/* "calc_blessian_mem" calculates the block Hessian. */ int calc_blessian_mem(PDB_File *PDB,dSparse_Matrix *PP1,int nres,int nblx, int elm,double **HB,double cut,double gam,double scl, double mlo,double mhi) { dSparse_Matrix *PP2; double **HR,***HT; int **CT,*BST1,*BST2; int ii,i,j,k,p,q,q1,q2,ti,tj,bi,bj,sb,nc,out; /* ------------------- INITIALIZE LOCAL VARIABLES ------------------- */ /* HR holde three rows (corresponding to 1 residue) of the full Hessian */ HR=zero_dmatrix(1,3*nres,1,3); /* CT is an array of contacts between blocks */ CT=unit_imatrix(0,nblx); /* Copy PP1 to PP2 and sort by second element */ PP2=(dSparse_Matrix *)malloc((size_t)sizeof(dSparse_Matrix)); PP2->IDX=imatrix(1,elm,1,2); PP2->X=dvector(1,elm); copy_dsparse(PP1,PP2,1,elm); dsort_PP2(PP2,elm,2); /* BST1: for all j: BST1[i]<=j<BST[i+1], PP1->IDX[j][1]=i */ /* BST2: for all j: BST2[i]<=j<BST2[i+1], PP2->IDX[j][2]=i */ BST1=ivector(1,3*nres+1); BST2=ivector(1,6*nblx+1); init_bst(BST1,PP1,elm,3*nres+1,1); init_bst(BST2,PP2,elm,6*nblx+1,2); /* ------------------- LOCAL VARIABLES INITIALIZED ------------------ */ /* ------------- FIND WHICH BLOCKS ARE IN CONTACT --------------- */ nc=find_contacts1(CT,PDB,nres,nblx,cut); /* Allocate a tensor for the block Hessian */ HT=zero_d3tensor(1,nc,1,6,1,6); /* Calculate each super-row of the full Hessian */ for(ii=1;ii<=nres;ii++){ if(PDB->atom[ii].model!=0){ /* ----------------- FIND SUPER-ROW OF FULL HESSIAN --------------- */ hess_superrow_mem(HR,CT,PDB,nres,ii,cut,gam,scl,mlo,mhi); /* Update elements of block hessian */ q1=BST1[3*(ii-1)+2]; q2=BST1[3*(ii-1)+3]; /* Sum over elements of projection matrix corresponding to residue ii: for each k in the following loop, PP1->IDX[k][1]==3*ii + 0,1,2 */ for(k=BST1[3*ii-2];k<BST1[3*ii+1];k++){ if(k<q1) q=1; else if(k<q2) q=2; else q=3; i=PP1->IDX[k][2]; bi=(i-1)/6+1; ti=i-6*(bi-1); /* Sum over all elements of projection matrix with column j>=i */ for(p=BST2[i];p<=elm;p++){ j=PP2->IDX[p][2]; bj=(j-1)/6+1; sb=CT[bi][bj]; if(i<=j && sb!=0){ /* the first condition should ALWAYS hold */ tj=j-6*(bj-1); HT[sb][ti][tj]+=(PP1->X[k]*PP2->X[p]*HR[PP2->IDX[p][1]][q]); } } } } } /* Print the block Hessian in sparse format */ out=bless_from_tensor(HB,HT,CT,nblx); /* Free up memory */ free_dmatrix(HR,1,3*nres,1,3); free_d3tensor(HT,1,nc,1,6,1,6); free_imatrix(CT,0,nblx,0,nblx); free_ivector(BST1,1,3*nres+1); free_ivector(BST2,1,6*nblx+1); free_imatrix(PP2->IDX,1,elm,1,2); free_dvector(PP2->X,1,elm); return out; }
/* creates matrices B and invB: BT = Power in the steady state. * NOTE: EXTRA nodes: 4 heat spreader peripheral nodes, 4 heat * sink inner peripheral nodes, 4 heat sink outer peripheral * nodes(north, south, east and west) and 1 ambient node. */ void populate_R_model_block(block_model_t *model, flp_t *flp) { /* shortcuts */ double **b = model->b; double *gx = model->gx, *gy = model->gy; double *gx_int = model->gx_int, *gy_int = model->gy_int; double *gx_sp = model->gx_sp, *gy_sp = model->gy_sp; double *gx_hs = model->gx_hs, *gy_hs = model->gy_hs; double *g_amb = model->g_amb; double **len = model->len, **g = model->g, **lu = model->lu; int **border = model->border; int *p = model->p; double t_chip = model->config.t_chip; double r_convec = model->config.r_convec; double s_sink = model->config.s_sink; double t_sink = model->config.t_sink; double s_spreader = model->config.s_spreader; double t_spreader = model->config.t_spreader; double t_interface = model->config.t_interface; double k_chip = model->config.k_chip; double k_sink = model->config.k_sink; double k_spreader = model->config.k_spreader; double k_interface = model->config.k_interface; int i, j, n = flp->n_units; double gn_sp=0, gs_sp=0, ge_sp=0, gw_sp=0; double gn_hs=0, gs_hs=0, ge_hs=0, gw_hs=0; double r_amb; double w_chip = get_total_width (flp); /* x-axis */ double l_chip = get_total_height (flp); /* y-axis */ /* sanity check on floorplan sizes */ if (w_chip > s_sink || l_chip > s_sink || w_chip > s_spreader || l_chip > s_spreader) { //print_flp(flp); //print_flp_fig(flp); //fatal("inordinate floorplan size!\n"); } if(model->flp != flp || model->n_units != flp->n_units || model->n_nodes != NL * flp->n_units + EXTRA) fatal("mismatch between the floorplan and the thermal model\n"); /* gx's and gy's of blocks */ for (i = 0; i < n; i++) { /* at the silicon layer */ if (model->config.block_omit_lateral) { gx[i] = gy[i] = 0; } else { gx[i] = 1.0/getr(k_chip, flp->units[i].width / 2.0, flp->units[i].height * t_chip); gy[i] = 1.0/getr(k_chip, flp->units[i].height / 2.0, flp->units[i].width * t_chip); } /* at the interface layer */ gx_int[i] = 1.0/getr(k_interface, flp->units[i].width / 2.0, flp->units[i].height * t_interface); gy_int[i] = 1.0/getr(k_interface, flp->units[i].height / 2.0, flp->units[i].width * t_interface); /* at the spreader layer */ gx_sp[i] = 1.0/getr(k_spreader, flp->units[i].width / 2.0, flp->units[i].height * t_spreader); gy_sp[i] = 1.0/getr(k_spreader, flp->units[i].height / 2.0, flp->units[i].width * t_spreader); /* at the heatsink layer */ gx_hs[i] = 1.0/getr(k_sink, flp->units[i].width / 2.0, flp->units[i].height * t_sink); gy_hs[i] = 1.0/getr(k_sink, flp->units[i].height / 2.0, flp->units[i].width * t_sink); } /* shared lengths between blocks */ for (i = 0; i < n; i++) for (j = i; j < n; j++) len[i][j] = len[j][i] = get_shared_len(flp, i, j); /* package R's */ populate_package_R(&model->pack, &model->config, w_chip, l_chip); /* short the R's from block centers to a particular chip edge */ for (i = 0; i < n; i++) { if (eq(flp->units[i].bottomy + flp->units[i].height, l_chip)) { gn_sp += gy_sp[i]; gn_hs += gy_hs[i]; border[i][2] = 1; /* block is on northern border */ } else border[i][2] = 0; if (eq(flp->units[i].bottomy, 0)) { gs_sp += gy_sp[i]; gs_hs += gy_hs[i]; border[i][3] = 1; /* block is on southern border */ } else border[i][3] = 0; if (eq(flp->units[i].leftx + flp->units[i].width, w_chip)) { ge_sp += gx_sp[i]; ge_hs += gx_hs[i]; border[i][1] = 1; /* block is on eastern border */ } else border[i][1] = 0; if (eq(flp->units[i].leftx, 0)) { gw_sp += gx_sp[i]; gw_hs += gx_hs[i]; border[i][0] = 1; /* block is on western border */ } else border[i][0] = 0; } /* initialize g */ zero_dmatrix(g, NL*n+EXTRA, NL*n+EXTRA); zero_dvector(g_amb, n+EXTRA); /* overall Rs between nodes */ for (i = 0; i < n; i++) { double area = (flp->units[i].height * flp->units[i].width); /* amongst functional units in the various layers */ for (j = 0; j < n; j++) { double part = 0, part_int = 0, part_sp = 0, part_hs = 0; if (is_horiz_adj(flp, i, j)) { part = gx[i] / flp->units[i].height; part_int = gx_int[i] / flp->units[i].height; part_sp = gx_sp[i] / flp->units[i].height; part_hs = gx_hs[i] / flp->units[i].height; } else if (is_vert_adj(flp, i,j)) { part = gy[i] / flp->units[i].width; part_int = gy_int[i] / flp->units[i].width; part_sp = gy_sp[i] / flp->units[i].width; part_hs = gy_hs[i] / flp->units[i].width; } g[i][j] = part * len[i][j]; g[IFACE*n+i][IFACE*n+j] = part_int * len[i][j]; g[HSP*n+i][HSP*n+j] = part_sp * len[i][j]; g[HSINK*n+i][HSINK*n+j] = part_hs * len[i][j]; } /* the 2.0 factor in the following equations is * explained during the calculation of the B matrix */ /* vertical g's in the silicon layer */ g[i][IFACE*n+i]=g[IFACE*n+i][i]=2.0/getr(k_chip, t_chip, area); /* vertical g's in the interface layer */ g[IFACE*n+i][HSP*n+i]=g[HSP*n+i][IFACE*n+i]=2.0/getr(k_interface, t_interface, area); /* vertical g's in the spreader layer */ g[HSP*n+i][HSINK*n+i]=g[HSINK*n+i][HSP*n+i]=2.0/getr(k_spreader, t_spreader, area); /* vertical g's in the heatsink core layer */ /* vertical R to ambient: divide r_convec proportional to area */ r_amb = r_convec * (s_sink * s_sink) / area; g_amb[i] = 1.0 / (getr(k_sink, t_sink, area) + r_amb); /* lateral g's from block center (spreader layer) to peripheral (n,s,e,w) spreader nodes */ g[HSP*n+i][NL*n+SP_N]=g[NL*n+SP_N][HSP*n+i]=2.0*border[i][2] / ((1.0/gy_sp[i])+model->pack.r_sp1_y*gn_sp/gy_sp[i]); g[HSP*n+i][NL*n+SP_S]=g[NL*n+SP_S][HSP*n+i]=2.0*border[i][3] / ((1.0/gy_sp[i])+model->pack.r_sp1_y*gs_sp/gy_sp[i]); g[HSP*n+i][NL*n+SP_E]=g[NL*n+SP_E][HSP*n+i]=2.0*border[i][1] / ((1.0/gx_sp[i])+model->pack.r_sp1_x*ge_sp/gx_sp[i]); g[HSP*n+i][NL*n+SP_W]=g[NL*n+SP_W][HSP*n+i]=2.0*border[i][0] / ((1.0/gx_sp[i])+model->pack.r_sp1_x*gw_sp/gx_sp[i]); /* lateral g's from block center (heatsink layer) to peripheral (n,s,e,w) heatsink nodes */ g[HSINK*n+i][NL*n+SINK_C_N]=g[NL*n+SINK_C_N][HSINK*n+i]=2.0*border[i][2] / ((1.0/gy_hs[i])+model->pack.r_hs1_y*gn_hs/gy_hs[i]); g[HSINK*n+i][NL*n+SINK_C_S]=g[NL*n+SINK_C_S][HSINK*n+i]=2.0*border[i][3] / ((1.0/gy_hs[i])+model->pack.r_hs1_y*gs_hs/gy_hs[i]); g[HSINK*n+i][NL*n+SINK_C_E]=g[NL*n+SINK_C_E][HSINK*n+i]=2.0*border[i][1] / ((1.0/gx_hs[i])+model->pack.r_hs1_x*ge_hs/gx_hs[i]); g[HSINK*n+i][NL*n+SINK_C_W]=g[NL*n+SINK_C_W][HSINK*n+i]=2.0*border[i][0] / ((1.0/gx_hs[i])+model->pack.r_hs1_x*gw_hs/gx_hs[i]); } /* g's