Implement a c++20 constexpr type demangler #2472

fwyzard · 2025-02-01T21:18:06Z

Replace the run-time type demangler based on Boost with a constexpr version based on embedding the type name in the compiler generated function name.

Based on https://www.reddit.com/r/cpp/comments/lfi6jt/finally_a_possibly_portable_way_to_convert_types/ and https://rodusek.com/posts/2021/03/09/getting-an-unmangled-type-name-at-compile-time/ .

fwyzard · 2025-02-01T21:30:42Z

include/alpaka/core/DemangleTypeNames.hpp

-#if BOOST_COMP_CLANG
-#    pragma clang diagnostic pop
-#endif
+    inline constexpr std::string demangled = std::string(demangle<T>());


This should be rewritten using an std::string_view rather than an std::string - but it requires changes in many other places, I'll take care in a second moment.

fwyzard · 2025-02-02T11:05:32Z

Looks like event with -std=c++20 many of the compilers we support did not provide std::format :-(

fwyzard · 2025-02-02T11:12:21Z

I really hate clang pointless warnings about perfectly valid C++ being treated as errors:

include/alpaka/acc/AccCpuSerial.hpp:162:42: error: declaration requires an exit-time destructor [-Werror,-Wexit-time-destructors]

fwyzard · 2025-02-02T12:31:35Z

Looks like event with -std=c++20 many of the compilers we support did not provide std::format :-(

Same for std::source_location:

2025-02-02T12:02:57.1072996Z /__w/alpaka/alpaka/include/alpaka/core/DemangleTypeNames.hpp:14:10: fatal error: 'source_location' file not found
2025-02-02T12:02:57.1073952Z #include <source_location>
2025-02-02T12:02:57.1074334Z ^~~~~~~~~~~~~~~~~

SimeonEhrig · 2025-02-03T08:12:37Z

include/alpaka/acc/AccCpuOmp2Blocks.hpp

+#    ifdef __cpp_lib_format
+                    std::format("AccCpuOmp2Blocks<{},{}>", TDim::value, core::demangled<TIdx>);
+#    else
+                    "AccCpuOmp2Blocks<"s + std::to_string(TDim::value) + ","s + std::string(core::demangled<TIdx>)


Is std::stringstream maybe a better solution?

I can give it a try.

I don't know, if std::stringstream has a performance disadvantage compare to std::format. I need to check it. But if not, I would suggest it.

I was a little bit searching. std::format should be faster by factors and concatenating std::string should be faster than std::stringstream. The benefit of std::stringstream is the robustness for different data types.

https://mariusbancila.ro/blog/2023/09/12/formatting-text-in-c-the-old-and-the-new-ways/

https://stackoverflow.com/questions/14741144/stdstringstream-vs-stdstring-for-concatenating-many-strings

https://lemire.me/blog/2023/10/19/for-processing-strings-streams-in-c-can-be-slow/

This should work:

static std::string const accName = # ifdef __cpp_lib_format std::format("AccCpuOmp2Blocks<{},{}>", TDim::value, core::demangled<TIdx>); # else (std::stringstream() << "AccCpuOmp2Blocks<" << TDim::value << "," << core::demangled<TIdx> << ">").str(); # endif

Do you prefer to keep the optional use of std::format, or go with std::stringstream in any case ?

I also thought about this idea, but than the PR does not make sense anymore, because the PR should remove a dependency and not replace it.

I don't mind having boost as a dependency, so for me improvement is the constexpr functionality.

However, unlike boost I think {fmt} is header only, and we could include it directly in alpaka as a third party library like catch and mdspan.

Anyway, it's just an idea, and in case we can follow up in a later PR.

I don't mind having boost as a dependency, so for me improvement is the constexpr functionality.

Ah okay. I thought this PR is pulling out a part of this PR #2409. But this was a misunderstanding from my side (to much different stuff in the morning ;-) ).

However, unlike boost I think {fmt} is header only, and we could include it directly in alpaka as a third party library like catch and mdspan.

Anyway, it's just an idea, and in case we can follow up in a later PR.

