xref: /petsc/src/ts/impls/implicit/irk/irk.c (revision 4e8208cbcbc709572b8abe32f33c78b69c819375)

/*
  Code for timestepping with implicit Runge-Kutta method

  Notes:
  The general system is written as

  F(t,U,Udot) = 0

*/
#include <petsc/private/tsimpl.h> /*I   "petscts.h"   I*/
#include <petscdm.h>
#include <petscdt.h>

static TSIRKType         TSIRKDefault = TSIRKGAUSS;
static PetscBool         TSIRKRegisterAllCalled;
static PetscBool         TSIRKPackageInitialized;
static PetscFunctionList TSIRKList;

struct _IRKTableau {
  PetscReal   *A, *b, *c;
  PetscScalar *A_inv, *A_inv_rowsum, *I_s;
  PetscReal   *binterp; /* Dense output formula */
};

typedef struct _IRKTableau *IRKTableau;

typedef struct {
  char        *method_name;
  PetscInt     order;   /* Classical approximation order of the method */
  PetscInt     nstages; /* Number of stages */
  PetscBool    stiffly_accurate;
  PetscInt     pinterp; /* Interpolation order */
  IRKTableau   tableau;
  Vec          U0;    /* Backup vector */
  Vec          Z;     /* Combined stage vector */
  Vec         *Y;     /* States computed during the step */
  Vec          Ydot;  /* Work vector holding time derivatives during residual evaluation */
  Vec          U;     /* U is used to compute Ydot = shift(Y-U) */
  Vec         *YdotI; /* Work vectors to hold the residual evaluation */
  Mat          TJ;    /* KAIJ matrix for the Jacobian of the combined system */
  PetscScalar *work;  /* Scalar work */
  TSStepStatus status;
  PetscBool    rebuild_completion;
  PetscReal    ccfl;
} TS_IRK;

/*@C
  TSIRKTableauCreate - create the tableau for `TSIRK` and provide the entries

  Not Collective

  Input Parameters:
+ ts           - timestepping context
. nstages      - number of stages, this is the dimension of the matrices below
. A            - stage coefficients (dimension `nstages` * `nstages`, row-major)
. b            - step completion table (dimension `nstages`)
. c            - abscissa (dimension `nstages`)
. binterp      - coefficients of the interpolation formula (dimension `nstages`), optional (use `NULL` to skip)
. A_inv        - inverse of `A` (dimension `nstages` * `nstages`, row-major), optional (use `NULL` to skip)
. A_inv_rowsum - row sum of the inverse of `A` (dimension `nstages`), optional (use `NULL` to skip)
- I_s          - identity matrix (dimension `nstages` * `nstages`), optional (use `NULL` to skip)

  Level: advanced

.seealso: [](ch_ts), `TSIRK`, `TSIRKRegister()`
@*/
PetscErrorCode TSIRKTableauCreate(TS ts, PetscInt nstages, const PetscReal *A, const PetscReal *b, const PetscReal *c, const PetscReal *binterp, const PetscScalar *A_inv, const PetscScalar *A_inv_rowsum, const PetscScalar *I_s)
{
  TS_IRK    *irk = (TS_IRK *)ts->data;
  IRKTableau tab = irk->tableau;

  PetscFunctionBegin;
  irk->order = nstages;
  PetscCall(PetscMalloc3(PetscSqr(nstages), &tab->A, PetscSqr(nstages), &tab->A_inv, PetscSqr(nstages), &tab->I_s));
  PetscCall(PetscMalloc4(nstages, &tab->b, nstages, &tab->c, nstages, &tab->binterp, nstages, &tab->A_inv_rowsum));
  PetscCall(PetscArraycpy(tab->A, A, PetscSqr(nstages)));
  PetscCall(PetscArraycpy(tab->b, b, nstages));
  PetscCall(PetscArraycpy(tab->c, c, nstages));
  /* optional coefficient arrays */
  if (binterp) PetscCall(PetscArraycpy(tab->binterp, binterp, nstages));
  if (A_inv) PetscCall(PetscArraycpy(tab->A_inv, A_inv, PetscSqr(nstages)));
  if (A_inv_rowsum) PetscCall(PetscArraycpy(tab->A_inv_rowsum, A_inv_rowsum, nstages));
  if (I_s) PetscCall(PetscArraycpy(tab->I_s, I_s, PetscSqr(nstages)));
  PetscFunctionReturn(PETSC_SUCCESS);
}

/* Arrays should be freed with PetscFree3(A,b,c) */
static PetscErrorCode TSIRKCreate_Gauss(TS ts)
{
  PetscInt     nstages;
  PetscReal   *gauss_A_real, *gauss_b, *b, *gauss_c;
  PetscScalar *gauss_A, *gauss_A_inv, *gauss_A_inv_rowsum, *I_s;
  PetscScalar *G0, *G1;
  PetscInt     i, j;
  Mat          G0mat, G1mat, Amat;

