real bump(real x, real y)
{
real cxs = cosr(x); cxs *= cxs;
real cys = cosr(y); cys *= cys;
return fabsr((cxs*cxs + cys*cys - 2*cxs*cys) / sqrtr(x*x+2*y*y));
}
void cullPoints (s32 n, f32 p[], s32 m, s32 i0, s32 iret[])
{
// compute the centroid of the polygon in cx,cy
s32 i,j;
f32 a,cx,cy,q;
f32 A[8];
s32 avail[8];
if (n==1) {
cx = p[0];
cy = p[1];
}
else if (n==2) {
cx = (f32)(0.5)*(p[0] + p[2]);
cy = (f32)(0.5)*(p[1] + p[3]);
}
else {
a = 0;
cx = 0;
cy = 0;
for (i=0; i<(n-1); i++) {
q = p[i*2]*p[i*2+3] - p[i*2+2]*p[i*2+1];
a += q;
cx += q*(p[i*2]+p[i*2+2]);
cy += q*(p[i*2+1]+p[i*2+3]);
}
q = p[n*2-2]*p[1] - p[0]*p[n*2-1];
a = 1.f/((f32)(3.0)*(a+q));
cx = a*(cx + q*(p[n*2-2]+p[0]));
cy = a*(cy + q*(p[n*2-1]+p[1]));
}
// compute the angle of each point w.r.t. the centroid
for (i=0; i<n; i++) A[i] = atan2r(p[i*2+1]-cy,p[i*2]-cx);
// search for points that have angles closest to A[i0] + i*(2*pi/m).
for (i=0; i<n; i++) avail[i] = 1;
avail[i0] = 0;
iret[0] = i0;
iret++;
for (j=1; j<m; j++) {
f32 maxdiff=1e9,diff;
a = (f32)(j)*(2*N3DPi/m) + A[i0];
if (a > N3DPi) a -= 2*N3DPi;
for (i=0; i<n; i++) {
if (avail[i]) {
diff = fabsr (A[i]-a);
if (diff > N3DPi) diff = 2*N3DPi - diff;
if (diff < maxdiff) {
maxdiff = diff;
*iret = i;
}
}
}
avail[*iret] = 0;
iret++;
}
}
/* n inner lobatto nodes (excluding -1,1) */
static void lobatto_nodes_aux(real *z, int n)
{
int i,j,np=n+1;
for(i=0; i<=n/2-1; ++i) {
real ox, x = cosr( (n-i)*PI/np );
do {
ox = x;
x -= legendre_d1(np,x)/legendre_d2(np,x);
} while(fabsr(x-ox)>-x*EPS);
z[i] = x - legendre_d1(np,x)/legendre_d2(np,x);
}
if(n&1) z[n/2]=0;
for(j=(n+1)/2,i=n/2-1; j<n; ++j,--i) z[j]=-z[i];
}
/* n nodes */
void gauss_nodes(real *z, int n)
{
int i,j;
for(i=0; i<=n/2-1; ++i) {
real ox, x = cosr( (2*n-2*i-1)*(PI/2)/n );
do {
ox = x;
x -= legendre(n,x)/legendre_d1(n,x);
} while(fabsr(x-ox)>-x*EPS);
z[i] = x - legendre(n,x)/legendre_d1(n,x);
}
if(n&1) z[n/2]=0;
for(j=(n+1)/2,i=n/2-1; j<n; ++j,--i) z[j]=-z[i];
}
real sa(unsigned seed)
{
srandom(seed);
real T;
if (initial_temp_method != constant)
T = INFINITY;
else
T = initial_temp;
real step_size_x = init_step_size, step_size_y = init_step_size;
real pos_x = 5, pos_y = 5;
real obj = bump(pos_x, pos_y);
real obj_pen = obj;
int samples_remaining = 5000-1;
real obj_d[5000];
int n_obj_d = 0;
real accepts[5000];
int n_accepts = 0;
int num_trials = 0, num_acceptances = 0;
int initial_trials = 500;
int max_trials = temp_length;
int max_acceptances = 0.6*temp_length;
real alpha = 0.1, omega = 2.1;
real best_obj = obj;
real best_x = pos_x, best_y = pos_y;
int best_time = samples_remaining;
while (samples_remaining > 0)
{
if (best_time - samples_remaining > 500)
{
best_time = samples_remaining;
pos_x = best_x;
pos_y = best_y;
obj = best_obj;
obj_pen = best_obj + penalty_weight * penalty(pos_x, pos_y) / T;
step_size_x = step_size_y = init_step_size;
}
real step_x, step_y;
if (step_method == gaussian)
{
step_x = step_size_x * randn();
step_y = step_size_y * randn();
}
else
{
step_x = step_size_x * (2*randf()-1);
step_y = step_size_y * (2*randf()-1);
}
real new_x = pos_x+step_x, new_y = pos_y+step_y;
real new_pen = penalty_weight * penalty(new_x, new_y);
if (T == INFINITY && new_pen != 0)
continue;
real new_obj = bump(new_x, new_y);
real new_obj_pen = new_obj + new_pen / T;
samples_remaining--;
num_trials++;
if (new_pen == 0)
{
if (new_obj > best_obj)
{
best_obj = new_obj;
best_x = new_x;
best_y = new_y;
best_time = samples_remaining;
}
}
real p;
if (step_method == parks)
{
real step_norm = sqrtr(step_x*step_x + step_y*step_y);
p = exp(- (obj_pen - new_obj_pen) / (T * step_norm));
}
else
{
p = exp(- (obj_pen - new_obj_pen) / T);
}
if (randf() < p)
{
num_acceptances++;
obj_d[n_obj_d++] = new_obj_pen - obj_pen;
pos_x = new_x;
pos_y = new_y;
obj = new_obj;
obj_pen = new_obj_pen;
accepts[n_accepts++] = new_obj_pen;
if (T != INFINITY && step_method == parks)
{
step_size_x = (1-alpha)*step_size_x + alpha*omega*fabsr(step_x);
step_size_y = (1-alpha)*step_size_y + alpha*omega*fabsr(step_y);
}
}
bool reduced_T = false;
if (T == INFINITY)
{
if (num_trials >= initial_trials)
{
if (initial_temp_method == kirkpatrick)
T = T_kirkpatrick(obj_d, n_obj_d);
else if (initial_temp_method == white)
T = std(obj_d, n_obj_d);
else
{
fprintf(stderr, "unknown temp method");
exit(1);
}
reduced_T = true;
}
}
else if (num_trials >= max_trials || num_acceptances >= max_acceptances)
{
if (temp_decay_method == huang)
{
real factor;
if (n_accepts < 2)
factor = 0.5;
else
{
factor = exp(-0.7*T/std(accepts, n_accepts));
if (factor < 0.5)
factor = 0.5;
}
T *= factor;
}
else
T *= temp_decay;
reduced_T = true;
}
if (reduced_T)
{
//printf("%g\n", T);
n_obj_d = 0;
num_trials = 0;
num_acceptances = 0;
n_accepts = 1;
accepts[0] = obj_pen;
}
}
return best_obj;
}
s32 xColBoxPlane(iCollisionObject* o1,iCollisionObject* o2)
{
X_Assert(o1->muClass==iCollisionObject::Col_BBox);
X_Assert(o2->muClass==iCollisionObject::Col_Plane);
cBBox *box = (cBBox*)o1;
cPlane *plane = (cPlane*)o2;
const f32* r = box->mpkWorld->rot;
const f32* n = plane->mkPlane;
vec4_t a,p;
f32 b1,b2,b3;
xMul0_344(a,n,r); // this is safe
xQ4Mul(a,a,box->mkHalfSide);
xQ4Scale(a,2.0f);
b1 = fabsr(a[0]);
b2 = fabsr(a[1]);
b3 = fabsr(a[2]);
// early exit test
f32 depth;
depth = n[PND] + 0.5f*(b1+b2+b3) - xVDotR(n,box->mpkWorld->pos);
if (X_IsFloatNeg(depth)) return 0;
// find number of contacts requested
const s32 maxc = 3;
// find deepest point
xQ4CpyMac(p,box->mpkWorld->pos);
#define P1(i,op) xQ4ScaleN##op(p,&r[i*4],box->mkHalfSide[i]);
#define PALL(i) if (a[i]>0) { P1(i,Sub) } else { P1(i,Add) }
PALL(0);
PALL(1);
PALL(2);
#undef P1
#undef PALL
// the deepest point is the first contact point
xQ4CpyMac(gpkContactHolder[0].pos,p);
xQ4CpyMac(gpkContactHolder[0].norm,n);
gpkContactHolder[0].depth = depth; //
s32 ret = 1; // ret is number of contact points found so far
// get the second and third contact points by starting from `p' and going
// along the two sides with the smallest projected length.
#define P1(i,j,op) \
xQ4CpyMac(gpkContactHolder[i].pos,p);\
xQ4ScaleN##op(gpkContactHolder[i].pos,&r[j*4],2*box->mkHalfSide[j]);
#define PALL(ctact,side,sideinc) \
depth -= b ## sideinc; \
if (depth < 0) goto done; \
if (a[sideinc-1] > 0) { P1(ctact,side,Add) } else { P1(ctact,side,Sub) } \
gpkContactHolder[ctact].depth = depth; \
ret++;
xQ4CpyMac(gpkContactHolder[1].norm,n);
xQ4CpyMac(gpkContactHolder[2].norm,n);
if (b1 < b2)
{
if (b3 < b1) goto use_side_3;
else
{
PALL(1,0,1); // use side 1
if (b2 < b3) goto contact2_2;
else goto contact2_3;
}
}
else
{
if (b3 < b2)
{
use_side_3: // use side 3
PALL(1,2,3);
if (b1 < b2) goto contact2_1;
else goto contact2_2;
}
else
{
PALL(1,1,2); // use side 2
if (b1 < b3) goto contact2_1;
else goto contact2_3;
}
}
contact2_1: PALL(2,0,1); goto done;
contact2_2: PALL(2,1,2); goto done;
contact2_3: PALL(2,2,3); goto done;
#undef P1
#undef PALL
done:
return ret;
}
s32 cBoxCollide(
const f32* p1,const f32* r1,const f32* sz1,
const f32* p2,const f32* r2,const f32* sz2,
cContact* c,f32* normal)
{
const f32 fudge_factor = 1.05f;
vec4_t p,pp1,pp2,nC; //secured_v
mat34_t r,q;
f32 s2,s;
s32 code;
const f32* nR;
//mat34_t r1,r2;
s32 invertnormal,i,j;
nC[3] = 0;
//xMTrans_34(r1,r1t);
//xMTrans_34(r2,r2t);
xQ4Sub(p,p2,p1);
xMul0_344(pp1,p,r1);
xMul0_344(pp2,p,r2);
// r(i,j) = row(r1,i). row(r2,j)
xMul0_34(r,r1,r2);
q[0] = fabsr(r[0]); q[1] = fabsr(r[1]); q[2] = fabsr(r[2]); q[3] = 0;
q[4] = fabsr(r[4]); q[5] = fabsr(r[5]); q[6] = fabsr(r[6]); q[7] = 0;
q[8] = fabsr(r[8]); q[9] = fabsr(r[9]); q[10] = fabsr(r[10]); q[11] = 0;
#define XSATEST(e1,e2,n,cc) \
s2 = fabsr(e1) - (e2); \
if (s2 > 0 ) return 0; \
if (s2 > s) \
{ \
s = s2; nR = (n); \
invertnormal = ((e1) < 0); \
code = (cc); \
}
s =-N3DInfinity;
invertnormal = 0;
code = 0;
// note that nC has not been used yet
// we use nC for sz1 rough calculation
// here
// separating axis = u1,u2,u3
xMul0_344(nC,sz2,q); // 3 multiplication with SSE
xQ4Add(nC,nC,sz1); // one addition with SSE
//TODO: HERE we need to take a descision, we can
//well over escape 3 multiplication for 1, in case
//we use xVDot for each test, but for average case
//I guess this will work better
XSATEST(pp1[0],nC[0],r1+0,1);
XSATEST(pp1[1],nC[1],r1+4,2);
