pico-8 cartridge // http://www.pico-8.com
version 16
__lua__
-- amico8 benchmark v1.1
-- stat(1)  = cpu fraction 0..1 (1 = full frame at 60fps), LIVE within a frame
-- timing uses stat(1) deltas around each test loop.
-- (note: stat(100) is the load() breadcrumb on amico8/pico-8, NOT a timer -
--  v1 used it as a cia timer, which broke once stat(100) became the breadcrumb.)
-- press x/z to pause/resume, up/dn to scroll

local tests = {}
local results = {}
local cur = 0
local phase = 0  -- 0=run, 1=done
local scroll = 0

-- test data
local ta = {}
local x1,x2,x3 = 1.5, 2.3, -0.7
local ents = {}
local ecls = {}

-- how many iterations each test runs
-- chosen so slow machines take ~20-80% of frame per test
local n = {
  500,   -- 1  empty loop (baseline)
  500,   -- 2  flr
  500,   -- 3  abs
  500,   -- 4  min
  500,   -- 5  max
  500,   -- 6  mid
  500,   -- 7  sin
  500,   -- 8  cos
  500,   -- 9  sqrt
  500,   -- 10 x+y
  500,   -- 11 x*y
  500,   -- 12 x/y
  50,    -- 13 all(t) 32 items
  50,    -- 14 pairs(t) 32 items
  100,   -- 15 spr(0,x,y)
  5,     -- 16 rectfill(0,0,127,127)
  20,    -- 17 oop e:update() x20 ents
  1,     -- 18 map(0,0,0,0,16,16)
  500,   -- 19 lerp(a,b,t) call
  500,   -- 20 clerp(a,b,t) call (with clamp)
}

local tname = {
  "empty x500",
  "flr   x500",
  "abs   x500",
  "min   x500",
  "max   x500",
  "mid   x500",
  "sin   x500",
  "cos   x500",
  "sqrt  x500",
  "x+y   x500",
  "x*y   x500",
  "x/y   x500",
  "all(t)x50 ",
  "pairs x50 ",
  "spr   x100",
  "rfill x5  ",
  "oop   x20 ",
  "map16x16  ",
  "lerp  x500",
  "clerp x500",
}

-- small function that AOT can compile (only ADD/SUB/MUL/RETURN)
local function lerp(a,b,t) return a+(b-a)*t end
-- slightly bigger: clamp + lerp
local function clerp(a,b,t)
  if t<0 then t=0 end
  if t>1 then t=1 end
  return a+(b-a)*t
end

local ntests = #tname

function _init()
  -- fill test array
  for i=1,32 do add(ta, i*0.1+0.05) end

  -- oop class + instances
  ecls = {
    update=function(self)
      self.x = self.x + self.vx
      if self.x > 120 then self.x = 0 end
      self.t = self.t + 1
    end
  }
  ecls.__index = ecls
  for i=1,20 do
    add(ents, setmetatable(
      {x=i*6,y=60,vx=0.5,t=0}, ecls))
  end

  -- draw a minimal sprite in slot 0 (solid 2x2 block)
  -- uses memset to poke sprite data into ram directly
  -- sprite 0 = address 0x0000, 8x8 nibble-packed = 32 bytes
  for i=0,31 do poke(i, 0x55) end -- color 5, all pixels

  cur = 1
  phase = 0
end

-- run one test, return us/iter + cpu fraction
function run_test(ti)
  local t0 = stat(1)   -- cpu fraction used so far this frame (live, counts UP)
  local ni = n[ti]
  local r = 0

