forked from moebiusband/NuSiF-Solver
		
	EnhancedSolver port complete
This commit is contained in:
		@@ -22,6 +22,7 @@ SRC       = $(filter-out $(wildcard $(SRC_DIR)/*-*.c),$(wildcard $(SRC_DIR)/*.c)
 | 
			
		||||
ASM       = $(patsubst $(SRC_DIR)/%.c, $(BUILD_DIR)/%.s, $(SRC))
 | 
			
		||||
OBJ       = $(patsubst $(SRC_DIR)/%.c, $(BUILD_DIR)/%.o, $(SRC))
 | 
			
		||||
OBJ      += $(BUILD_DIR)/comm-$(COMM_TYPE).o
 | 
			
		||||
OBJ      += $(BUILD_DIR)/solver-$(SOLVER).o
 | 
			
		||||
SOURCES   = $(SRC) $(wildcard $(SRC_DIR)/*.h)
 | 
			
		||||
CPPFLAGS := $(CPPFLAGS) $(DEFINES) $(OPTIONS) $(INCLUDES)
 | 
			
		||||
 | 
			
		||||
@@ -38,9 +39,20 @@ $(BUILD_DIR)/%.s:  %.c
 | 
			
		||||
	$(info ===>  GENERATE ASM  $@)
 | 
			
		||||
	$(CC) -S $(CPPFLAGS) $(CFLAGS) $< -o $@
 | 
			
		||||
 | 
			
		||||
.PHONY: clean distclean tags info asm format
 | 
			
		||||
.PHONY: clean distclean vis vis_clean tags info asm format
 | 
			
		||||
 | 
			
		||||
clean:
 | 
			
		||||
vis:
 | 
			
		||||
	$(info ===>  GENERATE VISUALIZATION)
 | 
			
		||||
	@gnuplot -e "filename='pressure.dat'" ./surface.plot
 | 
			
		||||
	@gnuplot -e "filename='velocity.dat'" ./vector.plot
 | 
			
		||||
	@gnuplot -e "filename='residual.dat'" ./residual.plot
 | 
			
		||||
 | 
			
		||||
vis_clean:
 | 
			
		||||
	$(info ===>  CLEAN VISUALIZATION)
 | 
			
		||||
	@rm -f *.dat
 | 
			
		||||
	@rm -f *.png
 | 
			
		||||
 | 
			
		||||
clean: vis_clean
 | 
			
		||||
	$(info ===>  CLEAN)
 | 
			
		||||
	@rm -rf $(BUILD_DIR)
 | 
			
		||||
	@rm -f tags
 | 
			
		||||
 
 | 
			
		||||
@@ -7,10 +7,10 @@
 | 
			
		||||
 | 
			
		||||
name canal             # name of flow setup
 | 
			
		||||
 | 
			
		||||
bcN     1              #  flags for boundary conditions
 | 
			
		||||
bcE     3              #  1 = no-slip      3 = outflow
 | 
			
		||||
bcS     1              #  2 = free-slip    4 = periodic
 | 
			
		||||
bcW     3              #
 | 
			
		||||
bcTop      1			#  flags for boundary conditions
 | 
			
		||||
bcBottom   1			#  1 = no-slip      3 = outflow
 | 
			
		||||
bcLeft     3			#  2 = free-slip    4 = periodic
 | 
			
		||||
bcRight    3			#
 | 
			
		||||
 | 
			
		||||
gx     0.0      # Body forces (e.g. gravity)
 | 
			
		||||
gy     0.0      #
 | 
			
		||||
@@ -27,15 +27,22 @@ p_init        0.0      # initial value for pressure
 | 
			
		||||
xlength       30.0     # domain size in x-direction
 | 
			
		||||
ylength       4.0	   # domain size in y-direction
 | 
			
		||||
imax          200      # number of interior cells in x-direction
 | 
			
		||||
jmax          50	   # number of interior cells in y-direction
 | 
			
		||||
jmax          40	   # number of interior cells in y-direction
 | 
			
		||||
 | 
			
		||||
# Time Data:
 | 
			
		||||
# ---------
 | 
			
		||||
 | 
			
		||||
te       100.0   # final time
 | 
			
		||||
te       60.0   # final time
 | 
			
		||||
dt       0.02    # time stepsize
 | 
			
		||||
tau      0.5     # safety factor for time stepsize control (<0 constant delt)
 | 
			