  PetscFunctionBegin;
  PetscCall(TSIRKGetNumStages(ts, &nstages));
  PetscCall(PetscMalloc3(PetscSqr(nstages), &gauss_A_real, nstages, &gauss_b, nstages, &gauss_c));
  PetscCall(PetscMalloc4(PetscSqr(nstages), &gauss_A, PetscSqr(nstages), &gauss_A_inv, nstages, &gauss_A_inv_rowsum, PetscSqr(nstages), &I_s));
  PetscCall(PetscMalloc3(nstages, &b, PetscSqr(nstages), &G0, PetscSqr(nstages), &G1));
  PetscCall(PetscDTGaussQuadrature(nstages, 0., 1., gauss_c, b));
  for (i = 0; i < nstages; i++) gauss_b[i] = b[i]; /* copy to possibly-complex array */

  /* A^T = G0^{-1} G1 */
  for (i = 0; i < nstages; i++) {
    for (j = 0; j < nstages; j++) {
      G0[i * nstages + j] = PetscPowRealInt(gauss_c[i], j);
      G1[i * nstages + j] = PetscPowRealInt(gauss_c[i], j + 1) / (j + 1);
    }
  }
  /* The arrays above are row-aligned, but we create dense matrices as the transpose */
  PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, nstages, nstages, G0, &G0mat));
  PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, nstages, nstages, G1, &G1mat));
  PetscCall(MatCreateSeqDense(PETSC_COMM_SELF, nstages, nstages, gauss_A, &Amat));
  PetscCall(MatLUFactor(G0mat, NULL, NULL, NULL));
  PetscCall(MatMatSolve(G0mat, G1mat, Amat));
  PetscCall(MatTranspose(Amat, MAT_INPLACE_MATRIX, &Amat));
  for (i = 0; i < nstages; i++)
    for (j = 0; j < nstages; j++) gauss_A_real[i * nstages + j] = PetscRealPart(gauss_A[i * nstages + j]);

  PetscCall(MatDestroy(&G0mat));
  PetscCall(MatDestroy(&G1mat));
  PetscCall(MatDestroy(&Amat));
  PetscCall(PetscFree3(b, G0, G1));

  { /* Invert A */
    /* PETSc does not provide a routine to calculate the inverse of a general matrix.
     * To get the inverse of A, we form a sequential BAIJ matrix from it, consisting of a single block with block size
     * equal to the dimension of A, and then use MatInvertBlockDiagonal(). */
    Mat                A_baij;
    PetscInt           idxm[1] = {0}, idxn[1] = {0};
    const PetscScalar *A_inv;

    PetscCall(MatCreateSeqBAIJ(PETSC_COMM_SELF, nstages, nstages, nstages, 1, NULL, &A_baij));
    PetscCall(MatSetOption(A_baij, MAT_ROW_ORIENTED, PETSC_FALSE));
    PetscCall(MatSetValuesBlocked(A_baij, 1, idxm, 1, idxn, gauss_A, INSERT_VALUES));
    PetscCall(MatAssemblyBegin(A_baij, MAT_FINAL_ASSEMBLY));
    PetscCall(MatAssemblyEnd(A_baij, MAT_FINAL_ASSEMBLY));
    PetscCall(MatInvertBlockDiagonal(A_baij, &A_inv));
    PetscCall(PetscArraycpy(gauss_A_inv, A_inv, nstages * nstages));
    PetscCall(MatDestroy(&A_baij));
  }

  /* Compute row sums A_inv_rowsum and identity I_s */
  for (i = 0; i < nstages; i++) {
    gauss_A_inv_rowsum[i] = 0;
    for (j = 0; j < nstages; j++) {
      gauss_A_inv_rowsum[i] += gauss_A_inv[i + nstages * j];
      I_s[i + nstages * j] = 1. * (i == j);
    }
  }
  PetscCall(TSIRKTableauCreate(ts, nstages, gauss_A_real, gauss_b, gauss_c, NULL, gauss_A_inv, gauss_A_inv_rowsum, I_s));
  PetscCall(PetscFree3(gauss_A_real, gauss_b, gauss_c));
  PetscCall(PetscFree4(gauss_A, gauss_A_inv, gauss_A_inv_rowsum, I_s));
  PetscFunctionReturn(PETSC_SUCCESS);
}

/*@C
  TSIRKRegister -  adds a `TSIRK` implementation

  Not Collective, No Fortran Support

  Input Parameters:
+ sname    - name of user-defined IRK scheme
- function - function to create method context

  Level: advanced

  Note:
  `TSIRKRegister()` may be called multiple times to add several user-defined families.

  Example Usage:
.vb
   TSIRKRegister("my_scheme", MySchemeCreate);
.ve

  Then, your scheme can be chosen with the procedural interface via
.vb
  TSIRKSetType(ts, "my_scheme")
.ve
  or at runtime via the option
.vb
  -ts_irk_type my_scheme
.ve

.seealso: [](ch_ts), `TSIRK`, `TSIRKRegisterAll()`
@*/
PetscErrorCode TSIRKRegister(const char sname[], PetscErrorCode (*function)(TS))
{
  PetscFunctionBegin;
  PetscCall(TSIRKInitializePackage());
  PetscCall(PetscFunctionListAdd(&TSIRKList, sname, function));
  PetscFunctionReturn(PETSC_SUCCESS);
}

/*@C
  TSIRKRegisterAll - Registers all of the implicit Runge-Kutta methods in `TSIRK`

  Not Collective, but should be called by all processes which will need the schemes to be registered

  Level: advanced

.seealso: [](ch_ts), `TSIRK`, `TSIRKRegisterDestroy()`
@*/
PetscErrorCode TSIRKRegisterAll(void)
{
  PetscFunctionBegin;
  if (TSIRKRegisterAllCalled) PetscFunctionReturn(PETSC_SUCCESS);
  TSIRKRegisterAllCalled = PETSC_TRUE;

  PetscCall(TSIRKRegister(TSIRKGAUSS, TSIRKCreate_Gauss));
  PetscFunctionReturn(PETSC_SUCCESS);
}

/*@C
  TSIRKRegisterDestroy - Frees the list of schemes that were registered by `TSIRKRegister()`.

