ref: c483a0f2077d7f1325a801aa1e0d5301bc6eb527
parent: affa540c607a2f2f5906e5897ab84643f430ff5a
author: Sigrid Solveig Haflínudóttir <sigrid@ftrv.se>
date: Sun Dec 31 23:49:47 EST 2023
-O3 -march=native; produce mip/fullbright/additive specific surface lighting funcs
--- a/Makefile
+++ b/Makefile
@@ -4,7 +4,7 @@
BIN=${DESTDIR}${PREFIX}/binEXTRA_CFLAGS=$$(pkg-config --cflags sdl2 openal)
EXTRA_LDFLAGS=$$(pkg-config --libs sdl2 openal)
-CFLAGS?=-O2 -g
+CFLAGS?=-O3 -g -march=native
CFLAGS+=-Wall -Wextra -Wno-unknown-pragmas -Wno-missing-field-initializers -Wno-implicit-fallthrough -Wno-microsoft-anon-tag
CFLAGS+=-fms-extensions
CFLAGS+=-I3rd/parg -Iunix -I. ${EXTRA_CFLAGS}--- a/model.h
+++ b/model.h
@@ -28,11 +28,12 @@
vec3_t position;
} mvertex_t;
-#define SIDE_FRONT 0
-#define SIDE_BACK 1
-#define SIDE_ON 2
+enum {+ SIDE_FRONT,
+ SIDE_BACK,
+ SIDE_ON,
+};
-
// plane_t structure
// !!! if this is changed, it must be changed in asm_i386.h too !!!
typedef struct mplane_s
@@ -44,9 +45,14 @@
byte pad[2];
} mplane_t;
+enum {+ DRAWSURF_FULLBRIGHT = 1<<0,
+};
+
typedef struct texture_s
{char name[16];
+ int drawsurf;
int width, height;
int anim_total; // total tenths in sequence ( 0 = no)
int anim_min, anim_max; // time for this frame min <=time< max
--- a/model_bsp.c
+++ b/model_bsp.c
@@ -133,6 +133,12 @@
tx->height = h;
// the pixels immediately follow the structures
torgbx(p, tx->pixels, pixels);
+ for(j = 0; j < pixels; j++){+ if(tx->pixels[j] != 0 && (tx->pixels[j] & 0xff000000) == 0){+ tx->drawsurf |= DRAWSURF_FULLBRIGHT;
+ break;
+ }
+ }
if(tx->name[0] == '{'){ for(j = 1; j < MIPLEVELS; j++){w /= 2;
--- a/model_bsp30.c
+++ b/model_bsp30.c
@@ -175,6 +175,7 @@
}
}
if(strchr(tx->name, '~') != nil){+ tx->drawsurf |= DRAWSURF_FULLBRIGHT;
/* last 32 colors are fullbright */
for(j = 0; j < pixels; j++){if(p[j] >= palsz-32)
--- a/r_surf.c
+++ b/r_surf.c
@@ -5,15 +5,10 @@
static int sourcetstep;
static int surfrowbytes; // used by ASM files
static int r_stepback;
-static int r_numhblocks, r_numvblocks;
static pixel_t *r_source, *r_sourcemax;
-static unsigned *r_lightptr[3];
-static int r_lightwidth;
static pixel_t *pbasesource;
static void *prowdestbase;
-static void R_DrawSurfaceBlock8(int mip);
-
static unsigned blocklights[3][18*18];
/*
@@ -193,7 +188,118 @@
return base;
}
+inline pixel_t
+addlight(pixel_t x, int lr, int lg, int lb)
+{+ int r, g, b;
+ if((x & 0xff000000U) == 0)
+ return x;
+
+ if(currententity != nil && (currententity->effects & EF_ADDITIVE) != 0)
+ return x;
+
+ r = (x>>16) & 0xff;
+ g = (x>>8) & 0xff;
+ b = (x>>0) & 0xff;
+
+ r = (r * ((64<<8)-(lr & 0xffff))) >> (8+VID_CBITS);
+ g = (g * ((64<<8)-(lg & 0xffff))) >> (8+VID_CBITS);
+ b = (b * ((64<<8)-(lb & 0xffff))) >> (8+VID_CBITS);
+ x = (x & 0xff000000) | r<<16 | g<<8 | b<<0;
+
+ return x;
+}
+
+#define fullbright 1