		||||
 | 
			
		||||
# Multigrid data:
 | 
			
		||||
# ---------
 | 
			
		||||
 | 
			
		||||
levels        2         # Multigrid levels
 | 
			
		||||
presmooth     5         # Pre-smoothning iterations
 | 
			
		||||
postsmooth    5         # Post-smoothning iterations
 | 
			
		||||
 | 
			
		||||
# Pressure Iteration Data:
 | 
			
		||||
# -----------------------
 | 
			
		||||
 | 
			
		||||
 
 | 
			
		||||
@@ -1,7 +1,11 @@
 | 
			
		||||
# Supported: GCC, CLANG, ICC
 | 
			
		||||
TAG ?= CLANG
 | 
			
		||||
TAG ?= ICC
 | 
			
		||||
# Supported: true, false
 | 
			
		||||
ENABLE_MPI ?= true
 | 
			
		||||
ENABLE_OPENMP ?= false
 | 
			
		||||
# Supported: rb, mg
 | 
			
		||||
SOLVER ?= mg
 | 
			
		||||
# Supported: v1, v2, v3
 | 
			
		||||
COMM_TYPE ?= v3
 | 
			
		||||
 | 
			
		||||
#Feature options
 | 
			
		||||
 
 | 
			
		||||
@@ -26,21 +26,28 @@ p_init    0.0		# initial value for pressure
 | 
			
		||||
 | 
			
		||||
xlength    1.0		# domain size in x-direction
 | 
			
		||||
ylength    1.0		# domain size in y-direction
 | 
			
		||||
imax       80		# number of interior cells in x-direction
 | 
			
		||||
jmax       80		# number of interior cells in y-direction
 | 
			
		||||
imax       128		# number of interior cells in x-direction
 | 
			
		||||
jmax       128		# number of interior cells in y-direction
 | 
			
		||||
 | 
			
		||||
# Time Data:
 | 
			
		||||
# ---------
 | 
			
		||||
 | 
			
		||||
te      10.0		# final time
 | 
			
		||||
dt     0.02	    # time stepsize
 | 
			
		||||
tau     0.5		# safety factor for time stepsize control (<0 constant delt)
 | 
			
		||||
dt      0.02	    # time stepsize
 | 
			
		||||
tau     0.5		    # safety factor for time stepsize control (<0 constant delt)
 | 
			
		||||
 | 
			
		||||
# Multigrid data:
 | 
			
		||||
# ---------
 | 
			
		||||
 | 
			
		||||
levels        2         # Multigrid levels
 | 
			
		||||
presmooth     20        # Pre-smoothning iterations
 | 
			
		||||
postsmooth    5         # Post-smoothning iterations
 | 
			
		||||
 | 
			
		||||
# Pressure Iteration Data:
 | 
			
		||||
# -----------------------
 | 
			
		||||
 | 
			
		||||
itermax  1000		# maximal number of pressure iteration in one time step
 | 
			
		||||
eps      0.001		# stopping tolerance for pressure iteration
 | 
			
		||||
omg      1.9		# relaxation parameter for SOR iteration
 | 
			
		||||
omg      1.7		# relaxation parameter for SOR iteration
 | 
			
		||||
gamma    0.9		# upwind differencing factor gamma
 | 
			
		||||
#===============================================================================
 | 
			
		||||
 
 | 
			
		||||
							
								
								
									
										9
									
								
								BasicSolver/2D-mpi/residual.plot
									
									
									
									
									
										Normal file
									
								
							
							
						
						
									
										9
									
								
								BasicSolver/2D-mpi/residual.plot
									
									
									
									
									
										Normal file
									
								
							@@ -0,0 +1,9 @@
 | 
			
		||||
set terminal png size 1800,768 enhanced font ,12
 | 
			
		||||
set output 'residual.png'
 | 
			
		||||
set datafile separator whitespace
 | 
			
		||||
set xlabel "Timestep"
 | 
			
		||||
set ylabel "Residual"
 | 
			
		||||
 | 
			
		||||
set logscale y 2
 | 
			
		||||
 | 
			
		||||
plot 'residual.dat' using 1:2 title "Residual"
 | 
			
		||||
@@ -14,7 +14,7 @@ static int sum(int* sizes, int position)
 | 
			
		||||
{
 | 
			
		||||
    int sum = 0;
 | 
			
		||||
 | 
			
		||||
    for (int i = 0; i < position; i++) {
 | 
			
		||||
    for (int i = 0; i < position; i += position) {
 | 
			
		||||
        sum += sizes[i];
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
@@ -67,6 +67,7 @@ int commIsBoundary(Comm* c, int direction)
 | 
			
