int main()
{
printf("\n");
printf("\n");
printf("\n");
printf(" HPMPC -- Library for High-Performance implementation of solvers for MPC.\n");
printf(" Copyright (C) 2014 by Technical University of Denmark. All rights reserved.\n");
printf("\n");
printf(" HPMPC is distributed in the hope that it will be useful,\n");
printf(" but WITHOUT ANY WARRANTY; without even the implied warranty of\n");
printf(" MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.\n");
printf(" See the GNU Lesser General Public License for more details.\n");
printf("\n");
printf("\n");
printf("\n");
printf("Riccati solver performance test - single precision\n");
printf("\n");
// maximum frequency of the processor
const float GHz_max = 2.9; //3.6; //2.9;
printf("Frequency used to compute theoretical peak: %5.1f GHz (edit test_dricposv.c to modify this value).\n", GHz_max);
printf("\n");
// maximum flops per cycle, single precision
#if defined(TARGET_X64_AVX)
const float flops_max = 16;
printf("Testing solvers for AVX instruction set, 64 bit: theoretical peak %5.1f Gflops\n", flops_max*GHz_max);
#elif defined(TARGET_X64_SSE3) || defined(TARGET_AMD_SSE3)
const float flops_max = 8;
printf("Testing solvers for SSE3 instruction set, 64 bit: theoretical peak %5.1f Gflops\n", flops_max*GHz_max);
#elif defined(TARGET_CORTEXA9)
const float flops_max = 4;
printf("Testing solvers for ARMv7a NEON instruction set: theoretical peak %5.1f Gflops\n", flops_max*GHz_max);
#elif defined(TARGET_X86_ATOM)
const float flops_max = 4;
printf("Testing solvers for SSE3 instruction set, 32 bit, optimized for Intel Atom: theoretical peak %5.1f Gflops\n", flops_max*GHz_max);
#elif defined(TARGET_POWERPC_G2)
const float flops_max = 2;
printf("Testing solvers for POWERPC instruction set, 32 bit: theoretical peak %5.1f Gflops\n", flops_max*GHz_max);
#elif defined(TARGET_C99_4X4)
const float flops_max = 2;
printf("Testing reference solvers, 4x4 kernel: theoretical peak %5.1f Gflops\n", flops_max*GHz_max);
#elif defined(TARGET_C99_2X2)
const float flops_max = 2;
printf("Testing reference solvers, 2x2 kernel: theoretical peak %5.1f Gflops\n", flops_max*GHz_max);
#endif
printf("\n");
printf("Tested solvers:\n");
printf("-sv : Riccati factorization and system solution (prediction step in IP methods)\n");
printf("-trs: system solution after a previous call to Riccati factorization (correction step in IP methods)\n");
printf("\n");
printf("\n");
#if defined(TARGET_X64_AVX) || defined(TARGET_X64_SSE3) || defined(TARGET_X86_ATOM) || defined(TARGET_AMD_SSE3)
printf("\nflush to zero on\n");
_MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON); // flush to zero subnormals !!! works only with one thread !!!