XSATEST(pp1[2],nC[2],r1+8,3);
xMul1_344(nC,sz1,q); // 3 multiplication with SSE
xQ4Add(nC,nC,sz2); // one addition with SSE
// separating axis = v1,v2,v3
XSATEST(pp2[0],nC[0],r2+0,4);
XSATEST(pp2[1],nC[1],r2+4,5);
XSATEST(pp2[2],nC[2],r2+8,6);
// note that nC has not been used yet
#undef XSATEST
#define XSATEST(expr1,e2,n1,n2,n3,cc) \
s2 = fabsr(expr1) - (e2); \
if (s2 > 0) return 0; \
l = sqrtr((n1)*(n1) + (n2)*(n2) + (n3)*(n3)); \
if (l > 0) { \
s2 /= l; \
if (s2*fudge_factor > s) { \
s = s2; \
nR = 0; \
nC[0] = (n1)/l; nC[1] = (n2)/l; nC[2] = (n3)/l; \
invertnormal = ((expr1) < 0); \
code = (cc); \
} \
}
{
// we need some temp vectors here
vec4_t tmp1,tmp2;
f32 l;
// with SSE these are effectively
xQ4ScaleS(tmp1,(f32*)&r[4],pp1[2]); // one multiplication
xQ4ScaleNSub(tmp1,(f32*)&r[8],pp1[1]); // one addition and one multiplication
xQ4ScaleS(tmp2,(f32*)&q[8],sz1[1]); // one multiplication
xQ4ScaleNAdd(tmp2,(f32*)&q[4],sz1[2]); // one addition and one multiplication
// separating axis = u1 x (v1,v2,v3)
XSATEST(tmp1[0],(tmp2[0]+sz2[1]*q[2]+sz2[2]*q[1]),0,-r[8],r[4],7);
XSATEST(tmp1[1],(tmp2[1]+sz2[0]*q[2]+sz2[2]*q[0]),0,-r[9],r[5],8);
XSATEST(tmp1[2],(tmp2[2]+sz2[0]*q[1]+sz2[1]*q[0]),0,-r[10],r[6],9);
xQ4ScaleS(tmp1,(f32*)&r[8],pp1[0]); // one multiplication
xQ4ScaleNSub(tmp1,(f32*)&r[0],pp1[2]); // one addition and one multiplication
xQ4ScaleS(tmp2,(f32*)&q[8],sz1[0]); // one multiplication
xQ4ScaleNAdd(tmp2,(f32*)&q[0],sz1[2]); // one addition and one multiplication
// separating axis = u2 x (v1,v2,v3)
XSATEST(tmp1[0],(tmp2[0]+sz2[1]*q[6]+sz2[2]*q[5]),r[8],0,-r[0],10);
XSATEST(tmp1[1],(tmp2[1]+sz2[0]*q[6]+sz2[2]*q[4]),r[9],0,-r[1],11);
XSATEST(tmp1[2],(tmp2[2]+sz2[0]*q[5]+sz2[1]*q[4]),r[10],0,-r[2],12);
xQ4ScaleS(tmp1,(f32*)&r[0],pp1[1]); // one multiplication
xQ4ScaleNSub(tmp1,(f32*)&r[4],pp1[0]); // one addition and one multiplication
xQ4ScaleS(tmp2,(f32*)&q[4],sz1[0]); // one multiplication
xQ4ScaleNAdd(tmp2,(f32*)&q[0],sz1[1]); // one addition and one multiplication
// separating axis = u3 x (v1,v2,v3)
XSATEST(tmp1[0],(tmp2[0]+sz2[1]*q[10]+sz2[2]*q[9]),-r[4],-r[0],0,13);
XSATEST(tmp1[1],(tmp2[1]+sz2[0]*q[10]+sz2[2]*q[8]),-r[5],-r[1],0,14);
XSATEST(tmp1[2],(tmp2[2]+sz2[0]*q[9]+sz2[1]*q[8]),-r[6],-r[2],0,15);
}
#undef XSATEST
if(!code) return 0;
if(nR)
{
xQ4CpyMac(normal,nR);
}
else
{
xMul1_344(normal,nC,r1);
}
if(invertnormal)
{
xQ4Scale(normal,-1.0f);
// normal[0] = -normal[0];
// normal[1] = -normal[1];
// normal[2] = -normal[2];
}
s = -s;
if (code > 6)
{
// an edge from box 1 touches an edge from box 2.
// find a point pa on the intersecting edge of box 1
vec4_t pa,pb,ua,ub;
f32 alpha,beta;
xQ4CpyMac(pa,p1);
xQ4CpyMac(pb,p2);
xMul0_344(ua,normal,r1);
for (j=0; j<3; j++)
{
// add sign
if( ISFLOATNEGETIVE(ua[j]) )
{
xQ4ScaleNSub(pa,&r1[j*4],sz1[j]);
}
else
{
xQ4ScaleNAdd(pa,&r1[j*4],sz1[j]);
}
}
// find a point pb on the intersecting edge of box 2
xMul0_344(ua,normal,r2);
for (j=0; j<3; j++)
{
// add sign
if( ISFLOATNEGETIVE(ua[j]) )
{
xQ4ScaleNSub(pb,&r2[j*4],sz2[j]);
}
else
{
xQ4ScaleNAdd(pb,&r2[j*4],sz2[j]);
}
}
// highly doubted
xQ4CpyMac(ua,&r1[(((code)-7)/3)*4]);
xQ4CpyMac(ub,&r2[(((code)-7)%3)*4]);
{
// line closest approach (pa,ua,pb,ub,&alpha,&beta);
vec4_t p;// secured_v
f32 uaub,q1,q2,d;
xQ4Sub(p,pb,pa);
xVDot(&uaub,ua,ub);
xVDot(&q1,ua,p);
xVDot(&q2,ub,p);
d = 1-uaub*uaub;
if (d <= 0.0001f)
{
alpha = 0;
beta = 0;
}
else
{
d = 1/d;
alpha = (q1 - uaub*q2)*d;
beta = (uaub*q1 - q2)*d;
}
}
xQ4ScaleNAdd(pa,ua,alpha);
xQ4ScaleNAdd(pb,ub,beta);
xQ4Add(c->pos,pa,pb);
xQ4Scale(c->pos,0.5f);
c->depth = s;
return 1;
}
// okay, we have sz1 face-something intersection (because the separating
// axis is perpendicular to sz1 face). define face 'sz1' to be the reference
// face (i.e. the normal vector is perpendicular to this) and face 'sz2' to be
// the incident face (the closest face of the other box).
{
vec4_t center; // secured_v
const f32 *ra,*rb;
const f32 *pa,*pb,*sa,*sb;
s32 lanr,a1,a2;
f32 quad[8]; // 2D coordinate of incident face (x,y pairs)
// nr = normal vector of reference face dotted with axes of incident box.