		||||
 | 
			
		||||
void commExchange(Comm* c, double* grid)
 | 
			
		||||
{
 | 
			
		||||
    // printf("Rank : %d In exchange \n", c->rank);
 | 
			
		||||
#ifdef _MPI
 | 
			
		||||
    MPI_Request requests[4] = { MPI_REQUEST_NULL,
 | 
			
		||||
        MPI_REQUEST_NULL,
 | 
			
		||||
@@ -173,9 +174,61 @@ void commPartition(Comm* c, int jmax, int imax)
 | 
			
		||||
{
 | 
			
		||||
#ifdef _MPI
 | 
			
		||||
    c->imaxLocal = imax;
 | 
			
		||||
    c->jmaxLocal = sizeOfRank(c->rank, c->size, jmax);
 | 
			
		||||
    c->jmaxLocal = sizeOfRank(c->coords[JDIM], c->size, jmax);
 | 
			
		||||
 | 
			
		||||
    c->neighbours[BOTTOM] = c->rank == 0 ? -1 : c->rank - 1;
 | 
			
		||||
    c->neighbours[TOP]    = c->rank == (c->size - 1) ? -1 : c->rank + 1;
 | 
			
		||||
    c->neighbours[LEFT]   = -1;
 | 
			
		||||
    c->neighbours[RIGHT]  = -1;
 | 
			
		||||
 | 
			
		||||
    c->coords[IDIM] = 0;
 | 
			
		||||
    c->coords[JDIM] = c->rank;
 | 
			
		||||
 | 
			
		||||
    c->dims[IDIM] = 1;
 | 
			
		||||
    c->dims[JDIM] = c->size;
 | 
			
		||||
#else
 | 
			
		||||
    c->imaxLocal = imax;
 | 
			
		||||
    c->jmaxLocal = jmax;
 | 
			
		||||
#endif
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void commUpdateDatatypes(Comm* oldcomm, Comm* newcomm, int imaxLocal, int jmaxLocal)
 | 
			
		||||
{
 | 
			
		||||
 | 
			
		||||
#if defined _MPI
 | 
			
		||||
    newcomm->comm = MPI_COMM_NULL;
 | 
			
		||||
    int result    = MPI_Comm_dup(MPI_COMM_WORLD, &newcomm->comm);
 | 
			
		||||
 | 
			
		||||
    if (result == MPI_ERR_COMM) {
 | 
			
		||||
        printf("\nNull communicator. Duplication failed !!\n");
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    newcomm->rank = oldcomm->rank;
 | 
			
		||||
    newcomm->size = oldcomm->size;
 | 
			
		||||
 | 
			
		||||
    newcomm->imaxLocal = imaxLocal / 2;
 | 
			
		||||
    newcomm->jmaxLocal = jmaxLocal / 2;
 | 
			
		||||
 | 
			
		||||
    newcomm->neighbours[BOTTOM] = newcomm->rank == 0 ? -1 : newcomm->rank - 1;
 | 
			
		||||
    newcomm->neighbours[TOP]    = newcomm->rank == (newcomm->size - 1) ? -1 : newcomm->rank + 1;
 | 
			
		||||
    newcomm->neighbours[LEFT]   = -1;
 | 
			
		||||
    newcomm->neighbours[RIGHT]  = -1;
 | 
			
		||||
 | 
			
		||||
    newcomm->coords[IDIM] = 0;
 | 
			
		||||
    newcomm->coords[JDIM] = newcomm->rank;
 | 
			
		||||
 | 
			
		||||
    newcomm->dims[IDIM] = 1;
 | 
			
		||||
    newcomm->dims[JDIM] = newcomm->size;
 | 
			
		||||
 | 
			
		||||
 | 
			
		||||
#endif
 | 
			
		||||
    newcomm->imaxLocal = imaxLocal;
 | 
			
		||||
    newcomm->jmaxLocal = jmaxLocal;
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void commFreeCommunicator(Comm* comm)
 | 
			