  Not Collective

  Level: advanced

.seealso: [](ch_ts), `TSIRK`, `TSIRKRegister()`, `TSIRKRegisterAll()`
@*/
PetscErrorCode TSIRKRegisterDestroy(void)
{
  PetscFunctionBegin;
  TSIRKRegisterAllCalled = PETSC_FALSE;
  PetscFunctionReturn(PETSC_SUCCESS);
}

/*@C
  TSIRKInitializePackage - This function initializes everything in the `TSIRK` package. It is called
  from `TSInitializePackage()`.

  Level: developer

.seealso: [](ch_ts), `TSIRK`, `PetscInitialize()`, `TSIRKFinalizePackage()`, `TSInitializePackage()`
@*/
PetscErrorCode TSIRKInitializePackage(void)
{
  PetscFunctionBegin;
  if (TSIRKPackageInitialized) PetscFunctionReturn(PETSC_SUCCESS);
  TSIRKPackageInitialized = PETSC_TRUE;
  PetscCall(TSIRKRegisterAll());
  PetscCall(PetscRegisterFinalize(TSIRKFinalizePackage));
  PetscFunctionReturn(PETSC_SUCCESS);
}

/*@C
  TSIRKFinalizePackage - This function destroys everything in the `TSIRK` package. It is
  called from `PetscFinalize()`.

  Level: developer

.seealso: [](ch_ts), `TSIRK`, `PetscFinalize()`, `TSInitializePackage()`
@*/
PetscErrorCode TSIRKFinalizePackage(void)
{
  PetscFunctionBegin;
  PetscCall(PetscFunctionListDestroy(&TSIRKList));
  TSIRKPackageInitialized = PETSC_FALSE;
  PetscFunctionReturn(PETSC_SUCCESS);
}

/*
 This function can be called before or after ts->vec_sol has been updated.
*/
static PetscErrorCode TSEvaluateStep_IRK(TS ts, PetscInt order, Vec U, PetscBool *done)
{
  TS_IRK      *irk   = (TS_IRK *)ts->data;
  IRKTableau   tab   = irk->tableau;
  Vec         *YdotI = irk->YdotI;
  PetscScalar *w     = irk->work;
  PetscReal    h;
  PetscInt     j;

  PetscFunctionBegin;
  switch (irk->status) {
  case TS_STEP_INCOMPLETE:
  case TS_STEP_PENDING:
    h = ts->time_step;
    break;
  case TS_STEP_COMPLETE:
    h = ts->ptime - ts->ptime_prev;
    break;
  default:
    SETERRQ(PetscObjectComm((PetscObject)ts), PETSC_ERR_PLIB, "Invalid TSStepStatus");
  }

  PetscCall(VecCopy(ts->vec_sol, U));
  for (j = 0; j < irk->nstages; j++) w[j] = h * tab->b[j];
  PetscCall(VecMAXPY(U, irk->nstages, w, YdotI));
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode TSRollBack_IRK(TS ts)
{
  TS_IRK *irk = (TS_IRK *)ts->data;

  PetscFunctionBegin;
  PetscCall(VecCopy(irk->U0, ts->vec_sol));
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode TSStep_IRK(TS ts)
{
  TS_IRK        *irk   = (TS_IRK *)ts->data;
  IRKTableau     tab   = irk->tableau;
  PetscScalar   *A_inv = tab->A_inv, *A_inv_rowsum = tab->A_inv_rowsum;
  const PetscInt nstages = irk->nstages;
  SNES           snes;
  PetscInt       i, j, its, lits, bs;
  TSAdapt        adapt;
  PetscInt       rejections     = 0;
  PetscBool      accept         = PETSC_TRUE;
  PetscReal      next_time_step = ts->time_step;

