mirror of
https://github.com/SEPPDROID/Digital-Research-Source-Code.git
synced 2025-10-23 08:24:18 +00:00
174 lines
7.9 KiB
ArmAsm
174 lines
7.9 KiB
ArmAsm
ttl fast floating point log (ffplog)
|
||
***************************************
|
||
* (c) copyright 1981 by motorola inc. *
|
||
***************************************
|
||
|
||
*************************************************
|
||
* ffplog *
|
||
* fast floating point logorithm *
|
||
* *
|
||
* input: d7 - input argument *
|
||
* *
|
||
* output: d7 - logorithmic result to base e *
|
||
* *
|
||
* all other registers totally transparent *
|
||
* *
|
||
* code size: 184 bytes stack work: 38 bytes *
|
||
* *
|
||
* condition codes: *
|
||
* z - set if the result is zero *
|
||
* n - set if result in is negative *
|
||
* v - set if invalid negative argument *
|
||
* or zero argument *
|
||
* c - undefined *
|
||
* x - undefined *
|
||
* *
|
||
* *
|
||
* notes: *
|
||
* 1) spot checks show errors bounded by *
|
||
* 5 x 10**-8. *
|
||
* 2) negative arguments are illegal and cause*
|
||
* the "v" bit to be set and the absolute *
|
||
* value used instead. *
|
||
* 3) a zero argument returns the largest *
|
||
* negative value possible with the "v" bit*
|
||
* set. *
|
||
* *
|
||
* time: (8mhz no wait states assumed) *
|
||
* *
|
||
* times are very data sensitive with *
|
||
* samples ranging from 170 to 556 *
|
||
* microseconds *
|
||
* *
|
||
*************************************************
|
||
page
|
||
ffplog idnt 1,2 ffp log
|
||
|
||
opt pcs
|
||
section 9
|
||
|
||
xdef ffplog entry point
|
||
|
||
xref ffphthet hypertangent table
|
||
xref ffpadd,ffpdiv,ffpsub,ffpmul2 arithmetic primitives
|
||
xref ffptnorm transcendental normalize routine
|
||
xref ffpcpyrt copyright stub
|
||
|
||
fpone equ $80000041 floating value for one
|
||
log2 equ $b1721840 log(2) = .6931471805
|
||
|
||
**************
|
||
* log entry *
|
||
**************
|
||
|
||
* insure argument positive
|
||
ffplog tst.b d7 ? test sign
|
||
beq.s fplzro branch argument zero
|
||
bpl.s fplok branch alright
|
||
|
||
* argument is negative - use the absolute value and set the "v" bit
|
||
and.b #$7f,d7 take absolute value
|
||
bsr.s fplok find log(abs(x))
|
||
*psetv or.b #$02,ccr set overflow bit
|
||
fpsetv dc.l $003c0002 ***assembler error***
|
||
rts return to caller
|
||
|
||
* argument is zero - return largest negative number with "v" bit
|
||
fplzro move.l #-1,d7 return largest negative
|
||
jmp fpsetv return with "v" bit set
|
||
|
||
* save work registers and strip exponent off
|
||
fplok movem.l d1-d6/a0,-(sp) save all work registers
|
||
move.b d7,-(sp) save original exponent
|
||
move.b #64+1,d7 force between 1 and 2
|
||
move.l #fpone,d6 load up a one
|
||
move.l d7,d2 copy argument
|
||
jsr ffpadd create arg+1
|
||
exg.l d7,d2 swap result with argument
|
||
jsr ffpsub create arg-1
|
||
move.l d2,d6 prepare for divide
|
||
jsr ffpdiv result is (arg-1)/(arg+1)
|
||
beq.s fplnocr zero so cordic not needed
|
||
* convert to bin(31,29) precision
|
||
sub.b #64+3,d7 adjust exponent
|
||
neg.b d7 for shift necessary
|
||
cmp.b #31,d7 ? insure not too small
|
||
bls.s fplshf no, go shift
|
||
move.l #0,d7 force to zero
|
||
fplshf lsr.l d7,d7 shift to bin(31,29) precision
|
||
|
||
*****************************************
|
||
* cordic calculation registers: *
|
||
* d1 - loop count a0 - table pointer *
|
||
* d2 - shift count *
|
||
* d3 - y' d5 - y *
|
||
* d4 - x' d6 - z *
|
||
* d7 - x *
|
||
*****************************************
|
||
|
||
move.l #0,d6 z=0
|
||
move.l #1<<29,d5 y=1
|
||
lea ffphthet,a0 to inverse hyperbolic tangent table
|
||
move.l #22,d1 loop 23 times
|
||
move.l #1,d2 prime shift counter
|
||
bra.s cordic enter cordic loop
|
||
|
||
* cordic loop
|
||
fplpls asr.l d2,d4 shift(x')
|
||
sub.l d4,d5 y = y - x'
|
||
add.l (a0),d6 z = z + hypertan(i)
|
||
cordic move.l d7,d4 x' = x
|
||
move.l d5,d3 y' = y
|
||
asr.l d2,d3 shift(y')
|
||
fplnlp sub.l d3,d7 x = x - y'
|
||
bpl.s fplpls branch negative
|
||
move.l d4,d7 restore x
|
||
add.l #4,a0 to next table entry
|
||
add.b #1,d2 increment shift count
|
||
lsr.l #1,d3 shift(y')
|
||
dbra d1,fplnlp and loop until done
|
||
|
||
* now convert to float and add exponent*log(2) for final result
|
||
move.l #0,d7 default zero if too small
|
||
jsr ffptnorm float z
|
||
beq.s fplnocr branch if too small
|
||
add.b #1,d6 times two
|
||
move.l d6,d7 setup in d7 in case exp=0
|
||
fplnocr move.l d7,d2 save result
|
||
move.l #0,d6 prepare original exponent load
|
||
move.b (sp)+,d6 load it back
|
||
sub.b #64+1,d6 convert exponent to binary
|
||
beq.s fplzpr branch zero partial here
|
||
move.b d6,d1 save sign byte
|
||
bpl.s fplpos branch positive value
|
||
neg.b d6 force positive
|
||
fplpos ror.l #8,d6 prepare to convert to integer
|
||
move.l #$47,d5 setup exponent mask
|
||
fplnorm add.l d6,d6 shift to left
|
||
dbmi d5,fplnorm exp-1 and branch if not normalized
|
||
move.b d5,d6 fix in exponent
|
||
and.b #$80,d1 extract sign
|
||
or.b d1,d6 insert sign in
|
||
move.l #log2,d7 multiply exponent by log(2)
|
||
jsr ffpmul2 multiply d6 and d7
|
||
move.l d2,d6 now add cordic result
|
||
jsr ffpadd for final answer
|
||
|
||
fplzpr movem.l (sp)+,d1-d6/a0 restore registers
|
||
rts return to caller
|
||
|
||
|
||
end
|
||
|
||
rts return to caller
|
||
|
||
|
||
end
|
||
|
||
rts return to caller
|
||
|
||
|
||
end
|
||
|
||
rts return to caller
|
||
|
||
|
||
end
|
||
|
||
|