"; multiple unrelated entities each should have a shot at customising operations involving each other. In this case, the magnitude is not a customisable object, so free functions should work fine. |
I tried as well and I couldn't make it for gcc-12 as well. So I cheated 😜 |
I just pushed another take at the
Both 3 & 4 may be difficult to fully express as C++ concepts; the problem with 3 is that we ideally would like to verify scaling against all possible magnitudes, and with 4 it may not even be clear what exactly should be allowed and what not. Thus, for now, I simply check 3 my testing for Note that I included two slightly different spellings of Finally, you may want to review the customisation rules of the scaling operation. The implementation of the two
In general, I believe that is more or less all that I would put into this PR, and therefore remove the "draft" mark as soon as all tests pass. |
…nt-multiplication-for-integral-conversion
While merging the latest changes, I ran into the following (now) failing test: |
Looks like I failed to correctly export Also, I presume |
example/measurement.cpp
Outdated
@@ -127,15 +149,23 @@ class measurement { | |||
private: | |||
value_type value_{}; | |||
value_type uncertainty_{}; | |||
|
|||
// NOLINTNEXTLINE(bugprone-easily-swappable-parameters) | |||
constexpr measurement(std::in_place_t, value_type val, value_type err) : |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This is not the right way of using in_place.
in_place
means that we pass a list of arguments for construction of one underlying type.
This is not the case here. We initialize two members with one argument each.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
If you would like to have in-place like construction for type then probably something like piecewise_construct is needed.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yeah, I just wanted a tag to select the private constructor, indicating "I know what I'm doing - I guarantee that err
is positive". It's mostly a performance optimisation though, so, for an example, we could just leave it out completely.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We can either remove it or use a new dedicated tag. Maybe something similar to:
mp-units/example/include/validated_type.h
Lines 43 to 44 in 727a898
inline constexpr struct validated_tag { | |
} validated; |
example/measurement.cpp
Outdated
} | ||
|
||
template<typename To, mp_units::Magnitude M> | ||
[[nodiscard]] constexpr measurement<To> scale(std::type_identity<measurement<To>>, M scaling_factor) const |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This type_identity
usage here is quite novel. I understand the rationale, but maybe we can find a more traditional way of doing it? @JohelEGP, do you have any ideas here?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The idea comes from boost::hana::type
. Would be great if the standard had its own dedicated "type tag".
if constexpr (std::is_integral_v<T>) { | ||
return static_cast<To>(value); | ||
} else { | ||
return value; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
If this is meant to be a floating point value, then !treat_as_floating_point
should be used.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Not really. The specification of cast_integral<To>(const From)&
(an implementation detail) is "cast to To
, if From
is integral". The point is that the default implementation of scale
basically has two separate cases:
- "use
operator*(float)
": Applied whenever at least one ofFrom
andTo
istreat_as_floating_point
. - "use a fixed-point multiplication": Can only be directly be applied to standard integral types, so is used just for those, plus custom types which are implicitly convertible and have an integral
value_type
.
So the first case is much broader. For standard types, this in particular includes the case where From
is integer and To
is floating-point. There, just multiplying the integer with some floating-point type may cause precision-loss warnigs to emitted. Those are "bogus", because whenever To
is explicitly specified, we're in fact in an "explicit cast" path. So we need to do an explicit cast here. But on the other hand, we cannot unilaterally do an explicit cast, because many types that should go through that path are not at all convertible to that floating-point type (e.g. std::complex
).
There are other ways to handle this. We could also just silence the warning locally, or branch-out the mixed int/float paths further above.
template<typename T> | ||
struct floating_point_scaling_factor_type { | ||
// fallback implementation for types with a `value_type` nested type | ||
using type = | ||
std::enable_if_t<!std::is_same_v<value_type_t<T>, T>, floating_point_scaling_factor_type<value_type_t<T>>>::type; | ||
}; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Probably, it is better to leave the primary template undefined and handle value_type_t
in a separate partial specialization. SFINAE is also much slower to compile than constraints and provides worse diagnostics.
template<std::floating_point T> | ||
struct floating_point_scaling_factor_type<T> : std::type_identity<T> {}; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I know type_identity
pattern, but it is actually more to type and requires one more dynamic allocation (class template instantiation) at compile time for each case. This is why I typically prefer to just type using
.