#endif
// to throw floating-point exception
/*#ifndef __APPLE__*/
/* feenableexcept(FE_DIVBYZERO | FE_INVALID | FE_OVERFLOW);*/
/*#endif*/
int err;
int i, j, ii, jj, idx;
const int bsd = D_MR; //d_get_mr();
const int bss = S_MR; //s_get_mr();
int info = 0;
int nn[] = {4, 6, 8, 10, 12, 16, 20, 24, 28, 32, 36, 40, 44, 48, 52, 56, 60, 64, 68, 72, 76, 80, 84, 88, 92, 96, 100, 104, 108, 112, 116, 120, 124, 128, 132, 136, 140, 144, 148, 152, 156, 160, 164, 168, 172, 176, 180, 184, 188, 192, 196, 200, 204, 208, 212, 216, 220, 224, 228, 232, 236, 240, 244, 248, 252, 256, 260, 264, 268, 272, 276, 280, 284, 288, 292, 296, 300};
int nnrep[] = {10000, 10000, 10000, 10000, 10000, 4000, 4000, 2000, 2000, 1000, 1000, 400, 400, 400, 200, 200, 200, 100, 100, 100, 100, 100, 100, 100, 100, 100, 100, 100, 100, 100, 100, 100, 100, 100, 40, 40, 40, 40, 40, 20, 20, 20, 20, 20, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10};
int vnx[] = {8, 12, 16, 24, 32, 48, 64, 96, 128, 192, 256, 512, 1024};
int vnrep[] = {100, 100, 100, 100, 100, 100, 50, 50, 50, 20, 10, 10};
int vN[] = {4, 8, 12, 16, 24, 32, 48, 64, 96, 128, 192, 256};
int ll;
for(ll=0; ll<77; ll++)
/* for(ll=0; ll<1; ll++)*/
{
int nx = nn[ll];//NX;//16;//nn[ll]; // number of states (it has to be even for the mass-spring system test problem)
int nu = 2;//NU;//5; // number of inputs (controllers) (it has to be at least 1 and at most nx/2 for the mass-spring system test problem)
int N = 10;//NN;//10; // horizon lenght
int nrep = nnrep[ll];
/* int nx = NX;//16;//nn[ll]; // number of states (it has to be even for the mass-spring system test problem)*/
/* int nu = NU;//5; // number of inputs (controllers) (it has to be at least 1 and at most nx/2 for the mass-spring system test problem)*/
/* int N = NN;//10; // horizon lenght*/
/* int nrep = NREP;*/
int rep;
int nz = nx+nu+1;
int pnz = bss*((nz+bss-nu%bss+bss-1)/bss);
/************************************************
* dynamical system
************************************************/
double *A; d_zeros(&A, nx, nx); // states update matrix
double *B; d_zeros(&B, nx, nu); // inputs matrix
double *b; d_zeros(&b, nx, 1); // states offset
double *x0; d_zeros(&x0, nx, 1); // initial state
double Ts = 0.5; // sampling time
mass_spring_system(Ts, nx, nu, N, A, B, b, x0);
for(jj=0; jj<nx; jj++)
b[jj] = 0.1;
for(jj=0; jj<nx; jj++)
x0[jj] = 0;
x0[0] = 3.5;
x0[1] = 3.5;
// d_print_mat(nx, nx, A, nx);
// d_print_mat(nx, nu, B, nx);
// d_print_mat(nx, 1, b, nx);
// d_print_mat(nx, 1, x0, nx);
/* packed */
double *BAb; d_zeros(&BAb, nx, nz);
dmcopy(nx, nu, B, nx, BAb, nx);
dmcopy(nx, nx, A, nx, BAb+nu*nx, nx);
dmcopy(nx, 1 , b, nx, BAb+(nu+nx)*nx, nx);
// d_print_mat(nx, nx+nu+1, BAb, nx);
/* transposed */
double *BAbt; d_zeros_align(&BAbt, pnz, pnz);
for(ii=0; ii<nx; ii++)
for(jj=0; jj<nz; jj++)
{
BAbt[jj+pnz*ii] = BAb[ii+nx*jj];
}
// d_print_mat(nz, nx+1, BAbt, pnz);
// s_print_mat(nz, nx+1, sBAbt, pnz);
// return 0;
/* packed into contiguous memory */
double *pBAbt; d_zeros_align(&pBAbt, pnz, pnz);
d_cvt_mat2pmat(nz, nx, 0, bsd, BAbt, pnz, pBAbt, pnz);
float *psBAbt; s_zeros_align(&psBAbt, pnz, pnz);
