path: root/init.c

/*
 * Copyright 2017 Anton Khirnov <anton@khirnov.net>
 *
 * This program is free software: you can redistribute it and/or modify
 * it under the terms of the GNU General Public License as published by
 * the Free Software Foundation, either version 3 of the License, or
 * (at your option) any later version.
 *
 * This program is distributed in the hope that it will be useful,
 * but WITHOUT ANY WARRANTY; without even the implied warranty of
 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
 * GNU General Public License for more details.
 *
 * You should have received a copy of the GNU General Public License
 * along with this program.  If not, see <http://www.gnu.org/licenses/>.
 */

#include "config.h"

#include <errno.h>
#include <float.h>
#include <math.h>
#include <stdlib.h>
#include <string.h>
#include <stdio.h>

#include <cblas.h>

#if HAVE_OPENCL
#include <cl.h>
#include <clBLAS.h>
#endif

#include "basis.h"
#include "common.h"
#include "cpu.h"
#include "log.h"
#include "nlsolve.h"
#include "td_constraints.h"
#include "teukolsky_data.h"
#include "threadpool.h"

#define NB_EQUATIONS 3

typedef struct TDPriv {
    unsigned int basis_order[NB_EQUATIONS][2];
    BasisSetContext *basis[NB_EQUATIONS][2];

    NLSolveContext *solver;

    ThreadPoolContext *tp;
    TDLogger logger;

    double *coeffs;
    double *coeffs_tmp;
    ptrdiff_t coeffs_stride;

    int cpu_flags;

    double (*scalarproduct_metric)(size_t len1, size_t len2, double *mat,
                                   double *vec1, double *vec2);
} TDPriv;

double tdi_scalarproduct_metric_fma3(size_t len1, size_t len2, double *mat,
                                     double *vec1, double *vec2);
double tdi_scalarproduct_metric_avx(size_t len1, size_t len2, double *mat,
                                    double *vec1, double *vec2);
double tdi_scalarproduct_metric_sse3(size_t len1, size_t len2, double *mat,
                                     double *vec1, double *vec2);
double tdi_scalarproduct_metric_c(size_t len1, size_t len2, double *mat,
                                  double *vec1, double *vec2);

static void init_opencl(TDPriv *s)
#if HAVE_OPENCL
{
    int err, count;
    cl_platform_id platform;
    cl_context_properties props[3];
    cl_device_id ocl_device;

    err = clGetPlatformIDs(1, &platform, &count);
    if (err != CL_SUCCESS || count < 1) {
        tdi_log(&s->logger, 0, "Could not get an OpenCL platform ID\n");
        return;
    }

    err = clGetDeviceIDs(platform, CL_DEVICE_TYPE_GPU, 1, &ocl_device, &count);
    if (err != CL_SUCCESS || count < 1) {
        tdi_log(&s->logger, 0, "Could not get an OpenCL device ID\n");
        return;
    }

    props[0] = CL_CONTEXT_PLATFORM;
    props[1] = (cl_context_properties)platform;
    props[2] = 0;

    s->ocl_ctx = clCreateContext(props, 1, &ocl_device, NULL, NULL, &err);
    if (err != CL_SUCCESS || !s->ocl_ctx) {
        tdi_log(&s->logger, 0, "Could not create an OpenCL context\n");
        return;
    }

    s->ocl_queue = clCreateCommandQueue(s->ocl_ctx, ocl_device, 0, &err);
    if (err != CL_SUCCESS || !s->ocl_queue) {
        tdi_log(&s->logger, 0, "Could not create an OpenCL command queue: %d\n", err);
        goto fail;
    }

    err = clblasSetup();
    if (err != CL_SUCCESS) {
        tdi_log(&s->logger, 0, "Error setting up clBLAS\n");
        goto fail;
    }

    return;
fail:
    if (s->ocl_queue)
        clReleaseCommandQueue(s->ocl_queue);
    s->ocl_queue = 0;

    if (s->ocl_ctx)
        clReleaseContext(s->ocl_ctx);
    s->ocl_ctx = 0;
}
#else
{
}
#endif

static const enum BasisFamily basis_sets[NB_EQUATIONS][2] = {
    { BASIS_FAMILY_SB_EVEN,  BASIS_FAMILY_COS_EVEN },
    { BASIS_FAMILY_SB_EVEN,  BASIS_FAMILY_COS_EVEN },
    { BASIS_FAMILY_SB_EVEN,  BASIS_FAMILY_COS_EVEN },
};