		||||
{
 | 
			
		||||
#ifdef _MPI
 | 
			
		||||
    MPI_Comm_free(&comm->comm);
 | 
			
		||||
#endif
 | 
			
		||||
}
 | 
			
		||||
@@ -4,18 +4,18 @@
 | 
			
		||||
 * Use of this source code is governed by a MIT style
 | 
			
		||||
 * license that can be found in the LICENSE file.
 | 
			
		||||
 */
 | 
			
		||||
#include "comm.h"
 | 
			
		||||
#include <stdio.h>
 | 
			
		||||
#include <stdlib.h>
 | 
			
		||||
 | 
			
		||||
#include "comm.h"
 | 
			
		||||
#include <string.h>
 | 
			
		||||
 | 
			
		||||
#ifdef _MPI
 | 
			
		||||
// subroutines local to this module
 | 
			
		||||
static int sum(int* sizes, int position)
 | 
			
		||||
static int sum(int* sizes, int init, int offset, int coord)
 | 
			
		||||
{
 | 
			
		||||
    int sum = 0;
 | 
			
		||||
 | 
			
		||||
    for (int i = 0; i < position; i++) {
 | 
			
		||||
    for (int i = init - offset; coord > 0; i -= offset, --coord) {
 | 
			
		||||
        sum += sizes[i];
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
@@ -79,8 +79,8 @@ static void assembleResult(Comm* c, double* src, double* dst, int imax, int jmax
 | 
			
		||||
            int newSizes[NDIMS] = { newSizesJ[i], newSizesI[i] };
 | 
			
		||||
            int coords[NDIMS];
 | 
			
		||||
            MPI_Cart_coords(c->comm, i, NDIMS, coords);
 | 
			
		||||
            int starts[NDIMS] = { sum(newSizesJ, coords[JDIM]),
 | 
			
		||||
                sum(newSizesI, coords[IDIM]) };
 | 
			
		||||
            int starts[NDIMS] = { sum(newSizesJ, i, 1, coords[JDIM]),
 | 
			
		||||
                sum(newSizesI, i, c->dims[JDIM], coords[IDIM]) };
 | 
			
		||||
            printf(
 | 
			
		||||
                "Rank: %d, Coords(i,j): %d %d, Size(i,j): %d %d, Target Size(i,j): %d %d "
 | 
			
		||||
                "Starts(i,j): %d %d\n",
 | 
			
		||||
@@ -252,8 +252,8 @@ void commPartition(Comm* c, int jmax, int imax)
 | 
			
		||||
    MPI_Cart_shift(c->comm, JDIM, 1, &c->neighbours[BOTTOM], &c->neighbours[TOP]);
 | 
			
		||||
    MPI_Cart_get(c->comm, NDIMS, c->dims, periods, c->coords);
 | 
			
		||||
 | 
			
		||||
    int imaxLocal = sizeOfRank(c->rank, dims[IDIM], imax);
 | 
			
		||||
    int jmaxLocal = sizeOfRank(c->rank, dims[JDIM], jmax);
 | 
			
		||||
    int imaxLocal = sizeOfRank(c->coords[IDIM], dims[IDIM], imax);
 | 
			
		||||
    int jmaxLocal = sizeOfRank(c->coords[JDIM], dims[JDIM], jmax);
 | 
			
		||||
 | 
			
		||||
    c->imaxLocal = imaxLocal;
 | 
			
		||||
    c->jmaxLocal = jmaxLocal;
 | 
			
		||||
@@ -285,3 +285,58 @@ void commPartition(Comm* c, int jmax, int imax)
 | 
			
		||||
    c->jmaxLocal = jmax;
 | 
			
		||||
#endif
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void commUpdateDatatypes(Comm* oldcomm, Comm* newcomm, int imaxLocal, int jmaxLocal)
 | 
			
		||||
{
 | 
			
		||||
#if defined _MPI
 | 
			
		||||
    newcomm->comm = MPI_COMM_NULL;
 | 
			
		||||
    int result    = MPI_Comm_dup(oldcomm->comm, &newcomm->comm);
 | 
			
		||||
 | 
			
		||||
    if (result == MPI_ERR_COMM) {
 | 
			
		||||
        printf("\nNull communicator. Duplication failed !!\n");
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    newcomm->rank = oldcomm->rank;
 | 
			
		||||
    newcomm->size = oldcomm->size;
 | 
			
		||||
 | 
			
		||||
    memcpy(&newcomm->neighbours, &oldcomm->neighbours, sizeof(oldcomm->neighbours));
 | 
			
