-
Notifications
You must be signed in to change notification settings - Fork 50
Commit
This commit does not belong to any branch on this repository, and may belong to a fork outside of the repository.
- Loading branch information
Showing
6 changed files
with
110 additions
and
110 deletions.
There are no files selected for viewing
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,106 @@ | ||
| ### Changes | ||
|
|
||
| Improved the performance of `c4/charconv.hpp` functions ([PR#77](https://github.com/biojppm/c4core/pull/77)): | ||
| - Added `digits_dec/hex/oct/bin()`. | ||
| - Optimized `write_dec/hex/oct/bin()`: | ||
| - these functions now return immediately without entering the loop if the output buffer is smaller than respectively `digits_dec/hex/oct/bin()`. This enables both: | ||
| - writing every character in its final position without having to revert the string at the end | ||
| - the need to check the buffer size on appending every character. | ||
| - `write_dec()` now writes two digits at once, thus halving the number of integer divisions. | ||
| - Added `write_dec/hex/oct/bin_unchecked()`, which receive precomputed `digits_dec/hex/oct/bin()`, thus speeding up the radix `itoa()/utoa()` overloads. | ||
| - Added `xtoa()` radix+digits overloads: | ||
| - `size_t xtoa(substr s, T v, T radix)` | ||
| - `size_t xtoa(substr s, T v, T radix, size_t num_digits)` | ||
| - `read_dec/hex/oct/bin()`: these functions no longer allow an empty input buffer. | ||
| - Use intrinsic functions `__builtin_clz()` (gcc) / `_BitScanReverse()` (msvc) in `c4::msb()` and `__builtin_ctz()` (gcc) / `_BitScanForward()` (msvc) in `c4::lsb()` when they are available. `msb()` is used by `digits_hex()/digits_bin()`. | ||
| - Refactored the charconv tests to improve consistency and thoroughness. | ||
| - Improved the charconv benchmarks to ensure full consistency across benchmarks. | ||
| - Special thanks and kudos to @fargies for being attentive and pinpointing several issues throughout the PR! | ||
| - Finding the best approach involved [writing a R&D benchmark for the several algorithm components](https://github.com/biojppm/c4core/tree/master/bm/bm_xtoa.cpp). This benchmark is disabled by default, and can be enabled with the flag `C4CORE_BM_XTOA_RND`. | ||
| - With the changes from this PR, the [charconv benchmark](https://github.com/biojppm/c4core/tree/master/bm_charconv.cpp) results show that on Linux/g++11.2, with integral types: | ||
| - `c4::to_chars()` can be expected to be roughly... | ||
| - ~40% to 2x faster than `std::to_chars()` | ||
| - ~10x-30x faster than `sprintf()` | ||
| - ~50x-100x faster than a naive `stringstream::operator<<()` followed by `stringstream::str()` | ||
| - `c4::from_chars()` can be expected to be roughly... | ||
| - ~10%-30% faster than `std::from_chars()` | ||
| - ~10x faster than `scanf()` | ||
| - ~30x-50x faster than a naive `stringstream::str()` followed by `stringstream::operator>>()` | ||
| - Here are the results from the run: | ||
| | Write throughput | | Read throughput | | | ||
| |:-------------------------|--------:|:-------------------------|---------:| | ||
| | **write `uint8_t`** | **MB/s**| **read `uint8_t`** | **MB/s**| | ||
| | `c4::to_chars<u8>` | 526.86 | `c4::from_chars<u8>` | 163.06 | | ||
| | `std::to_chars<u8>` | 379.03 | `std::from_chars<u8>` | 154.85 | | ||
| | `std::sprintf<u8>` | 20.49 | `std::scanf<u8>` | 15.75 | | ||
| | `std::stringstream<u8>` | 3.82 | `std::stringstream<u8>` | 3.83 | | ||
| | **write `int8_t`** | **MB/s**| **read `int8_t`** | **MB/s**| | ||
| | `c4::to_chars<i8>` | 599.98 | `c4::from_chars<i8>` | 184.20 | | ||
| | `std::to_chars<i8>` | 246.32 | `std::from_chars<i8>` | 156.40 | | ||
| | `std::sprintf<i8>` | 19.15 | `std::scanf<i8>` | 16.44 | | ||
| | `std::stringstream<i8>` | 3.83 | `std::stringstream<i8>` | 3.89 | | ||
