brooke/mandel-6502
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity

annotate cycle counts

Browse source
This commit is contained in:
Brooke Vibber 2022-12-30 17:33:18 -08:00
parent 6583dc3680
commit 4b1001bfdc
1 changed files with 38 additions and 34 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

72
mandel.s
View file

@ -57,19 +57,22 @@ plus:
neg 4, arg neg 4, arg
.endmacro .endmacro
; 2 + 9 * bytes cycles
.macro add bytes, arg1, arg2 .macro add bytes, arg1, arg2
clc clc ; 2 cyc
.repeat bytes, byte .repeat bytes, byte
lda arg1+byte lda arg1+byte ; 3 cyc
adc arg2+byte adc arg2+byte ; 3 cyc
sta arg1+byte sta arg1+byte ; 3 cyc
.endrepeat .endrepeat
.endmacro .endmacro
; 20 cycles
.macro add16 arg1, arg2 .macro add16 arg1, arg2
add 2, arg1, arg2 add 2, arg1, arg2
.endmacro .endmacro
; 38 cycles
.macro add32 arg1, arg2 .macro add32 arg1, arg2
add 4, arg1, arg2 add 4, arg1, arg2
.endmacro .endmacro
@ -112,6 +115,7 @@ plus:
shr 4, arg shr 4, arg
.endmacro .endmacro
; 24 to 49 cycles
.macro bitmul16 arg1, arg2, result, bitnum .macro bitmul16 arg1, arg2, result, bitnum
.local next .local next
@ -119,79 +123,79 @@ plus:
; arg1 must be 0 or positive ; arg1 must be 0 or positive
; arg2 must be 0 or positive ; arg2 must be 0 or positive
clc clc ; 2 cyc
; check if arg1 has 0 or 1 bit in this place ; check if arg1 has 0 or 1 bit in this place
; 5 cycles either way
.if bitnum < 8 .if bitnum < 8
lda arg1 lda arg1 ; 3 cyc
and #(1 << bitnum) and #(1 << bitnum) ; 2 cyc
.else .else
lda arg1 + 1 lda arg1 + 1 ; 3 cyc
and #(1 << (bitnum - 8)) and #(1 << (bitnum - 8)) ; 2 cyc
.endif .endif
beq next beq next ; 2 cyc
; 16-bit add on the top bits ; 16-bit add on the top bits
lda result + 2 add16 result + 2, arg2 ; 20 cyc
adc arg2
sta result + 2
lda result + 3
adc arg2 + 1
sta result + 3
next: next:
; Shift the 32-bit result down by one bit, ; Shift the 32-bit result down by one bit,
; saving the previous carry. ; saving the previous carry.
ror result + 3 ror result + 3 ; 5 cyc
ror result + 2 ror result + 2 ; 5 cyc
ror result + 1 ror result + 1 ; 5 cyc
.if bitnum >= 8 .if bitnum >= 8
; we can save 5 cycles * 8 bits = 40 cycles total by skipping this byte ; we can save 5 cycles * 8 bits = 40 cycles total by skipping this byte
; when it's all uninitialized data ; when it's all uninitialized data
ror result ror result ; 5 cyc
.endif .endif
.endmacro .endmacro
; 5 to 25 cycles
.macro check_sign arg .macro check_sign arg
; Check sign bit and flip argument to postive, ; Check sign bit and flip argument to postive,
; keeping a count of sign bits in the X register. ; keeping a count of sign bits in the X register.
.local positive .local positive
lda arg + 1 lda arg + 1 ; 3 cyc
bpl positive bpl positive ; 2 cyc
neg16 arg neg16 arg ; 18 cyc
inx inx ; 2 cyc
positive: positive:
.endmacro .endmacro
; 579 to 725 cycles
.proc imul16 .proc imul16
arg1 = FR0 ; 16-bit arg (clobbered) arg1 = FR0 ; 16-bit arg (clobbered)
arg2 = FR1 ; 16-bit arg (clobbered) arg2 = FR1 ; 16-bit arg (clobbered)
result = FR2 ; 32-bit result result = FR2 ; 32-bit result
ldx #0 ldx #0 ; 2 cyc
; counts the number of sign bits in X ; counts the number of sign bits in X
check_sign arg1 check_sign arg1 ; 5 to 25 cyc
check_sign arg2 check_sign arg2 ; 5 to 25 cyc
; zero out the 32-bit temp's top 16 bits ; zero out the 32-bit temp's top 16 bits
lda #0 lda #0 ; 2 cyc
sta result + 2 sta result + 2 ; 3 cyc
sta result + 3 sta result + 3 ; 3 cyc
; the bottom two bytes will get cleared by the shifts ; the bottom two bytes will get cleared by the shifts
; unrolled loop for maximum speed, at the cost ; unrolled loop for maximum speed, at the cost
; of a larger routine ; of a larger routine
.repeat 16, bitnum .repeat 16, bitnum
; first half: 24 to 45 cycles
; second half: 29 to 49 cycles
bitmul16 arg1, arg2, result, bitnum bitmul16 arg1, arg2, result, bitnum
.endrepeat .endrepeat
; In case of mixed input signs, return a negative result. ; In case of mixed input signs, return a negative result.
cpx #1 cpx #1 ; 2 cyc
bne positive_result bne positive_result ; 2 cyc
neg32 result neg32 result ; 34 cyc
positive_result: positive_result:
rts rts ; 6 cyc
.endproc .endproc
.proc iter .proc iter