#!/usr/bin/env python3
import os
import time
import numpy as np
from cereal import log
from iqdbc.car.interfaces import ACCEL_MIN, ACCEL_MAX
from openpilot.common.realtime import DT_MDL
from openpilot.common.swaglog import cloudlog
# WARNING: imports outside of constants will not trigger a rebuild
from openpilot.selfdrive.modeld.constants import index_function, ModelConstants
from openpilot.selfdrive.controls.radard import _LEAD_ACCEL_TAU  # legacy lead extrapolation (newLeadMpc=False)
from openpilot.common.params import Params, UnknownKeyName

LEAD_T_IDXS_MODEL = np.array(ModelConstants.LEAD_T_IDXS)  # [0, 2, 4, 6, 8, 10]s

if __name__ == '__main__':  # generating code
  from openpilot.third_party.acados.acados_template import AcadosModel, AcadosOcp, AcadosOcpSolver
else:
  from openpilot.selfdrive.controls.lib.longitudinal_mpc_lib.c_generated_code.acados_ocp_solver_pyx import AcadosOcpSolverCython

from casadi import SX, vertcat

MODEL_NAME = 'long'
LONG_MPC_DIR = os.path.dirname(os.path.abspath(__file__))
EXPORT_DIR = os.path.join(LONG_MPC_DIR, "c_generated_code")
JSON_FILE = os.path.join(LONG_MPC_DIR, "acados_ocp_long.json")

LongitudinalPlanSource = log.LongitudinalPlan.LongitudinalPlanSource
MPC_SOURCES = (LongitudinalPlanSource.lead0, LongitudinalPlanSource.lead1)

X_DIM = 3
U_DIM = 1
PARAM_DIM = 5
COST_E_DIM = 4
COST_DIM = COST_E_DIM + 1
CONSTR_DIM = 4

X_EGO_OBSTACLE_COST = 3.
X_EGO_COST = 0.
V_EGO_COST = 0.
A_EGO_COST = 0.
J_EGO_COST = 20.
DANGER_ZONE_COST = 100.
CRASH_DISTANCE = .25
LEAD_DANGER_FACTOR = 0.75
LIMIT_COST = 1e6
ACADOS_SOLVER_TYPE = 'SQP_RTI'

# Fewer timestamps don't hurt performance and lead to
# much better convergence of the MPC with low iterations
N = 12
MAX_T = 10.0
T_IDXS_LST = [index_function(idx, max_val=MAX_T, max_idx=N) for idx in range(N+1)]

T_IDXS = np.array(T_IDXS_LST)
FCW_IDXS = T_IDXS < 5.0
T_DIFFS = np.diff(T_IDXS, prepend=[0.])
COMFORT_BRAKE = 2.5
STOP_DISTANCE = 3.0
MIN_X_LEAD_FACTOR = 0.5
LEAD_PULLAWAY_VREL = 0.5
LEAD_PULLAWAY_ABRAKE = -0.5

def get_jerk_factor(personality=log.LongitudinalPersonality.standard):
  if personality==log.LongitudinalPersonality.relaxed:
    return 1.0
  elif personality==log.LongitudinalPersonality.standard:
    return 1.0
  elif personality==log.LongitudinalPersonality.aggressive:
    return 0.5
  else:
    raise NotImplementedError("Longitudinal personality not supported")


def get_T_FOLLOW(personality=log.LongitudinalPersonality.standard):
  if personality==log.LongitudinalPersonality.relaxed:
    return 1.75
  elif personality==log.LongitudinalPersonality.standard:
    return 1.45
  elif personality==log.LongitudinalPersonality.aggressive:
    return 1.25
  else:
    raise NotImplementedError("Longitudinal personality not supported")

def get_stopped_equivalence_factor(v_lead):
  return (v_lead**2) / (2 * COMFORT_BRAKE)

def get_safe_obstacle_distance(v_ego, t_follow):
  return (v_ego**2) / (2 * COMFORT_BRAKE) + t_follow * v_ego + STOP_DISTANCE

def gen_long_model():
  model = AcadosModel()
  model.name = MODEL_NAME

  # states
  x_ego, v_ego, a_ego = SX.sym('x_ego'), SX.sym('v_ego'), SX.sym('a_ego')
  model.x = vertcat(x_ego, v_ego, a_ego)