| | **write `uint16_t`** | **MB/s**| **read `uint16_t`** | **MB/s**| | ||
| | `c4::to_chars<u16>` | 486.40 | `c4::from_chars<u16>` | 349.48 | | ||
| | `std::to_chars<u16>` | 454.24 | `std::from_chars<u16>` | 319.13 | | ||
| | `std::sprintf<u16>` | 38.74 | `std::scanf<u16>` | 28.12 | | ||
| | `std::stringstream<u16>` | 7.08 | `std::stringstream<u16>`| 6.73 | | ||
| | **write `int16_t`** | **MB/s**| **read `int16_t`** | **MB/s**| | ||
| | `c4::to_chars<i16>` | 507.44 | `c4::from_chars<i16>` | 282.95 | | ||
| | `std::to_chars<i16>` | 297.49 | `std::from_chars<i16>` | 186.18 | | ||
| | `std::sprintf<i16>` | 39.03 | `std::scanf<i16>` | 28.45 | | ||
| | `std::stringstream<i16>` | 6.98 | `std::stringstream<i16>`| 6.49 | | ||
| | **write `uint32_t`** | **MB/s**| **read `uint32_t`** | **MB/s**| | ||
| | `c4::to_chars<u32>` | 730.12 | `c4::from_chars<u32>` | 463.95 | | ||
| | `std::to_chars<u32>` | 514.76 | `std::from_chars<u32>` | 329.42 | | ||
| | `std::sprintf<u32>` | 71.19 | `std::scanf<u32>` | 44.97 | | ||
| | `std::stringstream<u32>` | 14.05 | `std::stringstream<u32>`| 12.57 | | ||
| | **write `int32_t`** | **MB/s**| **read `int32_t`** | **MB/s**| | ||
| | `c4::to_chars<i32>` | 618.76 | `c4::from_chars<i32>` | 345.53 | | ||
| | `std::to_chars<i32>` | 394.72 | `std::from_chars<i32>` | 224.46 | | ||
| | `std::sprintf<i32>` | 71.14 | `std::scanf<i32>` | 43.49 | | ||
| | `std::stringstream<i32>` | 13.91 | `std::stringstream<i32>`| 12.03 | | ||
| | **write `uint64_t`** | **MB/s**| **read `uint64_t`** | **MB/s**| | ||
| | `c4::to_chars<u64>` | 1118.87 | `c4::from_chars<u64>` | 928.49 | | ||
| | `std::to_chars<u64>` | 886.58 | `std::from_chars<u64>` | 759.03 | | ||
| | `std::sprintf<u64>` | 140.96 | `std::scanf<u64>` | 91.60 | | ||
| | `std::stringstream<u64>` | 28.01 | `std::stringstream<u64>`| 25.00 | | ||
| | **write `int64_t`** | **MB/s**| **read `int64_t`** | **MB/s**| | ||
| | `c4::to_chars<i64>` | 1198.78 | `c4::from_chars<i64>` | 713.76 | | ||
| | `std::to_chars<i64>` | 882.17 | `std::from_chars<i64>` | 646.18 | | ||
| | `std::sprintf<i64>` | 138.79 | `std::scanf<i64>` | 90.07 | | ||
| | `std::stringstream<i64>` | 27.62 | `std::stringstream<i64>`| 25.12 | | ||
|
|
||
| If you feel suspicious about these bold claims, you can browse through [c4core's CI benchmark results](https://github.com/biojppm/c4core/actions/workflows/benchmarks.yml) which will hopefully give these more substance. | ||
|
|
||
|
|
||
| ### New features | ||
|
|
||
| - Added `bool c4::overflows<T>(csubstr s)` for detecting whether a string overflows a given integral type. See [PR#78](https://github.com/biojppm/c4core/pull/78). | ||
| - Also, added `c4::fmt::overflow_checked()` (and the corresponding `from_chars()` overload) to enable a check for overflow before parsing from string: | ||
| ```c++ | ||
| c4::from_chars(str, &val); // no overflow check | ||
| c4::from_chars(str, c4::fmt::overflow_checked(val)); // enable overflow check | ||
| // as an example, the implementation looks like: | ||
| template<class T> | ||
| bool c4::from_chars(c4::csubstr str, c4::fmt::overflow_checked<T> oc) | ||
| { | ||
| if(overflows<T>(str)) | ||
| return false; | ||
| return c4::from_chars(str, oc.val); | ||
| } | ||
| ``` | ||
| ### Fixes | ||
| - Fix missing endianess macro on windows arm/arm64 compilations [PR #76](https://github.com/biojppm/c4core/pull/76) | ||
| - Add missing `#define` for the include guard of the amalgamated header (see [rapidyaml#246](https://github.com/biojppm/rapidyaml/issues/246)). | ||
| - Fix CPU detection with ARMEL [PR #86](https://github.com/biojppm/c4core/pull/86). | ||
| - Fix GCC version detection [PR #87](https://github.com/biojppm/c4core/pull/87). | ||
| - Fix [cmake#8](https://github.com/biojppm/cmake/issues/8): `SOVERSION` missing from shared libraries. | ||