// anr = absolute values of nr.
vec4_t normal2,nr,anr; // secured_v
if (code <= 3)
{
ra = r1;
rb = r2;
pa = (f32*)p1;
pb = (f32*)p2;
sa = (f32*)sz1;
sb = (f32*)sz2;
xQ4CpyMac(normal2,normal);
}
else
{
ra = r2;
rb = r1;
pa = (f32*)p2;
pb = (f32*)p1;
sa = (f32*)sz2;
sb = (f32*)sz1;
normal2[0] = -normal[0];
normal2[1] = -normal[1];
normal2[2] = -normal[2];
normal2[3] = 0;
}
xMul0_344(nr,normal2,rb);
anr[0] = fabsr (nr[0]);
anr[1] = fabsr (nr[1]);
anr[2] = fabsr (nr[2]);
anr[3] = 0; // secured_v
// find the largest compontent of anr: this corresponds to the normal
// for the indident face. the other axis numbers of the indicent face
// are stored in a1,a2.
if (anr[1] > anr[0])
{
if (anr[1] > anr[2])
{
a1 = 0;
lanr = 1;
a2 = 2;
}
else
{
a1 = 0;
a2 = 1;
lanr = 2;
}
}
else
{
if (anr[0] > anr[2])
{
lanr = 0;
a1 = 1;
a2 = 2;
}
else
{
a1 = 0;
a2 = 1;
lanr = 2;
}
}
// compute center point of incident face, in reference-face coordinates
xQ4Sub(center,pb,pa);
if (nr[lanr] < 0)
{
xQ4ScaleNAdd(center,&rb[4*lanr],sb[lanr]);
}
else
{
xQ4ScaleNSub(center,&rb[4*lanr],sb[lanr]);
}
// find the normal and non-normal axis numbers of the reference box
{
vec4_t point[8]; // penetrating contact points
vec4_t m_;
s32 codeN,code1,code2;
f32 c1,c2;
f32 rect[2];
f32 ret[16];
s32 inr;
s32 cnum = 0; // number of penetrating contact points found
f32 dep[8]; // depths for those points
f32 det1;
if (code <= 3)
codeN = code-1;
else
codeN = code-4;
if (codeN==0)
{
code1 = 1;
code2 = 2;
}
else if (codeN==1)
{
code1 = 0;
code2 = 2;
}
else
{
code1 = 0;
code2 = 1;
}
// find the four corners of the incident face, in reference-face coordinates
{
xVDot(&c1,center,&ra[code1*4]);
xVDot(&c2,center,&ra[code1*4]);
}
// r(i,j) = col(r1,i). col(r2,j)
if(code <= 3)
{
m_[0] = r[code1*4+a1];
m_[1] = r[code1*4+a2];
m_[2] = r[code2*4+a1];
m_[3] = r[code2*4+a2];
}
else
{
m_[0] = r[a1*4+code1];
m_[1] = r[a2*4+code1];
m_[2] = r[a1*4+code2];
m_[3] = r[a2*4+code2];
}
{
f32 k1 = m_[0]*sb[a1];
f32 k2 = m_[2]*sb[a1];
f32 k3 = m_[1]*sb[a2];
f32 k4 = m_[3]*sb[a2];
quad[0] = c1 - k1 - k3;
quad[1] = c2 - k2 - k4;
quad[2] = c1 - k1 + k3;
quad[3] = c2 - k2 + k4;
quad[4] = c1 + k1 + k3;
quad[5] = c2 + k2 + k4;
quad[6] = c1 + k1 - k3;
quad[7] = c2 + k2 - k4;
}
// find the size of the reference face
rect[0] = sa[code1];
rect[1] = sa[code2];
// intersect the incident and reference faces
{
// s32 n = intersectRectQuad (rect,quad,ret);
//(rect[2],p[8],ret[16])
// q (and r) contain nq (and nr) coordinate points for the current (and
// chopped) polygons
s32 nq=4;
f32 buffer[16];
f32 *q = quad;
f32 *r = ret;
s32 dir,sign;
for (dir=0; dir <= 1; dir++)
{
// direction notation: xy[0] = x axis, xy[1] = y axis
for (sign=-1; sign <= 1; sign += 2)
{
// chop q along the line xy[dir] = sign*rect[dir]
f32 *pq = q;
f32 *pr = r;
inr = 0;
for (i=nq; i > 0; i--)
{
f32 *nextq;
// go through all points in q and all lines between adjacent points
if (sign*pq[dir] < rect[dir])
{
// this point is inside the chopping line
pr[0] = pq[0];
pr[1] = pq[1];
pr += 2;
inr++;
if (inr & 8)
{
q = r;
goto done;
}
}
nextq = (i > 1) ? pq+2 : q;
if ((sign*pq[dir] < rect[dir]) ^ (sign*nextq[dir] < rect[dir]))
{
// this line crosses the chopping line
pr[1-dir] = pq[1-dir] + (nextq[1-dir]-pq[1-dir]) /
(nextq[dir]-pq[dir]) * (sign*rect[dir]-pq[dir]);
pr[dir] = sign*rect[dir];
pr += 2;
inr++;
if (inr & 8)
{
q = r;
goto done;
}
}
pq += 2;
}
q = r;
r = (q==ret) ? buffer : ret;
nq = inr;
}
}
done:
if (q != ret)
copyDwords(ret,q,inr*2);
}
if (inr < 1) return 0; // this should never happen
// convert the intersection points into reference-face coordinates,
// and compute the contact position and depth for each point. only keep
// those points that have sz1 positive (penetrating) depth. delete points in
// the 'ret' array as necessary so that 'point' and 'ret' correspond.
det1 = 1/(m_[0]*m_[3] - m_[1]*m_[2]);
xQ4Scale(m_,det1);
for (j=0; j < inr; j++)
{
f32 k1 = m_[3]*(ret[j*2]-c1) - m_[1]*(ret[j*2+1]-c2);
f32 k2 = -m_[2]*(ret[j*2]-c1) + m_[0]*(ret[j*2+1]-c2);
xQ4CpyMac(point[cnum],center);
xQ4ScaleNAdd(point[cnum],&rb[a1*4],k1);
xQ4ScaleNAdd(point[cnum],&rb[a2*4],k2);
dep[cnum] = sa[codeN] - xVDotR(normal2,point[cnum]);
if(dep[cnum] >= 0)
{
ret[cnum*2] = ret[j*2];
ret[cnum*2+1] = ret[j*2+1];
cnum++;
}
}
if (cnum < 1)
return 0; // this should never happen
// we can't generate more contacts than we actually have
if (cnum <= N3DMaxContact)
{
// we have less contacts than we need, so we use them all
for (j=0; j < cnum; j++)
{
cContact *con = c+j;
xQ4Add(con->pos,point[j],pa);
con->depth = dep[j]; // this seems ok
}
}
else
{
// we have more contacts than are wanted, some of them must be culled.