  PetscFunctionBegin;
  if (!ts->steprollback) PetscCall(VecCopy(ts->vec_sol, irk->U0));
  PetscCall(VecGetBlockSize(ts->vec_sol, &bs));
  for (i = 0; i < nstages; i++) PetscCall(VecStrideScatter(ts->vec_sol, i * bs, irk->Z, INSERT_VALUES));

  irk->status = TS_STEP_INCOMPLETE;
  while (!ts->reason && irk->status != TS_STEP_COMPLETE) {
    PetscCall(VecCopy(ts->vec_sol, irk->U));
    PetscCall(TSGetSNES(ts, &snes));
    PetscCall(SNESSolve(snes, NULL, irk->Z));
    PetscCall(SNESGetIterationNumber(snes, &its));
    PetscCall(SNESGetLinearSolveIterations(snes, &lits));
    ts->snes_its += its;
    ts->ksp_its += lits;
    PetscCall(VecStrideGatherAll(irk->Z, irk->Y, INSERT_VALUES));
    for (i = 0; i < nstages; i++) {
      PetscCall(VecZeroEntries(irk->YdotI[i]));
      for (j = 0; j < nstages; j++) PetscCall(VecAXPY(irk->YdotI[i], A_inv[i + j * nstages] / ts->time_step, irk->Y[j]));
      PetscCall(VecAXPY(irk->YdotI[i], -A_inv_rowsum[i] / ts->time_step, irk->U));
    }
    irk->status = TS_STEP_INCOMPLETE;
    PetscCall(TSEvaluateStep_IRK(ts, irk->order, ts->vec_sol, NULL));
    irk->status = TS_STEP_PENDING;
    PetscCall(TSGetAdapt(ts, &adapt));
    PetscCall(TSAdaptChoose(adapt, ts, ts->time_step, NULL, &next_time_step, &accept));
    irk->status = accept ? TS_STEP_COMPLETE : TS_STEP_INCOMPLETE;
    if (!accept) {
      PetscCall(TSRollBack_IRK(ts));
      ts->time_step = next_time_step;
      goto reject_step;
    }

    ts->ptime += ts->time_step;
    ts->time_step = next_time_step;
    break;
  reject_step:
    ts->reject++;
    accept = PETSC_FALSE;
    if (!ts->reason && ++rejections > ts->max_reject && ts->max_reject >= 0) {
      ts->reason = TS_DIVERGED_STEP_REJECTED;
      PetscCall(PetscInfo(ts, "Step=%" PetscInt_FMT ", step rejections %" PetscInt_FMT " greater than current TS allowed, stopping solve\n", ts->steps, rejections));
    }
  }
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode TSInterpolate_IRK(TS ts, PetscReal itime, Vec U)
{
  TS_IRK          *irk     = (TS_IRK *)ts->data;
  PetscInt         nstages = irk->nstages, pinterp = irk->pinterp, i, j;
  PetscReal        h;
  PetscReal        tt, t;
  PetscScalar     *bt;
  const PetscReal *B = irk->tableau->binterp;

  PetscFunctionBegin;
  PetscCheck(B, PetscObjectComm((PetscObject)ts), PETSC_ERR_SUP, "TSIRK %s does not have an interpolation formula", irk->method_name);
  switch (irk->status) {
  case TS_STEP_INCOMPLETE:
  case TS_STEP_PENDING:
    h = ts->time_step;
    t = (itime - ts->ptime) / h;
    break;
  case TS_STEP_COMPLETE:
    h = ts->ptime - ts->ptime_prev;
    t = (itime - ts->ptime) / h + 1; /* In the interval [0,1] */
    break;
  default:
    SETERRQ(PetscObjectComm((PetscObject)ts), PETSC_ERR_PLIB, "Invalid TSStepStatus");
  }
  PetscCall(PetscMalloc1(nstages, &bt));
  for (i = 0; i < nstages; i++) bt[i] = 0;
  for (j = 0, tt = t; j < pinterp; j++, tt *= t) {
    for (i = 0; i < nstages; i++) bt[i] += h * B[i * pinterp + j] * tt;
  }
  PetscCall(VecMAXPY(U, nstages, bt, irk->YdotI));
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode TSIRKTableauReset(TS ts)
{
  TS_IRK    *irk = (TS_IRK *)ts->data;
  IRKTableau tab = irk->tableau;

  PetscFunctionBegin;
  if (!tab) PetscFunctionReturn(PETSC_SUCCESS);
  PetscCall(PetscFree3(tab->A, tab->A_inv, tab->I_s));
  PetscCall(PetscFree4(tab->b, tab->c, tab->binterp, tab->A_inv_rowsum));
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode TSReset_IRK(TS ts)
{
  TS_IRK *irk = (TS_IRK *)ts->data;

  PetscFunctionBegin;
  PetscCall(TSIRKTableauReset(ts));
  if (irk->tableau) PetscCall(PetscFree(irk->tableau));
  if (irk->method_name) PetscCall(PetscFree(irk->method_name));
  if (irk->work) PetscCall(PetscFree(irk->work));
  PetscCall(VecDestroyVecs(irk->nstages, &irk->Y));
  PetscCall(VecDestroyVecs(irk->nstages, &irk->YdotI));
  PetscCall(VecDestroy(&irk->Ydot));
  PetscCall(VecDestroy(&irk->Z));
  PetscCall(VecDestroy(&irk->U));
  PetscCall(VecDestroy(&irk->U0));
  PetscCall(MatDestroy(&irk->TJ));
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode TSIRKGetVecs(TS ts, DM dm, Vec *U)
{
  TS_IRK *irk = (TS_IRK *)ts->data;

  PetscFunctionBegin;
  if (U) {
    if (dm && dm != ts->dm) PetscCall(DMGetNamedGlobalVector(dm, "TSIRK_U", U));
    else *U = irk->U;
  }
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode TSIRKRestoreVecs(TS ts, DM dm, Vec *U)
{
  PetscFunctionBegin;
  if (U) {
    if (dm && dm != ts->dm) PetscCall(DMRestoreNamedGlobalVector(dm, "TSIRK_U", U));
  }
  PetscFunctionReturn(PETSC_SUCCESS);
}

