I encountered a mysterious case where the same floating-point computation experiences a massive slowdown depending on the data — a factor of 20, or even 60 if I use @turbo from LoopVectorization.jl. I’m guessing it has to do with floating-point exceptions arising from underflow, since the slowdown …

Massive data-dependent floating-point slowdown

PeterSimon May 28, 2021, 9:36pm 4

set_zero_subnormals(true)

prevents the slowdown. See, e.g. this performance tip

4 Likes

Topic		Replies	Views
Slowdown due to subnormal float, coming from neural net training Performance	20	931	October 27, 2022
50x speed difference in gemv for different values in vector General Usage	8	5005	August 16, 2017
Float32 performance fluctuating over time Performance	1	306	October 22, 2021
Why are calls to this function slow, compared to Python+Numpy? General Usage fast-math	14	2171	August 13, 2017
Will decreasing the precision of intermediate variables improve performance of code? Performance performance	12	1233	May 9, 2020