		||||
    memcpy(&newcomm->coords, &oldcomm->coords, sizeof(oldcomm->coords));
 | 
			
		||||
    memcpy(&newcomm->dims, &oldcomm->dims, sizeof(oldcomm->dims));
 | 
			
		||||
 | 
			
		||||
    newcomm->imaxLocal = imaxLocal/2;
 | 
			
		||||
    newcomm->jmaxLocal = jmaxLocal/2;
 | 
			
		||||
 | 
			
		||||
    MPI_Datatype jBufferType;
 | 
			
		||||
    MPI_Type_contiguous(imaxLocal, MPI_DOUBLE, &jBufferType);
 | 
			
		||||
    MPI_Type_commit(&jBufferType);
 | 
			
		||||
 | 
			
		||||
    MPI_Datatype iBufferType;
 | 
			
		||||
    MPI_Type_vector(jmaxLocal, 1, imaxLocal + 2, MPI_DOUBLE, &iBufferType);
 | 
			
		||||
    MPI_Type_commit(&iBufferType);
 | 
			
		||||
 | 
			
		||||
    newcomm->bufferTypes[LEFT]   = iBufferType;
 | 
			
		||||
    newcomm->bufferTypes[RIGHT]  = iBufferType;
 | 
			
		||||
    newcomm->bufferTypes[BOTTOM] = jBufferType;
 | 
			
		||||
    newcomm->bufferTypes[TOP]    = jBufferType;
 | 
			
		||||
 | 
			
		||||
    newcomm->sdispls[LEFT]   = (imaxLocal + 2) + 1;
 | 
			
		||||
    newcomm->sdispls[RIGHT]  = (imaxLocal + 2) + imaxLocal;
 | 
			
		||||
    newcomm->sdispls[BOTTOM] = (imaxLocal + 2) + 1;
 | 
			
		||||
    newcomm->sdispls[TOP]    = jmaxLocal * (imaxLocal + 2) + 1;
 | 
			
		||||
 | 
			
		||||
    newcomm->rdispls[LEFT]   = (imaxLocal + 2);
 | 
			
		||||
    newcomm->rdispls[RIGHT]  = (imaxLocal + 2) + (imaxLocal + 1);
 | 
			
		||||
    newcomm->rdispls[BOTTOM] = 1;
 | 
			
		||||
    newcomm->rdispls[TOP]    = (jmaxLocal + 1) * (imaxLocal + 2) + 1;
 | 
			
		||||
#else
 | 
			
		||||
    newcomm->imaxLocal = imaxLocal;
 | 
			
		||||
    newcomm->jmaxLocal = jmaxLocal;
 | 
			
		||||
#endif
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void commFreeCommunicator(Comm* comm)
 | 
			
		||||
{
 | 
			
		||||
#ifdef _MPI
 | 
			
		||||
    MPI_Comm_free(&comm->comm);
 | 
			
		||||
#endif
 | 
			
		||||
}
 | 
			
		||||
@@ -11,11 +11,11 @@
 | 
			
		||||
 | 
			
		||||
#ifdef _MPI
 | 
			
		||||
// subroutines local to this module
 | 
			
		||||
static int sum(int* sizes, int position)
 | 
			
		||||
static int sum(int* sizes, int init, int offset, int coord)
 | 
			
		||||
{
 | 
			
		||||
    int sum = 0;
 | 
			
		||||
 | 
			
		||||
    for (int i = 0; i < position; i++) {
 | 
			
		||||
    for (int i = init - offset; coord > 0; i -= offset, --coord) {
 | 
			
		||||
        sum += sizes[i];
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
@@ -79,8 +79,8 @@ static void assembleResult(Comm* c, double* src, double* dst, int imax, int jmax
 | 
			
		||||
            int newSizes[NDIMS] = { newSizesJ[i], newSizesI[i] };
 | 
			
		||||
            int coords[NDIMS];
 | 
			
		||||
            MPI_Cart_coords(c->comm, i, NDIMS, coords);
 | 
			
		||||
            int starts[NDIMS] = { sum(newSizesJ, coords[JDIM]),
 | 
			
		||||
                sum(newSizesI, coords[IDIM]) };
 | 
			
		||||
            int starts[NDIMS] = { sum(newSizesJ, i, 1, coords[JDIM]),
 | 
			
		||||
                sum(newSizesI, i, c->dims[JDIM], coords[IDIM]) };
 | 
			
		||||
            printf(
 | 
			
		||||
                "Rank: %d, Coords(i,j): %d %d, Size(i,j): %d %d, Target Size(i,j): %d %d "
 | 
			
