RE: Regarding Items 15 and 17 on the agenda today (decimal floating-point)

["Tsen Jr, S Charles" <s.charles.tsen.jr@xxxxxxxxx>]

Michel,

> More information is required for Addition and Comparison...

Thanks for raising these points.  It is important to know the mix of
data, when looking at these results.  I can give you the breakdown of
the data set we used for 64-bit and 128-bit comparisons.  Marius will
soon reply with data for Addition.

> So I would like to know what the proportion of special cases was
> (zero, Inf, NaNs), what the range of significance was, and how the
> exponents were distributed...

In the 64-bit case, we have 120 test cases.

- Of the 120 cases, 14 cases had a flavor NaN or INFINITY in one of the
operands.  It is very easy to test these two special cases by examining
the G-field.  These test cases take minimal clock cycles.

- Of the 120 cases, 4 cases had zero (and not NaN or INFINITY) in one of
the operands. Bear in mind that the value zero includes all values of
non-canonicals, so it is not a straightforward as NaN or INFINITY.
Hence, these 4 test cases will take more clock cycles.

In the 128-bit case, we had 262 test cases.

- Of the 262 cases, 25 had a flavor of NaN or INFINITY in one of the
operands.

- Of the 262 test cases, 6 had zero (and not NaN or INFINITY) in one of
the operands.

In our numerical testcases, we tried to use values to stress boundary
conditions and represent all ranges of exponents and significants.  We
tried to make it fairly uniform.  Let me know if there is more
information you are curious about. 

Thanks,
Chuck Tsen
University of Wisconsin
MESA Lab, Professor Schulte
Intel Intern