From 2266c7aac5e6e3a3fb1d363526b9a4cd38b64bd3 Mon Sep 17 00:00:00 2001 From: Eric Kunze Date: Fri, 27 Oct 2023 14:38:56 -0700 Subject: Accuracy requirements for RESIZE operator Initial accuracy requirements for RESIZE calculations. Signed-off-by: Eric Kunze Change-Id: I44f997df857915ca77b04dd0b87dee9ccac6e7b0 --- chapters/introduction.adoc | 8 ++++++++ 1 file changed, 8 insertions(+) diff --git a/chapters/introduction.adoc b/chapters/introduction.adoc index a580472..0a94ef3 100644 --- a/chapters/introduction.adoc +++ b/chapters/introduction.adoc @@ -321,6 +321,14 @@ Let `out_ref` be the result of the fp64_t reference implementation of `sin(x)`. Let `err_bnd = abs(out_ref) * exp2(-normal_frac/2)` + Then `tosa_reference_check_fp_bnd(out_imp, out_ref, err_bnd)` must be true +| <> +| The result corresponds to a sequence of floating-point calculations. + +The allowable error bound for the result of a resize is based on the maximum value of an element in the input tensor. + +Let `out_imp` be the implementation output. + +Let `out_ref` be the result of the fp64_t reference implementation. + +Let `err_bnd = max(abs(input)) * 0.006`. + +Then `tosa_reference_check_fp_bnd(out_imp, out_ref, err_bnd)` must be true. + |=== ===== Operator sequence precision requirement -- cgit v1.2.1