  # controls
  j_ego = SX.sym('j_ego')
  model.u = vertcat(j_ego)

  # xdot
  x_ego_dot = SX.sym('x_ego_dot')
  v_ego_dot = SX.sym('v_ego_dot')
  a_ego_dot = SX.sym('a_ego_dot')
  model.xdot = vertcat(x_ego_dot, v_ego_dot, a_ego_dot)

  # live parameters
  a_min = SX.sym('a_min')
  a_max = SX.sym('a_max')
  x_obstacle = SX.sym('x_obstacle')
  lead_t_follow = SX.sym('lead_t_follow')
  lead_danger_factor = SX.sym('lead_danger_factor')
  model.p = vertcat(a_min, a_max, x_obstacle, lead_t_follow, lead_danger_factor)

  # dynamics model
  f_expl = vertcat(v_ego, a_ego, j_ego)
  model.f_impl_expr = model.xdot - f_expl
  model.f_expl_expr = f_expl
  return model

def gen_long_ocp():
  ocp = AcadosOcp()
  ocp.model = gen_long_model()

  Tf = T_IDXS[-1]

  # set dimensions
  ocp.dims.N = N

  # set cost module
  ocp.cost.cost_type = 'NONLINEAR_LS'
  ocp.cost.cost_type_e = 'NONLINEAR_LS'

  QR = np.zeros((COST_DIM, COST_DIM))
  Q = np.zeros((COST_E_DIM, COST_E_DIM))

  ocp.cost.W = QR
  ocp.cost.W_e = Q

  x_ego, v_ego, a_ego = ocp.model.x[0], ocp.model.x[1], ocp.model.x[2]
  j_ego = ocp.model.u[0]

  a_min, a_max = ocp.model.p[0], ocp.model.p[1]
  x_obstacle = ocp.model.p[2]
  lead_t_follow = ocp.model.p[3]
  lead_danger_factor = ocp.model.p[4]

  ocp.cost.yref = np.zeros((COST_DIM, ))
  ocp.cost.yref_e = np.zeros((COST_E_DIM, ))

  desired_dist_comfort = get_safe_obstacle_distance(v_ego, lead_t_follow)

  # The main cost in normal operation is how close you are to the "desired" distance
  # from an obstacle at every timestep. This obstacle can be a lead car
  # or other object. In e2e mode we can use x_position targets as a cost
  # instead.
  costs = [((x_obstacle - x_ego) - (desired_dist_comfort)) / (v_ego + 10.),
           x_ego,
           v_ego,
           a_ego,
           j_ego]
  ocp.model.cost_y_expr = vertcat(*costs)
  ocp.model.cost_y_expr_e = vertcat(*costs[:-1])

  # Constraints on speed, acceleration and desired distance to
  # the obstacle, which is treated as a slack constraint so it
  # behaves like an asymmetrical cost.
  constraints = vertcat(v_ego,
                        (a_ego - a_min),
                        (a_max - a_ego),
                        ((x_obstacle - x_ego) - lead_danger_factor * (desired_dist_comfort)) / (v_ego + 10.))
  ocp.model.con_h_expr = constraints

  x0 = np.zeros(X_DIM)
  ocp.constraints.x0 = x0
  ocp.parameter_values = np.array([-1.2, 1.2, 0.0, get_T_FOLLOW(), LEAD_DANGER_FACTOR])


  # We put all constraint cost weights to 0 and only set them at runtime
  cost_weights = np.zeros(CONSTR_DIM)
  ocp.cost.zl = cost_weights
  ocp.cost.Zl = cost_weights
  ocp.cost.Zu = cost_weights
  ocp.cost.zu = cost_weights

  ocp.constraints.lh = np.zeros(CONSTR_DIM)
  ocp.constraints.uh = 1e4*np.ones(CONSTR_DIM)
  ocp.constraints.idxsh = np.arange(CONSTR_DIM)