static void log_callback(TDLogger *log, int level, const char *fmt, va_list vl)
{
    TDContext *td = log->opaque;
    td->log_callback(td, level, fmt, vl);
}

static int teukolsky_init_check_options(TDContext *td)
{
    TDPriv *s = td->priv;
    int ret;

    s->cpu_flags = tdi_init_cpu_flags();

    if (!td->nb_threads) {
        td->nb_threads = tdi_cpu_count();
        if (!td->nb_threads)
            td->nb_threads = 1;
    }

    ret = tdi_threadpool_init(&s->tp, td->nb_threads);
    if (ret < 0)
        return ret;

    init_opencl(s);

    s->logger.log    = log_callback;
    s->logger.opaque = td;

    //s->scalarproduct_metric = tdi_scalarproduct_metric_c;
    //if (EXTERNAL_SSE3(s->cpu_flags))
    //    s->scalarproduct_metric = tdi_scalarproduct_metric_sse3;
    //if (EXTERNAL_AVX(s->cpu_flags))
    //    s->scalarproduct_metric = tdi_scalarproduct_metric_avx;
    //if (EXTERNAL_FMA3(s->cpu_flags))
    //    s->scalarproduct_metric = tdi_scalarproduct_metric_fma3;
//    for (int i = 0; i < ctx->nb_equations; i++)
//        for (int j = 0; j < 2; j++) {
//            s->ps_ctx->basis[i][j]       = ctx->basis[i][j];
//            s->ps_ctx->solve_order[i][j] = basis_order[i][j];
//            max_order = MAX(max_order, basis_order[i][j]);
//        }
//
    //for (int i = 0; i < max_order; i++) {
    //    tdi_log(&s->logger, 2, "%d ", i);
    //    for (int j = 0; j < ctx->nb_equations; j++)
    //        for (int k = 0; k < 2; k++) {
    //            if (i < s->ps_ctx->solve_order[j][k])
    //                tdi_log(&s->logger, 2, "%8.8g\t", s->ps_ctx->colloc_grid[j][k][i]);
    //            else
    //                tdi_log(&s->logger, 2, "        ");
    //        }
    //    tdi_log(&s->logger, 2, "\n");
    //}

    s->basis_order[0][0] = td->nb_coeffs[0];
    s->basis_order[0][1] = td->nb_coeffs[1];
    s->basis_order[1][0] = td->nb_coeffs[0];
    s->basis_order[1][1] = td->nb_coeffs[1];
    s->basis_order[2][0] = td->nb_coeffs[0];
    s->basis_order[2][1] = td->nb_coeffs[1];

    ret = posix_memalign((void**)&s->coeffs, 32,
                         sizeof(*s->coeffs) * NB_EQUATIONS * td->nb_coeffs[0] * td->nb_coeffs[1]);
    if (ret)
        return -ENOMEM;
    memset(s->coeffs, 0, sizeof(*s->coeffs) * NB_EQUATIONS * td->nb_coeffs[0] * td->nb_coeffs[1]);

    if (td->solution_branch > 0) {
        ret = posix_memalign((void**)&s->coeffs_tmp, 32,
                             sizeof(*s->coeffs_tmp) * NB_EQUATIONS * td->nb_coeffs[0] * td->nb_coeffs[1]);
        if (ret)
            return -ENOMEM;
        memset(s->coeffs_tmp, 0, sizeof(*s->coeffs_tmp) * NB_EQUATIONS * td->nb_coeffs[0] * td->nb_coeffs[1]);
    }

    for (int i = 0; i < NB_EQUATIONS; i++)
        td->coeffs[i] = s->coeffs + i * td->nb_coeffs[0] * td->nb_coeffs[1];

    for (int i = 0; i < ARRAY_ELEMS(basis_sets); i++)
        for (int j = 0; j < ARRAY_ELEMS(basis_sets[i]); j++) {
            double sf;

            ret = tdi_basis_init(&s->basis[i][j], basis_sets[i][j], td->basis_scale_factor[j]);
            if (ret < 0)
                return ret;
        }


    ret = tdi_nlsolve_context_alloc(&s->solver, ARRAY_ELEMS(basis_sets));
    if (ret < 0) {
        tdi_log(&s->logger, 0, "Error allocating the non-linear solver\n");
        return ret;
    }

    s->solver->logger  = s->logger;
    s->solver->tp      = s->tp;
    s->solver->maxiter = td->max_iter;
    s->solver->atol    = td->atol;

    memcpy(s->solver->basis,       s->basis,       sizeof(s->basis));
    memcpy(s->solver->solve_order, s->basis_order, sizeof(s->basis_order));