		||||
                "Starts(i,j): %d %d\n",
 | 
			
		||||
@@ -139,7 +139,6 @@ void commExchange(Comm* c, double* grid)
 | 
			
		||||
{
 | 
			
		||||
#ifdef _MPI
 | 
			
		||||
    int counts[NDIRS] = { 1, 1, 1, 1 };
 | 
			
		||||
 | 
			
		||||
    MPI_Neighbor_alltoallw(grid,
 | 
			
		||||
        counts,
 | 
			
		||||
        c->sdispls,
 | 
			
		||||
@@ -233,8 +232,8 @@ void commPartition(Comm* c, int jmax, int imax)
 | 
			
		||||
    MPI_Cart_shift(c->comm, JDIM, 1, &c->neighbours[BOTTOM], &c->neighbours[TOP]);
 | 
			
		||||
    MPI_Cart_get(c->comm, NDIMS, c->dims, periods, c->coords);
 | 
			
		||||
 | 
			
		||||
    int imaxLocal = sizeOfRank(c->rank, dims[IDIM], imax);
 | 
			
		||||
    int jmaxLocal = sizeOfRank(c->rank, dims[JDIM], jmax);
 | 
			
		||||
    int imaxLocal = sizeOfRank(c->coords[IDIM], dims[IDIM], imax);
 | 
			
		||||
    int jmaxLocal = sizeOfRank(c->coords[JDIM], dims[JDIM], jmax);
 | 
			
		||||
 | 
			
		||||
    c->imaxLocal = imaxLocal;
 | 
			
		||||
    c->jmaxLocal = jmaxLocal;
 | 
			
		||||
@@ -267,3 +266,55 @@ void commPartition(Comm* c, int jmax, int imax)
 | 
			
		||||
    c->jmaxLocal = jmax;
 | 
			
		||||
#endif
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void commUpdateDatatypes(Comm* oldcomm, Comm* newcomm, int imaxLocal, int jmaxLocal)
 | 
			
		||||
{
 | 
			
		||||
#if defined _MPI
 | 
			
		||||
 | 
			
		||||
    int result = MPI_Comm_dup(oldcomm->comm, &newcomm->comm);
 | 
			
		||||
 | 
			
		||||
    if (result == MPI_ERR_COMM) {
 | 
			
		||||
        printf("\nNull communicator. Duplication failed !!\n");
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    newcomm->rank = oldcomm->rank;
 | 
			
		||||
    newcomm->size = oldcomm->size;
 | 
			
		||||
    
 | 
			
		||||
 | 
			
		||||
    newcomm->imaxLocal = imaxLocal / 2;
 | 
			
		||||
    newcomm->jmaxLocal = jmaxLocal / 2;
 | 
			
		||||
 | 
			
		||||
    MPI_Datatype jBufferType;
 | 
			
		||||
    MPI_Type_contiguous(imaxLocal, MPI_DOUBLE, &jBufferType);
 | 
			
		||||
    MPI_Type_commit(&jBufferType);
 | 
			
		||||
 | 
			
		||||
    MPI_Datatype iBufferType;
 | 
			
		||||
    MPI_Type_vector(jmaxLocal, 1, imaxLocal + 2, MPI_DOUBLE, &iBufferType);
 | 
			
		||||
    MPI_Type_commit(&iBufferType);
 | 
			
		||||
 | 
			
		||||
    newcomm->bufferTypes[LEFT]   = iBufferType;
 | 
			
		||||
    newcomm->bufferTypes[RIGHT]  = iBufferType;
 | 
			
		||||
    newcomm->bufferTypes[BOTTOM] = jBufferType;
 | 
			
		||||
    newcomm->bufferTypes[TOP]    = jBufferType;
 | 
			
		||||
 | 
			
		||||
    newcomm->sdispls[LEFT]   = (imaxLocal + 2) + 1;
 | 
			
		||||
    newcomm->sdispls[RIGHT]  = (imaxLocal + 2) + imaxLocal;
 | 
			
		||||
    newcomm->sdispls[BOTTOM] = (imaxLocal + 2) + 1;
 | 
			
		||||
    newcomm->sdispls[TOP]    = jmaxLocal * (imaxLocal + 2) + 1;
 | 
			
		||||
 | 
			
		||||
    newcomm->rdispls[LEFT]   = (imaxLocal + 2);
 | 
			
		||||
    newcomm->rdispls[RIGHT]  = (imaxLocal + 2) + (imaxLocal + 1);
 | 
			
		||||
    newcomm->rdispls[BOTTOM] = 1;
 | 
			
		||||
    newcomm->rdispls[TOP]    = (jmaxLocal + 1) * (imaxLocal + 2) + 1;
 | 
			
		||||
#else
 | 
			
		||||
    newcomm->imaxLocal = imaxLocal;
 | 
			
		||||
    newcomm->jmaxLocal = jmaxLocal;
 | 
			
		||||
#endif
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void commFreeCommunicator(Comm* comm)
 | 
			