+#define additive 1
+#define addlight addlight_f1_a1
+#define DrawSurfaceBlock_m0 DrawSurfaceBlock_f1_a1_m0
+#define DrawSurfaceBlock_m1 DrawSurfaceBlock_f1_a1_m1
+#define DrawSurfaceBlock_m2 DrawSurfaceBlock_f1_a1_m2
+#define DrawSurfaceBlock_m3 DrawSurfaceBlock_f1_a1_m3
+#include "r_surf_x.h"
+#undef fullbright
+#undef additive
+#undef addlight
+#undef DrawSurfaceBlock_m0
+#undef DrawSurfaceBlock_m1
+#undef DrawSurfaceBlock_m2
+#undef DrawSurfaceBlock_m3
+
+#define fullbright 0
+#define additive 1
+#define addlight addlight_f0_a1
+#define DrawSurfaceBlock_m0 DrawSurfaceBlock_f0_a1_m0
+#define DrawSurfaceBlock_m1 DrawSurfaceBlock_f0_a1_m1
+#define DrawSurfaceBlock_m2 DrawSurfaceBlock_f0_a1_m2
+#define DrawSurfaceBlock_m3 DrawSurfaceBlock_f0_a1_m3
+#include "r_surf_x.h"
+#undef fullbright
+#undef additive
+#undef addlight
+#undef DrawSurfaceBlock_m0
+#undef DrawSurfaceBlock_m1
+#undef DrawSurfaceBlock_m2
+#undef DrawSurfaceBlock_m3
+
+#define fullbright 1
+#define additive 0
+#define addlight addlight_f1_a0
+#define DrawSurfaceBlock_m0 DrawSurfaceBlock_f1_a0_m0
+#define DrawSurfaceBlock_m1 DrawSurfaceBlock_f1_a0_m1
+#define DrawSurfaceBlock_m2 DrawSurfaceBlock_f1_a0_m2
+#define DrawSurfaceBlock_m3 DrawSurfaceBlock_f1_a0_m3
+#include "r_surf_x.h"
+#undef fullbright
+#undef additive
+#undef addlight
+#undef DrawSurfaceBlock_m0
+#undef DrawSurfaceBlock_m1
+#undef DrawSurfaceBlock_m2
+#undef DrawSurfaceBlock_m3
+
+#define fullbright 0
+#define additive 0
+#define addlight addlight_f0_a0
+#define DrawSurfaceBlock_m0 DrawSurfaceBlock_f0_a0_m0
+#define DrawSurfaceBlock_m1 DrawSurfaceBlock_f0_a0_m1
+#define DrawSurfaceBlock_m2 DrawSurfaceBlock_f0_a0_m2
+#define DrawSurfaceBlock_m3 DrawSurfaceBlock_f0_a0_m3
+#include "r_surf_x.h"
+#undef fullbright
+#undef additive
+#undef addlight
+#undef DrawSurfaceBlock_m0
+#undef DrawSurfaceBlock_m1
+#undef DrawSurfaceBlock_m2
+#undef DrawSurfaceBlock_m3
+
+typedef void (*drawfunc)(unsigned *lp[4], unsigned lw, int nb);
+
+static const drawfunc drawsurf[2/*fullbright*/][2/*additive*/][4/*mipmap*/] = {+ {+ {+ DrawSurfaceBlock_f0_a0_m0, DrawSurfaceBlock_f0_a0_m1,
+ DrawSurfaceBlock_f0_a0_m2, DrawSurfaceBlock_f0_a0_m3,
+ },
+ {+ DrawSurfaceBlock_f0_a1_m0, DrawSurfaceBlock_f0_a1_m1,
+ DrawSurfaceBlock_f0_a1_m2, DrawSurfaceBlock_f0_a1_m3,
+ },
+ },
+ {+ {+ DrawSurfaceBlock_f1_a0_m0, DrawSurfaceBlock_f1_a0_m1,
+ DrawSurfaceBlock_f1_a0_m2, DrawSurfaceBlock_f1_a0_m3,
+ },
+ {+ DrawSurfaceBlock_f1_a1_m0, DrawSurfaceBlock_f1_a1_m1,
+ DrawSurfaceBlock_f1_a1_m2, DrawSurfaceBlock_f1_a1_m3,
+ },
+ },
+};
+
/*
===============
R_DrawSurface
@@ -202,12 +308,15 @@
void R_DrawSurface (void)
{pixel_t *basetptr;
- int smax, tmax, twidth;
+ int smax, tmax, twidth, lightwidth;
int u, blockdivshift, blocksize;
int soffset, basetoffset, texwidth;
int horzblockstep;