#if HAVE_OPENCL
    s->solver->ocl_ctx   = s->ocl_ctx;
    s->solver->ocl_queue = s->ocl_queue;
#endif

    ret = tdi_nlsolve_context_init(s->solver);
    if (ret < 0) {
        tdi_log(&s->logger, 0, "Error initializing the non-linear solver\n");
        return ret;
    }

    return 0;
}

static int constraint_eval_alloc(const TDContext *td, double amplitude,
                                 TDConstraintEvalContext **pce)
{
    TDPriv *priv = td->priv;
    TDConstraintEvalContext *ce;
    int ret;

    ret = tdi_constraint_eval_alloc(&ce, td->family);
    if (ret < 0) {
        tdi_log(&priv->logger, 0, "Error allocating the constraints evaluator\n");
        return ret;
    }

    ce->logger       = priv->logger;
    ce->amplitude    = amplitude;
    ce->nb_coords[0] = td->nb_coeffs[0];
    ce->nb_coords[1] = td->nb_coeffs[1];
    ce->coords[0]    = priv->solver->colloc_grid[0][0];
    ce->coords[1]    = priv->solver->colloc_grid[0][1];

    ret = tdi_constraint_eval_init(ce);
    if (ret < 0) {
        tdi_constraint_eval_free(&ce);
        tdi_log(&priv->logger, 0, "Error initializing the constraints evaluator\n");
        return ret;
    }

    *pce = ce;
    return 0;
}

static int teukolsky_constraint_eval(void *opaque, unsigned int eq_idx,
                                     const unsigned int *colloc_grid_order,
                                     const double * const *colloc_grid,
                                     const double * const (*vars)[PSSOLVE_DIFF_ORDER_NB],
                                     double *dst)
{
    TDConstraintEvalContext *ce = opaque;
    const double *amplitude = opaque;

    return tdi_constraint_eval(ce, eq_idx, vars, dst);
}

int td_solve(TDContext *td, double *coeffs_init[3])
{
    TDPriv *s = td->priv;
    TDConstraintEvalContext *ce;
    double a0;
    int ret;

    ret = teukolsky_init_check_options(td);
    if (ret < 0)
        return ret;

    ret = constraint_eval_alloc(td, 0.0, &ce);
    if (ret < 0)
        return ret;
    if (fabs(td->amplitude) >= ce->a_diverge) {
        tdi_log(&s->logger, 0,
                "Amplitude A=%16.16g is above the point A_{max}=%g, no solutions "
                "are known to exist there. Set solution_branch=1 to get to the "
                "second branch where mass increases with decreasing amplitude\n",
                td->amplitude, ce->a_converge);
    }
    a0 = ce->a_converge;
    tdi_constraint_eval_free(&ce);

    if (coeffs_init) {
        for (int i = 0; i < 3; i++) {
            memcpy(td->coeffs[i], coeffs_init[i], sizeof(*td->coeffs[i]) *
                   td->nb_coeffs[0] * td->nb_coeffs[1]);
        }
    }

    if (td->solution_branch == 0 || coeffs_init) {
        // direct solve with default (flat space) or user-provided initial guess
        ret = constraint_eval_alloc(td, td->amplitude, &ce);
        if (ret < 0)
            return ret;

        ret = tdi_nlsolve_solve(s->solver, teukolsky_constraint_eval, NULL,
                                ce, s->coeffs, 0);
        tdi_constraint_eval_free(&ce);
        if (ret < 0) {
            tdi_log(&s->logger, 0, "tdi_nlsolve_solve() failed: %d", ret);
            return ret;
        }
    } else {
        // second branch requested and no user-provided initial guess
        // execute two lower-branch solutions and extrapolate to get to the upper branch
        const int N = NB_EQUATIONS * td->nb_coeffs[0] * td->nb_coeffs[1];
        const double delta = 1e-5;
        const double a1 = a0 - delta;

        double cur_amplitude, new_amplitude, inverse_step;
        int dir;

        ret = constraint_eval_alloc(td, a0, &ce);
        if (ret < 0)
            return ret;

        ret = tdi_nlsolve_solve(s->solver, teukolsky_constraint_eval, NULL,
                                ce, s->coeffs, 0);
        tdi_constraint_eval_free(&ce);
        if (ret < 0)
            return ret;

        ret = constraint_eval_alloc(td, a1, &ce);
        if (ret < 0)
            return ret;
        ret = tdi_nlsolve_solve(s->solver, teukolsky_constraint_eval, NULL,
                                ce, s->coeffs_tmp, 0);
        tdi_constraint_eval_free(&ce);
        if (ret < 0)
            return ret;

        cblas_daxpy(N, -1.0, s->coeffs,     1, s->coeffs_tmp, 1);
        cblas_daxpy(N, -1.0, s->coeffs_tmp, 1, s->coeffs,     1);