		||||
{
 | 
			
		||||
#ifdef _MPI
 | 
			
		||||
    MPI_Comm_free(&comm->comm);
 | 
			
		||||
#endif
 | 
			
		||||
}
 | 
			
		||||
@@ -41,6 +41,8 @@ extern void commExchange(Comm*, double*);
 | 
			
		||||
extern void commShift(Comm* c, double* f, double* g);
 | 
			
		||||
extern void commReduction(double* v, int op);
 | 
			
		||||
extern int commIsBoundary(Comm* c, int direction);
 | 
			
		||||
extern void commUpdateDatatypes(Comm*, Comm*, int, int);
 | 
			
		||||
extern void commFreeCommunicator(Comm*);
 | 
			
		||||
extern void commCollectResult(Comm* c,
 | 
			
		||||
    double* ug,
 | 
			
		||||
    double* vg,
 | 
			
		||||
 
 | 
			
		||||
@@ -273,17 +273,18 @@ void setSpecialBoundaryCondition(Discretization* s)
 | 
			
		||||
        if (commIsBoundary(&s->comm, LEFT)) {
 | 
			
		||||
            double ylength = s->grid.ylength;
 | 
			
		||||
            double dy      = s->grid.dy;
 | 
			
		||||
            int rest       = s->grid.jmax % s->comm.size;
 | 
			
		||||
            int yc         = s->comm.rank * (s->grid.jmax / s->comm.size) +
 | 
			
		||||
            int rest       = s->grid.jmax % s->comm.dims[JDIM];
 | 
			
		||||
            int yc         = s->comm.rank * (s->grid.jmax / s->comm.dims[JDIM]) +
 | 
			
		||||
                     MIN(rest, s->comm.rank);
 | 
			
		||||
            double ys = dy * (yc + 0.5);
 | 
			
		||||
            double y;
 | 
			
		||||
 | 
			
		||||
            /* printf("RANK %d yc: %d ys: %f\n", solver->rank, yc, ys); */
 | 
			
		||||
            // printf("RANK %d yc: %d ys: %f\n", s->comm.rank, yc, ys);
 | 
			
		||||
 | 
			
		||||
            for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                y       = ys + dy * (j - 0.5);
 | 
			
		||||
                U(0, j) = y * (ylength - y) * 4.0 / (ylength * ylength);
 | 
			
		||||
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 
 | 
			
		||||
@@ -26,7 +26,9 @@ static void writeResults(Discretization* s)
 | 
			
		||||
    double* pg = allocate(64, bytesize);
 | 
			
		||||
 | 
			
		||||
    commCollectResult(&s->comm, ug, vg, pg, s->u, s->v, s->p, s->grid.imax, s->grid.jmax);
 | 
			
		||||
    writeResult(s, ug, vg, pg);
 | 
			
		||||
    if (commIsMaster(&s->comm)) {
 | 
			
		||||
        writeResult(s, ug, vg, pg);
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    free(ug);
 | 
			
