11.001001000011111101101010100010001000 Arithmazium
Home

Underflow tests with scaling

In order to compare the tiny values from two sources, Paranoia scales the values up by a large power of the radix. This defends against faulty comparisons, again on the CDD 6000 class. In this case, the values are not so small that they behave like zero in multiplication, but they do all compare equal to each other. 

milestone = 120  # ==============================
# less_tiny_B and less_tiny_x result from similar loops, one starting
# with C, the other with (1 + SAFE_ULPS_OF_ONE) * C.
# Check to see who is the tiniest of them all, but scale up
# to avoid faulty comparisons as on CDC 6000.
# If the number encoding is such that what "looks like" the smallest power
# of the radix is interpreted as zero, then the smallest nonzero values
# will not be power of B. This test catches that case. The value h_fact
# records that the smallest power of B is too big by nearly a factor of H.
if ONE_OVER_C * less_tiny_B > ONE_OVER_C * less_tiny_x:
    h_fact = H
    min_positive = tiny_x
else:
    h_fact = ONE
    min_positive = tiny_B
Home