+ int r_numhblocks, r_numvblocks;
pixel_t *pcolumndest;
texture_t *mt;
+ drawfunc draw;
+ unsigned *lp[3];
// calculate the lightings
R_BuildLightMap ();
@@ -216,6 +325,8 @@
mt = r_drawsurf.texture;
+ draw = drawsurf[mt->drawsurf][currententity != nil && (currententity->effects & EF_ADDITIVE) != 0][r_drawsurf.surfmip];
+
r_source = mt->pixels + mt->offsets[r_drawsurf.surfmip];
// the fractional light values should range from 0 to (VID_GRADES - 1) << 16
@@ -226,7 +337,7 @@
blocksize = 16 >> r_drawsurf.surfmip;
blockdivshift = 4 - r_drawsurf.surfmip;
- r_lightwidth = (r_drawsurf.surf->extents[0]>>4)+1;
+ lightwidth = (r_drawsurf.surf->extents[0]>>4)+1;
r_numhblocks = r_drawsurf.surfwidth >> blockdivshift;
r_numvblocks = r_drawsurf.surfheight >> blockdivshift;
@@ -247,22 +358,19 @@
// << 16 components are to guarantee positive values for %
soffset = ((soffset >> r_drawsurf.surfmip) + (smax << 16)) % smax;
- basetptr = &r_source[((((basetoffset >> r_drawsurf.surfmip)
- + (tmax << 16)) % tmax) * twidth)];
+ basetptr = &r_source[((((basetoffset >> r_drawsurf.surfmip) + (tmax << 16)) % tmax) * twidth)];
pcolumndest = r_drawsurf.surfdat;
- for (u=0 ; u<r_numhblocks; u++)
- {- r_lightptr[0] = blocklights[0] + u;
- r_lightptr[1] = blocklights[1] + u;
- r_lightptr[2] = blocklights[2] + u;
-
+ for (u=0 ; u<r_numhblocks; u++){prowdestbase = pcolumndest;
pbasesource = basetptr + soffset;
+ lp[0] = blocklights[0]+u;
+ lp[1] = blocklights[1]+u;
+ lp[2] = blocklights[2]+u;
- R_DrawSurfaceBlock8(r_drawsurf.surfmip);
+ draw(lp, lightwidth, r_numvblocks);
soffset = soffset + blocksize;
if (soffset >= smax)
@@ -269,78 +377,5 @@
soffset = 0;
pcolumndest += horzblockstep;
- }
-}
-
-
-//=============================================================================
-
-inline pixel_t
-addlight(pixel_t x, int lr, int lg, int lb)
-{- int r, g, b;
-
- if((x & 0xff000000U) == 0)
- return x;
-
- if(currententity != nil && (currententity->effects & EF_ADDITIVE) != 0)
- return x;
-
- r = (x>>16) & 0xff;
- g = (x>>8) & 0xff;
- b = (x>>0) & 0xff;
-
- r = (r * ((64<<8)-(lr & 0xffff))) >> (8+VID_CBITS);
- g = (g * ((64<<8)-(lg & 0xffff))) >> (8+VID_CBITS);
- b = (b * ((64<<8)-(lb & 0xffff))) >> (8+VID_CBITS);
- x = (x & 0xff000000) | r<<16 | g<<8 | b<<0;
-
- return x;
-}
-
-static void
-R_DrawSurfaceBlock8(int mip)
-{- int b, v, i, j, lightstep[3], light[3], lightleft[3], lightright[3];
- int lightleftstep[3], lightrightstep[3];
- pixel_t *psource, *prowdest;
-
- psource = pbasesource;
- prowdest = prowdestbase;
-
- for (v=0 ; v<r_numvblocks ; v++)
- {- for(j = 0; j < 3; j++){- lightleft[j] = r_lightptr[j][0];
- lightright[j] = r_lightptr[j][1];
- r_lightptr[j] += r_lightwidth;
- lightleftstep[j] = (r_lightptr[j][0] - lightleft[j]) >> (4-mip);
- lightrightstep[j] = (r_lightptr[j][1] - lightright[j]) >> (4-mip);