// find the deepest point, it is always the first contact.
s32 i1 = 0;
f32 maxdepth = dep[0];
s32 iret[8];
for (i=1; i<cnum; i++)
{
if (dep[i] > maxdepth)
{
maxdepth = dep[i];
i1 = i;
}
}
{
// compute the centroid of the polygon in cx,cy
f32 a,cx,cy,q;
f32 ang[8];
s32 avail[8];
if (cnum==1)
{
cx = ret[0];
cy = ret[1];
}
else if (cnum==2)
{
cx = 0.5f*(ret[0] + ret[2]);
cy = 0.5f*(ret[1] + ret[3]);
}
else
{
a = 0;
cx = 0;
cy = 0;
for (i=0;i<(cnum-1); i++)
{
q = ret[i*2]*ret[i*2+3] - ret[i*2+2]*ret[i*2+1];
a += q;
cx += q*(ret[i*2]+ret[i*2+2]);
cy += q*(ret[i*2+1]+ret[i*2+3]);
}
q = ret[cnum*2-2]*ret[1] - ret[0]*ret[cnum*2-1];
a = 1/((3*(a+q)));
cx = a*(cx + q*(ret[cnum*2-2]+ret[0]));
cy = a*(cy + q*(ret[cnum*2-1]+ret[1]));
}
// compute the angle of each point w.r.t. the centroid
for (i=0; i<cnum; i++)
ang[i] = atan2r(p[i*2+1]-cy,p[i*2]-cx);
// search for points that have angles closest to sz1[i1] + i*(2*pi/m).
for (i=0; i<cnum; i++) avail[i] = 1;
avail[i1] = 0;
iret[0] = i1;
{
s32* piret = iret;
piret++;
for (j=1; j<N3DMaxContact; j++)
{
f32 maxdiff=1e9,diff;
a = (f32)j*(2*N3DPi/N3DMaxContact) + ang[i1];
if(a > N3DPi) a -= 2*N3DPi;
for (i=0; i<cnum; i++)
{
if (avail[i])
{
diff = fabsr (ang[i]-a);
if (diff > N3DPi) diff = 2*N3DPi - diff;
if (diff < maxdiff)
{
maxdiff = diff;
*piret = i;
}
}
}
avail[*piret] = 0;
piret++;
}
}
}
for (j=0; j < N3DMaxContact; j++)
{
cContact *con = c+j;
xQ4Add(con->pos,point[iret[j]],pa);
con->depth = dep[iret[j]]; // this seems ok
}
cnum = N3DMaxContact;
}
return cnum;
}
}
// return_code = code;
}
int main(int argc, char** argv)
{
int iter_max = 1000;
const real tol = 1.0e-5;
memset(A, 0, NY * NX * sizeof(real));
// set rhs
for (int iy = 1; iy < NY-1; iy++)
{
for( int ix = 1; ix < NX-1; ix++ )
{
const real x = -1.0 + (2.0*ix/(NX-1));
const real y = -1.0 + (2.0*iy/(NY-1));
rhs[iy][ix] = expr(-10.0*(x*x + y*y));
}
}
printf("Jacobi relaxation Calculation: %d x %d mesh\n", NY, NX);
StartTimer();
int iter = 0;
real error = 1.0;
#pragma acc data copy(A) copyin(rhs) create(Anew)
while ( error > tol && iter < iter_max )
{
error = 0.0;
#pragma acc kernels
for (int iy = 1; iy < NY-1; iy++)
{
for( int ix = 1; ix < NX-1; ix++ )
{
Anew[iy][ix] = -0.25 * (rhs[iy][ix] - ( A[iy][ix+1] + A[iy][ix-1]
+ A[iy-1][ix] + A[iy+1][ix] ));
error = fmaxr( error, fabsr(Anew[iy][ix]-A[iy][ix]));
}
}
#pragma acc kernels
for (int iy = 1; iy < NY-1; iy++)
{
for( int ix = 1; ix < NX-1; ix++ )
{
A[iy][ix] = Anew[iy][ix];
}
}
//Periodic boundary conditions
#pragma acc kernels
for( int ix = 1; ix < NX-1; ix++ )
{
A[0][ix] = A[(NY-2)][ix];
A[(NY-1)][ix] = A[1][ix];
}
#pragma acc kernels
for (int iy = 1; iy < NY-1; iy++)
{
A[iy][0] = A[iy][(NX-2)];
A[iy][(NX-1)] = A[iy][1];
}
if((iter % 100) == 0) printf("%5d, %0.6f\n", iter, error);
iter++;
}
double runtime = GetTimer();
printf( "%dx%d: 1 GPU: %8.4f s\n", NY,NX, runtime/ 1000.0 );
return 0;
}
int main(int argc, char** argv)
{
int iter_max = 1000;
const real tol = 1.0e-5;
int rank = 0;
int size = 1;
//Initialize MPI and determine rank and size
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
MPI_Comm_size(MPI_COMM_WORLD, &size);
if ( size > MAX_MPI_SIZE )
{
if ( 0 == rank )
{
fprintf(stderr,"ERROR: Only up to %d MPI ranks are supported.\n",MAX_MPI_SIZE);
}
return -1;
}
dim2 size2d = size_to_2Dsize(size);
int sizex = size2d.x;
int sizey = size2d.y;
assert(sizex*sizey == size);
int rankx = rank%sizex;
int ranky = rank/sizex;
memset(A, 0, NY * NX * sizeof(real));
memset(Aref, 0, NY * NX * sizeof(real));
// set rhs
for (int iy = 1; iy < NY-1; iy++)
{
for( int ix = 1; ix < NX-1; ix++ )
{
const real x = -1.0 + (2.0*ix/(NX-1));
const real y = -1.0 + (2.0*iy/(NY-1));
rhs[iy][ix] = expr(-10.0*(x*x + y*y));
}
}
#if _OPENACC
acc_device_t device_type = acc_get_device_type();
if ( acc_device_nvidia == device_type )
{
int ngpus=acc_get_num_devices(acc_device_nvidia);
int devicenum=rank%ngpus;
acc_set_device_num(devicenum,acc_device_nvidia);
}
// Call acc_init after acc_set_device_num to avoid multiple contexts on device 0 in multi GPU systems
acc_init(device_type);
#endif /*_OPENACC*/
// Ensure correctness if NX%sizex != 0
int chunk_sizex = ceil( (1.0*NX)/sizex );
int ix_start = rankx * chunk_sizex;
int ix_end = ix_start + chunk_sizex;
// Do not process boundaries
ix_start = max( ix_start, 1 );
ix_end = min( ix_end, NX - 1 );
// Ensure correctness if NY%sizey != 0
int chunk_sizey = ceil( (1.0*NY)/sizey );
int iy_start = ranky * chunk_sizey;
int iy_end = iy_start + chunk_sizey;
// Do not process boundaries
iy_start = max( iy_start, 1 );
iy_end = min( iy_end, NY - 1 );
if ( rank == 0) printf("Jacobi relaxation Calculation: %d x %d mesh\n", NY, NX);
if ( rank == 0) printf("Calculate reference solution and time serial execution.\n");
StartTimer();
poisson2d_serial( rank, iter_max, tol );
double runtime_serial = GetTimer();