/*
  This defines the nonlinear equations that is to be solved with SNES
    G[e\otimes t + C*dt, Z, Zdot] = 0
    Zdot = (In \otimes S)*Z - (In \otimes Se) U
  where S = 1/(dt*A)
*/
static PetscErrorCode SNESTSFormFunction_IRK(SNES snes, Vec ZC, Vec FC, TS ts)
{
  TS_IRK            *irk     = (TS_IRK *)ts->data;
  IRKTableau         tab     = irk->tableau;
  const PetscInt     nstages = irk->nstages;
  const PetscReal   *c       = tab->c;
  const PetscScalar *A_inv = tab->A_inv, *A_inv_rowsum = tab->A_inv_rowsum;
  DM                 dm, dmsave;
  Vec                U, *YdotI = irk->YdotI, Ydot = irk->Ydot, *Y = irk->Y;
  PetscReal          h = ts->time_step;
  PetscInt           i, j;

  PetscFunctionBegin;
  PetscCall(SNESGetDM(snes, &dm));
  PetscCall(TSIRKGetVecs(ts, dm, &U));
  PetscCall(VecStrideGatherAll(ZC, Y, INSERT_VALUES));
  dmsave = ts->dm;
  ts->dm = dm;
  for (i = 0; i < nstages; i++) {
    PetscCall(VecZeroEntries(Ydot));
    for (j = 0; j < nstages; j++) PetscCall(VecAXPY(Ydot, A_inv[j * nstages + i] / h, Y[j]));
    PetscCall(VecAXPY(Ydot, -A_inv_rowsum[i] / h, U)); /* Ydot = (S \otimes In)*Z - (Se \otimes In) U */
    PetscCall(TSComputeIFunction(ts, ts->ptime + ts->time_step * c[i], Y[i], Ydot, YdotI[i], PETSC_FALSE));
  }
  PetscCall(VecStrideScatterAll(YdotI, FC, INSERT_VALUES));
  ts->dm = dmsave;
  PetscCall(TSIRKRestoreVecs(ts, dm, &U));
  PetscFunctionReturn(PETSC_SUCCESS);
}

/*
   For explicit ODE, the Jacobian is
     JC = I_n \otimes S - J \otimes I_s
   For DAE, the Jacobian is
     JC = M_n \otimes S - J \otimes I_s
*/
static PetscErrorCode SNESTSFormJacobian_IRK(SNES snes, Vec ZC, Mat JC, Mat JCpre, TS ts)
{
  TS_IRK          *irk     = (TS_IRK *)ts->data;
  IRKTableau       tab     = irk->tableau;
  const PetscInt   nstages = irk->nstages;
  const PetscReal *c       = tab->c;
  DM               dm, dmsave;
  Vec             *Y = irk->Y, Ydot = irk->Ydot;
  Mat              J;
  PetscScalar     *S;
  PetscInt         i, j, bs;

  PetscFunctionBegin;
  PetscCall(SNESGetDM(snes, &dm));
  /* irk->Ydot has already been computed in SNESTSFormFunction_IRK (SNES guarantees this) */
  dmsave = ts->dm;
  ts->dm = dm;
  PetscCall(VecGetBlockSize(Y[nstages - 1], &bs));
  PetscCheck(ts->equation_type <= TS_EQ_ODE_EXPLICIT, PetscObjectComm((PetscObject)ts), PETSC_ERR_SUP, "TSIRK %s does not support implicit formula", irk->method_name); /* TODO: need the mass matrix for DAE  */
  /* Support explicit formulas only */
  PetscCall(VecStrideGather(ZC, (nstages - 1) * bs, Y[nstages - 1], INSERT_VALUES));
  PetscCall(MatKAIJGetAIJ(JC, &J));
  PetscCall(TSComputeIJacobian(ts, ts->ptime + ts->time_step * c[nstages - 1], Y[nstages - 1], Ydot, 0, J, J, PETSC_FALSE));
  PetscCall(MatKAIJGetS(JC, NULL, NULL, &S));
  for (i = 0; i < nstages; i++)
    for (j = 0; j < nstages; j++) S[i + nstages * j] = tab->A_inv[i + nstages * j] / ts->time_step;
  PetscCall(MatKAIJRestoreS(JC, &S));
  ts->dm = dmsave;
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode DMCoarsenHook_TSIRK(DM fine, DM coarse, PetscCtx ctx)
{
  PetscFunctionBegin;
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode DMRestrictHook_TSIRK(DM fine, Mat restrct, Vec rscale, Mat inject, DM coarse, PetscCtx ctx)
{
  TS  ts = (TS)ctx;
  Vec U, U_c;

