devlinzhou/deterministic_float

fast soft float-point for deterministic computing，高性能、一致性计算的软件浮点数

deterministic_float

• Cross-platform deterministic float point, fast soft float point.

• Fast soft float point for deterministic computing, That you can make deterministic plugin by use GFloat to replace float-point. like：

  • Physics Engine (Bullet、PhysX)
  • AI Path Finding (Recast Navigation)
  • AI Engine

• Compare the overall performance of hard float IEEE-754 and GFloat numbers, They differ by no more than 500%, Refer to the multiply-add instructions for measuring the computing power of the graphics card. compare multiply and add function:

• 跨平台的确定性浮点数，高性能的软件浮点数

• 你可以使用GFloat替代各种中间件的浮点数，从而实现带确定性的计算，实现帧同步游戏逻辑，也可以实现加密货币计算逻辑，例如：

  • 物理引擎（Bullet、PhysX）
  • AI 寻路（Recast Navigation）
  • AI 引擎

• 整体上GFloat与IEEE754的硬件浮点数float，性能差距在500%以内。参考显卡衡量计算能力的乘加指令，主要比较乘法和加法函数的性能：

IEEE-754 float & GFloat

• IEEE-754 float

  sign	exponents（8 bits）					fractions（23 bits）
  31	30	29	...	24	23	22	21	...	1	0

  • $\mathbf{X}_{IEEE754} = (-1)^\mathbf{sign} \times (1.\mathbf{fraction}) \times 2 ^{\mathbf{exponent} - 127}$

• GFloat

  sign	fractions（23 bits）					exponents（8 bits）
  31	30	29	...	9	8	7	6	...	1	0

  • $\mathbf{X}_{GFloat} = (-1)^\mathbf{sign} \times (\mathbf{fraction}) \times 2 ^{\mathbf{exponent} - 127}$

Performance float（IEEE-754） vs GFloat

• Several important functions compare, for more information to view TestAndBenchMark
• Call every function 1000000 Times, get the float vs GFloat time

Platform	Win64	Win64	iOS	OSX	Android
CPU	Ryzen-5950X	i9-12900H	Apple A15	Apple M1Pro	HiSilicon-KIRIN-990
	float vs GFloat	float vs GFloat	float vs GFloat	float vs GFloat	float vs GFloat
Add	0.63 vs 2.63 ms	0.44 vs 1.48ms	0.43 vs 1.85 ms	0.34 vs 2.24 ms	0.89 vs 8.76 ms
Mul	0.70 vs 1.14 ms	0.52 vs 0.79 ms	0.42 vs 0.78 ms	0.34 vs 0.93 ms	0.62 vs 2.10 ms
Sin	10.20 vs 8.39 ms	8.6 vs 3.15 ms	3.6 vs 4.9 ms	2.29 vs 5.28 ms	11.58 vs 15.75 ms
Sqrt	1.08 vs 9.29 ms	0.7 vs 6.2 ms	0.83 vs 5.64 ms	0.64 vs 6.06 ms	0.65 vs 15.77 ms
Exp	2.06 vs 10.83ms	2.15 vs 24.52ms	5.52 vs 7.03 ms	5.75 vs 6.52 ms	6.06 vs 20.90 ms

How to start

• Four OS platform projects for fast and easy to start test&benchmark

Target Platform	Compiler	Project
Win	Visual Studio 2022	deterministic/deterministic_float.sln
MacOS	XCode 13.4	deterministic/deterministic_float.xcodeproj
iOS	XCode 13.4	deterministic/deterministic_float_ios.xcodeproj
Android	Android studio 4.0	deterministic/deterministic_float_android

Current Support Platform For Test&Benchmark

Operation System	Win64	MacOS	iOS	Android
	$\checkmark$	$\checkmark$	$\checkmark$	$\checkmark$

deterministic_float's Issues

Is it possible to improve performance by reducing precision?

Just a question: if we limit the precision to 2 or 3 decimals, could it be a trade-off to improve performance?

Add overflow assert and outlog

• overflow Test
• overflow OutPutlog
• Normalize Test
• Normalize OutPutLog

Sincos table use defined in the code， instead of compute when initalize

Optimize Power logarithm

• Exp
• Log
• Power

Optmize Sqrt and InvSqrt

• Sqrt
• InvSqrt

Optimize Trigonometric function

• Sin
• Cos
• Tan
• ASin
• ACos
• ATan

Use cmake generate projects