  # The HPIPM solver can give decent solutions even when it is stopped early
  # Which is critical for our purpose where compute time is strictly bounded
  # We use HPIPM in the SPEED_ABS mode, which ensures fastest runtime. This
  # does not cause issues since the problem is well bounded.
  ocp.solver_options.qp_solver = 'PARTIAL_CONDENSING_HPIPM'
  ocp.solver_options.hessian_approx = 'GAUSS_NEWTON'
  ocp.solver_options.integrator_type = 'ERK'
  ocp.solver_options.nlp_solver_type = ACADOS_SOLVER_TYPE
  ocp.solver_options.qp_solver_cond_N = 1

  # More iterations take too much time and less lead to inaccurate convergence in
  # some situations. Ideally we would run just 1 iteration to ensure fixed runtime.
  ocp.solver_options.qp_solver_iter_max = 10
  ocp.solver_options.qp_tol = 1e-3

  # set prediction horizon
  ocp.solver_options.tf = Tf
  ocp.solver_options.shooting_nodes = T_IDXS

  ocp.code_export_directory = EXPORT_DIR
  return ocp


class LongitudinalMpc:
  def __init__(self, dt=DT_MDL):
    self.dt = dt
    self._params = Params()
    self.new_lead_mpc = self._read_new_lead_mpc()
    self.solver = AcadosOcpSolverCython(MODEL_NAME, ACADOS_SOLVER_TYPE, N)
    self.reset()
    self.source = LongitudinalPlanSource.cruise

  def _read_new_lead_mpc(self) -> bool:
    try:
      return self._params.get_bool("newLeadMpc")
    except UnknownKeyName:
      return True

  def reset(self):
    self.solver.reset()

    self.x_sol = np.zeros((N+1, X_DIM))
    self.u_sol = np.zeros((N, 1))
    self.v_solution = np.zeros(N+1)
    self.a_solution = np.zeros(N+1)
    self.j_solution = np.zeros(N)
    self.yref = np.zeros((N+1, COST_DIM))

    for i in range(N):
      self.solver.cost_set(i, "yref", self.yref[i])
    self.solver.cost_set(N, "yref", self.yref[N][:COST_E_DIM])

    self.params = np.zeros((N+1, PARAM_DIM))
    for i in range(N+1):
      self.solver.set(i, 'x', np.zeros(X_DIM))

    self.last_cloudlog_t = 0
    self.status = False
    self.crash_cnt = 0.0
    self.solution_status = 0
    # timers
    self.solve_time = 0.0
    self.time_qp_solution = 0.0
    self.time_linearization = 0.0
    self.time_integrator = 0.0
    self.x0 = np.zeros(X_DIM)
    self.lead_xv_0 = np.zeros((N+1, 2))
    self.lead_xv_1 = np.zeros((N+1, 2))
    self.set_weights()

  def set_cost_weights(self, cost_weights, constraint_cost_weights):
    W = np.asfortranarray(np.diag(cost_weights))
    for i in range(N):
      self.solver.cost_set(i, 'W', W)
    # Setting the slice without the copy make the array not contiguous,
    # causing issues with the C interface.
    self.solver.cost_set(N, 'W', np.copy(W[:COST_E_DIM, :COST_E_DIM]))

    # Set L2 slack cost on lower bound constraints
    Zl = np.array(constraint_cost_weights)
    for i in range(N):
      self.solver.cost_set(i, 'Zl', Zl)

  def set_weights(self, personality=log.LongitudinalPersonality.standard):
    jerk_factor = get_jerk_factor(personality)
    cost_weights = [X_EGO_OBSTACLE_COST, X_EGO_COST, V_EGO_COST, A_EGO_COST, jerk_factor * J_EGO_COST]
    constraint_cost_weights = [LIMIT_COST, LIMIT_COST, LIMIT_COST, DANGER_ZONE_COST]
    self.set_cost_weights(cost_weights, constraint_cost_weights)

  def set_cur_state(self, v, a):
    v_prev = self.x0[1]
    self.x0[1] = v
    self.x0[2] = a
    if abs(v_prev - v) > 2.:  # probably only helps if v < v_prev
      for i in range(N+1):
        self.solver.set(i, 'x', self.x0)