s_cvt_d2s_pmat(nz, nx, bsd, pBAbt, pnz, bss, psBAbt, pnz);
// d_print_pmat(nz, nx, bsd, pBAbt, pnz);
// s_print_pmat(nz, nx, bss, spBAbt, pnz);
/************************************************
* cost function
************************************************/
double *Q; d_zeros_align(&Q, pnz, pnz);
for(ii=0; ii<nu; ii++) Q[ii*(pnz+1)] = 2.0;
for(; ii<pnz; ii++) Q[ii*(pnz+1)] = 1.0;
for(ii=0; ii<nz; ii++) Q[nx+nu+ii*pnz] = 1.0;
Q[(nx+nu)*(pnz+1)] = 1e6;
/* packed into contiguous memory */
float *pQ; s_zeros_align(&pQ, pnz, pnz);
cvt_d2s_mat2pmat(nz, nz, 0, bss, Q, pnz, pQ, pnz);
/* matrices series */
float *(hpQ[N+1]);
float *(hq[N+1]);
float *(hux[N+1]);
float *(hpi[N+1]);
float *(hpBAbt[N]);
float *(hrb[N]);
float *(hrq[N+1]);
for(jj=0; jj<N; jj++)
{
s_zeros_align(&hpQ[jj], pnz, pnz);
s_zeros_align(&hq[jj], pnz, 1);
s_zeros_align(&hux[jj], pnz, 1);
s_zeros_align(&hpi[jj], nx, 1);
hpBAbt[jj] = psBAbt;
s_zeros_align(&hrb[jj], nx, 1);
s_zeros_align(&hrq[jj], nx+nu, 1);
}
s_zeros_align(&hpQ[N], pnz, pnz);
s_zeros_align(&hq[N], pnz, 1);
s_zeros_align(&hux[N], pnz, 1);
s_zeros_align(&hpi[N], nx, 1);
s_zeros_align(&hrq[N], nx+nu, 1);
// starting guess
for(jj=0; jj<nx; jj++) hux[0][nu+jj] = (float) x0[jj];
float *pL; s_zeros_align(&pL, pnz, pnz);
float *pBAbtL; s_zeros_align(&pBAbtL, pnz, pnz);
/************************************************
* riccati-like iteration
************************************************/
// predictor
// restore cost function
for(ii=0; ii<N; ii++)
{
for(jj=0; jj<pnz*pnz; jj++) hpQ[ii][jj]=pQ[jj];
}
for(jj=0; jj<pnz*pnz; jj++) hpQ[N][jj]=pQ[jj];
// call the solver
sricposv_mpc(nx, nu, N, pnz, hpBAbt, hpQ, hux, pL, pBAbtL, COMPUTE_MULT, hpi, &info);
if(PRINTRES==1)
{
/* print result */
printf("\n\nsv\n\n");
for(ii=0; ii<N; ii++)
s_print_mat(1, nu, hux[ii], 1);
}
if(PRINTRES==1 && COMPUTE_MULT==1)
{
// print result
printf("\n\nsv\n\n");
for(ii=0; ii<N; ii++)
s_print_mat(1, nx, hpi[ii+1], 1);
}
// corrector
// clear solution
for(ii=0; ii<N; ii++)
{
for(jj=0; jj<nu; jj++) hux[ii][jj] = 0;
for(jj=0; jj<nx; jj++) hux[ii+1][nu+jj] = 0;
}
// restore linear part of cost function
for(ii=0; ii<N; ii++)
{
for(jj=0; jj<nx+nu; jj++) hq[ii][jj] = Q[nx+nu+pnz*jj];
}
for(jj=0; jj<nx+nu; jj++) hq[N][jj] = Q[nx+nu+pnz*jj];
// call the solver
sricpotrs_mpc(nx, nu, N, pnz, hpBAbt, hpQ, hq, hux, pBAbtL, COMPUTE_MULT, hpi);
if(PRINTRES==1)
{
// print result
printf("\n\ntrs\n\n");
for(ii=0; ii<N; ii++)
s_print_mat(1, nu, hux[ii], 1);
}
if(PRINTRES==1 && COMPUTE_MULT==1)
{
// print result
printf("\n\ntrs\n\n");
for(ii=0; ii<N; ii++)
s_print_mat(1, nx, hpi[ii+1], 1);
}
// restore cost function
for(ii=0; ii<N; ii++)
{
for(jj=0; jj<pnz*pnz; jj++) hpQ[ii][jj]=pQ[jj];
}
for(jj=0; jj<pnz*pnz; jj++) hpQ[N][jj]=pQ[jj];
// restore linear part of cost function
for(ii=0; ii<N; ii++)
{
for(jj=0; jj<nx+nu; jj++) hq[ii][jj] = Q[nx+nu+pnz*jj];
}
for(jj=0; jj<nx+nu; jj++) hq[N][jj] = Q[nx+nu+pnz*jj];
// residuals computation
sres(nx, nu, N, pnz, hpBAbt, hpQ, hq, hux, hpi, hrq, hrb);
if(PRINTRES==1 && COMPUTE_MULT==1)
{
// print result
printf("\n\nres\n\n");
for(ii=0; ii<+N; ii++)
s_print_mat(1, nx+nu, hrq[ii], 1);
for(ii=0; ii<N; ii++)
s_print_mat(1, nx, hrb[ii], 1);
}
// timing
struct timeval tv0, tv1, tv2;
gettimeofday(&tv0, NULL); // start
// double precision
for(rep=0; rep<nrep; rep++)
{
// restore cost function
for(ii=0; ii<N; ii++)
{
for(jj=0; jj<pnz*pnz; jj++) hpQ[ii][jj]=pQ[jj];
}
for(jj=0; jj<pnz*pnz; jj++) hpQ[N][jj]=pQ[jj];
// call the solver
sricposv_mpc(nx, nu, N, pnz, hpBAbt, hpQ, hux, pL, pBAbtL, COMPUTE_MULT, hpi, &info);
}
gettimeofday(&tv1, NULL); // start
for(rep=0; rep<nrep; rep++)
{
// clear solution
for(ii=0; ii<N; ii++)
{
for(jj=0; jj<nu; jj++) hux[ii][jj] = 0;
for(jj=0; jj<nx; jj++) hux[ii+1][nu+jj] = 0;
}
// restore linear part of cost function
for(ii=0; ii<N; ii++)
{
for(jj=0; jj<nx+nu; jj++) hq[ii][jj] = Q[nx+nu+pnz*jj];
}
for(jj=0; jj<nx+nu; jj++) hq[N][jj] = Q[nx+nu+pnz*jj];
// call the solver
sricpotrs_mpc(nx, nu, N, pnz, hpBAbt, hpQ, hq, hux, pBAbtL, COMPUTE_MULT, hpi);
}
gettimeofday(&tv2, NULL); // start
float time_sv = (float) (tv1.tv_sec-tv0.tv_sec)/(nrep+0.0)+(tv1.tv_usec-tv0.tv_usec)/(nrep*1e6);
float flop_sv = (1.0/3.0*nx*nx*nx+3.0/2.0*nx*nx) + N*(7.0/3.0*nx*nx*nx+4.0*nx*nx*nu+2.0*nx*nu*nu+1.0/3.0*nu*nu*nu+13.0/2.0*nx*nx+9.0*nx*nu+5.0/2.0*nu*nu);
if(COMPUTE_MULT==1)
flop_sv += N*2*nx*nx;
float Gflops_sv = 1e-9*flop_sv/time_sv;
float time_trs = (float) (tv2.tv_sec-tv1.tv_sec)/(nrep+0.0)+(tv2.tv_usec-tv1.tv_usec)/(nrep*1e6);
float flop_trs = N*(8.0*nx*nx+8.0*nx*nu+2.0*nu*nu);
if(COMPUTE_MULT==1)
flop_trs += N*2*nx*nx;
float Gflops_trs = 1e-9*flop_trs/time_trs;
float Gflops_max = flops_max * GHz_max;
if(ll==0)
printf("\nnx\tnu\tN\tsv time\t\tsv Gflops\tsv \%\t\ttrs time\ttrs Gflops\ttrs \%\n\n");
printf("%d\t%d\t%d\t%e\t%f\t%f\t%e\t%f\t%f\n", nx, nu, N, time_sv, Gflops_sv, 100.0*Gflops_sv/Gflops_max, time_trs, Gflops_trs, 100.0*Gflops_trs/Gflops_max);
/************************************************
* return
************************************************/
free(A);
free(B);
free(b);
free(x0);
free(BAb);
free(BAbt);
free(pBAbt);
free(Q);
free(pQ);
free(pL);
free(pBAbtL);
for(jj=0; jj<N; jj++)
{
free(hpQ[jj]);
free(hq[jj]);
free(hux[jj]);
free(hpi[jj]);
}
free(hpQ[N]);
free(hq[N]);
free(hux[N]);
free(hpi[N]);
} // increase size
printf("\n");
printf("\n");
printf("\n");
return 0;
}
int main()
{
printf("\n");
printf("\n");
printf("\n");
printf(" HPMPC -- Library for High-Performance implementation of solvers for MPC.\n");
printf(" Copyright (C) 2014 by Technical University of Denmark. All rights reserved.\n");
printf("\n");
printf(" HPMPC is distributed in the hope that it will be useful,\n");
printf(" but WITHOUT ANY WARRANTY; without even the implied warranty of\n");
printf(" MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.\n");
printf(" See the GNU Lesser General Public License for more details.\n");
printf("\n");
printf("\n");
printf("\n");
#if defined(TARGET_X64_AVX2) || defined(TARGET_X64_AVX) || defined(TARGET_X64_SSE3) || defined(TARGET_X86_ATOM) || defined(TARGET_AMD_SSE3)
/* printf("\nflush subnormals to zero\n\n");*/
_MM_SET_FLUSH_ZERO_MODE(_MM_FLUSH_ZERO_ON); // flush to zero subnormals !!! works only with one thread !!!
#endif
int ii, jj, idx;
int rep, nrep=NREP;
int nx = NX; // number of states (it has to be even for the mass-spring system test problem)
int nu = NU; // number of inputs (controllers) (it has to be at least 1 and at most nx/2 for the mass-spring system test problem)
int N = NN; // horizon lenght
int nb = NB; // number of box constrained inputs and states
printf(" Test problem: mass-spring system with %d masses and %d controls.\n", nx/2, nu);
printf("\n");
printf(" MPC problem size: %d states, %d inputs, %d horizon length, %d two-sided box constraints.\n", nx, nu, N, nb);
printf("\n");
printf(" ADMM method parameters: single precision, %d maximum iterations, %5.1e exit tolerance in primal and duality measure (edit file test_admm_ip_box.c to change them).\n", K_MAX_ADMM, TOL);
int info = 0;
const int bs = S_MR; //d_get_mr();
const int ncl = S_NCL;
const int nal = bs*ncl; // number of doubles per cache line
const int nz = nx+nu+1;
const int pnz = bs*((nz+bs-1)/bs);
const int pnx = bs*((nx+bs-1)/bs);
const int cnz = ncl*((nx+nu+1+ncl-1)/ncl);
const int cnx = ncl*((nx+ncl-1)/ncl);
const int pnb = bs*((2*nb+bs-1)/bs); // packed number of box constraints
const int anz = nal*((nz+nal-1)/nal);
const int anx = nal*((nx+nal-1)/nal);
const int anb = nal*((2*nb+nal-1)/nal); // cache aligned number of box constraints
const int pad = (ncl-nx%ncl)%ncl; // packing between BAbtL & P
const int cnl = cnz<cnx+ncl ? nx+pad+cnx+ncl : nx+pad+cnz;
/************************************************
* dynamical system
************************************************/
double *A; d_zeros(&A, nx, nx); // states update matrix
double *B; d_zeros(&B, nx, nu); // inputs matrix
double *b; d_zeros(&b, nx, 1); // states offset
double *x0; d_zeros(&x0, nx, 1); // initial state
double Ts = 0.5; // sampling time
mass_spring_system(Ts, nx, nu, N, A, B, b, x0);
for(jj=0; jj<nx; jj++)
b[jj] = 0.1;
for(jj=0; jj<nx; jj++)
x0[jj] = 0;
x0[0] = 3.5;
x0[1] = 3.5;
// d_print_mat(nx, nx, A, nx);
// d_print_mat(nx, nu, B, nx);
// d_print_mat(nx, 1, b, nx);
// d_print_mat(nx, 1, x0, nx);
/* packed */
/* double *BAb; d_zeros(&BAb, nx, nz);*/
/* dmcopy(nx, nu, B, nx, BAb, nx);*/
/* dmcopy(nx, nx, A, nx, BAb+nu*nx, nx);*/
/* dmcopy(nx, 1 , b, nx, BAb+(nu+nx)*nx, nx);*/
/* transposed */
/* double *BAbt; d_zeros_align(&BAbt, pnz, pnz);*/
/* for(ii=0; ii<nx; ii++)*/
/* for(jj=0; jj<nz; jj++)*/
/* {*/
/* BAbt[jj+pnz*ii] = BAb[ii+nx*jj];*/
/* }*/
/* packed into contiguous memory */
float *pBAbt; s_zeros_align(&pBAbt, pnz, cnx);
/* d_cvt_mat2pmat(nz, nx, 0, bs, BAbt, pnz, pBAbt, cnx);*/
/* d_cvt_tran_mat2pmat(nx, nz, 0, bs, BAb, nx, pBAbt, cnx);*/
