Lunatic: more commands, timing test for xmath's sine functions.

git-svn-id: https://svn.eduke32.com/eduke32@3489 1a8010ca-5511-0410-912e-c29ae57300e0

polymer/eduke32/Makefile

@@ -157,7 +157,8 @@ ifneq (0,$(LUNATIC))
                $(OBJ)/luaJIT_BC_bitar.$o \
                $(OBJ)/luaJIT_BC_control.$o \
                $(OBJ)/luaJIT_BC_bcarray.$o \
-               $(OBJ)/luaJIT_BC_bcheck.$o
+               $(OBJ)/luaJIT_BC_bcheck.$o \
+               $(OBJ)/luaJIT_BC_xmath.$o
 
     # now, take care of having the necessary symbols (sector, wall, etc.) in the
     # executable no matter what the debugging level

polymer/eduke32/build/src/engine.c

@@ -11036,6 +11036,13 @@ int32_t krecip(int32_t num)
     return(krecipasm(num));
 }
 
+#ifdef LUNATIC
+int32_t Mulscale(int32_t a, int32_t b, int32_t sh)
+{
+    return mulscale(a, b, sh);
+}
+#endif
+
 // Gets the BUILD unit height and z offset of a sprite.
 // Returns the z offset, 'height' may be NULL.
 int32_t spriteheightofsptr(const spritetype *spr, int32_t *height, int32_t alsotileyofs)

polymer/eduke32/source/common.c

@@ -387,7 +387,7 @@ int32_t maybe_append_ext(char *wbuf, int32_t wbufsiz, const char *fn, const char
 }
 
 
-// Approximations to the 2D and 3D euclidean distances. Initial EDuke32 SVN import says
+// Approximations to 2D and 3D Euclidean distances. Initial EDuke32 SVN import says
 // in jmact/mathutil.c: "Ken's reverse-engineering job".
 // Note that jmact/mathutil.c contains practically the same code, but where the
 // individual x/y(/z) distances are passed instead.

polymer/eduke32/source/lunatic/control.lua

@@ -765,12 +765,10 @@ function _awayfromwall(spr, d)
     return true
 end
 
-local BANG2RAD = math.pi/1024
+local xmath = require("xmath")
 
 local function cossinb(bang)
-    -- XXX: better use the precalc'd arrays instead?
+    return xmath.cosb(bang), xmath.sinb(bang)
-    local ang = BANG2RAD*bang
-    return 16384*math.cos(ang), 16384*math.sin(ang)
 end
 
 local function manhatdist(v1, v2)
@@ -844,6 +842,10 @@ function _cansee(aci, ps)
     return can
 end
 
+function _canseespr(s1, s2)
+    return cansee(sprite[s1], sprite[s1].sectnum, sprite[s2], sprite[s2].sectnum) and 1 or 0
+end
+
 function _sleepcheck(aci, dist)
     local acs = actor[aci]
     if (dist > MAXSLEEPDIST and acs.timetosleep == 0) then

polymer/eduke32/source/lunatic/defs.ilua

@@ -1091,6 +1091,7 @@ local allowed_modules = {
     geom = geom,
     stat = require("stat"),
     bitar = require("bitar"),
+    xmath = require("xmath"),
 
     con = con,
 }

polymer/eduke32/source/lunatic/defs_common.lua

@@ -274,6 +274,7 @@ double gethitickms(void);
 int32_t krand(void);
 int32_t ksqrt(uint32_t num);
 int32_t __fastcall getangle(int32_t xvect, int32_t yvect);
+int32_t Mulscale(int32_t a, int32_t b, int32_t sh);
 ]]
 
 local bcheck = require("bcheck")
@@ -750,12 +751,8 @@ function hitscan(pos, sectnum, vx,vy,vz, cliptype)
 end
 
 function cansee(pos1,sect1, pos2,sect2)
-    if (sect1 >= ffiC.numsectors+0ULL) then
+    check_sector_idx(sect1)
-        error("passed out-of-bounds first sector number "..sect1, 2)
+    check_sector_idx(sect2)
-    end
-    if (sect2 >= ffiC.numsectors+0ULL) then
-        error("passed out-of-bounds second sector number "..sect2, 2)
-    end
 