//Wait for all processes to ensure correct timing of the parallel version
MPI_Barrier( MPI_COMM_WORLD );
if ( rank == 0) printf("Parallel execution.\n");
StartTimer();
int iter = 0;
real error = 1.0;
#pragma acc data copy(A) copyin(rhs) create(Anew,to_left,from_left,to_right,from_right)
while ( error > tol && iter < iter_max )
{
error = 0.0;
#pragma acc kernels
for (int iy = iy_start; iy < iy_end; iy++)
{
for( int ix = ix_start; ix < ix_end; ix++ )
{
Anew[iy][ix] = -0.25 * (rhs[iy][ix] - ( A[iy][ix+1] + A[iy][ix-1]
+ A[iy-1][ix] + A[iy+1][ix] ));
error = fmaxr( error, fabsr(Anew[iy][ix]-A[iy][ix]));
}
}
real globalerror = 0.0;
MPI_Allreduce( &error, &globalerror, 1, MPI_REAL_TYPE, MPI_MAX, MPI_COMM_WORLD );
error = globalerror;
#pragma acc kernels
for (int iy = iy_start; iy < iy_end; iy++)
{
for( int ix = ix_start; ix < ix_end; ix++ )
{
A[iy][ix] = Anew[iy][ix];
}
}
//Periodic boundary conditions
int topy = (ranky == 0) ? (sizey-1) : ranky-1;
int bottomy = (ranky == (sizey-1)) ? 0 : ranky+1;
int top = topy * sizex + rankx;
int bottom = bottomy * sizex + rankx;
#pragma acc host_data use_device( A )
{
//1. Sent row iy_start (first modified row) to top receive lower boundary (iy_end) from bottom
MPI_Sendrecv( &A[iy_start][ix_start], (ix_end-ix_start), MPI_REAL_TYPE, top , 0, &A[iy_end][ix_start], (ix_end-ix_start), MPI_REAL_TYPE, bottom, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE );
//2. Sent row (iy_end-1) (last modified row) to bottom receive upper boundary (iy_start-1) from top
MPI_Sendrecv( &A[(iy_end-1)][ix_start], (ix_end-ix_start), MPI_REAL_TYPE, bottom, 0, &A[(iy_start-1)][ix_start], (ix_end-ix_start), MPI_REAL_TYPE, top , 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE );
}
int leftx = (rankx == 0) ? (sizex-1) : rankx-1;
int rightx = (rankx == (sizex-1)) ? 0 : rankx+1;
int left = ranky * sizex + leftx;
int right = ranky * sizex + rightx;
#pragma acc kernels
for( int iy = iy_start; iy < iy_end; iy++ )
{
to_left[iy] = A[iy][ix_start];
to_right[iy] = A[iy][ix_end-1];
}
#pragma acc host_data use_device( to_left, from_left, to_right, from_right )
{
//1. Sent to_left starting from first modified row (iy_start) to last modified row to left and receive the same rows into from_right from right
MPI_Sendrecv( to_left+iy_start, (iy_end-iy_start), MPI_REAL_TYPE, left , 0, from_right+iy_start, (iy_end-iy_start), MPI_REAL_TYPE, right, 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE );
//2. Sent to_right starting from first modified row (iy_start) to last modified row to left and receive the same rows into from_left from left
MPI_Sendrecv( to_right+iy_start, (iy_end-iy_start), MPI_REAL_TYPE, right , 0, from_left+iy_start, (iy_end-iy_start), MPI_REAL_TYPE, left , 0, MPI_COMM_WORLD, MPI_STATUS_IGNORE );
}
#pragma acc kernels
for( int iy = iy_start; iy < iy_end; iy++ )
{
A[iy][ix_start-1] = from_left[iy];
A[iy][ix_end] = from_right[iy];
}
if(rank == 0 && (iter % 100) == 0) printf("%5d, %0.6f\n", iter, error);
iter++;
}
MPI_Barrier( MPI_COMM_WORLD );
double runtime = GetTimer();
if (check_results( rank, ix_start, ix_end, iy_start, iy_end, tol ) && rank == 0)
{
printf( "Num GPUs: %d with a (%d,%d) layout.\n", size, sizey,sizex );
printf( "%dx%d: 1 GPU: %8.4f s, %d GPUs: %8.4f s, speedup: %8.2f, efficiency: %8.2f%\n", NY,NX, runtime_serial/ 1000.0, size, runtime/ 1000.0, runtime_serial/runtime, runtime_serial/(size*runtime)*100 );
}
MPI_Finalize();
return 0;
}
s32 dBoxBox (const f32* p1, const f32* R1, const f32* side1,
const f32* p2, const f32* R2, const f32* side2,
cContact* contact,f32* normal)
{
s32 maxc = 4;
const f32 fudge_factor = (f32)(1.05);
vec4_t p,pp,normalC;
const f32 *normalR = 0;
f32 depth;
f32 A[3],B[3],R11,R12,R13,R21,R22,R23,R31,R32,R33,
Q11,Q12,Q13,Q21,Q22,Q23,Q31,Q32,Q33,s,s2,l;
s32 i,j,invert_normal,code;
const f32 *Ra,*Rb,*pa,*pb,*Sa,*Sb;
vec4_t normal2,nr,anr;vec4_t center;
s32 lanr,a1,a2;
s32 codeN,code1,code2;
// get vector from centers of box 1 to box 2, relative to box 1
p[0] = p2[0] - p1[0];
p[1] = p2[1] - p1[1];
p[2] = p2[2] - p1[2];
p[3] = 0;
xMul0_344 (pp,p,R1); // get pp = p relative to body 1
// get side lengths / 2
A[0] = side1[0]*(f32)(0.5);
A[1] = side1[1]*(f32)(0.5);
A[2] = side1[2]*(f32)(0.5);
B[0] = side2[0]*(f32)(0.5);
B[1] = side2[1]*(f32)(0.5);
B[2] = side2[2]*(f32)(0.5);
// Rij is R1'*R2, i.e. the relative rotation between R1 and R2
R11 = dDOT(R1+0,R2+0); R12 = dDOT(R1+0,R2+4); R13 = dDOT(R1+0,R2+8);
R21 = dDOT(R1+4,R2+0); R22 = dDOT(R1+4,R2+4); R23 = dDOT(R1+4,R2+8);
R31 = dDOT(R1+8,R2+0); R32 = dDOT(R1+8,R2+4); R33 = dDOT(R1+8,R2+8);
Q11 = fabsr(R11); Q12 = fabsr(R12); Q13 = fabsr(R13);
Q21 = fabsr(R21); Q22 = fabsr(R22); Q23 = fabsr(R23);
Q31 = fabsr(R31); Q32 = fabsr(R32); Q33 = fabsr(R33);
// for all 15 possible separating axes:
// * see if the axis separates the boxes. if so, return 0.