  @staticmethod
  def extrapolate_lead(x_lead, v_lead, a_lead, a_lead_tau):
    a_lead_traj = a_lead * np.exp(-a_lead_tau * (T_IDXS**2)/2.)
    v_lead_traj = np.clip(v_lead + np.cumsum(T_DIFFS * a_lead_traj), 0.0, 1e8)
    x_lead_traj = x_lead + np.cumsum(T_DIFFS * v_lead_traj)
    lead_xv = np.column_stack((x_lead_traj, v_lead_traj))
    return lead_xv

  def process_lead_legacy(self, lead):
    # behavior before PR #37824 (newLeadMpc=False): one immediate radar lead prediction
    # extrapolated forward with acceleration decaying to 0
    v_ego = self.x0[1]
    if lead is not None and lead.status:
      x_lead = lead.dRel
      v_lead = lead.vLead
      a_lead = lead.aLeadK
      a_lead_tau = lead.aLeadTau
    else:
      # Fake a fast lead car, so mpc can keep running in the same mode
      x_lead = 50.0
      v_lead = v_ego + 10.0
      a_lead = 0.0
      a_lead_tau = _LEAD_ACCEL_TAU

    # MPC will not converge if immediate crash is expected
    # Clip lead distance to what is still possible to brake for
    min_x_lead = MIN_X_LEAD_FACTOR * (v_ego + v_lead) * (v_ego - v_lead) / (-ACCEL_MIN * 2)
    x_lead = np.clip(x_lead, min_x_lead, 1e8)
    v_lead = np.clip(v_lead, 0.0, 1e8)
    a_lead = np.clip(a_lead, -10., 5.)
    lead_xv = self.extrapolate_lead(x_lead, v_lead, a_lead, a_lead_tau)
    return lead_xv

  def process_lead(self, model_lead, radar_lead):
    v_ego = self.x0[1]
    if model_lead.prob > 0.5 and radar_lead.status:
      # Anchor at radar's trusted h=0, use model's delta for h>0. On radarless, radarState
      # is synthesized from the model (radard.get_RadarState_from_vision), so this collapses
      # to `x - RADAR_TO_CAMERA` and `v_ego + (model.v - model_v_ego)` — identical to the
      # prior formula. On radar cars, real radar measurements anchor the trajectory.
      x_lead_traj = float(radar_lead.dRel) + (np.asarray(model_lead.x, dtype=np.float64) - model_lead.x[0])
      v_lead_traj = float(radar_lead.vLead) + (np.asarray(model_lead.v, dtype=np.float64) - model_lead.v[0])
    else:
      # Fake a fast lead so MPC stays in the same mode.
      x_lead_traj = 50.0 + (v_ego + 10.0) * LEAD_T_IDXS_MODEL
      v_lead_traj = np.full_like(LEAD_T_IDXS_MODEL, v_ego + 10.0)

    # MPC won't converge on immediate crashes; lift h=0 to the minimum braking distance.
    v_lead_0 = v_lead_traj[0]
    min_x_lead = MIN_X_LEAD_FACTOR * (v_ego + v_lead_0) * (v_ego - v_lead_0) / (-ACCEL_MIN * 2)
    x_lead_traj[0] = max(x_lead_traj[0], min_x_lead)
    v_lead_traj = np.clip(v_lead_traj, 0.0, 1e8)

    x_lead_mpc = np.maximum.accumulate(np.interp(T_IDXS, LEAD_T_IDXS_MODEL, x_lead_traj))
    v_lead_mpc = np.interp(T_IDXS, LEAD_T_IDXS_MODEL, v_lead_traj)
    if radar_lead.status and radar_lead.vRel > LEAD_PULLAWAY_VREL and radar_lead.aLeadK > LEAD_PULLAWAY_ABRAKE:
      # ty spysyweeb for lead pull away fix phantom launch braking so you don't ram the lead in edge cases.
      radar_velocity_floor = np.full_like(T_IDXS, float(radar_lead.vLead))
      radar_distance_floor = float(radar_lead.dRel) + float(radar_lead.vLead) * T_IDXS
      v_lead_mpc = np.maximum(v_lead_mpc, radar_velocity_floor)
      x_lead_mpc = np.maximum(x_lead_mpc, radar_distance_floor)
    return np.column_stack((x_lead_mpc, v_lead_mpc))

