Microchip Technology MA330020 Data Sheet

 2008-2014 Microchip Technology Inc.

DS70000318G-page 41

The SA and SB bits are modified each time data
passes through the adder/subtracter, but can only be
cleared by the user application. When set, they indicate
that the accumulator has overflowed its maximum
range (bit 31 for 32-bit saturation or bit 39 for 40-bit
saturation) and will be saturated (if saturation is
enabled). When saturation is not enabled, SA and SB
default to bit 39 overflow and thus, indicate that a cata-
strophic overflow has occurred. If the COVTE bit in the
INTCON1 register is set, SA and SB bits will generate
an arithmetic warning trap when saturation is disabled.

The Overflow and Saturation Status bits can optionally
be viewed in the STATUS Register (SR) as the logical
OR of OA and OB (in bit OAB) and the logical OR of SA
and SB (in bit SAB). Programmers can check one bit
in the STATUS Register to determine if either
accumulator has overflowed, or one bit to determine if
either accumulator has saturated. This is useful for
complex number arithmetic, which typically uses both
accumulators.

The device supports three Saturation and Overflow
modes:

• Bit 39 Overflow and Saturation:

When bit 39 overflow and saturation occurs, the 
saturation logic loads the maximally positive 
9.31 (0x7FFFFFFFFF) or maximally negative 
9.31 value (0x8000000000) into the target accumu-
lator. The SA or SB bit is set and remains set until 
cleared by the user application. This condition is 
referred to as ‘super saturation’ and provides 
protection against erroneous data or unexpected 
algorithm problems (such as gain calculations).

• Bit 31 Overflow and Saturation:

When bit 31 overflow and saturation occurs, the 
saturation logic then loads the maximally positive 
1.31 value (0x007FFFFFFF) or maximally nega-
tive 1.31 value (0x0080000000) into the target 
accumulator. The SA or SB bit is set and remains 
set until cleared by the user application. When 
this Saturation mode is in effect, the guard bits are 
not used, so the OA, OB or OAB bits are never 
set.

• Bit 39 Catastrophic Overflow:

The bit 39 Overflow Status bit from the adder is 
used to set the SA or SB bit, which remains set 
until cleared by the user application. No saturation 
operation is performed, and the accumulator is 
allowed to overflow, destroying its sign. If the 
COVTE bit in the INTCON1 register is set, a 
catastrophic overflow can initiate a trap exception.

3.6.3

ACCUMULATOR ‘WRITE BACK’

The  MAC class of instructions (with the exception of
MPY

,  MPY.N,  ED and EDAC) can optionally write a

rounded version of the high word (bits 31 through 16)
of the accumulator that is not targeted by the instruction
into data space memory. The write is performed across
the X bus into combined X and Y address space. The
following addressing modes are supported:

• W13, Register Direct:

The rounded contents of the non-target accumulator 
are written into W13 as a 1.15 fraction.

• [W13] + = 2, Register Indirect with Post-Increment:

The rounded contents of the non-target 
accumulator are written into the address pointed 
to by W13 as a 1.15 fraction. W13 is then 
incremented by 2 (for a word write).

3.6.3.1

Round Logic

The round logic is a combinational block that performs
a conventional (biased) or convergent (unbiased)
round function during an accumulator write (store). The
Round mode is determined by the state of the RND bit
in the CORCON register. It generates a 16-bit,
1.15 data value that is passed to the data space write
saturation logic. If rounding is not indicated by the
instruction, a truncated 1.15 data value is stored and
the least significant word is simply discarded.

Conventional rounding zero-extends bit 15 of the accu-
mulator and adds it to the ACCxH word (bits 16 through
31 of the accumulator).

• If the ACCxL word (bits 0 through 15 of the 

accumulator) is between 0x8000 and 0xFFFF 
(0x8000 included), ACCxH is incremented.

• If ACCxL is between 0x0000 and 0x7FFF, ACCxH 

is left unchanged.

A consequence of this algorithm is that over a
succession of random rounding operations, the value
tends to be biased slightly positive. 

Convergent (or unbiased) rounding operates in the same
manner as conventional rounding, except when ACCxL
equals 0x8000. In this case, the Least Significant bit
(bit 16 of the accumulator) of ACCxH is examined:

• If it is ‘1’, ACCxH is incremented.
• If it is ‘0’, ACCxH is not modified.

Assuming that bit 16 is effectively random in nature,
this scheme removes any rounding bias that may
accumulate.