[raspi-internals] Re: Review: 80 bit vector instructions - scalar register update & accumulator
- From: Herman Hermitage <hermanhermitage@xxxxxxxxxxx>
- To: "raspi-internals@xxxxxxxxxxxxx" <raspi-internals@xxxxxxxxxxxxx>
- Date: Sat, 13 Jul 2013 18:44:25 +1200
Corrections:
Accumulator
====
There is a signed 48 bit Accumulator associated with each lane. It saturates
to MAX 0x7fffffffffff and MIN 0x800000000000.
f_i:
0 1----- ENA Enable Normal Accumulation function.
0 -1---- HIGH Apply the accumulation to the top 32 bits of the
Accumulator.
0 --1--- SIGN Accumulate values as if they were signed quantities.
0 ---1-- CLRA Clear accumulator before initial instruction execution
(doesn't apply to repeated instruction phases).
0 ----1- WBA Write back new value to accumulator.
0 -----1 SUB Sub ALU result from Accumulator.
Clearing accmulator:
====
CLRA Clear accumulator on initially entering
instruction.
Accumulation/Decumulation
====
UACC (ENA|WBA) Accumulate with unsigned value.
UDEC (ENA|WBA|SUB) Decumulate with unsigned value.
SACC (ENA|SIGN|WBA) Accumulate with signed value.
SDEC (ENA|SIGN|WBA|SUB) Decumulate with unsigned value.
UACCH (ENA|HIGH|WBA) Accumulate with unsigned value to high
word of accumulator.
UDECH (ENA|HIGH|WBA|SUB) Decumulate with unsigned value to high
word of accumulator.
SACCH (ENA|HIGH|SIGN|WBA) Accumulate with signed value to high word
of accumulator.
SDECH (ENA|HIGH|SIGN|WBA|SUB) Decumulate with signed value to high word
of accumulator.
The parts of the names are:
U | S -> Unsigned, Signed value
ACC -> Accumulate: ie ACC[i] += unsigned(D[i]), or ACC[i] += signed(D[i]))
DEC -> Decumulate: ie ACC[i] -= unsigned(D[i]), or ACC[i] -= signed(D[i]))
H -> Use high 32 bits of accumulator.
Add/Sub Accumulator to result value (without accumulator update):
===
(The key difference being the ACC[i] values are _not_ updated by the operation).
UADDA (ENA) D'[i] = ACC[i]+unsigned(D[i]).
USUBA (ENA|SUB) D'[i] = ACC[i]-unsigned(D[i]).
SADDA (ENA|SIGN) D'[i] = ACC[i]+signed(D[i]).
SSUBA (ENA|SIGN|SUB) D'[i] = ACC[i]-signed(D[i]).
UADDAH (ENA|HIGH) D'[i] = ACCH[i]+unsigned(D[i]).
USUBAH (ENA|HIGH|SUB) D'[i] = ACCH[i]-unsigned(D[i]).
SADDAH (ENA|HIGH|SIGN) D'[i] = ACCH[i]+signed(D[i]).
SSUBAH (ENA|HIGH|SIGN|SUB) D'[i] = ACCH[i]-signed(D[i]).
These names are speculative (I dont have a solid reference point).
The parts of the names are:
U, S -> Unsigned or signed
ADD -> Add accumulator to result
SUB -> Sub result from accumulator
H -> Use high 32 bits of accumulator.
I'll try and post the equivalent C code to make the computations clearer.
We still have an unknown when ENA bit is clear. From inspection in the blob it
may relate to the SETF/IFxx flags (Phire's idea).
I'm still poking around.
Cheers
Herman
Other related posts:
