Portable SIMD-based C++ option pricing library (Black-Scholes).
I was curious to see how much performance there is to be gained by using a relatively straightforward SIMD implementation compared to the "naive" scalar implementation. I wanted a portable solution as I'd never tried writing SIMD code on an Apple chip (M2), hence the choice of highway (https://github.com/google/highway) library. It currently lacks some of the necessary maths operations though, most notably an std::erf equivalent for standard normal CDF.
I made some basic optimizations to the data layout and the order of arithmetic operations, but much more work needs to be done for the SIMD implementation.
For now, calculates option prices (call/put), as well as the main greeks (Delta, Gamma, Vega, Theta, Rho).
Written in C++20, compiled with Apple clang 14.0.3. Dependencies (through vcpkg):
- highway (1.0.7)
Additional dependencies for tests:
- gtest (1.14.0)
- benchmark (1.8.3)
10,000,000 randomized call and put options for both floats and doubles.
Can use at most Vec128, giving us 2 doubles or 4 floats per vector.
----------------------------------------------------------------
Benchmark Time CPU Iterations
----------------------------------------------------------------
BM_FastPrice<double> 886793667 ns 886623000 ns 1
BM_NaivePrice<double> 1127674666 ns 1127539000 ns 1
BM_FastPrice<float> 673567875 ns 673450000 ns 1
BM_NaivePrice<float> 1162894792 ns 1162567000 ns 1Should be possible to use at most Vec256, giving us 4 doubles or 8 floats per vector.
TODO