        // obtain solution for a1 in the upper branch
        ret = constraint_eval_alloc(td, a1, &ce);
        if (ret < 0)
            return ret;

        ret = tdi_nlsolve_solve(s->solver, teukolsky_constraint_eval, NULL,
                                ce, s->coeffs, 0);
        tdi_constraint_eval_free(&ce);
        if (ret < 0) {
            tdi_log(&s->logger, 0, "Failed to get into the upper branch\n");
            return ret;
        }

        cur_amplitude = a1;
        dir = SGN(td->amplitude - cur_amplitude);
        inverse_step = 1.0 / td->amplitude - 1.0 / cur_amplitude;
        while (fabs(cur_amplitude - td->amplitude) > DBL_EPSILON) {
            //double scale_factor;
            //scale_factor  = cur_amplitude;
            //cur_amplitude = MIN(td->amplitude, 2 * cur_amplitude);
            //scale_factor  = cur_amplitude / scale_factor;

            new_amplitude = 1.0 / ((1.0 / cur_amplitude) + inverse_step);
            if (dir * (td->amplitude - new_amplitude) < 0)
                new_amplitude = td->amplitude;
            tdi_log(&s->logger, 2, "Trying amplitude %g\n", new_amplitude);

            ret = constraint_eval_alloc(td, new_amplitude, &ce);
            if (ret < 0)
                return ret;

            memcpy(s->coeffs_tmp, s->coeffs, sizeof(*s->coeffs) * N);
            ret = tdi_nlsolve_solve(s->solver, teukolsky_constraint_eval, NULL,
                                    ce, s->coeffs_tmp, 1);
            tdi_constraint_eval_free(&ce);
            if (ret == -EDOM) {
                inverse_step = 0.5 * inverse_step;
                //if (fabs(inverse_step) < 1e-2)
                //    return ret;
                continue;
            } else if (ret < 0)
                return ret;
            cur_amplitude = new_amplitude;
            memcpy(s->coeffs, s->coeffs_tmp, sizeof(*s->coeffs) * N);
        }
        //while (1) {


        //    if (fabs(cur_amplitude - td->amplitude) < DBL_EPSILON)
        //        goto finish;

        //    cblas_dscal(td->nb_coeffs[0] * td->nb_coeffs[1], SQR(scale_factor), td->coeffs[0], 1);
        //    cblas_dscal(td->nb_coeffs[0] * td->nb_coeffs[1], scale_factor, td->coeffs[1], 1);
        //    cblas_dscal(td->nb_coeffs[0] * td->nb_coeffs[1], scale_factor, td->coeffs[2], 1);

        //}

    }
finish:
    tdi_nlsolve_print_stats(s->solver);

    return 0;
}

static void log_default_callback(const TDContext *td, int level, const char *fmt, va_list vl)
{
    vfprintf(stderr, fmt, vl);
}

TDContext *td_context_alloc(void)
{
    TDContext *td = calloc(1, sizeof(*td));

    if (!td)
        return NULL;

    td->nb_threads = 1;

    td->family          = TD_FAMILY_SIMPLE_TIME_ANTISYM;
    td->solution_branch = 0;
    td->amplitude       = 1.0;

    td->nb_coeffs[0] = 32;
    td->nb_coeffs[1] = 16;

    td->basis_scale_factor[0] = 3.0;
    td->basis_scale_factor[1] = 3.0;

    td->max_iter = 16;
    td->atol     = 1e-12;

    td->log_callback = log_default_callback;

    td->priv = calloc(1, sizeof(TDPriv));
    if (!td->priv) {
        free(td);
        return NULL;
    }

    return td;
}

void td_context_free(TDContext **ptd)
{
    TDContext *td = *ptd;
    TDPriv *s;

    if (!td)
        return;

    s = td->priv;

    tdi_nlsolve_context_free(&s->solver);
    tdi_threadpool_free(&s->tp);