// placeholder to report failure | ||
return std::false_type{}; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
This should be equivalent to the lack of definition for scaling_traits
.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Not sure I understand correctly: you suggest select_scaling_traits
should not be defined for those types? Indeed, that is what I wanted, but somehow, I was unable to make it work. Let me try again, maybe it was due to another bug that I have fixed in the meantime...
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Yes, I think that having scaling_traits
specialization for treat_as_floating_point
and !treat_as_floating_point
would be a better solution here. It would use interfaces that also the user has to use to make things like this work. We do not need any additional dispatch here.
In case of problems with constraints subsumption, we could expose proper concepts to make it work.
If there is no way to do it as above, then we could implement this in a similar way to CPOs (Niebloids).
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
The latest stint here now indeed works with a scaling_traits<From,To>
, where To
is optional. I provided an implementation for treat_as_floating_point
(assuming that multiplication with a standard floating-point value is correct for all of those) and another one for "integer-likes" (standard integers, plus those which have a nested value_type
that is a standard integer, and which further interconverts with that). Note that this does not cover all representation types (those which are neither treat_as_floating_point
nor integer). That is on purpose. Any non-floating-point representation needs precise knowledge about range and precision, which we don't have. So requiring the user to provide a specialisation is the only right path here IMHO.
template<typename To, Magnitude M, typename From> | ||
requires detail::HasScalingTraits<To, From> || | ||
requires(const From& value) { value.scale(std::type_identity<To>{}, M{}); } | ||
constexpr To scale(std::type_identity<To> to_type, M scaling_factor, const From& value) | ||
{ | ||
if constexpr (requires { | ||
{ value.scale(to_type, scaling_factor) } -> std::convertible_to<To>; | ||
}) { | ||
return value.scale(to_type, scaling_factor); | ||
} else { | ||
return detail::select_scaling_traits<To, From>.scale_from(scaling_factor, value); | ||
} | ||
} |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
With the above suggestions, this will simplify to:
template<typename To, Magnitude M, typename From> | |
requires detail::HasScalingTraits<To, From> || | |
requires(const From& value) { value.scale(std::type_identity<To>{}, M{}); } | |
constexpr To scale(std::type_identity<To> to_type, M scaling_factor, const From& value) | |
{ | |
if constexpr (requires { | |
{ value.scale(to_type, scaling_factor) } -> std::convertible_to<To>; | |
}) { | |
return value.scale(to_type, scaling_factor); | |
} else { | |
return detail::select_scaling_traits<To, From>.scale_from(scaling_factor, value); | |
} | |
} | |
template<typename To, Magnitude M, typename From> | |
requires requires { typename scaling_traits<To, From>; } || | |
requires(const From& value) { value.scale(std::type_identity<To>{}, M{}); } | |
constexpr To scale(std::type_identity<To> to_type, M scaling_factor, const From& value) | |
{ | |
if constexpr (requires { | |
{ value.scale(to_type, scaling_factor) } -> std::convertible_to<To>; | |
}) { | |
return value.scale(to_type, scaling_factor); | |
} else { | |
return scaling_traits<To, From>.scale_from(scaling_factor, value); | |
} | |
} |
if constexpr (requires { value.scale(scaling_factor); }) { | ||
return value.scale(scaling_factor); |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I do not think that T::scale
is a good idea. I do not know of any representation type that provides such a member function today. Also, requiring representation types to add it tightly couples them with mp-units, which is not good. I would strongly prefer to have some non-member or type_traits-based customization only.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Fine with me. Explicit specialisations are sometimes a bit painful to do because of the namespaces. That said, there are plenty of precedents in the standard library. ADL is a bit more convenient, and modern approaches would go through a CPO. On the other hand, I would actually prefer not to dispatch on the magnitude type at all, because if a different implementation is chosen for different magnitudes by accident, that would be very confusing. Instead, all paths for different magnitudes should remain close and under the control of the same "customiser". Thus, I think I'll give it another shot just explicit specialisation of mp_units::scaling_traits<From, To=UnspecifiedRepresentation>
.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
ADL is not a good choice here as someone may want to provide a customization for a third-party existing type, and this would require injecting some stuff into the third-party namespace, which may not be allowed, like in the case of std
.
