Lines Matching refs:nearest

182 fraction, rounding to nearest.
214 ### Round-to-nearest rounding: unbiased in principle but not in practice
216 The simplest fix is to avoid the bias in 2. by rounding-to-nearest instead of
226 to achieve rounding-to-nearest (while the above rounding towards 0 corresponded
229 In principle, rounding-to-nearest is unbiased and optimal in various ways.
240 different from 0. That rounding-to-nearest is "unbiased" only means that over a
251 instance if we have to round the value 3.8 to the nearest integer, we can round
265 rounding-to-nearest --- one can check that it is 2x higher.
271 Still, for large enough GEMM depth, that is better than rounding-to-nearest
274 ### Switching between rounding-to-nearest and probabilistic rounding
278 rounding-to-nearest will be more accurate for smaller GEMM depths.