cvt_tran_d2s_mat2pmat(nx, nu, 0, bs, B, nx, pBAbt, cnx);
cvt_tran_d2s_mat2pmat(nx, nx, nu, bs, A, nx, pBAbt+nu/bs*cnx*bs+nu%bs, cnx);
for (jj = 0; jj<nx; jj++)
pBAbt[(nx+nu)/bs*cnx*bs+(nx+nu)%bs+jj*bs] = (float) b[jj];
/* s_print_pmat (nz, nx, bs, pBAbt, cnx);*/
/* exit(1);*/
/************************************************
* box constraints
************************************************/
/* double *db; d_zeros_align(&db, 2*nb, 1);*/
/* for(jj=0; jj<2*nu; jj++)*/
/* db[jj] = - 0.5; // umin*/
/* for(; jj<2*nb; jj++)*/
/* db[jj] = - 4.0; // xmin*/
float *lb; s_zeros_align(&lb, nx+nu, 1);
for(jj=0; jj<nu; jj++)
lb[jj] = - 0.5; // umin
for(; jj<nu+nx; jj++)
lb[jj] = - 4.0; // xmin
float *ub; s_zeros_align(&ub, nx+nu, 1);
for(jj=0; jj<nu; jj++)
ub[jj] = 0.5; // uman
for(; jj<nu+nx; jj++)
ub[jj] = 4.0; // xman
/************************************************
* cost function
************************************************/
float *Q; s_zeros_align(&Q, pnz, pnz);
for(ii=0; ii<nu; ii++) Q[ii*(pnz+1)] = 2.0;
for(; ii<pnz; ii++) Q[ii*(pnz+1)] = 1.0;
for(ii=0; ii<nz; ii++) Q[nx+nu+ii*pnz] = 0.1;
/* Q[(nx+nu)*(pnz+1)] = 1e35; // large enough (not needed any longer) */
/* packed into contiguous memory */
float *pQ; s_zeros_align(&pQ, pnz, cnz);
s_cvt_mat2pmat(nz, nz, 0, bs, Q, pnz, pQ, cnz);
/************************************************
* matrices series
************************************************/
float *(hpQ[N+1]);
float *(hq[N+1]);
float *(hux[N+1]);
float *(hpi[N+1]);
float *(hlam[N+1]);
float *(ht[N+1]);
float *(hpBAbt[N]);
float *(hlb[N+1]);
float *(hub[N+1]);
float *(hrb[N]);
float *(hrq[N+1]);
float *(hrd[N+1]);
float *(hux_v[N+1]);
float *(hux_w[N+1]);
for(jj=0; jj<N; jj++)
{
s_zeros_align(&hpQ[jj], pnz, cnz);
}
s_zeros_align(&hpQ[N], pnz, pnz);
for(jj=0; jj<N; jj++)
{
s_zeros_align(&hq[jj], anz, 1);
s_zeros_align(&hux[jj], anz, 1);
s_zeros_align(&hpi[jj], anx, 1);
s_zeros_align(&hlam[jj],anb, 1); // TODO pnb
s_zeros_align(&ht[jj], anb, 1); // TODO pnb
hpBAbt[jj] = pBAbt;
hlb[jj] = lb;
hub[jj] = ub;
s_zeros_align(&hrb[jj], anx, 1);
s_zeros_align(&hrq[jj], anz, 1);
s_zeros_align(&hrd[jj], anb, 1); // TODO pnb
s_zeros_align(&hux_v[jj], anz, 1);
s_zeros_align(&hux_w[jj], anz, 1);
}
s_zeros_align(&hq[N], anz, 1);
s_zeros_align(&hux[N], anz, 1);
s_zeros_align(&hpi[N], anx, 1);
s_zeros_align(&hlam[N], anb, 1); // TODO pnb
s_zeros_align(&ht[N], anb, 1); // TODO pnb
hlb[N] = lb;
hub[N] = ub;
s_zeros_align(&hrq[N], anz, 1);
s_zeros_align(&hrd[N], anb, 1); // TODO pnb
s_zeros_align(&hux_v[N], anz, 1);
s_zeros_align(&hux_w[N], anz, 1);
// starting guess
// for(jj=0; jj<nx; jj++) hux[0][nu+jj]=x0[jj];