* return an instance of @c Rep that approximates `scaling_factor * value`, another element of $\mathcal{V}$. | ||
* This needs to be defined at least for `From = Rep`, as well as any other representation | ||
* types for which interoperability is desired. | ||
* - `template <typename To, Magnitude M> static constexpr auto scale(M scaling_factor, const Rep &value)`: |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I do not clearly understand what is the difference here. It both examples in scaling.h
you have exactly the same implementation. Do we need two functions here? Maybe one function would be enough with proper template parameters?
If we need both, then shouldn't those be named scale
and scale_to
instead?
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We just use one of them so far, but I think we should later be using both. See the comment above: #615 (comment)
As for naming: naming is hard. Within the previous arrangement, where Rep
in the scaling_traits<Rep>
could take either the From
or the To
role, I concluded that when there is no To
specified, scale
is a good name, but when Rep
is the To
, the From
type comes from the argument, so scale_from
is correct because the Rep
"scales from" the argument. I agree though, it's confusing. With scaling_traits<From, To=Unspecified>
, scale
is the only reasonable name.
|
||
template<typename T> | ||
concept MagnitudeScalable = detail::WeaklyRegular<T> && requires(T a, T b, std::type_identity<T> to_type) { | ||
{ mp_units::scale(mag<1>, a) }; |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
We could check return type here as well. For example, WeaklyRegular
could be better than nothing here.
https://github.com/mpusz/mp-units/blob/master/src/core/include/mp-units/bits/module_macros.h |
Yes, |
@@ -201,6 +199,8 @@ concept HasScalingTraits = !std::convertible_to<decltype(select_scaling_traits<T | |||
|
|||
} // namespace detail | |||
|
|||
MP_UNITS_EXPORT_BEGIN |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
As mentioned before, we should not provide public API in .../bits/...
.
namespace mp_units { | ||
|
||
namespace detail { |
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
namespace mp_units { | |
namespace detail { | |
namespace mp_units::detail { |
The above is shorter and makes it harder to export public APIs from .../bits/...
.
…t.cpp as per c++ core guidelines)
…cel previous runs on the same branch
…d; also, fix freestanding
…ion-for-integral-conversion
This PR provides an implementation of what I played with in this comment to issue 580. Basically, it completely distinguishes floating-point and integral paths. For floating-point paths, the conversion implementation remains the same (a single multiplication with a floating-point representation of the conversion factor). For the integral paths, it now also uses a single multiplication, but now with a fixed-point representation. (mixed conversions take the floating-point path). Fixed-point multiplications have the advantage that they are comparatively cheap (much cheaper than a division) and can accurately describe all reasonable conversion factors between two n-bit numbers using a n.n (2n total) fixed-point representation. Unreasonable conversion factors are those larger than 2^n or smaller than 2^-n, i.e. those which either overflow for all input values or underflow for all input values. The cost of such a fixed-point multiplication is at most that of 4x n-bit integer multiplication plus some bookkeeping if their output is restricted to n-bit (as in the C++ standard; unlike typical hardware); if the output of a n-bit multiplication can be 2n (most instruction-sets provide those), it will just take two of them.
With the implementation change here, the computation of conversions between two quantity types of the same dimension and both integer type stay tightly within the types mandated by the source and destination type, without expanding to intmax unless necessary. In general, this will never cause overflows unless the result type actually cannot carry the result (the fixed-point multiplication is guaranteed to be sufficient). However, this practice triggered the weakly related #614, where it turned a silent overflow into a compilation failure. For now, this PR just disables those two test. Once we have a fix for that one, we should re-enable those tests and rebase this PR.