  if ti==1 then
    for i=1,ni do r=r+1 end
  elseif ti==2 then
    for i=1,ni do r=flr(x1) end
  elseif ti==3 then
    for i=1,ni do r=abs(x1) end
  elseif ti==4 then
    for i=1,ni do r=min(x1,x2) end
  elseif ti==5 then
    for i=1,ni do r=max(x1,x2) end
  elseif ti==6 then
    for i=1,ni do r=mid(x1,x2,x3) end
  elseif ti==7 then
    for i=1,ni do r=sin(x1) end
  elseif ti==8 then
    for i=1,ni do r=cos(x1) end
  elseif ti==9 then
    for i=1,ni do r=sqrt(x1) end
  elseif ti==10 then
    for i=1,ni do r=x1+x2 end
  elseif ti==11 then
    for i=1,ni do r=x1*x2 end
  elseif ti==12 then
    for i=1,ni do r=x1/x2 end
  elseif ti==13 then
    for i=1,ni do for v in all(ta) do r=v end end
  elseif ti==14 then
    for i=1,ni do for k,v in pairs(ta) do r=v end end
  elseif ti==15 then
    for i=1,ni do spr(0, (i%16)*8, flr(i/16)*8) end
  elseif ti==16 then
    for i=1,ni do rectfill(0,0,127,127, i%15+1) end
  elseif ti==17 then
    for i=1,ni do for e in all(ents) do e:update() end end
  elseif ti==18 then
    for i=1,ni do map(0,0,0,0,16,16) end
  elseif ti==19 then
    for i=1,ni do r=lerp(x1,x2,0.5) end
  elseif ti==20 then
    for i=1,ni do r=clerp(x1,x2,0.5) end
  end

  -- cpu fraction consumed by the test loop (stat(1) counts UP, 1.0 = full 60fps frame)
  local t1 = stat(1)
  local dt = t1 - t0
  if dt < 0   then dt = 0   end   -- guard (shouldn't happen inside one _update)
  if dt > 1.9 then dt = 1.9 end   -- clamp so dt*16667 stays under fix32 max (32767)
  -- us/iter = frac-of-frame * frame-budget(16667us at 60fps) / iterations
  local us = dt * 16667 / ni
  return us, dt  -- us/iter, cpu fraction (of this test loop)
end

local paused = false

function _update60()
  if btn(4) then paused = not paused end
  if btn(2) then scroll = max(0, scroll-1) end
  if btn(3) then scroll = min(ntests-8, scroll+1) end

  if paused or phase==1 then return end

  if cur <= ntests then
    local us, cpu = run_test(cur)
    results[cur] = {us=us, cpu=cpu, name=tname[cur]}
    cur = cur + 1
  else
    phase = 1
  end
end

function _draw()
  cls(1)

  -- title
  print("amico8 benchmark", 2, 2, 7)
  local st = phase==1 and "done" or ("test "..cur.."/"..ntests)
  print(st, 90, 2, 6)

  -- progress bar
  local pct = (cur-1) / ntests
  rectfill(0, 9, flr(127*pct), 10, 11)
  rectfill(flr(127*pct)+1, 9, 127, 10, 5)

  -- column headers
  print("test           us/it  cpu%", 0, 13, 6)
  line(0, 20, 127, 20, 5)

  -- results
  local row = 0
  for i=1,ntests do
    if i >= scroll+1 and row < 10 then
      local y = 22 + row*10
      local r = results[i]
      if r then
        -- color: green<50% frame, yellow<90%, red>90%
        local col = r.cpu < 0.5 and 11 or (r.cpu < 0.9 and 10 or 8)
        print(r.name, 0, y, 7)
        -- us/iter
        local us_s = tostr(r.us)
        print(us_s, 80, y, col)
        -- cpu%
        local cpu_pct = flr(r.cpu * 100)
        print(cpu_pct.."%", 108, y, col)
      else
        local col = i==cur and 7 or 5
        print(tname[i], 0, y, col)
        print("...", 80, y, 5)
      end
      row = row + 1
    end
  end

  -- footer
  line(0, 118, 127, 118, 5)
  if phase==1 then
    print("z=restart  up/dn=scroll", 0, 121, 6)
  else
    print("running...  up/dn=scroll", 0, 121, 5)
  end
end