slightly faster handling of signed mul

previously we were flipping the inputs if negative, and then the
output if both inputs were negative

turns out you can just treat the whole thing as an unsigned mul
and then subtract each term from the high word if the other term
is negative.

https://stackoverflow.com/a/28827013

this saves a handful of cycles, reducing our runtime to a merge
14.211 ms/px \o/

mandel.s

@@ -26,7 +26,6 @@ z_buffer_start  = $b1 ; u8: index into z_buffer
 z_buffer_end    = $b2 ; u8: index into z_buffer
 temp            = $b4 ; u16
 temp2           = $b6 ; u16
-
 pixel_ptr       = $b8 ; u16
 pixel_color     = $ba ; u8
 pixel_mask      = $bb ; u8
@@ -345,68 +344,6 @@ fill_masks:
     neg 4, arg
 .endmacro
 
-; inner loop for imul16
-; bitnum < 8: 25 or 41 cycles
-; bitnum >= 8: 30 or 46 cycles
-.macro bitmul16 arg1, arg2, result, bitnum
-    .local zero
-    .local one
-    .local next
-
-    ; does 16-bit adds
-    ; arg1 and arg2 are treated as unsigned
-    ; negative signed inputs must be flipped first
-
-    ; 7 cycles up to the branch
-
-    ; check if arg1 has 0 or 1 bit in this place
-    ; 5 cycles either way
-    .if bitnum < 8
-        lda arg1                 ; 3 cyc
-        and #(1 << (bitnum))       ; 2 cyc
-    .else
-        lda arg1 + 1             ; 3 cyc
-        and #(1 << ((bitnum) - 8)) ; 2 cyc
-    .endif
-    bne one ; 2 cyc
-
-zero: ; 18 cyc, 23 cyc
-    lsr result + 3 ; 5 cyc
-    jmp next       ; 3 cyc
-
-one: ; 32 cyc, 37 cyc
-    ; 16-bit add on the top bits
-    clc            ; 2 cyc
-    lda result + 2 ; 3 cyc
-    adc arg2       ; 3 cyc
-    sta result + 2 ; 3 cyc
-    lda result + 3 ; 3 cyc
-    adc arg2 + 1   ; 3 cyc
-    ror a          ; 2 cyc - get a jump on the shift
-    sta result + 3 ; 3 cyc
-next:
-    ror result + 2 ; 5 cyc
-    ror result + 1 ; 5 cyc
-    .if bitnum >= 8
-        ; we can save 5 cycles * 8 bits = 40 cycles total by skipping this byte
-        ; when it's all uninitialized data
-        ror result ; 5 cyc
-    .endif
-
-.endmacro
-
-; 5 to 25 cycles
-.macro check_sign arg
-    ; Check sign bit and flip argument to postive,
-    ; keeping a count of sign bits in the Y register.
-    .local positive
-    lda arg + 1   ; 3 cyc
-    bpl positive  ; 2 cyc
-    neg16 arg     ; 18 cyc
-    iny           ; 2 cyc
-positive:
-.endmacro
-
 ; 518 - 828 cyc
 .macro imul16 dest, arg1, arg2
     copy16 FR0, arg1  ; 12 cyc
@@ -430,42 +367,6 @@ positive:
     copy16 dest, FR2 + 2  ; 12 cyc
 .endmacro
 