/************************************************
* riccati-like iteration
************************************************/
float *work; s_zeros_align(&work, (N+1)*(pnz*cnl + 4*anz + 2*anx) + 3*anz, 1); // work space
int kk = 0; // acutal number of iterations
/* char prec = PREC; // double/single precision*/
/* float sp_thr = SP_THR; // threshold to switch between double and single precision*/
int k_max = K_MAX_ADMM; // maximum number of iterations in the ADMM method
float tol = TOL*sqrt(N*(nx+nu));//TOL; // tolerance in the duality measure
/* float sigma[] = {0.4, 0.3, 0.01}; // control primal-dual IP behaviour*/
float rho = 2.0; // penalty parameter
float alpha = 1.5; // relaxation parameter
float *stat; s_zeros(&stat, 5, k_max); // stats from the ADMM routine
int compute_mult = COMPUTE_MULT_ADMM;
int warm_start = 0;//WARM_START;
/* float mu = -1.0;*/
/* initizile the cost function */
for(ii=0; ii<N; ii++)
{
for(jj=0; jj<pnz*cnz; jj++) hpQ[ii][jj]=pQ[jj];
}
for(jj=0; jj<pnz*cnz; jj++) hpQ[N][jj]=pQ[jj];
// initial states
float xx0[] = {3.5, 3.5, 3.66465, 2.15833, 1.81327, -0.94207, 1.86531, -2.35760, 2.91534, 1.79890, -1.49600, -0.76600, -2.60268, 1.92456, 1.66630, -2.28522, 3.12038, 1.83830, 1.93519, -1.87113};
/* warm up */
// initialize states and inputs
for(ii=0; ii<=N; ii++)
for(jj=0; jj<nx+nu; jj++)
hux[ii][jj] = 0;
hux[0][nu+0] = xx0[0];
hux[0][nu+1] = xx0[1];
// call the ADMM solver
// if(FREE_X0==0)
// {
s_admm_box_mpc(&kk, k_max, tol, tol, warm_start, 1, rho, alpha, stat, nx, nu, N, hpBAbt, hpQ, hlb, hub, hux, hux_v, hux_w, compute_mult, hpi, work);
// }
// else
// {
///* d_ip_box_mhe(&kk, k_max, tol, warm_start, sigma, stat, nx, nu, N, nb, hpBAbt, hpQ, hdb, hux, compute_mult, hpi, hlam, ht, work);*/
// }
int kk_avg = 0;
/* timing */
struct timeval tv0, tv1;
gettimeofday(&tv0, NULL); // start
for(rep=0; rep<nrep; rep++)
{
idx = rep%10;
x0[0] = xx0[2*idx];
x0[1] = xx0[2*idx+1];
// initialize states and inputs
for(ii=0; ii<=N; ii++)
for(jj=0; jj<nx+nu; jj++)
hux[ii][jj] = 0;
hux[0][nu+0] = xx0[2*idx];
hux[0][nu+1] = xx0[2*idx+1];
// call the ADMM solver
// if(FREE_X0==0)
// {
s_admm_box_mpc(&kk, k_max, tol, tol, warm_start, 0, rho, alpha, stat, nx, nu, N, hpBAbt, hpQ, hlb, hub, hux, hux_v, hux_w, compute_mult, hpi, work);
// }
// else
// {
///* d_ip_box_mhe(&kk, k_max, tol, warm_start, sigma, stat, nx, nu, N, nb, hpBAbt, hpQ, hdb, hux, compute_mult, hpi, hlam, ht, work);*/
// }
kk_avg += kk;
}
gettimeofday(&tv1, NULL); // stop
float time = (tv1.tv_sec-tv0.tv_sec)/(nrep+0.0)+(tv1.tv_usec-tv0.tv_usec)/(nrep*1e6);
/* printf("\nnx\tnu\tN\tkernel\n\n");*/
/* printf("\n%d\t%d\t%d\t%e\n\n", nx, nu, N, time);*/