| - Update fastfloat to 3.5.1. | ||
| ### Thanks | ||
| - @fargies | ||
| - @daichifukui | ||
| - @janisozaur |
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -1,106 +0,0 @@ | ||
| ### Changes | ||
|
|
||
| Improved the performance of `c4/charconv.hpp` functions ([PR#77](https://github.com/biojppm/c4core/pull/77)): | ||
| - Added `digits_dec/hex/oct/bin()`. | ||
| - Optimized `write_dec/hex/oct/bin()`: | ||
| - these functions now return immediately without entering the loop if the output buffer is smaller than respectively `digits_dec/hex/oct/bin()`. This enables both: | ||
| - writing every character in its final position without having to revert the string at the end | ||
| - the need to check the buffer size on appending every character. | ||
| - `write_dec()` now writes two digits at once, thus halving the number of integer divisions. | ||
| - Added `write_dec/hex/oct/bin_unchecked()`, which receive precomputed `digits_dec/hex/oct/bin()`, thus speeding up the radix `itoa()/utoa()` overloads. | ||
| - Added `xtoa()` radix+digits overloads: | ||
| - `size_t xtoa(substr s, T v, T radix)` | ||
| - `size_t xtoa(substr s, T v, T radix, size_t num_digits)` | ||
| - `read_dec/hex/oct/bin()`: these functions no longer allow an empty input buffer. | ||
| - Use intrinsic functions `__builtin_clz()` (gcc) / `_BitScanReverse()` (msvc) in `c4::msb()` and `__builtin_ctz()` (gcc) / `_BitScanForward()` (msvc) in `c4::lsb()` when they are available. `msb()` is used by `digits_hex()/digits_bin()`. | ||
| - Refactored the charconv tests to improve consistency and thoroughness. | ||
| - Improved the charconv benchmarks to ensure full consistency across benchmarks. | ||
| - Special thanks and kudos to @fargies for being attentive and pinpointing several issues throughout the PR! | ||
| - Finding the best approach involved [writing a R&D benchmark for the several algorithm components](https://github.com/biojppm/c4core/tree/master/bm/bm_xtoa.cpp). This benchmark is disabled by default, and can be enabled with the flag `C4CORE_BM_XTOA_RND`. | ||
| - With the changes from this PR, the [charconv benchmark](https://github.com/biojppm/c4core/tree/master/bm_charconv.cpp) results show that on Linux/g++11.2, with integral types: | ||
| - `c4::to_chars()` can be expected to be roughly... | ||
| - ~40% to 2x faster than `std::to_chars()` | ||
| - ~10x-30x faster than `sprintf()` | ||
| - ~50x-100x faster than a naive `stringstream::operator<<()` followed by `stringstream::str()` | ||
| - `c4::from_chars()` can be expected to be roughly... | ||
| - ~10%-30% faster than `std::from_chars()` | ||
| - ~10x faster than `scanf()` | ||
| - ~30x-50x faster than a naive `stringstream::str()` followed by `stringstream::operator>>()` | ||
| - Here are the results from the run: | ||
| | Write throughput | | Read throughput | | | ||
| |:-------------------------|--------:|:-------------------------|---------:| | ||
| | **write `uint8_t`** | **MB/s**| **read `uint8_t`** | **MB/s**| | ||
| | `c4::to_chars<u8>` | 526.86 | `c4::from_chars<u8>` | 163.06 | | ||
| | `std::to_chars<u8>` | 379.03 | `std::from_chars<u8>` | 154.85 | | ||
| | `std::sprintf<u8>` | 20.49 | `std::scanf<u8>` | 15.75 | | ||
| | `std::stringstream<u8>` | 3.82 | `std::stringstream<u8>` | 3.83 | | ||
| | **write `int8_t`** | **MB/s**| **read `int8_t`** | **MB/s**| | ||
| | `c4::to_chars<i8>` | 599.98 | `c4::from_chars<i8>` | 184.20 | | ||
| | `std::to_chars<i8>` | 246.32 | `std::from_chars<i8>` | 156.40 | | ||
| | `std::sprintf<i8>` | 19.15 | `std::scanf<i8>` | 16.44 | | ||
| | `std::stringstream<i8>` | 3.83 | `std::stringstream<i8>` | 3.89 | | ||
| | **write `uint16_t`** | **MB/s**| **read `uint16_t`** | **MB/s**| | ||
| | `c4::to_chars<u16>` | 486.40 | `c4::from_chars<u16>` | 349.48 | | ||
