ref: a2f781fe0e786384ed36edb93aceb42b3b1b76ab
parent: a53e14793a11a6af0b83bdbd94772c47ee41e44b
author: Sigrid Solveig Haflínudóttir <sigrid@ftrv.se>
date: Tue Jan 9 18:12:46 EST 2024
first experiment in making drawing more parallel
--- a/d_edge.c
+++ b/d_edge.c
@@ -1,3 +1,9 @@
+#define _GNU_SOURCE
+#include <pthread.h>
+#include <sys/syscall.h>
+#include <unistd.h>
+#include <limits.h>
+#include "fast_barrier.h"
#include "quakedef.h"
float scale_for_mip;
@@ -23,7 +29,7 @@
}
static void
-D_DrawSolidSurface(surf_t *surf, pixel_t color)
+D_DrawSolidSurface(surf_t *surf, pixel_t color, int first, int end)
{espan_t *span;
pixel_t *pdest;
@@ -31,6 +37,8 @@
int u, u2;
for(span = surf->spans; span; span=span->pnext){+ if(span->v < first || span->v >= end)
+ continue;
pdest = dvars.fb + span->v*dvars.w;
pz = dvars.zb + span->v*dvars.w;
memset(pz, 0xfe, span->count*sizeof(*pz));
@@ -80,9 +88,31 @@
tv->t.bbextent = ((pface->extents[1] << 16) >> miplevel) - 1;
}
-void
-D_DrawSurfaces(view_t *v0)
+static fast_barrier_t spansgobrr, spansgohome;
+static pthread_spinlock_t spancache;
+
+typedef struct span_thread_t span_thread_t;
+
+struct span_thread_t {+ pthread_t tid;
+ int n;
+ int first;
+ int end;
+};
+
+static int nthreads = 8;
+static bool spawned = false;
+static view_t *v0;
+
+static void
+spancachelock(int n)
{+ (n ? pthread_spin_lock : pthread_spin_unlock)(&spancache);
+}
+
+static void
+spannothread(view_t *v0, int first, int end)
+{vec3_t local_modelorg, transformed_modelorg, world_transformed_modelorg;
surfcache_t *pcurrentcache;
drawsurf_t ds = {0};@@ -90,23 +120,35 @@
int miplevel;
entity_t *e;
texvars_t t;
- surf_t *s;
byte alpha;
bool blend;
+ surf_t *s;
view_t v;
+ espan_t *sp;
+ bool yes;
+ ///uvlong t0 = nanosec();
+
memmove(&v, v0, sizeof(v));
TransformVector(v.modelorg, transformed_modelorg, &v);
VectorCopy(transformed_modelorg, world_transformed_modelorg);
// TODO: could preset a lot of this at mode set time
- for(s = &surfaces[1]; s < surface_p; s++){- if(!s->spans)
+ for(s = surfaces+1; s < surface_p; s++){+ e = s->entity;
+ if(!s->spans || ((surfdrawflags(s->flags) | entdrawflags(e)) ^ r_drawflags))
continue;
- e = s->entity;
- if((surfdrawflags(s->flags) | entdrawflags(e)) ^ r_drawflags)
+ yes = false;
+ for(sp = s->spans; sp != nil; sp = sp->pnext){+ if(sp->v >= first && sp->v < end){+ yes = true;
+ break;
+ }
+ }
+ if(!yes)
continue;
+
alpha = 255;
if(enthasalpha(e) && e->alpha != 255)
alpha = e->alpha;
@@ -127,20 +169,20 @@
pface = s->data;
if(s->flags & SURF_DRAWSKY){- D_DrawSkyScans8(s->spans);
+ D_DrawSkyScans8(s->spans, first, end);
}else if(s->flags & SURF_DRAWBACKGROUND){- D_DrawSolidSurface(s, q1pal[(int)r_clearcolor.value & 0xFF]);
+ D_DrawSolidSurface(s, q1pal[(int)r_clearcolor.value & 0xFF], first, end);
}else if(s->flags & SURF_DRAWTURB){t.p = pface->texinfo->texture->pixels;
t.w = 64;
D_CalcGradients(0, pface, transformed_modelorg, &v, &t);
- D_DrawSpans(s->spans, &t, alpha, SPAN_TURB);
+ D_DrawSpans(s->spans, &t, alpha, SPAN_TURB, first, end);