// * find the depth of the penetration along the separating axis (s2)
// * if this is the largest depth so far, record it.
// the normal vector will be set to the separating axis with the smallest
// depth. note: normalR is set to point to a column of R1 or R2 if that is
// the smallest depth normal so far. otherwise normalR is 0 and normalC is
// set to a vector relative to body 1. invert_normal is 1 if the sign of
// the normal should be flipped.
#define TST(expr_1,expr2,norm,cc) \
expr1 = expr_1;\
s2 = fabsr(expr1) - (expr2); \
if (s2 > 0) return 0; \
if (s2 > s) { \
s = s2; \
normalR = norm; \
invert_normal = ((expr1) < 0); \
code = (cc); \
}
f32 expr1;
s = -N3DInfinity;
invert_normal = 0;
code = 0;
// separating axis = u1,u2,u3
TST (pp[0],(A[0] + B[0]*Q11 + B[1]*Q12 + B[2]*Q13),R1+0,1);
TST (pp[1],(A[1] + B[0]*Q21 + B[1]*Q22 + B[2]*Q23),R1+4,2);
TST (pp[2],(A[2] + B[0]*Q31 + B[1]*Q32 + B[2]*Q33),R1+8,3);
// separating axis = v1,v2,v3
TST (dDOT(R2+0,p),(A[0]*Q11 + A[1]*Q21 + A[2]*Q31 + B[0]),R2+0,4);
TST (dDOT(R2+4,p),(A[0]*Q12 + A[1]*Q22 + A[2]*Q32 + B[1]),R2+4,5);
TST (dDOT(R2+8,p),(A[0]*Q13 + A[1]*Q23 + A[2]*Q33 + B[2]),R2+8,6);
// note: cross product axes need to be scaled when s is computed.
// normal (n1,n2,n3) is relative to box 1.
#undef TST
#define TST(expr1,expr2,n1,n2,n3,cc) \
s2 = fabsr(expr1) - (expr2); \
if (s2 > 0) return 0; \
l = fsqrtr((n1)*(n1) + (n2)*(n2) + (n3)*(n3)); \
if (l > 0) { \
s2 /= l; \
if (s2*fudge_factor > s) { \
s = s2; \
normalR = 0; \
normalC[0] = (n1)/l; normalC[1] = (n2)/l; normalC[2] = (n3)/l; \
invert_normal = ((expr1) < 0); \
code = (cc); \
} \
}
// separating axis = u1 x (v1,v2,v3)
TST(pp[2]*R21-pp[1]*R31,(A[1]*Q31+A[2]*Q21+B[1]*Q13+B[2]*Q12),0,-R31,R21,7);
TST(pp[2]*R22-pp[1]*R32,(A[1]*Q32+A[2]*Q22+B[0]*Q13+B[2]*Q11),0,-R32,R22,8);
TST(pp[2]*R23-pp[1]*R33,(A[1]*Q33+A[2]*Q23+B[0]*Q12+B[1]*Q11),0,-R33,R23,9);
// separating axis = u2 x (v1,v2,v3)
TST(pp[0]*R31-pp[2]*R11,(A[0]*Q31+A[2]*Q11+B[1]*Q23+B[2]*Q22),R31,0,-R11,10);
TST(pp[0]*R32-pp[2]*R12,(A[0]*Q32+A[2]*Q12+B[0]*Q23+B[2]*Q21),R32,0,-R12,11);
TST(pp[0]*R33-pp[2]*R13,(A[0]*Q33+A[2]*Q13+B[0]*Q22+B[1]*Q21),R33,0,-R13,12);
// separating axis = u3 x (v1,v2,v3)
TST(pp[1]*R11-pp[0]*R21,(A[0]*Q21+A[1]*Q11+B[1]*Q33+B[2]*Q32),-R21,R11,0,13);
TST(pp[1]*R12-pp[0]*R22,(A[0]*Q22+A[1]*Q12+B[0]*Q33+B[2]*Q31),-R22,R12,0,14);
TST(pp[1]*R13-pp[0]*R23,(A[0]*Q23+A[1]*Q13+B[0]*Q32+B[1]*Q31),-R23,R13,0,15);
#undef TST
if (!code) return 0;
// if we get to this point, the boxes interpenetrate. compute the normal
// in global coordinates.
if (normalR) {
normal[0] = normalR[0];
normal[1] = normalR[1];
normal[2] = normalR[2];
}
else {
xMul1_344 (normal,normalC,R1);
}
if (invert_normal) {
normal[0] = -normal[0];
normal[1] = -normal[1];
normal[2] = -normal[2];
}
depth = -s;
// compute contact point(s)
if (code > 6) {
// an edge from box 1 touches an edge from box 2.