| | `std::to_chars<u16>` | 454.24 | `std::from_chars<u16>` | 319.13 | | ||
| | `std::sprintf<u16>` | 38.74 | `std::scanf<u16>` | 28.12 | | ||
| | `std::stringstream<u16>` | 7.08 | `std::stringstream<u16>`| 6.73 | | ||
| | **write `int16_t`** | **MB/s**| **read `int16_t`** | **MB/s**| | ||
| | `c4::to_chars<i16>` | 507.44 | `c4::from_chars<i16>` | 282.95 | | ||
| | `std::to_chars<i16>` | 297.49 | `std::from_chars<i16>` | 186.18 | | ||
| | `std::sprintf<i16>` | 39.03 | `std::scanf<i16>` | 28.45 | | ||
| | `std::stringstream<i16>` | 6.98 | `std::stringstream<i16>`| 6.49 | | ||
| | **write `uint32_t`** | **MB/s**| **read `uint32_t`** | **MB/s**| | ||
| | `c4::to_chars<u32>` | 730.12 | `c4::from_chars<u32>` | 463.95 | | ||
| | `std::to_chars<u32>` | 514.76 | `std::from_chars<u32>` | 329.42 | | ||
| | `std::sprintf<u32>` | 71.19 | `std::scanf<u32>` | 44.97 | | ||
| | `std::stringstream<u32>` | 14.05 | `std::stringstream<u32>`| 12.57 | | ||
| | **write `int32_t`** | **MB/s**| **read `int32_t`** | **MB/s**| | ||
| | `c4::to_chars<i32>` | 618.76 | `c4::from_chars<i32>` | 345.53 | | ||
| | `std::to_chars<i32>` | 394.72 | `std::from_chars<i32>` | 224.46 | | ||
| | `std::sprintf<i32>` | 71.14 | `std::scanf<i32>` | 43.49 | | ||
| | `std::stringstream<i32>` | 13.91 | `std::stringstream<i32>`| 12.03 | | ||
| | **write `uint64_t`** | **MB/s**| **read `uint64_t`** | **MB/s**| | ||
| | `c4::to_chars<u64>` | 1118.87 | `c4::from_chars<u64>` | 928.49 | | ||
| | `std::to_chars<u64>` | 886.58 | `std::from_chars<u64>` | 759.03 | | ||
| | `std::sprintf<u64>` | 140.96 | `std::scanf<u64>` | 91.60 | | ||
| | `std::stringstream<u64>` | 28.01 | `std::stringstream<u64>`| 25.00 | | ||
| | **write `int64_t`** | **MB/s**| **read `int64_t`** | **MB/s**| | ||
| | `c4::to_chars<i64>` | 1198.78 | `c4::from_chars<i64>` | 713.76 | | ||
| | `std::to_chars<i64>` | 882.17 | `std::from_chars<i64>` | 646.18 | | ||
| | `std::sprintf<i64>` | 138.79 | `std::scanf<i64>` | 90.07 | | ||
| | `std::stringstream<i64>` | 27.62 | `std::stringstream<i64>`| 25.12 | | ||
|
|
||
| If you feel suspicious about these bold claims, you can browse through [c4core's CI benchmark results](https://github.com/biojppm/c4core/actions/workflows/benchmarks.yml) which will hopefully give these more substance. | ||
|
|
||
|
|
||
| ### New features | ||
|
|
||
| - Added `bool c4::overflows<T>(csubstr s)` for detecting whether a string overflows a given integral type. See [PR#78](https://github.com/biojppm/c4core/pull/78). | ||
| - Also, added `c4::fmt::overflow_checked()` (and the corresponding `from_chars()` overload) to enable a check for overflow before parsing from string: | ||
| ```c++ | ||
| c4::from_chars(str, &val); // no overflow check | ||
| c4::from_chars(str, c4::fmt::overflow_checked(val)); // enable overflow check | ||
| // as an example, the implementation looks like: | ||
| template<class T> | ||
| bool c4::from_chars(c4::csubstr str, c4::fmt::overflow_checked<T> oc) | ||
| { | ||
| if(overflows<T>(str)) | ||
| return false; | ||
| return c4::from_chars(str, oc.val); | ||
| } | ||
| ``` | ||
| ### Fixes | ||
| - Fix missing endianess macro on windows arm/arm64 compilations [PR #76](https://github.com/biojppm/c4core/pull/76) | ||
| - Add missing `#define` for the include guard of the amalgamated header (see [rapidyaml#246](https://github.com/biojppm/rapidyaml/issues/246)). | ||
| - Fix CPU detection with ARMEL [PR #86](https://github.com/biojppm/c4core/pull/86). | ||
| - Fix GCC version detection [PR #87](https://github.com/biojppm/c4core/pull/87). | ||
| - Fix [cmake#8](https://github.com/biojppm/cmake/issues/8): `SOVERSION` missing from shared libraries. | ||
| - Update fastfloat to 3.5.1. | ||
| ### Thanks | ||
| - @fargies | ||
| - @daichifukui | ||
| - @janisozaur | ||
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters