Login
You're viewing the mstdn.social public feed.
  • Jul 4, 2026, 8:26 PM

    I'm doing this thing ATM and was reminded of just how cursed fast-math is. SRLY: How can enough people think that changing the control word to flush denormals to zero isn't an idea that needs to be burned to the ground. Ashes scattered. Build a monument of warning to those that come up with such horrendous ideas in the future.

    💬 1🔄 2⭐ 0

Replies

  • Jul 4, 2026, 8:39 PM

    @mbr Maybe it is stockholm syndrome or ignorance on my part but isn't dividing by a denormal likely to introduce an inf or quantized firefly unless your expressions ranges are carefully clamped?

    💬 1🔄 0⭐ 0
  • Jul 4, 2026, 8:53 PM

    @vethanis The evil part is changing the control word since it effects all computations. Wanting to flush denormals is fine, but you change, do the operations and restore unless you're the only thing on that thread.

    💬 2🔄 0⭐ 0
  • 💬 0🔄 0⭐ 0
  • Jul 5, 2026, 8:15 AM

    @mbr Ah, yeah multiple libraries of code with different decisions leaking between one another. Not fun.

    💬 0🔄 0⭐ 0