// find a point pa on the intersecting edge of box 1
vec4_t pa,pb;
f32 sign;f32 alpha,beta;
vec4_t ua,ub;
for (i=0; i<3; i++) pa[i] = p1[i];
for (j=0; j<3; j++) {
sign = (dDOT(normal,(R1+j*4)) > 0) ? (f32)(1.0) : (f32)(-1.0);
xQ4ScaleNAdd(pa,&R1[j*4],sign * A[j]);
}
// find a point pb on the intersecting edge of box 2
for (i=0; i<3; i++) pb[i] = p2[i];
for (j=0; j<3; j++) {
sign = (dDOT(normal,(R2+j*4)) > 0) ? (f32)(-1.0) : (f32)(1.0);
xQ4ScaleNAdd(pb,&R2[j*4],sign * B[j]);
}
for (i=0; i<3; i++) ua[i] = R1[(((code)-7)/3)*4 + i];
for (i=0; i<3; i++) ub[i] = R2[(((code)-7)%3)*4 + i];
dLineClosestApproach (pa,ua,pb,ub,&alpha,&beta);
for (i=0; i<3; i++) pa[i] += ua[i]*alpha;
for (i=0; i<3; i++) pb[i] += ub[i]*beta;
for (i=0; i<3; i++) contact[0].pos[i] = (f32)(0.5)*(pa[i]+pb[i]);
contact[0].depth = depth;
return 1;
}
// okay, we have a face-something intersection (because the separating
// axis is perpendicular to a face). define face 'a' to be the reference
// face (i.e. the normal vector is perpendicular to this) and face 'b' to be
// the incident face (the closest face of the other box).
if (code <= 3) {
Ra = R1;
Rb = R2;
pa = p1;
pb = p2;
Sa = A;
Sb = B;
}
else {
Ra = R2;
Rb = R1;
pa = p2;
pb = p1;
Sa = B;
Sb = A;
}
// nr = normal vector of reference face dotted with axes of incident box.
// anr = absolute values of nr.
if (code <= 3) {
normal2[0] = normal[0];
normal2[1] = normal[1];
normal2[2] = normal[2];
}
else {
normal2[0] = -normal[0];
normal2[1] = -normal[1];
normal2[2] = -normal[2];
}
xMul0_344 (nr,normal2,Rb);
anr[0] = fabsr (nr[0]);
anr[1] = fabsr (nr[1]);
anr[2] = fabsr (nr[2]);
// find the largest compontent of anr: this corresponds to the normal
// for the indident face. the other axis numbers of the indicent face
// are stored in a1,a2.
if (anr[1] > anr[0]) {
if (anr[1] > anr[2]) {
a1 = 0;
lanr = 1;
a2 = 2;
}
else {
a1 = 0;
a2 = 1;
lanr = 2;
}
}
else {
if (anr[0] > anr[2]) {
lanr = 0;
a1 = 1;
a2 = 2;
}
else {
a1 = 0;
a2 = 1;
lanr = 2;
}
}
// compute center point of incident face, in reference-face coordinates
if (nr[lanr] < 0) {
for (i=0; i<3; i++) center[i] = pb[i] - pa[i] + Sb[lanr] * Rb[i+lanr*4];
}
else {
for (i=0; i<3; i++) center[i] = pb[i] - pa[i] - Sb[lanr] * Rb[i+lanr*4];
}
// find the normal and non-normal axis numbers of the reference box
if (code <= 3) codeN = code-1; else codeN = code-4;
if (codeN==0) {
code1 = 1;
code2 = 2;
}
else if (codeN==1) {
code1 = 0;
code2 = 2;
}
else {
code1 = 0;
code2 = 1;
}
// find the four corners of the incident face, in reference-face coordinates
f32 quad[8]; // 2D coordinate of incident face (x,y pairs)
f32 c1,c2,m11,m12,m21,m22;
c1 = dDOT (center,(Ra+code1*4));
c2 = dDOT (center,(Ra+code2*4));
// optimize this? - we have already computed this data above, but it is not
// stored in an easy-to-index format. for now it's quicker just to recompute
// the four dot products.
m11 = dDOT (Ra+code1*4,Rb+a1*4);
m12 = dDOT (Ra+code1*4,Rb+a2*4);
m21 = dDOT (Ra+code2*4,Rb+a1*4);
m22 = dDOT (Ra+code2*4,Rb+a2*4);
{
f32 k1 = m11*Sb[a1];
f32 k2 = m21*Sb[a1];
f32 k3 = m12*Sb[a2];
f32 k4 = m22*Sb[a2];
quad[0] = c1 - k1 - k3;
quad[1] = c2 - k2 - k4;
quad[2] = c1 - k1 + k3;
quad[3] = c2 - k2 + k4;
quad[4] = c1 + k1 + k3;
quad[5] = c2 + k2 + k4;
quad[6] = c1 + k1 - k3;
quad[7] = c2 + k2 - k4;
}
// find the size of the reference face
f32 rect[2];
rect[0] = Sa[code1];
rect[1] = Sa[code2];
// intersect the incident and reference faces
f32 ret[16];
s32 n = intersectRectQuad (rect,quad,ret);
if (n < 1) return 0; // this should never happen
// convert the intersection points into reference-face coordinates,
// and compute the contact position and depth for each point. only keep
// those points that have a positive (penetrating) depth. delete points in
// the 'ret' array as necessary so that 'point' and 'ret' correspond.
f32 point[3*8]; // penetrating contact points
f32 dep[8]; // depths for those points
f32 det1 = 1.f/(m11*m22 - m12*m21);
m11 *= det1;
m12 *= det1;
m21 *= det1;
m22 *= det1;
s32 cnum = 0; // number of penetrating contact points found
for (j=0; j < n; j++) {
f32 k1 = m22*(ret[j*2]-c1) - m12*(ret[j*2+1]-c2);
f32 k2 = -m21*(ret[j*2]-c1) + m11*(ret[j*2+1]-c2);
for (i=0; i<3; i++) point[cnum*3+i] =
center[i] + k1*Rb[i+a1*4] + k2*Rb[i+a2*4];
dep[cnum] = Sa[codeN] - dDOT(normal2,point+cnum*3);
if (dep[cnum] >= 0) {
ret[cnum*2] = ret[j*2];
ret[cnum*2+1] = ret[j*2+1];
cnum++;
}
}
if (cnum < 1) return 0; // this should never happen
// we can't generate more contacts than we actually have
if (maxc > cnum) maxc = cnum;
if (maxc < 1) maxc = 1;
if (cnum <= maxc) {
// we have less contacts than we need, so we use them all
for (j=0; j < cnum; j++) {
cContact *con = &contact[j];
for (i=0; i<3; i++) con->pos[i] = point[j*3+i] + pa[i];
con->depth = dep[j];
}
}
else {
// we have more contacts than are wanted, some of them must be culled.
// find the deepest point, it is always the first contact.
s32 i1 = 0;
f32 maxdepth = dep[0];
for (i=1; i<cnum; i++) {
if (dep[i] > maxdepth) {
maxdepth = dep[i];
i1 = i;
}
}
//s32 iret[8];
//cullPoints (cnum,ret,maxc,i1,iret);
for (j=0; j < maxc; j++) {
cContact *con = &contact[j];
for (i=0; i<3; i++) con->pos[i] = point[j*3+i] + pa[i];
con->depth = dep[j];
}
cnum = maxc;
}
return cnum;
}