     local ret = ffiC.cansee(pos1.x,pos1.y,pos1.z, sect1,
                             pos2.x,pos2.y,pos2.z, sect2)

polymer/eduke32/source/lunatic/dynsymlist

@@ -119,6 +119,9 @@ luaJIT_BC_randgen;
 luaJIT_BC_stat;
 luaJIT_BC_bitar;
 luaJIT_BC_control;
+luaJIT_BC_bcarray;
+luaJIT_BC_bcheck;
+luaJIT_BC_xmath;
 
 rand_jkiss_u32;
 rand_jkiss_dbl;

polymer/eduke32/source/lunatic/lunacon.lua

@@ -137,13 +137,15 @@ local function getlinecol(pos) end -- fwd-decl
 local function new_initial_codetab()
     return {
         "local _con, _bit, _math = require'con', require'bit', require'math';",
+        "local _xmath, _geom = require'xmath', require'geom';",
         "local sector, sprite, actor, player = sector, sprite, actor, player;",
-        "local gameactor, gameevent, _gv = gameactor, gameevent, gv;"
+        "local gameactor, gameevent, _gv = gameactor, gameevent, gv;",
+        "local cansee = cansee;"
            }
 end
 
 -- CON global system gamevar
-local function CSV(var) return "gv._csv"..var end
+local function CSV(var) return "_gv._csv"..var end
 
 -- Creates the table of predefined game variables.
 -- KEEPINSYNC gamevars.c: Gv_AddSystemVars()
@@ -1401,14 +1403,17 @@ local Cinner = {
 
     --- 2. Math operations
     sqrt = cmd(R,W)
-        / "%2=gv.ksqrt(%1)",
+        / "%2=_gv.ksqrt(%1)",
     calchypotenuse = cmd(W,R,R)
         / "%1=_con._hypot(%2,%3)",
-    sin = cmd(W,R),
+    sin = cmd(W,R)
-    cos = cmd(W,R),
+        / "%1=_xmath.ksin(%2)",
-    mulscale = cmd(W,R,R,R),
+    cos = cmd(W,R)
+        / "%1=_xmath.kcos(%2)",
+    mulscale = cmd(W,R,R,R)
+        / "%1=_gv.Mulscale(%1,%2,%3)",
     getangle = cmd(W,R,R)
-        / "%1=gv.getangle(%2,%3)",
+        / "%1=_gv.getangle(%2,%3)",
     getincangle = cmd(W,R,R)
         / "%1=_con._angdiff(%2,%3)",
 
@@ -1460,7 +1465,7 @@ local Cinner = {
     lockplayer = cmd(R)
         / PLS".transporter_hold=%1",
     quake = cmd(R)
-        / "gv.doQuake(%1,81)",  -- EARTHQUAKE
+        / "_gv.doQuake(%1,81)",  -- EARTHQUAKE
     jump = cmd(R)
         / handle.NYI,  -- will never be
     cmenu = cmd(R)
@@ -1474,7 +1479,6 @@ local Cinner = {
     echo = cmd(R)
         / "_con._echo(%1)",
     starttrackvar = cmd(R),
-    clearmapstate = cmd(R),
     activatecheat = cmd(R)
         / handle.NYI,
     setgamepalette = cmd(R),
@@ -1650,18 +1654,25 @@ local Cinner = {
         / handle.addlog,
     addweaponvar = cmd(R,R)  -- NLCF
         / handle.addweapon,
-    cansee = cmd(R,R,R,R,R,R,R,R,W),
+    cansee = cmd(R,R,R,R,R,R,R,R,W)
-    canseespr = cmd(R,R,W),
+        / "%9=cansee(_geom.ivec3(%1,%2,%3),%4, _geom.ivec3(%5,%6,%7),%8) and 1 or 0",
+    canseespr = cmd(R,R,W)
+        / "%3=_con._canseespr(%1,%2)",
     changespritesect = cmd(R,R)
         / "sprite.changesect(%1,%2)",
     changespritestat = cmd(R,R)
         / "sprite.changestat(%1,%2)",
-    clipmove = cmd(W,W,W,R,W,R,R,R,R,R,R),
+    clipmove = cmd(W,W,W,R,W,R,R,R,R,R,R)
-    clipmovenoslide = cmd(W,W,W,R,W,R,R,R,R,R,R),
+        / handle.NYI,
+    clipmovenoslide = cmd(W,W,W,R,W,R,R,R,R,R,R)
+        / handle.NYI,
     displayrand = cmd(W),
     displayrandvar = cmd(W,D),
     displayrandvarvar = cmd(W,R),
-    dist = cmd(W,R,R),
+    dist = cmd(W,R,R)
+        / "%1=_xmath.dist(sprite[%1],sprite[%2])",
+    ldist = cmd(W,R,R)
+        / "%1=_xmath.ldist(sprite[%1],sprite[%2])",
     dragpoint = cmd(R,R,R),
     hitscan = cmd(R,R,R,R,R,R,R,W,W,W,W,W,W,R), -- 7R 6W 1R
 