}else{miplevel = D_MipLevelForScale(s->nearzi * scale_for_mip * pface->texinfo->mipadjust);
if(s->flags & SURF_FENCE)
miplevel = max(miplevel-1, 0);
- pcurrentcache = D_CacheSurface(s->entity, pface, &ds, miplevel);
+ pcurrentcache = D_CacheSurface(s->entity, pface, &ds, miplevel, spancachelock);
t.p = pcurrentcache->pixels;
t.w = pcurrentcache->width;
D_CalcGradients(miplevel, pface, transformed_modelorg, &v, &t);
@@ -148,7 +190,9 @@
D_DrawSpans(s->spans, &t, alpha,
(alpha == 255 && (s->flags & SURF_FENCE))
? SPAN_FENCE
- : (blend ? SPAN_BLEND : SPAN_SOLID)
+ : (blend ? SPAN_BLEND : SPAN_SOLID),
+ first,
+ end
);
}
@@ -156,5 +200,182 @@
VectorCopy(world_transformed_modelorg, transformed_modelorg);
memmove(&v, v0, sizeof(v));
}
+ }
+
+ ///uvlong t1 = nanosec();
+ ///if(first != 0 || end != vid.height)
+ /// fprintf(stderr, "@%d %llu\n", 0, t1-t0);
+}
+
+static void *
+spanthread(void *th_)
+{+ vec3_t local_modelorg, transformed_modelorg, world_transformed_modelorg;
+ surfcache_t *pcurrentcache;
+ span_thread_t *th = th_;
+ drawsurf_t ds = {0};+ msurface_t *pface;
+ int miplevel, ns;
+ entity_t *e;
+ texvars_t t;
+ byte alpha;
+ bool blend;
+ surf_t *s;
+ espan_t *sp;
+ bool yes;
+ view_t v;
+
+ for(;;){+ fast_barrier_wait(&spansgobrr);
+
+ //uvlong t0 = nanosec();
+ memmove(&v, v0, sizeof(v));
+ TransformVector(v.modelorg, transformed_modelorg, &v);
+ VectorCopy(transformed_modelorg, world_transformed_modelorg);
+ ns = 0;
+
+ // TODO: could preset a lot of this at mode set time
+ for(s = surfaces+1; s < surface_p; s++){+ e = s->entity;
+ if(!s->spans || ((surfdrawflags(s->flags) | entdrawflags(e)) ^ r_drawflags))
+ continue;
+ yes = false;
+ for(sp = s->spans; sp != nil; sp = sp->pnext){+ if(sp->v >= th->first && sp->v < th->end){+ yes = true;
+ break;
+ }
+ }
+ if(!yes)
+ continue;
+ ns++;
+ alpha = 255;
+ if(enthasalpha(e) && e->alpha != 255)
+ alpha = e->alpha;
+ else if(s->flags & SURF_TRANS)
+ alpha *= alphafor(s->flags);
+ if(alpha < 1)
+ alpha = 255;
+
+ t.z.stepu = s->d_zistepu;
+ t.z.stepv = s->d_zistepv;
+ t.z.origin = s->d_ziorigin;
+
+ if(insubmodel(s)){+ VectorSubtract(v.org, e->origin, local_modelorg);
+ TransformVector(local_modelorg, transformed_modelorg, &v);
+ R_RotateBmodel(e, &v);
+ }
+
+ pface = s->data;
+ if(s->flags & SURF_DRAWSKY){+ D_DrawSkyScans8(s->spans, th->first, th->end);
+ }else if(s->flags & SURF_DRAWBACKGROUND){+ D_DrawSolidSurface(s, q1pal[(int)r_clearcolor.value & 0xFF], th->first, th->end);
+ }else if(s->flags & SURF_DRAWTURB){+ t.p = pface->texinfo->texture->pixels;
+ t.w = 64;
+ D_CalcGradients(0, pface, transformed_modelorg, &v, &t);
+ D_DrawSpans(s->spans, &t, alpha, SPAN_TURB, th->first, th->end);
+ }else{+ miplevel = D_MipLevelForScale(s->nearzi * scale_for_mip * pface->texinfo->mipadjust);
+ if(s->flags & SURF_FENCE)
+ miplevel = max(miplevel-1, 0);
+
+ pcurrentcache = D_CacheSurface(s->entity, pface, &ds, miplevel, spancachelock);
+ t.p = pcurrentcache->pixels;
+ t.w = pcurrentcache->width;
+ D_CalcGradients(miplevel, pface, transformed_modelorg, &v, &t);
+ blend = (s->flags & SURF_FENCE) || (r_drawflags & DRAW_BLEND);
+ D_DrawSpans(s->spans, &t, alpha,
+ (alpha == 255 && (s->flags & SURF_FENCE))
+ ? SPAN_FENCE
+ : (blend ? SPAN_BLEND : SPAN_SOLID),
+ th->first,
+ th->end
+ );
+ }
+
+ if(insubmodel(s)){+ VectorCopy(world_transformed_modelorg, transformed_modelorg);
+ memmove(&v, v0, sizeof(v));
+ }
+ }
+
+ ///uvlong t1 = nanosec();
+ ///fprintf(stderr, "@%d %llu\n", th->n, t1-t0);
+ //fprintf(stderr, "!%d %d\n", th->n, ns);
+ fast_barrier_wait(&spansgohome);
+ }
+
+ return nil;
+}
+
+static span_thread_t *threads;
+
+void
+D_DrawSurfaces(view_t *v0_)
+{+ static int lastheight = -1;
+ span_thread_t *t;
+ int i, split, dt, n, y;
+
+ if(lastheight < 0)
+ pthread_spin_init(&spancache, PTHREAD_PROCESS_PRIVATE);
+
+ if(nthreads > 1 && threads == nil){+ pthread_barrierattr_t battr;
+ cpu_set_t set;
+
+ pthread_barrierattr_setpshared(&battr, PTHREAD_PROCESS_PRIVATE);
+ fast_barrier_init(&spansgobrr, &battr, nthreads);
+ fast_barrier_init(&spansgohome, &battr, nthreads);
+
+ threads = calloc(1, sizeof(*threads) * nthreads);
+ for(t = threads, i = 0; i < nthreads; i++, t++){+ t->n = i;
+ CPU_ZERO(&set);
+ CPU_SET(2*i, &set);
+ if(i == 0){+ sched_setaffinity(getpid(), sizeof(set), &set);
+ }else{+ pthread_create(&t->tid, nil, spanthread, t);
+ pthread_setaffinity_np(t->tid, sizeof(set), &set);
+ }
+ }
+ spawned = true;
+ }
+ if(threads != nil && lastheight != vid.height){+ lastheight = vid.height;
+ split = (nthreads+2)*nthreads/8;
+ dt = vid.height/2 / split;
+ n = dt*nthreads/2;
+ y = 0;
+ for(t = threads, i = 0; i < nthreads; i++, t++){+ t->first = y;
+ t->end = y = y + n;
+ if((n -= dt) == 0){+ dt = -dt;
+ n = -dt;
+ }
+ ///fprintf(stderr, "# %d: %d...%d\n", i, t->first, t->end);
+ }
+ t[-1].end = vid.height;
+ }
+
+ v0 = v0_;
+ if(nthreads < 2 || (r_drawflags & DRAW_BLEND) != 0){+ // overhead (lots of small objects + synchronization)
+ // not worth it - run it all in the same thread
+ spannothread(v0, 0, vid.height);
+ }else{+ ///uvlong t0 = nanosec();
+ fast_barrier_wait(&spansgobrr);
+ ///uvlong t1 = nanosec();
+ spannothread(v0, threads[0].first, threads[0].end);
+ ///uvlong t2 = nanosec();
+ fast_barrier_wait(&spansgohome);
+ ///uvlong t3 = nanosec();
+ ///fprintf(stderr, "---------- total=%llu start_barrier=%llu end_barrier=%llu\n", t3-t0, t1-t0, t3-t2);
}
}
--- a/d_local.h
+++ b/d_local.h
@@ -3,7 +3,7 @@
enum {DS_SPAN_LIST_END = -128,
- SURFCACHE_SIZE_AT_320X200 = 600*1024,
+ SURFCACHE_SIZE_AT_320X200 = 8*1024*1024,
};
typedef struct {@@ -79,11 +79,11 @@
SPAN_TURB,
};
-void D_DrawSpans(espan_t *pspan, texvars_t *t, byte alpha, int spanfunc);
+void D_DrawSpans(espan_t *pspan, texvars_t *t, byte alpha, int spanfunc, int first, int end);
-void D_DrawSkyScans8 (espan_t *pspan);
+void D_DrawSkyScans8 (espan_t *pspan, int first, int end);
-surfcache_t *D_CacheSurface(entity_t *e, msurface_t *ms, drawsurf_t *ds, int miplevel);
+surfcache_t *D_CacheSurface(entity_t *e, msurface_t *ms, drawsurf_t *ds, int miplevel, void (*lock)(int n));
extern int *d_pscantable;
extern int d_scantable[MAXHEIGHT];
--- a/d_scan.c
+++ b/d_scan.c
@@ -69,7 +69,7 @@
}
void
-D_DrawSpans(espan_t *pspan, texvars_t *tv, byte alpha, int spanfunc)
+D_DrawSpans(espan_t *pspan, texvars_t *tv, byte alpha, int spanfunc, int first, int end)
{int count, spancount, izistep, spancountminus1, spanshift, spanmax;
pixel_t *pdest;
@@ -95,6 +95,8 @@
fogenabled = isfogged();
do{+ if(pspan->v < first || pspan->v >= end)
+ continue;
pdest = dvars.fb + pspan->v*dvars.w + pspan->u;
pz = dvars.zb + pspan->v*dvars.w + pspan->u;
zi = tv->z.origin + pspan->v*tv->z.stepv + pspan->u*tv->z.stepu;
--- a/d_sky.c
+++ b/d_sky.c
@@ -36,7 +36,8 @@
D_DrawSkyScans8
=================
*/
-void D_DrawSkyScans8 (espan_t *pspan)
+void
+D_DrawSkyScans8(espan_t *pspan, int first, int end)
{int count, spancount, u, v, spancountminus1;
pixel_t *pdest, pix;
@@ -57,6 +58,8 @@
do
{+ if(pspan->v < first || pspan->v >= end)
+ continue;
pdest = dvars.fb + pspan->v*dvars.w + pspan->u;
count = pspan->count;
pz = dvars.zb + pspan->v*dvars.w + pspan->u;
--- a/d_surf.c
+++ b/d_surf.c
@@ -181,7 +181,7 @@
================
*/
surfcache_t *
-D_CacheSurface(entity_t *e, msurface_t *ms, drawsurf_t *ds, int miplevel)
+D_CacheSurface(entity_t *e, msurface_t *ms, drawsurf_t *ds, int miplevel, void (*lock)(int))
{surfcache_t *cache;
@@ -192,6 +192,8 @@
ds->lightadj[2] = d_lightstylevalue[ms->styles[2]];
ds->lightadj[3] = d_lightstylevalue[ms->styles[3]];
+ if(lock != nil)
+ lock(1);
// see if the cache holds apropriate data
cache = ms->cachespots[miplevel];
@@ -201,8 +203,11 @@
&& cache->lightadj[0] == ds->lightadj[0]
&& cache->lightadj[1] == ds->lightadj[1]
&& cache->lightadj[2] == ds->lightadj[2]
- && cache->lightadj[3] == ds->lightadj[3] )
+ && cache->lightadj[3] == ds->lightadj[3] ){+ if(lock != nil)
+ lock(0);
return cache;
+ }
// determine shape of surface
surfscale = 1.0 / (1<<miplevel);
@@ -228,6 +233,8 @@
cache->lightadj[3] = ds->lightadj[3];
ds->m = ms;
R_DrawSurface(e, ds);
+ if(lock != nil)
+ lock(0);
return ms->cachespots[miplevel];
}
--- a/r_local.h
+++ b/r_local.h
@@ -101,8 +101,6 @@
void R_DrawSolidClippedSubmodelPolygons (model_t *pmodel, view_t *v);
void R_DrawSubmodelPolygons (model_t *pmodel, view_t *v, int clipflags);
-void R_AddPolygonEdges (emitpoint_t *pverts, int numverts, int miplevel);
-surf_t *R_GetSurf (void);
void R_AliasDrawModel (alight_t *plighting, view_t *v);
void R_BeginEdgeFrame (void);
void R_ScanEdges(view_t *v);
@@ -110,13 +108,6 @@
void R_InsertNewEdges (edge_t *edgestoadd, edge_t *edgelist);
void R_StepActiveU (edge_t *pedge);
void R_RemoveEdges (edge_t *pedge);
-
-extern void R_Surf8Start (void);
-extern void R_Surf8End (void);
-extern void R_Surf16Start (void);
-extern void R_Surf16End (void);
-extern void R_EdgeCodeStart (void);
-extern void R_EdgeCodeEnd (void);
extern void R_RotateBmodel (entity_t *e, view_t *v);
--- a/screen.c
+++ b/screen.c
@@ -321,6 +321,7 @@
fps = host_framecount - lastcnt;
lastcnt = host_framecount;
lastframetime = t;
+fprintf(stderr, "%d\n", fps);
}
n = snprint(s, sizeof(s), "%d", fps);
Draw_String(vid.width - n*8, 0, s);