realtimechris

No, you only require division for digit-counts greater than or equal to 10: https://github.com/RealTimeChris/Jsonifier/blob/dev/Include/...

So we can properly dispatch the correct length-based function with minimal branching to detect which length to serialize for: https://github.com/RealTimeChris/Jsonifier/blob/dev/Include/...

*Optimizing uint64_t Digit Counting: A New Method that Beats Lemire's by Up to 27%*

In the quest to improve the performance of my high-speed JSON library, JSONIFIER, I recently stumbled upon a breakthrough in optimizing the calculation of digit counts for `uint64_t` values. While Lemire’s method of using the `lzcnt` instruction for determining the number of digits in a 32-bit unsigned integer has been widely regarded as efficient, I’ve developed a new method that achieves even faster results for 64-bit unsigned integers (i.e., `uint64_t`), with significant gains across different compilers and platforms.

### The Existing Method: Lemire’s Approach

Lemire’s method, known for its efficiency, calculates the number of digits in a `uint32_t` by leveraging the `lzcnt` instruction, which finds the index of the most significant bit set to 1. This is combined with a static lookup table to map the result to the corresponding number of digits.

Here’s the code for Lemire’s method:

```cpp JSONIFIER_INLINE int int_log2(uint32_t x) { return 31 - simd_internal::lzcnt(x | 1); }

JSONIFIER_INLINE int fast_digit_count(uint32_t x) { static uint64_t table[] = { ... }; return (x + table[int_log2(x)]) >> 32; } ```

While this approach works well for 32-bit integers, the need for a faster and more efficient solution for `uint64_t` led me to create an alternative method, which uses a more streamlined approach without the overhead of a lookup table.

### My New Method: RTC-64-Bit Digit Counting

I’ve designed a new approach for 64-bit unsigned integers that leverages a similar logic but optimizes the process by storing precomputed digit counts for specific ranges and applying simple threshold checks. The result is faster execution with reduced computational overhead.

Here's the code for the new method:

```cpp JSONIFIER_INLINE_VARIABLE uint8_t digitCounts[]{ ... };

JSONIFIER_INLINE_VARIABLE uint64_t digitCountThresholds[]{ ... };

JSONIFIER_INLINE uint64_t fastDigitCount(const uint64_t inputValue) { const uint64_t originalDigitCount{ digitCounts[simd_internal::lzcnt(inputValue)] }; return originalDigitCount + static_cast<uint64_t>(inputValue > digitCountThresholds[originalDigitCount]); } ```

This method works by using a static array to hold the precomputed digit counts and another array for threshold values that determine the exact number of digits in a `uint64_t`. The key optimization lies in the efficient use of a bit manipulation technique and direct threshold checking to avoid unnecessary computations.

### [Benchmark Results](https://github.com/RealTimeChris/BenchmarkSuite/blob/digit-c...)

I ran performance benchmarks comparing my new RTC-64-bit method with Lemire’s approach and the traditional `log10` method across various platforms and compilers. The results were consistently impressive:

#### GCC/Ubuntu: - *RTC-64-bit* outperforms *Lemire-32-bit* by *27.33%*. - *Lemire-32-bit* beats *Log10-32-bit* by a massive *814.16%*.

#### Clang/Ubuntu: - *RTC-64-bit* outperforms *Lemire-32-bit* by *143.34%*. - *Lemire-32-bit* beats *Log10-32-bit* by *522.01%*.

#### MSVC/Windows: - *RTC-64-bit* is *12.50%* faster than *Lemire-32-bit*. - *Lemire-32-bit* beats *Log10-32-bit* by *515.90%*.

#### Clang/MacOS: - *RTC-64-bit* is *25.37%* faster than *Lemire-32-bit*. - *Lemire-32-bit* beats *Log10-32-bit* by *343.97%*.

### Key Takeaways

The RTC-64-bit method not only delivers improved performance on modern hardware but also significantly reduces the overhead of traditional methods by eliminating the need for a large lookup table. This is especially beneficial for high-performance applications where every cycle counts, such as in JSON serialization and parsing, where speed is critical.

It has reflection for the member names as well as compile-time hash maps for storing the locations to be parsed into/serialized from.

Here is a json parser/serializer that I've been working on for about a year or so now. There is benchmarks available at https://github.com/RealTimeChris/Json-Performance comparing it to the other 2 most performant C++ parsers/serializers. Let me know what you think! Cheers!

As you can see - the parser/serializer I've developed uses the minimal possible memory allocations. Nice parser though.

Hey everyone - so I've taken simdjson's algorithm and improved upon it by resaturating the CPU registers with data instead of only operating on 64-bytes worth of string data at a time, as well as incorporated some compile-time hashmaps for the keys of data/memory locations being parsed. Let me know what you think! Also here's some benchmarks: https://github.com/RealTimeChris/Json-Performance