@@ -1672,11 +1683,10 @@ local Cinner = {
     digitalnumberz = cmd(R,R,R,R,R,R,R,R,R,R,R,R),  -- 12R
     minitext = cmd(R,R,R,R,R),
 
-    ldist = cmd(W,R,R),
+    lineintersect = cmd(R,R,R,R,R,R,R,R,R,R,W,W,W,W)  -- 10R 4W
-    lineintersect = cmd(R,R,R,R,R,R,R,R,R,R,W,W,W,W),  -- 10R 4W
+        / handle.NYI,
-    rayintersect = cmd(R,R,R,R,R,R,R,R,R,R,W,W,W,W),  -- 10R 4W
+    rayintersect = cmd(R,R,R,R,R,R,R,R,R,R,W,W,W,W)  -- 10R 4W
-    loadmapstate = cmd(),
+        / handle.NYI,
-    savemapstate = cmd(),
     movesprite = cmd(R,R,R,R,R,W),
     neartag = cmd(R,R,R,R,R,W,W,W,W,R,R),
     palfrom = (sp1 * tok.define)^-4
@@ -1703,6 +1713,10 @@ local Cinner = {
     myospalx = cmd(R,R,R,R,R,R)
         / "_con._myos(%1,%2,32768,%3,%4,%5,%6)",
 
+    clearmapstate = cmd(R),
+    loadmapstate = cmd(),
+    savemapstate = cmd(),
+
     headspritesect = cmd(W,R),
     headspritestat = cmd(W,R),
     nextspritesect = cmd(W,R),
@@ -1770,7 +1784,7 @@ local Cinner = {
     getkeyname = cmd(R,R,R),
     getpname = cmd(R,R),
     getticks = cmd(W)
-        / "%1=gv.getticks()",
+        / "%1=_gv.getticks()",
     gettimedate = cmd(W,W,W,W,W,W,W,W)
         / "%1,%2,%3,%4,%5,%6,%7,%8=_con._gettimedate()",
     getzrange = cmd(R,R,R,R,W,W,W,W,R,R),

polymer/eduke32/source/lunatic/test.elua

@@ -400,6 +400,34 @@ bittest.sieve()
 require("test/test_geom")
 require("test/test_rotspr")
 
+do
+    -- Test ksin vs. sinb
+    local xmath = require "xmath"
+    local sum = 0
+
+    local N = 1000
+    local t = gv.gethitickms()
+    for i=0,N*2048-1 do
+        sum = sum+xmath.ksin(i)
+    end
+    t = gv.gethitickms()-t
+    sum = sum*1e12
+    printf("ksin: sum*1e12=%.03f, %.03fus per 0-2047 cycle", sum, t)
+
+    sum = 0
+    local t = gv.gethitickms()
+    for i=0,N*2048-1 do
+        sum = sum+xmath.sinb(i)
+    end
+    t = gv.gethitickms()-t
+    sum = sum*1e12
+    printf("sinb: sum*1e12=%.03f, %.03fus per 0-2047 cycle", sum, t)
+
+    -- Results (helixhorned x86):
+    -- ksin: sum*1e12=0.000, 3.801us per 0-2047 cycle
+    -- sinb: sum*1e12=0.052, 2.886us per 0-2047 cycle
+end
+
 print('---=== END TEST SCRIPT ===---')
 
 --[[