from peripheral(n,s,e,w) nodes */ /* vertical g's between peripheral spreader nodes and center peripheral heatsink nodes */ g[NL*n+SP_N][NL*n+SINK_C_N]=g[NL*n+SINK_C_N][NL*n+SP_N]=2.0/model->pack.r_sp_per_y; g[NL*n+SP_S][NL*n+SINK_C_S]=g[NL*n+SINK_C_S][NL*n+SP_S]=2.0/model->pack.r_sp_per_y; g[NL*n+SP_E][NL*n+SINK_C_E]=g[NL*n+SINK_C_E][NL*n+SP_E]=2.0/model->pack.r_sp_per_x; g[NL*n+SP_W][NL*n+SINK_C_W]=g[NL*n+SINK_C_W][NL*n+SP_W]=2.0/model->pack.r_sp_per_x; /* lateral g's between peripheral outer sink nodes and center peripheral sink nodes */ g[NL*n+SINK_C_N][NL*n+SINK_N]=g[NL*n+SINK_N][NL*n+SINK_C_N]=2.0/(model->pack.r_hs + model->pack.r_hs2_y); g[NL*n+SINK_C_S][NL*n+SINK_S]=g[NL*n+SINK_S][NL*n+SINK_C_S]=2.0/(model->pack.r_hs + model->pack.r_hs2_y); g[NL*n+SINK_C_E][NL*n+SINK_E]=g[NL*n+SINK_E][NL*n+SINK_C_E]=2.0/(model->pack.r_hs + model->pack.r_hs2_x); g[NL*n+SINK_C_W][NL*n+SINK_W]=g[NL*n+SINK_W][NL*n+SINK_C_W]=2.0/(model->pack.r_hs + model->pack.r_hs2_x); /* vertical g's between inner peripheral sink nodes and ambient */ g_amb[n+SINK_C_N] = g_amb[n+SINK_C_S] = 1.0 / (model->pack.r_hs_c_per_y+model->pack.r_amb_c_per_y); g_amb[n+SINK_C_E] = g_amb[n+SINK_C_W] = 1.0 / (model->pack.r_hs_c_per_x+model->pack.r_amb_c_per_x); /* vertical g's between outer peripheral sink nodes and ambient */ g_amb[n+SINK_N] = g_amb[n+SINK_S] = g_amb[n+SINK_E] = g_amb[n+SINK_W] = 1.0 / (model->pack.r_hs_per+model->pack.r_amb_per); /* calculate matrix B such that BT = POWER in steady state */ /* non-diagonal elements */ for (i = 0; i < NL*n+EXTRA; i++) for (j = 0; j < i; j++) if ((g[i][j] == 0.0) || (g[j][i] == 0.0)) b[i][j] = b[j][i] = 0.0; else /* here is why the 2.0 factor comes when calculating g[][] */ b[i][j] = b[j][i] = -1.0/((1.0/g[i][j])+(1.0/g[j][i])); /* diagonal elements */ for (i = 0; i < NL*n+EXTRA; i++) { /* functional blocks in the heat sink layer */ if (i >= HSINK*n && i < NL*n) b[i][i] = g_amb[i%n]; /* heat sink peripheral nodes */ else if (i >= NL*n+SINK_C_W) b[i][i] = g_amb[n+i-NL*n]; /* all other nodes that are not connected to the ambient */ else b[i][i] = 0.0; /* sum up the conductances */ for(j=0; j < NL*n+EXTRA; j++) if (i != j) b[i][i] -= b[i][j]; } /* compute the LUP decomposition of B and store it too */ copy_dmatrix(lu, b, NL*n+EXTRA, NL*n+EXTRA); /* * B is a symmetric positive definite matrix. It is * symmetric because if a node A is connected to B, * then B is also connected to A with the same R value. * It is positive definite because of the following * informal argument from Professor Lieven Vandenberghe's * lecture slides for the spring 2004-2005 EE 103 class * at UCLA: http://www.ee.ucla.edu/~vandenbe/103/chol.pdf * x^T*B*x = voltage^T * (B*x) = voltage^T * current * = total power dissipated in the resistors > 0 * for x != 0. */ lupdcmp(lu, NL*n+EXTRA, p, 1); /* done */ model->flp = flp; model->r_ready = TRUE; }