#if HAVE_OPENCL
    if (s->ocl_queue)
        clReleaseCommandQueue(s->ocl_queue);
    if (s->ocl_ctx)
        clReleaseContext(s->ocl_ctx);
#endif

    for (int i = 0; i < ARRAY_ELEMS(s->basis); i++)
        for (int j = 0; j < ARRAY_ELEMS(s->basis[i]); j++)
            tdi_basis_free(&s->basis[i][j]);

    free(s->coeffs);
    free(s->coeffs_tmp);

    free(td->priv);
    free(td);
    *ptd = NULL;
}

static double scalarproduct_metric_c(size_t len1, size_t len2, double *mat,
                                     double *vec1, double *vec2)
{
    double val = 0.0;
    for (int m = 0; m < len2; m++) {
        double tmp = 0.0;
        for (int n = 0; n < len1; n++)
            tmp += mat[m * len1 + n] * vec1[n];
        val += tmp * vec2[m];
    }
    return val;
}

static int eval_var(const TDContext *td, unsigned int var_idx,
                    size_t nb_coords, const double *r, const double *theta,
                    const unsigned int diff_order[2],
                    double *out)
{
    TDPriv *s = td->priv;
    double *basis_val[2] = { NULL };
    int ret = 0;

    if (diff_order[0] > 0 || diff_order[1] > 0) {
        tdi_log(&s->logger, 0, "Derivatives of higher order than 2 are not supported.\n");
        return -ENOSYS;
    }

    posix_memalign((void**)&basis_val[0], 32, sizeof(*basis_val[0]) * td->nb_coeffs[0]);
    posix_memalign((void**)&basis_val[1], 32, sizeof(*basis_val[1]) * td->nb_coeffs[1]);
    if (!basis_val[0] || !basis_val[1]) {
        ret = -ENOMEM;
        goto fail;
    }
    memset(basis_val[0], 0, sizeof(*basis_val[0]) * td->nb_coeffs[0]);
    memset(basis_val[1], 0, sizeof(*basis_val[1]) * td->nb_coeffs[1]);

    for (int i = 0; i < nb_coords; i++) {
        double theta_val = theta[i];
        double r_val     = r[i];

        double val = (var_idx == 0) ? 1.0 : 0.0;

        for (int k = 0; k < td->nb_coeffs[0]; k++)
            basis_val[0][k] = tdi_basis_eval(s->basis[var_idx][0], BS_EVAL_TYPE_VALUE, r_val, k);
        for (int k = 0; k < td->nb_coeffs[1]; k++)
            basis_val[1][k] = tdi_basis_eval(s->basis[var_idx][1], BS_EVAL_TYPE_VALUE, theta_val, k);

        val += scalarproduct_metric_c(td->nb_coeffs[0], td->nb_coeffs[1], td->coeffs[var_idx],
                                      basis_val[0], basis_val[1]);

        out[i] = val;
    }


fail:
    free(basis_val[0]);
    free(basis_val[1]);

    return ret;
}

int td_eval_psi(const TDContext *td,
                size_t nb_coords, const double *r, const double *theta,
                const unsigned int diff_order[2],
                double *out)
{
    return eval_var(td, 0, nb_coords, r, theta, diff_order, out);
}

int td_eval_krr(const TDContext *td,
                size_t nb_coords, const double *r, const double *theta,
                const unsigned int diff_order[2],
                double *out)
{
    return eval_var(td, 1, nb_coords, r, theta, diff_order, out);
}
int td_eval_kpp(const TDContext *td,
                size_t nb_coords, const double *r, const double *theta,
                const unsigned int diff_order[2],
                double *out)
{
    return eval_var(td, 2, nb_coords, r, theta, diff_order, out);
}

int td_eval_krt(const TDContext *td,
                size_t nb_coords, const double *r, const double *theta,
                const unsigned int diff_order[2],
                double *out)
{
    TDConstraintEvalContext *ce;
    int ret;

    if (diff_order[0] || diff_order[1])
        return -ENOSYS;

    ret = constraint_eval_alloc(td, td->amplitude, &ce);
    if (ret < 0)
        return ret;

    for (int i = 0; i < nb_coords; i++) {
        double theta_val = theta[i];
        double r_val     = r[i];

        out[i] = tdi_constraint_eval_k_rtheta(ce, r_val, theta_val);
    }

    tdi_constraint_eval_free(&ce);

    return 0;
}