Fair enough. We need only to support one specific version, which we can ship via sub directory or via CMake. We can discuss this tomorrow in the VC.

During the VC we discussed, if you should use the fmt library. @psychocoderHPC is against it, because the advantage of using it in a debug function is not big enough for the extra work caused by maintaining a dependency.

Unfortunately gcc 10 does not like the ostringstream-based syntax (it works with gcc 11 and later, and all clang+ versions I tried).

I'll revert to

... = "AccCpuOmp2Blocks<"s + std::to_string(TDim::value) + ","s + std::string(core::demangled<TIdx>);

SimeonEhrig · 2025-02-03T09:52:15Z

include/alpaka/acc/AccCpuOmp2Blocks.hpp

@@ -162,7 +166,21 @@ namespace alpaka
        {
            ALPAKA_FN_HOST static auto getAccName() -> std::string


The PR headline mentioned a constexpr type demangler. Is it possible to make the function constexpr or consteval? In this case the performance of std::format and std::stringstream does not matter.

I didn't find a way.

The performance doesn't matter much anyway, because the name is computed only once, the first time it's used.

At the moment I play a little bit around with it but it is not easy that compiler really optimize.
But if the performance does not matter, we simply keep the simple implementation and optimize it, if become an bottleneck.

Okay, I found a major problem, if we want to make it constexpr. C++23 is required for static variables in constexpr functions.

Yes, it's the same issue I had to work around here:

// Store the demangled type name as a null-terminated array of bytes. // Note: this could be a function-static constexpr variable in c++23. template<typename T> inline constexpr auto storage = demangleAsArray<T>();

Maybe we can use a consteval function and store it as a static constexpr data member.

During the VC today we talked about the subject. It is not performance critical. Therefore we can keep it as it is.

psychocoderHPC · 2025-02-04T09:15:46Z

The fallback source location implementation for demange equal to #2405?

psychocoderHPC · 2025-02-04T14:15:35Z

include/alpaka/acc/AccGpuUniformCudaHipRt.hpp

+#    ifdef __cpp_lib_format
+                    std::format("AccGpu{}Rt<{},{}>", TApi::name, TDim::value, core::demangled<TIdx>);
+#    else
+                    (std::ostringstream()


To workaround the nvcc error here

error: class "std::basic_ostream<char, std::char_traits<char>>" has no member "str" .str(); ^

I would try if std::string(<ostringstream>) works

psychocoderHPC · 2025-02-04T14:17:13Z

include/alpaka/core/DemangleTypeNames.hpp

+    // Store the demangled type name as a null-terminated array of bytes.
+    // Note: this could be a function-static constexpr variable in c++23.
+    template<typename T>
+    inline constexpr auto storage = demangleAsArray<T>();


ci error

/builds/hzdr/crp/alpaka/include/alpaka/core/DemangleTypeNames.hpp(78): error: call to consteval function "<unnamed>::demangleAsArray<T>() [with T=uint32_t]" did not produce a valid constant expression /builds/hzdr/crp/alpaka/include/alpaka/core/DemangleTypeNames.hpp(72): note #2751-D: access to uninitialized object

Maybe using inline ALPAKA_FN_HOST constexpr auto storage = demangleAsArray<T>(); solve the issue

fwyzard changed the title ~~Implement a c++20 constext type demangler~~ Implement a c++20 constexpr type demangler Feb 1, 2025

fwyzard force-pushed the constexpr_demangle branch from 925d205 to 51449bb Compare February 1, 2025 21:18

fwyzard requested a review from SimeonEhrig February 1, 2025 21:20

fwyzard added the Type:Enhancement label Feb 1, 2025

fwyzard added this to the 2.0.0 milestone Feb 1, 2025

fwyzard linked an issue Feb 1, 2025 that may be closed by this pull request

remove boost demangle #2405

Open

fwyzard self-assigned this Feb 1, 2025

fwyzard force-pushed the constexpr_demangle branch from 51449bb to 2ae434c Compare February 1, 2025 21:28

fwyzard commented Feb 1, 2025

View reviewed changes

fwyzard force-pushed the constexpr_demangle branch 3 times, most recently from 6c62bb8 to 5b621d2 Compare February 2, 2025 11:04

fwyzard force-pushed the constexpr_demangle branch 2 times, most recently from 355b6a6 to ffd5c5d Compare February 2, 2025 12:00

fwyzard force-pushed the constexpr_demangle branch 3 times, most recently from ebde117 to 2824b49 Compare February 3, 2025 08:08

