// GEN-2 extended benchmark: BOX vs COMPOUND factory-part scenes, larger N (up to 262144), // real RTX-class GPU vs single-thread CPU. Answers the goal's literal ask: how does the // GPU-vs-CPU speedup + correctness differ on harder (compound/heavy) objects vs simple boxes. // Same math as bench.cu; adds a `shapeType` dimension. HONEST measured only. // Build: nvcc -O3 -arch=native -o bench2 bench/bench2.cu #include #include #include #include #include #include #include struct Body { float px,py,pz, vx,vy,vz, hx,hy,hz, invMass; }; static unsigned lcg(unsigned &s){ s = s*1664525u + 1013904223u; return s; } static float frand(unsigned &s){ return (lcg(s) & 0xFFFFFF) / float(0x1000000); } // compound=1 -> irregular non-uniform extents + heavy varied mass (factory parts) // compound=0 -> uniform simple boxes, uniform light mass void initScene(std::vector& b, int n, unsigned seed, int compound){ b.resize(n); unsigned s=seed; for(int i=0;i&b,int iters,float dt,double*penOut){ int n=b.size(); for(int i=0;i0) b[i].vy+=-9.81f*dt; double penSum=0;int penCnt=0; float cell=1.4f; for(int it=0;it head(n*2,-1), nxt(n,-1); auto H=[&](int x,int y,int z){ unsigned h=(x*73856093)^(y*19349663)^(z*83492791); return h%(unsigned)(n*2); }; for(int i=0;i0&&oy>0&&oz>0){solvePairCPU(b[i],b[j]); if(it==iters-1){penSum+=fminf(fminf(ox,oy),oz);penCnt++;}}}}} } for(int i=0;i0){x.vy+=0.2f*gp*60.0f;if(x.vy<0)x.vy=0;}x.px+=x.vx*dt;x.py+=x.vy*dt;x.pz+=x.vz*dt;} if(penOut)*penOut=penCnt?penSum/penCnt:0.0; return 0; } __global__ void integrateVel(Body*b,int n,float dt){int i=blockIdx.x*blockDim.x+threadIdx.x;if(i>=n)return;if(b[i].invMass>0)b[i].vy+=-9.81f*dt;} __global__ void solveJacobiGrid(const Body*in,Body*out,int n,float cell){ int i=blockIdx.x*blockDim.x+threadIdx.x;if(i>=n)return;Body a=in[i];float ax=0,ay=0,az=0; // brute over neighbors within a spatial window (bounded scan for GPU simplicity at these N) for(int j=0;jcell*cell)continue; float ox=(a.hx+b.hx)-fabsf(dx),oy=(a.hy+b.hy)-fabsf(dy),oz=(a.hz+b.hz)-fabsf(dz); if(ox<=0||oy<=0||oz<=0)continue; float nx=0,ny=0,nz=0,depth; if(ox<=oy&&ox<=oz){nx=dx<0?-1:1;depth=ox;}else if(oy<=oz){ny=dy<0?-1:1;depth=oy;}else{nz=dz<0?-1:1;depth=oz;} float vn=(b.vx-a.vx)*nx+(b.vy-a.vy)*ny+(b.vz-a.vz)*nz;float im=a.invMass+b.invMass;if(im<1e-9f)continue; float bias=0.2f*fmaxf(depth-0.001f,0.0f)*60.0f;float jm=-(vn-bias)/im;if(jm<0)jm=0; ax-=nx*jm*a.invMass*0.5f;ay-=ny*jm*a.invMass*0.5f;az-=nz*jm*a.invMass*0.5f;} a.vx+=ax;a.vy+=ay;a.vz+=az;out[i]=a; } __global__ void integratePos(Body*b,int n,float dt){int i=blockIdx.x*blockDim.x+threadIdx.x;if(i>=n)return;Body&x=b[i];if(x.invMass<=0)return;float gp=x.hy-x.py;if(gp>0){x.vy+=0.2f*gp*60.0f;if(x.vy<0)x.vy=0;}x.px+=x.vx*dt;x.py+=x.vy*dt;x.pz+=x.vz*dt;} void stepGPU(Body*dA,Body*dB,int n,int iters,float dt){int T=128,G=(n+T-1)/T;float cell=1.4f;integrateVel<<>>(dA,n,dt);for(int it=0;it>>(dA,dB,n,cell);cudaMemcpy(dA,dB,n*sizeof(Body),cudaMemcpyDeviceToDevice);}integratePos<<>>(dA,n,dt);} int main(int argc,char**argv){ const char*jsonPath="results/gen2_results.json"; for(int i=1;i cpu; initScene(cpu,n,seed,comp); std::vector gi=cpu; double pen=0; for(int i=0;i(t1-t0).count()/STEPS; Body*dA,*dB;cudaMalloc(&dA,n*sizeof(Body));cudaMalloc(&dB,n*sizeof(Body));cudaMemcpy(dA,gi.data(),n*sizeof(Body),cudaMemcpyHostToDevice); for(int i=0;i gb(n);cudaMemcpy(gb.data(),dA,n*sizeof(Body),cudaMemcpyDeviceToHost); double div=0;for(int i=0;i