  PetscFunctionBegin;
  PetscCall(TSIRKGetVecs(ts, fine, &U));
  PetscCall(TSIRKGetVecs(ts, coarse, &U_c));
  PetscCall(MatRestrict(restrct, U, U_c));
  PetscCall(VecPointwiseMult(U_c, rscale, U_c));
  PetscCall(TSIRKRestoreVecs(ts, fine, &U));
  PetscCall(TSIRKRestoreVecs(ts, coarse, &U_c));
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode DMSubDomainHook_TSIRK(DM dm, DM subdm, PetscCtx ctx)
{
  PetscFunctionBegin;
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode DMSubDomainRestrictHook_TSIRK(DM dm, VecScatter gscat, VecScatter lscat, DM subdm, PetscCtx ctx)
{
  TS  ts = (TS)ctx;
  Vec U, U_c;

  PetscFunctionBegin;
  PetscCall(TSIRKGetVecs(ts, dm, &U));
  PetscCall(TSIRKGetVecs(ts, subdm, &U_c));

  PetscCall(VecScatterBegin(gscat, U, U_c, INSERT_VALUES, SCATTER_FORWARD));
  PetscCall(VecScatterEnd(gscat, U, U_c, INSERT_VALUES, SCATTER_FORWARD));

  PetscCall(TSIRKRestoreVecs(ts, dm, &U));
  PetscCall(TSIRKRestoreVecs(ts, subdm, &U_c));
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode TSSetUp_IRK(TS ts)
{
  TS_IRK        *irk = (TS_IRK *)ts->data;
  IRKTableau     tab = irk->tableau;
  DM             dm;
  Mat            J;
  Vec            R;
  const PetscInt nstages = irk->nstages;
  PetscInt       vsize, bs;

  PetscFunctionBegin;
  if (!irk->work) PetscCall(PetscMalloc1(irk->nstages, &irk->work));
  if (!irk->Y) PetscCall(VecDuplicateVecs(ts->vec_sol, irk->nstages, &irk->Y));
  if (!irk->YdotI) PetscCall(VecDuplicateVecs(ts->vec_sol, irk->nstages, &irk->YdotI));
  if (!irk->Ydot) PetscCall(VecDuplicate(ts->vec_sol, &irk->Ydot));
  if (!irk->U) PetscCall(VecDuplicate(ts->vec_sol, &irk->U));
  if (!irk->U0) PetscCall(VecDuplicate(ts->vec_sol, &irk->U0));
  if (!irk->Z) {
    PetscCall(VecCreate(PetscObjectComm((PetscObject)ts->vec_sol), &irk->Z));
    PetscCall(VecGetSize(ts->vec_sol, &vsize));
    PetscCall(VecSetSizes(irk->Z, PETSC_DECIDE, vsize * irk->nstages));
    PetscCall(VecGetBlockSize(ts->vec_sol, &bs));
    PetscCall(VecSetBlockSize(irk->Z, irk->nstages * bs));
    PetscCall(VecSetFromOptions(irk->Z));
  }
  PetscCall(TSGetDM(ts, &dm));
  PetscCall(DMCoarsenHookAdd(dm, DMCoarsenHook_TSIRK, DMRestrictHook_TSIRK, ts));
  PetscCall(DMSubDomainHookAdd(dm, DMSubDomainHook_TSIRK, DMSubDomainRestrictHook_TSIRK, ts));

  PetscCall(TSGetSNES(ts, &ts->snes));
  PetscCall(VecDuplicate(irk->Z, &R));
  PetscCall(SNESSetFunction(ts->snes, R, SNESTSFormFunction, ts));
  PetscCall(TSGetIJacobian(ts, &J, NULL, NULL, NULL));
  if (!irk->TJ) {
    /* Create the KAIJ matrix for solving the stages */
    PetscCall(MatCreateKAIJ(J, nstages, nstages, tab->A_inv, tab->I_s, &irk->TJ));
  }
  PetscCall(SNESSetJacobian(ts->snes, irk->TJ, irk->TJ, SNESTSFormJacobian, ts));
  PetscCall(VecDestroy(&R));
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode TSSetFromOptions_IRK(TS ts, PetscOptionItems PetscOptionsObject)
{
  TS_IRK *irk        = (TS_IRK *)ts->data;
  char    tname[256] = TSIRKGAUSS;

  PetscFunctionBegin;
  PetscOptionsHeadBegin(PetscOptionsObject, "IRK ODE solver options");
  {
    PetscBool flg1, flg2;
    PetscCall(PetscOptionsInt("-ts_irk_nstages", "Stages of the IRK method", "TSIRKSetNumStages", irk->nstages, &irk->nstages, &flg1));
    PetscCall(PetscOptionsFList("-ts_irk_type", "Type of IRK method", "TSIRKSetType", TSIRKList, irk->method_name[0] ? irk->method_name : tname, tname, sizeof(tname), &flg2));
    if (flg1 || flg2 || !irk->method_name[0]) { /* Create the method tableau after nstages or method is set */
      PetscCall(TSIRKSetType(ts, tname));
    }
  }
  PetscOptionsHeadEnd();
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode TSView_IRK(TS ts, PetscViewer viewer)
{
  TS_IRK   *irk = (TS_IRK *)ts->data;
  PetscBool isascii;

  PetscFunctionBegin;
  PetscCall(PetscObjectTypeCompare((PetscObject)viewer, PETSCVIEWERASCII, &isascii));
  if (isascii) {
    IRKTableau tab = irk->tableau;
    TSIRKType  irktype;
    char       buf[512];