SimeonEhrig reviewed Feb 3, 2025

View reviewed changes

fwyzard force-pushed the constexpr_demangle branch from 2824b49 to b8b0985 Compare February 3, 2025 09:21

SimeonEhrig reviewed Feb 3, 2025

View reviewed changes

fwyzard removed their assignment Feb 3, 2025

fwyzard force-pushed the constexpr_demangle branch from b8b0985 to fe19cda Compare February 3, 2025 10:16

psychocoderHPC requested changes Feb 4, 2025

View reviewed changes

fwyzard force-pushed the constexpr_demangle branch 3 times, most recently from 29a8abc to 17523d9 Compare February 18, 2025 11:14

fwyzard added 3 commits February 18, 2025 15:59

Use std::format for building accelerator names

b156921

Implement a c++20 constexpr type demangler

925b90b

Avoid a false positive compiler error with GCC 12

4a223b3

fwyzard force-pushed the constexpr_demangle branch from 17523d9 to 4a223b3 Compare February 18, 2025 15:04

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Implement a c++20 constexpr type demangler #2472

Implement a c++20 constexpr type demangler #2472

fwyzard commented Feb 1, 2025 •

edited

Loading

fwyzard Feb 1, 2025 •

edited

Loading

fwyzard commented Feb 2, 2025 •

edited

Loading

fwyzard commented Feb 2, 2025 •

edited

Loading

fwyzard commented Feb 2, 2025

SimeonEhrig Feb 3, 2025

fwyzard Feb 3, 2025 •

edited

Loading

SimeonEhrig Feb 3, 2025

SimeonEhrig Feb 3, 2025

fwyzard Feb 3, 2025

SimeonEhrig Feb 3, 2025

fwyzard Feb 3, 2025 •

edited

Loading

SimeonEhrig Feb 3, 2025

SimeonEhrig Feb 4, 2025

fwyzard Feb 18, 2025

SimeonEhrig Feb 3, 2025

fwyzard Feb 3, 2025

SimeonEhrig Feb 3, 2025

SimeonEhrig Feb 3, 2025

fwyzard Feb 3, 2025

SimeonEhrig Feb 4, 2025

psychocoderHPC commented Feb 4, 2025

psychocoderHPC Feb 4, 2025

psychocoderHPC Feb 4, 2025

		@@ -162,7 +166,21 @@ namespace alpaka
		{
		ALPAKA_FN_HOST static auto getAccName() -> std::string

Implement a c++20 constexpr type demangler #2472

Are you sure you want to change the base?

Implement a c++20 constexpr type demangler #2472

Conversation

fwyzard commented Feb 1, 2025 • edited Loading

fwyzard Feb 1, 2025 • edited Loading

Choose a reason for hiding this comment

fwyzard commented Feb 2, 2025 • edited Loading

fwyzard commented Feb 2, 2025 • edited Loading

fwyzard commented Feb 2, 2025

Choose a reason for hiding this comment

fwyzard Feb 3, 2025 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

fwyzard Feb 3, 2025 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

psychocoderHPC commented Feb 4, 2025

Choose a reason for hiding this comment

Choose a reason for hiding this comment

fwyzard commented Feb 1, 2025 •

edited

Loading

fwyzard Feb 1, 2025 •

edited

Loading

fwyzard commented Feb 2, 2025 •

edited

Loading

fwyzard commented Feb 2, 2025 •

edited

Loading

fwyzard Feb 3, 2025 •

edited

Loading

fwyzard Feb 3, 2025 •

edited

Loading