numba-dpex icon indicating copy to clipboard operation
numba-dpex copied to clipboard

Published 20 hours ago • verified publisher icon IntelPython

Lower than expected performance in blackscholes numpy implementation

Open adarshyoga opened this issue 1 year ago • 1 comments

The blackscholes numpy implementation in dpbench is ~26X slower than the corresponding kernel and prange implementations.

How to reproduce:

  1. Follow instructions to setup dpbench.
  2. Run blackscholes - python -c "import dpbench; dpbench.run_benchmark(\"black_scholes\")"

adarshyoga avatar Mar 20 '23 20:03 adarshyoga

The slowdown maybe related to kernel launch overhead in the JitKernel custom dispatcher class. Overhead is especially noticeable with small problem sizes. The experimental.dispatcher.KernelDispatcher fixes the launch overhead.

Can you please reevaluate with the new dispatcher?

diptorupd avatar Dec 20 '23 02:12 diptorupd