- }
-
- for (i=0 ; i<16>>mip ; i++)
- {- for(j = 0; j < 3; j++){- lightstep[j] = (lightleft[j] - lightright[j]) >> (4-mip);
- light[j] = lightright[j];
- }
-
- for(b = (16>>mip)-1; b >= 0; b--){- prowdest[b] = addlight(psource[b], light[0], light[1], light[2]);
- light[0] += lightstep[0];
- light[1] += lightstep[1];
- light[2] += lightstep[2];
- }
-
- psource += sourcetstep;
- prowdest += surfrowbytes;
- for(j = 0; j < 3; j++){- lightright[j] += lightrightstep[j];
- lightleft[j] += lightleftstep[j];
- }
- }
-
- if (psource >= r_sourcemax)
- psource -= r_stepback;
}
}
--- /dev/null
+++ b/r_surf_block.h
@@ -1,0 +1,48 @@
+#define N 3
+static void
+DrawSurfaceBlock(unsigned *lp[N], unsigned lw, int nb)
+{+ int b, v, i, j, lightstep[N], light[N], lightleft[N], lightright[N];
+ int lightleftstep[N], lightrightstep[N];
+ pixel_t *psource, *prowdest;
+
+ psource = pbasesource;
+ prowdest = prowdestbase;
+
+ for (v=0 ; v<nb ; v++)
+ {+ for(j = 0; j < N; j++){+ lightleft[j] = lp[j][0];
+ lightright[j] = lp[j][1];
+ lp[j] += lw;
+ lightleftstep[j] = (lp[j][0] - lightleft[j]) >> (4-mip);
+ lightrightstep[j] = (lp[j][1] - lightright[j]) >> (4-mip);
+ }
+
+ for (i=0 ; i<16>>mip; i++)
+ {+ for(j = 0; j < N; j++){+ lightstep[j] = (lightleft[j] - lightright[j]) >> (4-mip);
+ light[j] = lightright[j];
+ }
+
+ for(b = (16>>mip)-1; b >= 0; b--){+ prowdest[b] = addlight(psource[b], light[0], light[1], light[2]);
+ for(j = 0; j < N; j++)
+ light[j] += lightstep[j];
+ }
+
+ psource += sourcetstep;
+ prowdest += surfrowbytes;
+
+ for(j = 0; j < N; j++){+ lightright[j] += lightrightstep[j];
+ lightleft[j] += lightleftstep[j];
+ }
+ }
+
+ if (psource >= r_sourcemax)
+ psource -= r_stepback;
+ }
+}
+#undef N
--- /dev/null
+++ b/r_surf_light.h
@@ -1,0 +1,22 @@
+static inline pixel_t
+addlight(pixel_t x, int lr, int lg, int lb)
+{+ int r, g, b;
+
+ if(fullbright && (x & 0xff000000U) == 0)
+ return x;
+
+ if(additive)
+ return x;
+
+ r = (x>>16) & 0xff;
+ g = (x>>8) & 0xff;
+ b = (x>>0) & 0xff;
+
+ r = (r * ((64<<8)-(lr & 0xffff))) >> (8+VID_CBITS);
+ g = (g * ((64<<8)-(lg & 0xffff))) >> (8+VID_CBITS);
+ b = (b * ((64<<8)-(lb & 0xffff))) >> (8+VID_CBITS);
+ x = (x & 0xff000000) | r<<16 | g<<8 | b<<0;
+
+ return x;
+}
--- /dev/null
+++ b/r_surf_x.h
@@ -1,0 +1,25 @@
+#include "r_surf_light.h"
+
+#define mip 0
+#define DrawSurfaceBlock DrawSurfaceBlock_m0
+#include "r_surf_block.h"
+#undef mip
+#undef DrawSurfaceBlock
+
+#define mip 1
+#define DrawSurfaceBlock DrawSurfaceBlock_m1
+#include "r_surf_block.h"
+#undef mip
+#undef DrawSurfaceBlock
+
+#define mip 2
+#define DrawSurfaceBlock DrawSurfaceBlock_m2
+#include "r_surf_block.h"
+#undef mip
+#undef DrawSurfaceBlock
+
+#define mip 3
+#define DrawSurfaceBlock DrawSurfaceBlock_m3
+#include "r_surf_block.h"
+#undef mip
+#undef DrawSurfaceBlock