When using float and double floating-point types for standard math library operations, there are detailed ULP error standards documented (as referenced in the CUDA C Programming Guide). [image] [image] However, when I use the half floating-point type for standard math library operations, I lose …

Does the ulp error standard for half precision floating point mathematical functions seem to be missing?

Robert_Crovella December 6, 2024, 4:23am 2

Topic		Replies	Views
Precision issue about Single-Precision Floating-Point Function's table in programming guide CUDA Programming and Performance	2	526	July 14, 2023
round-to-double on GPUs? precision issue on the GPU CUDA Programming and Performance	10	10413	November 17, 2009
Accuracy: CUDA 4.0 math vs C math CUDA math functions have lower accuracy than C math ones CUDA Programming and Performance	2	5414	July 18, 2011
create a fp16(half) value directly CUDA Programming and Performance	2	1462	April 14, 2016
When using the __hfma function, the Device and Host results differ by 1 ULP, which should be 0ulp in theory CUDA Programming and Performance cuda , kernel	8	127	December 12, 2024
precision CUDA Programming and Performance	3	2624	December 16, 2008
FP16 constants in math_constants.h? CUDA Programming and Performance	3	1316	December 19, 2013
Does CUDA support double precision? CUDA Programming and Performance	1	1406	February 5, 2009
Half precision (fp16) min and max limits in CUDA 9 CUDA Programming and Performance	1	1973	November 14, 2017
IEEE-754 compliant division CUDA Programming and Performance	5	10125	November 26, 2008