    PetscCall(TSIRKGetType(ts, &irktype));
    PetscCall(PetscViewerASCIIPrintf(viewer, "  IRK type %s\n", irktype));
    PetscCall(PetscFormatRealArray(buf, sizeof(buf), "% 8.6f", irk->nstages, tab->c));
    PetscCall(PetscViewerASCIIPrintf(viewer, "  Abscissa       c = %s\n", buf));
    PetscCall(PetscViewerASCIIPrintf(viewer, "Stiffly accurate: %s\n", irk->stiffly_accurate ? "yes" : "no"));
    PetscCall(PetscFormatRealArray(buf, sizeof(buf), "% 8.6f", PetscSqr(irk->nstages), tab->A));
    PetscCall(PetscViewerASCIIPrintf(viewer, "  A coefficients       A = %s\n", buf));
  }
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode TSLoad_IRK(TS ts, PetscViewer viewer)
{
  SNES    snes;
  TSAdapt adapt;

  PetscFunctionBegin;
  PetscCall(TSGetAdapt(ts, &adapt));
  PetscCall(TSAdaptLoad(adapt, viewer));
  PetscCall(TSGetSNES(ts, &snes));
  PetscCall(SNESLoad(snes, viewer));
  /* function and Jacobian context for SNES when used with TS is always ts object */
  PetscCall(SNESSetFunction(snes, NULL, NULL, ts));
  PetscCall(SNESSetJacobian(snes, NULL, NULL, NULL, ts));
  PetscFunctionReturn(PETSC_SUCCESS);
}

/*@
  TSIRKSetType - Set the type of `TSIRK` scheme to use

  Logically Collective

  Input Parameters:
+ ts      - timestepping context
- irktype - type of `TSIRK` scheme

  Options Database Key:
. -ts_irk_type <gauss> - set irk type

  Level: intermediate

.seealso: [](ch_ts), `TSIRKGetType()`, `TSIRK`, `TSIRKType`, `TSIRKGAUSS`
@*/
PetscErrorCode TSIRKSetType(TS ts, TSIRKType irktype)
{
  PetscFunctionBegin;
  PetscValidHeaderSpecific(ts, TS_CLASSID, 1);
  PetscAssertPointer(irktype, 2);
  PetscTryMethod(ts, "TSIRKSetType_C", (TS, TSIRKType), (ts, irktype));
  PetscFunctionReturn(PETSC_SUCCESS);
}

/*@
  TSIRKGetType - Get the type of `TSIRK` IMEX scheme being used

  Logically Collective

  Input Parameter:
. ts - timestepping context

  Output Parameter:
. irktype - type of `TSIRK` IMEX scheme

  Level: intermediate

.seealso: [](ch_ts), `TSIRK`, `TSIRKType`, `TSIRKGAUSS`
@*/
PetscErrorCode TSIRKGetType(TS ts, TSIRKType *irktype)
{
  PetscFunctionBegin;
  PetscValidHeaderSpecific(ts, TS_CLASSID, 1);
  PetscUseMethod(ts, "TSIRKGetType_C", (TS, TSIRKType *), (ts, irktype));
  PetscFunctionReturn(PETSC_SUCCESS);
}

/*@
  TSIRKSetNumStages - Set the number of stages of `TSIRK` scheme to use

  Logically Collective

  Input Parameters:
+ ts      - timestepping context
- nstages - number of stages of `TSIRK` scheme

  Options Database Key:
. -ts_irk_nstages <int> - set number of stages

  Level: intermediate

.seealso: [](ch_ts), `TSIRKGetNumStages()`, `TSIRK`
@*/
PetscErrorCode TSIRKSetNumStages(TS ts, PetscInt nstages)
{
  PetscFunctionBegin;
  PetscValidHeaderSpecific(ts, TS_CLASSID, 1);
  PetscTryMethod(ts, "TSIRKSetNumStages_C", (TS, PetscInt), (ts, nstages));
  PetscFunctionReturn(PETSC_SUCCESS);
}

/*@
  TSIRKGetNumStages - Get the number of stages of `TSIRK` scheme

  Logically Collective

  Input Parameters:
+ ts      - timestepping context
- nstages - number of stages of `TSIRK` scheme

  Level: intermediate

.seealso: [](ch_ts), `TSIRKSetNumStages()`, `TSIRK`
@*/
PetscErrorCode TSIRKGetNumStages(TS ts, PetscInt *nstages)
{
  PetscFunctionBegin;
  PetscValidHeaderSpecific(ts, TS_CLASSID, 1);
  PetscAssertPointer(nstages, 2);
  PetscTryMethod(ts, "TSIRKGetNumStages_C", (TS, PetscInt *), (ts, nstages));
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode TSIRKGetType_IRK(TS ts, TSIRKType *irktype)
{
  TS_IRK *irk = (TS_IRK *)ts->data;

  PetscFunctionBegin;
  *irktype = irk->method_name;
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode TSIRKSetType_IRK(TS ts, TSIRKType irktype)
{
  TS_IRK *irk = (TS_IRK *)ts->data;
  PetscErrorCode (*irkcreate)(TS);