  def update(self, modelV2, radarstate, personality=log.LongitudinalPersonality.standard):
    self.new_lead_mpc = self._read_new_lead_mpc()
    t_follow = get_T_FOLLOW(personality)
    model_leads = modelV2.leadsV3

    if self.new_lead_mpc:
      # PR #37824: use the model's full predicted lead horizon
      self.status = model_leads[0].prob > 0.5 or model_leads[1].prob > 0.5
      lead_xv_0 = self.process_lead(model_leads[0], radarstate.leadOne)
      lead_xv_1 = self.process_lead(model_leads[1], radarstate.leadTwo)
    else:
      # pre-PR behavior: radar lead extrapolated with accel decay
      self.status = radarstate.leadOne.status or radarstate.leadTwo.status
      lead_xv_0 = self.process_lead_legacy(radarstate.leadOne)
      lead_xv_1 = self.process_lead_legacy(radarstate.leadTwo)
    self.lead_xv_0 = lead_xv_0
    self.lead_xv_1 = lead_xv_1

    # To estimate a safe distance from a moving lead, we calculate how much stopping
    # distance that lead needs as a minimum. We can add that to the current distance
    # and then treat that as a stopped car/obstacle at this new distance.
    lead_0_obstacle = lead_xv_0[:,0] + get_stopped_equivalence_factor(lead_xv_0[:,1])
    lead_1_obstacle = lead_xv_1[:,0] + get_stopped_equivalence_factor(lead_xv_1[:,1])

    x_obstacles = np.column_stack([lead_0_obstacle, lead_1_obstacle])
    self.source = MPC_SOURCES[np.argmin(x_obstacles[0])]

    self.yref[:,:] = 0.0
    for i in range(N):
      self.solver.set(i, "yref", self.yref[i])
    self.solver.set(N, "yref", self.yref[N][:COST_E_DIM])

    self.params[:,0] = ACCEL_MIN
    self.params[:,1] = ACCEL_MAX
    self.params[:,2] = np.min(x_obstacles, axis=1)
    self.params[:,3] = t_follow
    self.params[:,4] = LEAD_DANGER_FACTOR

    self.run()
    lead_crash_prob = model_leads[0].prob if self.new_lead_mpc else radarstate.leadOne.modelProb
    if (np.any(lead_xv_0[FCW_IDXS,0] - self.x_sol[FCW_IDXS,0] < CRASH_DISTANCE) and
            lead_crash_prob > 0.9):
      self.crash_cnt += 1
    else:
      self.crash_cnt = 0

  def run(self):
    for i in range(N+1):
      self.solver.set(i, 'p', self.params[i])
    self.solver.constraints_set(0, "lbx", self.x0)
    self.solver.constraints_set(0, "ubx", self.x0)

    self.solution_status = self.solver.solve()
    self.solve_time = float(self.solver.get_stats('time_tot')[0])
    self.time_qp_solution = float(self.solver.get_stats('time_qp')[0])
    self.time_linearization = float(self.solver.get_stats('time_lin')[0])
    self.time_integrator = float(self.solver.get_stats('time_sim')[0])

    for i in range(N+1):
      self.x_sol[i] = self.solver.get(i, 'x')
    for i in range(N):
      self.u_sol[i] = self.solver.get(i, 'u')

    self.v_solution = self.x_sol[:,1]
    self.a_solution = self.x_sol[:,2]
    self.j_solution = self.u_sol[:,0]

    t = time.monotonic()
    if self.solution_status != 0:
      if t > self.last_cloudlog_t + 5.0:
        self.last_cloudlog_t = t
        cloudlog.warning(f"Long mpc reset, solution_status: {self.solution_status}")
      self.reset()


if __name__ == "__main__":
  ocp = gen_long_ocp()
  AcadosOcpSolver.generate(ocp, json_file=JSON_FILE)