Per section F.1.4 of the textbook, it would be good to understand the root cause of this discrepancy:
After this section was written, TestFloat revealed an FMA bug (cvw issue #578) related to subtraction of a small value. Z_m^shifted had to be widened to Nf + 4 integer bits and 3Nf + 2 fractional bits to obtain a correct guard bit. The extra integer and fractional bit propagate through the rest of the FMA and postprocessing. The root cause is not yet fully explained.
Per section F.1.4 of the textbook, it would be good to understand the root cause of this discrepancy:
After this section was written, TestFloat revealed an FMA bug (cvw issue #578) related to subtraction of a small value. Z_m^shifted had to be widened to Nf + 4 integer bits and 3Nf + 2 fractional bits to obtain a correct guard bit. The extra integer and fractional bit propagate through the rest of the FMA and postprocessing. The root cause is not yet fully explained.