turning json option into macro parameter

[lemire]

We have added a few options (such as json_fmt). These trigger a short series of branches in our parse_number_string function. Unfortunately, this is costing us a bit of performance. We can get it back by turning the runtime parameters into template parameters... there is still a branch, but it is more direct and likely extremely cheap.

Building:

cmake -B build -D FASTFLOAT_BENCHMARKS=ON
cmake --build build

Let us benchmark on an Apple M2 processor (LLVM 16). These benchmarks are part of our library.

Before:

⚡  ./build/benchmarks/realbenchmark
####
# reading /Users/dlemire/CVS/github/fast_float/bui/benchmarks/data/canada.txt
####
# read 111126 lines
ASCII volume = 1.93374 MB
fastfloat (64)                          :  1166.05 MB/s (+/- 8.5 %)    67.01 Mfloat/s
fastfloat (32)                          :  1211.52 MB/s (+/- 7.5 %)    69.62 Mfloat/s
UTF-16 volume = 3.86749 MB
fastfloat (64)                          :  2187.60 MB/s (+/- 5.6 %)    62.86 Mfloat/s
fastfloat (32)                          :  2269.15 MB/s (+/- 4.6 %)    65.20 Mfloat/s
####
# reading /Users/dlemire/CVS/github/fast_float/bui/benchmarks/data/mesh.txt
####
# read 73019 lines
ASCII volume = 0.536009 MB
fastfloat (64)                          :   773.74 MB/s (+/- 4.7 %)   105.40 Mfloat/s
fastfloat (32)                          :   686.86 MB/s (+/- 4.3 %)    93.57 Mfloat/s
UTF-16 volume = 1.07202 MB
fastfloat (64)                          :  1551.87 MB/s (+/- 4.8 %)   105.70 Mfloat/s
fastfloat (32)                          :  1396.92 MB/s (+/- 6.0 %)    95.15 Mfloat/s

After:

⚡  ./build/benchmarks/realbenchmark
####
# reading /Users/dlemire/CVS/github/fast_float/build/benchmarks/data/canada.txt
####
# read 111126 lines
ASCII volume = 1.93374 MB
fastfloat (64)                          :  1175.62 MB/s (+/- 4.9 %)    67.56 Mfloat/s
fastfloat (32)                          :  1224.92 MB/s (+/- 4.1 %)    70.39 Mfloat/s
UTF-16 volume = 3.86749 MB
fastfloat (64)                          :  2220.99 MB/s (+/- 4.4 %)    63.82 Mfloat/s
fastfloat (32)                          :  2292.35 MB/s (+/- 4.4 %)    65.87 Mfloat/s
####
# reading /Users/dlemire/CVS/github/fast_float/build/benchmarks/data/mesh.txt
####
# read 73019 lines
ASCII volume = 0.536009 MB
fastfloat (64)                          :   873.69 MB/s (+/- 2.0 %)   119.02 Mfloat/s
fastfloat (32)                          :   785.65 MB/s (+/- 4.3 %)   107.03 Mfloat/s
UTF-16 volume = 1.07202 MB
fastfloat (64)                          :  1638.44 MB/s (+/- 1.8 %)   111.60 Mfloat/s
fastfloat (32)                          :  1548.04 MB/s (+/- 4.2 %)   105.44 Mfloat/s

On the mesh dataset, we get a performance boost of about 15%.

This might be related to PR #307 by @IRainman

[dalle]

Looks good. Do you think there be a macro for if-constexpr for C++17 enabled compilers?

Something like this?

#if __cplusplus >= 201703L || (defined(_MSVC_LANG) && _MSVC_LANG >= 201703L)
#define FASTFLOAT_IF_CONSTEXPR17(x) if constexpr (x)
#else
#define FASTFLOAT_IF_CONSTEXPR17(x) if (x)
#endif

...
FASTFLOAT_IF_CONSTEXPR17(basic_json_fmt)

[dalle]

Could you move the #include "event_counter.h" at the top of the file outside of the ifdef? That way it will compile on Windows too.

[lemire]

Ok

[IRainman]

Nice! Also the code definitely more cache friendly, it's shown by more constant time results on output.

[dalle]

Looks good. I commented somewhere else that we could explore using if-constexpr, but it's up to you. I'm a bit worried of "constant expression in if-statement" warnings appearing.

[IRainman]

Also about std::from_chars and compatibility, as you can remember, the standard regexp isn't ideal and the fmt library, especially with compile time parsing macros
fmt::format(FMT_COMPILE("some value: {}"), value));
much faster than std implementation and generates very compact code.

[IRainman]

Looks good. I commented somewhere else that we could explore using if-constexpr, but it's up to you. I'm a bit worried of "constant expression in if-statement" warnings appearing.

It's probably fine because it's old code and in the std function should be consteval in used in constexpr if. Maybe I should write a defect report for standard std::is_constant_evaluated. What do you think?

[lemire]

@dalle

I am adding

#if defined(__cpp_if_constexpr) && __cpp_if_constexpr >= 201606L
#define FASTFLOAT_IF_CONSTEXPR17(x) if constexpr (x)
#else
#define FASTFLOAT_IF_CONSTEXPR17(x) if (x)
#endif

[IRainman]

Maybe it's better to convert all options into a template parameter?

[lemire]

Maybe it's better to convert all options into a template parameter?

Well. That's an empirical question (i.e., see benchmarks).