printf("\n");
printf(" Average number of iterations over %d runs: %5.1f\n", nrep, kk_avg / (float) nrep);
/* printf(" Average number of iterations over %d runs: %d\n", nrep, kk);*/
printf("\n");
printf(" Average solution time over %d runs: %5.2e seconds\n", nrep, time);
printf("\n");
// restore linear part of cost function
for(ii=0; ii<N; ii++)
{
for(jj=0; jj<nx+nu; jj++) hq[ii][jj] = Q[nx+nu+pnz*jj];
}
for(jj=0; jj<nx+nu; jj++) hq[N][jj] = Q[nx+nu+pnz*jj];
// residuals computation
/* if(FREE_X0==0)*/
/* d_res_ip_box_mpc(nx, nu, N, nb, hpBAbt, hpQ, hq, hux, hdb, hpi, hlam, ht, hrq, hrb, hrd, &mu);*/
/* else*/
/* d_res_ip_box_mhe(nx, nu, N, nb, hpBAbt, hpQ, hq, hux, hdb, hpi, hlam, ht, hrq, hrb, hrd, &mu);*/
if(PRINTSTAT==1)
{
printf("\n");
printf("\n");
printf(" Print ADMM statistics of the last run\n");
printf("\n");
for(jj=0; jj<kk; jj++)
printf("k = %d\t\tp_res = %f\td_res = %f\n", jj, stat[5*jj], stat[5*jj+1]);
printf("\n");
}
if(PRINTRES==1)
{
printf("\n");
printf("\n");
printf(" Print solution\n");
printf("\n");
printf("\nu = \n\n");
for(ii=0; ii<N; ii++)
s_print_mat(1, nu, hux[ii], 1);
}
if(0 && PRINTRES==1 && COMPUTE_MULT_ADMM==1)
{
// print result
// print result
printf("\n");
printf("\n");
printf(" Print residuals\n\n");
printf("\n");
printf("\n");
printf("rq = \n\n");
// if(FREE_X0==0)
// {
s_print_mat(1, nu, hrq[0], 1);
for(ii=1; ii<=N; ii++)
/* s_print_mat_e(1, nx+nu, hrq[ii], 1);*/
s_print_mat(1, nx+nu, hrq[ii], 1);
// }
// else
// {
// for(ii=0; ii<=N; ii++)
///* s_print_mat_e(1, nx+nu, hrq[ii], 1);*/
// s_print_mat(1, nx+nu, hrq[ii], 1);
// }
printf("\n");
printf("\n");
printf("rb = \n\n");
for(ii=0; ii<N; ii++)
/* s_print_mat_e(1, nx, hrb[ii], 1);*/
s_print_mat(1, nx, hrb[ii], 1);
printf("\n");
printf("\n");
printf("rd = \n\n");
for(ii=0; ii<=N; ii++)
/* s_print_mat_e(1, 2*nb, hrd[ii], 1);*/
s_print_mat(1, 2*nb, hrd[ii], 1);
printf("\n");
printf("\n");
/* printf("mu = %e\n\n", mu);*/
}
/* printf("\nnx\tnu\tN\tkernel\n\n");*/
/* printf("\n%d\t%d\t%d\t%e\n\n", nx, nu, N, time);*/
/************************************************
* free memory and return
************************************************/
free(A);
free(B);
free(b);
free(x0);
/* free(BAb);*/
/* free(BAbt);*/
free(pBAbt);
free(lb);
free(ub);
free(Q);
free(pQ);
free(work);
free(stat);
for(jj=0; jj<N; jj++)
{
free(hpQ[jj]);
free(hq[jj]);
free(hux[jj]);
free(hpi[jj]);
free(hlam[jj]);
free(ht[jj]);
free(hrb[jj]);
free(hrq[jj]);
free(hrd[jj]);
free(hux_v[jj]);
free(hux_w[jj]);
}
free(hpQ[N]);
free(hq[N]);
free(hux[N]);
free(hpi[N]);
free(hlam[N]);
free(ht[N]);
free(hrq[N]);
free(hrd[N]);
free(hux_v[N]);
free(hux_w[N]);
return 0;
}