  PetscFunctionBegin;
  if (irk->method_name) {
    PetscCall(PetscFree(irk->method_name));
    PetscCall(TSIRKTableauReset(ts));
  }
  PetscCall(PetscFunctionListFind(TSIRKList, irktype, &irkcreate));
  PetscCheck(irkcreate, PetscObjectComm((PetscObject)ts), PETSC_ERR_ARG_UNKNOWN_TYPE, "Unknown TSIRK type \"%s\" given", irktype);
  PetscCall((*irkcreate)(ts));
  PetscCall(PetscStrallocpy(irktype, &irk->method_name));
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode TSIRKSetNumStages_IRK(TS ts, PetscInt nstages)
{
  TS_IRK *irk = (TS_IRK *)ts->data;

  PetscFunctionBegin;
  PetscCheck(nstages > 0, PETSC_COMM_SELF, PETSC_ERR_ARG_OUTOFRANGE, "input argument, %" PetscInt_FMT ", out of range", nstages);
  irk->nstages = nstages;
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode TSIRKGetNumStages_IRK(TS ts, PetscInt *nstages)
{
  TS_IRK *irk = (TS_IRK *)ts->data;

  PetscFunctionBegin;
  PetscAssertPointer(nstages, 2);
  *nstages = irk->nstages;
  PetscFunctionReturn(PETSC_SUCCESS);
}

static PetscErrorCode TSDestroy_IRK(TS ts)
{
  PetscFunctionBegin;
  PetscCall(TSReset_IRK(ts));
  if (ts->dm) {
    PetscCall(DMCoarsenHookRemove(ts->dm, DMCoarsenHook_TSIRK, DMRestrictHook_TSIRK, ts));
    PetscCall(DMSubDomainHookRemove(ts->dm, DMSubDomainHook_TSIRK, DMSubDomainRestrictHook_TSIRK, ts));
  }
  PetscCall(PetscFree(ts->data));
  PetscCall(PetscObjectComposeFunction((PetscObject)ts, "TSIRKSetType_C", NULL));
  PetscCall(PetscObjectComposeFunction((PetscObject)ts, "TSIRKGetType_C", NULL));
  PetscCall(PetscObjectComposeFunction((PetscObject)ts, "TSIRKSetNumStages_C", NULL));
  PetscCall(PetscObjectComposeFunction((PetscObject)ts, "TSIRKGetNumStages_C", NULL));
  PetscFunctionReturn(PETSC_SUCCESS);
}

/*MC
  TSIRK - ODE and DAE solver using Implicit Runge-Kutta schemes

  Level: beginner

  Notes:
  `TSIRK` uses the sparse Kronecker product matrix implementation of `MATKAIJ` to achieve good arithmetic intensity.

  Gauss-Legrendre methods are currently supported. These are A-stable symplectic methods with an arbitrary number of stages. The order of accuracy is 2s
  when using s stages. The default method uses three stages and thus has an order of six. The number of stages (thus order) can be set with
  `-ts_irk_nstages` or `TSIRKSetNumStages()`.

.seealso: [](ch_ts), `TSCreate()`, `TS`, `TSSetType()`, `TSIRKSetType()`, `TSIRKGetType()`, `TSIRKGAUSS`, `TSIRKRegister()`, `TSIRKSetNumStages()`, `TSType`
M*/
PETSC_EXTERN PetscErrorCode TSCreate_IRK(TS ts)
{
  TS_IRK *irk;

  PetscFunctionBegin;
  PetscCall(TSIRKInitializePackage());

  ts->ops->reset          = TSReset_IRK;
  ts->ops->destroy        = TSDestroy_IRK;
  ts->ops->view           = TSView_IRK;
  ts->ops->load           = TSLoad_IRK;
  ts->ops->setup          = TSSetUp_IRK;
  ts->ops->step           = TSStep_IRK;
  ts->ops->interpolate    = TSInterpolate_IRK;
  ts->ops->evaluatestep   = TSEvaluateStep_IRK;
  ts->ops->rollback       = TSRollBack_IRK;
  ts->ops->setfromoptions = TSSetFromOptions_IRK;
  ts->ops->snesfunction   = SNESTSFormFunction_IRK;
  ts->ops->snesjacobian   = SNESTSFormJacobian_IRK;

  ts->usessnes = PETSC_TRUE;

  PetscCall(PetscNew(&irk));
  ts->data = (void *)irk;

  PetscCall(PetscObjectComposeFunction((PetscObject)ts, "TSIRKSetType_C", TSIRKSetType_IRK));
  PetscCall(PetscObjectComposeFunction((PetscObject)ts, "TSIRKGetType_C", TSIRKGetType_IRK));
  PetscCall(PetscObjectComposeFunction((PetscObject)ts, "TSIRKSetNumStages_C", TSIRKSetNumStages_IRK));
  PetscCall(PetscObjectComposeFunction((PetscObject)ts, "TSIRKGetNumStages_C", TSIRKGetNumStages_IRK));
  /* 3-stage IRK_Gauss is the default */
  PetscCall(PetscNew(&irk->tableau));
  irk->nstages = 3;
  PetscCall(TSIRKSetType(ts, TSIRKDefault));
  PetscFunctionReturn(PETSC_SUCCESS);
}