-; min 470 cycles
-; max 780 cycles
-.proc imul16_func_orig
-    arg1 = FR0   ; 16-bit arg (clobbered)
-    arg2 = FR1   ; 16-bit arg (clobbered)
-    result = FR2 ; 32-bit result
-
-    ldy #0          ; 2 cyc
-    ; counts the number of sign bits in Y
-    check_sign arg1 ; 5 to 25 cyc
-    check_sign arg2 ; 5 to 25 cyc
-    
-    ; zero out the 32-bit temp's top 16 bits
-    lda #0          ; 2 cyc
-    sta result + 2  ; 3 cyc
-    sta result + 3  ; 3 cyc
-    ; the bottom two bytes will get cleared by the shifts
-
-    ; unrolled loop for maximum speed, at the cost
-    ; of a larger routine
-    ; 440 to 696 cycles
-    .repeat 16, bitnum
-        ; bitnum < 8: 25 or 41 cycles
-        ; bitnum >= 8: 30 or 46 cycles
-        bitmul16 arg1, arg2, result, bitnum
-    .endrepeat
-
-    ; In case of mixed input signs, return a negative result.
-    cpy #1              ; 2 cyc
-    bne positive_result ; 2 cyc
-    neg32 result        ; 34 cyc
-positive_result:
-
-    rts ; 6 cyc
-.endproc
-
 ; Adapted from https://everything2.com/title/Fast+6502+multiplication
 .macro imul8 dest, arg1, arg2
     .local under256
@@ -512,6 +413,7 @@ positive_result:
     small_product:
         sec                   ; - x^2/2: 25 cycles
         sbc mul_lobyte256,x
+        sta mul_product_lo
         lda mul_product_hi
         sbc mul_hibyte256,x
         sta mul_product_hi
@@ -524,27 +426,19 @@ positive_result:
     result = FR2 ; 32-bit result
     inter = temp2
 
-    ldy #0          ; 2 cyc
-    ; counts the number of sign bits in Y
-    check_sign arg1 ; 5 to 25 cyc
-    check_sign arg2 ; 5 to 25 cyc
-
     ; h1l1 * h2l2
     ; (h1*256 + l1) * (h2*256 + l2)
     ; h1*256*(h2*256 + l2) + l1*(h2*256 + l2)
     ; h1*h2*256*256 + h1*l2*256 + h2*l1*256 + l1*l2
 
+    imul8 result, arg1, arg2
     lda #0
-    sta result + 0
-    sta result + 1
     sta result + 2
     sta result + 3
 
-    imul8 inter, arg1, arg2
-    add16 result, result, inter
-
     imul8 inter, arg1 + 1, arg2
     add16 result + 1, result + 1, inter
+    add_carry result + 3
 
     imul8 inter, arg1, arg2 + 1
     add16 result + 1, result + 1, inter
@@ -553,11 +447,16 @@ positive_result:
     imul8 inter, arg1 + 1, arg2 + 1
     add16 result + 2, result + 2, inter
 
-    ; In case of mixed input signs, return a negative result.
-    cpy #1              ; 2 cyc
-    bne positive_result ; 2 cyc
-    neg32 result        ; 34 cyc
-positive_result:
+    ; In case of negative inputs, adjust high word
+    ; https://stackoverflow.com/a/28827013
+    lda arg1 + 1
+    bpl arg1_pos
+    sub16 result + 2, result + 2, arg2
+arg1_pos:
+    lda arg2 + 1
+    bpl arg2_pos
+    sub16 result + 2, result + 2, arg1
+arg2_pos:
 
     rts ; 6 cyc
 .endproc

readme.md

@@ -37,6 +37,7 @@ Add a running counter of ms/px using the vertical blank interrupts as a timer. T
 Check for cycles in (zx,zy) output when in the 'lake'; if values repeat, they cannot escape. This is a big time saver in fractint.
 
 I may be able to do a faster multiply using tables of squares for 8-bit component multiplication.
+(done)
 
 ## Deps and build instructions
 

tables.js

@@ -11,6 +11,11 @@ function db(func) {
     return lines.join('\n');
 }
 
+let squares = [];
+for (let i = 0; i < 512; i++) {
+    squares.push(Math.trunc((i * i + 1) / 2));
+}
+
 console.log(
 `.segment "TABLES"
 
@@ -20,14 +25,14 @@ console.log(
 
 .align 256
 mul_lobyte256:
-${db((x) => Math.round(x * x / 2) & 0xff)}
+${db((i) => squares[i] & 0xff)}
 
 .align 256
 mul_hibyte256:
-${db((x) => (Math.round(x * x / 2) >> 8) & 0xff)}
+${db((i) => (squares[i] >> 8) & 0xff)}
 
 .align 256
 mul_hibyte512:
-${db((x) => (Math.round((x + 256) * (x + 256) / 2) >> 8) & 0xff)}
+${db((i) => (squares[i + 256] >> 8) & 0xff)}
 
 `);