		||||
    free(vg);
 | 
			
		||||
@@ -47,6 +49,9 @@ int main(int argc, char** argv)
 | 
			
		||||
    commInit(&d.comm, argc, argv);
 | 
			
		||||
    initParameter(&p);
 | 
			
		||||
 | 
			
		||||
    FILE* fp;
 | 
			
		||||
    if (commIsMaster(&d.comm)) fp = initResidualWriter();
 | 
			
		||||
 | 
			
		||||
    if (argc != 2) {
 | 
			
		||||
        printf("Usage: %s <configFile>\n", argv[0]);
 | 
			
		||||
        exit(EXIT_SUCCESS);
 | 
			
		||||
@@ -77,16 +82,21 @@ int main(int argc, char** argv)
 | 
			
		||||
    double tau = d.tau;
 | 
			
		||||
    double te  = d.te;
 | 
			
		||||
    double t   = 0.0;
 | 
			
		||||
    double res = 0.0;
 | 
			
		||||
 | 
			
		||||
    timeStart = getTimeStamp();
 | 
			
		||||
    while (t <= te) {
 | 
			
		||||
        
 | 
			
		||||
        if (tau > 0.0) computeTimestep(&d);
 | 
			
		||||
        setBoundaryConditions(&d);
 | 
			
		||||
        setSpecialBoundaryCondition(&d);
 | 
			
		||||
        computeFG(&d);
 | 
			
		||||
        computeRHS(&d);
 | 
			
		||||
        solve(&s, d.p, d.rhs);
 | 
			
		||||
        res = solve(&s, d.p, d.rhs);
 | 
			
		||||
        adaptUV(&d);
 | 
			
		||||
 | 
			
		||||
        if (commIsMaster(&d.comm)) writeResidual(fp, t, res);
 | 
			
		||||
 | 
			
		||||
        t += d.dt;
 | 
			
		||||
 | 
			
		||||
#ifdef VERBOSE
 | 
			
		||||
@@ -104,7 +114,7 @@ int main(int argc, char** argv)
 | 
			
		||||
    if (commIsMaster(s.comm)) {
 | 
			
		||||
        printf("Solution took %.2fs\n", timeStop - timeStart);
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    if (commIsMaster(&d.comm)) fclose(fp);
 | 
			
		||||
    writeResults(&d);
 | 
			
		||||
    commFinalize(s.comm);
 | 
			
		||||
    return EXIT_SUCCESS;
 | 
			
		||||
 
 | 
			
		||||
@@ -14,13 +14,16 @@
 | 
			
		||||
 | 
			
		||||
void initParameter(Parameter* param)
 | 
			
		||||
{
 | 
			
		||||
    param->xlength = 1.0;
 | 
			
		||||
    param->ylength = 1.0;
 | 
			
		||||
    param->imax    = 100;
 | 
			
		||||
    param->jmax    = 100;
 | 
			
		||||
    param->itermax = 1000;
 | 
			
		||||
    param->eps     = 0.0001;
 | 
			
		||||
    param->omg     = 1.8;
 | 
			
		||||
    param->xlength    = 1.0;
 | 
			
		||||
    param->ylength    = 1.0;
 | 
			
		||||
    param->imax       = 100;
 | 
			
		||||
    param->jmax       = 100;
 | 
			
		||||
    param->itermax    = 1000;
 | 
			
		||||
    param->eps        = 0.0001;
 | 
			
		||||
    param->omg        = 1.8;
 | 
			
		||||
    param->levels     = 5;
 | 
			
		||||
    param->presmooth  = 5;
 | 
			
		||||
    param->postsmooth = 5;
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void readParameter(Parameter* param, const char* filename)
 | 
			
		||||
@@ -72,6 +75,9 @@ void readParameter(Parameter* param, const char* filename)
 | 
			
		||||
            PARSE_INT(bcRight);
 | 
			
		||||
            PARSE_INT(bcBottom);
 | 
			
		||||
            PARSE_INT(bcTop);
 | 
			
		||||
            PARSE_INT(levels);
 | 
			
		||||
            PARSE_INT(presmooth);
 | 
			
		||||
            PARSE_INT(postsmooth);
 | 
			
		||||
            PARSE_REAL(u_init);
 | 
			
		||||
            PARSE_REAL(v_init);
 | 
			
		||||
            PARSE_REAL(p_init);
 | 
			
		||||
 
 | 
			
		||||
@@ -18,6 +18,7 @@ typedef struct {
 | 
			
		||||
    char* name;
 | 
			
		||||
    int bcLeft, bcRight, bcBottom, bcTop;
 | 
			
		||||
    double u_init, v_init, p_init;
 | 
			
		||||
    int levels, presmooth, postsmooth;
 | 
			
		||||
} Parameter;
 | 
			
		||||
 | 
			
		||||
void initParameter(Parameter*);
 | 
			
		||||
 
 | 
			
		||||
@@ -8,7 +8,7 @@
 | 
			
		||||
#include <mpi.h>
 | 
			
		||||
#include <stdio.h>
 | 
			
		||||
#include <string.h>
 | 
			
		||||
 | 
			
		||||
#include <stdlib.h>
 | 
			
		||||
#include "progress.h"
 | 
			
		||||
 | 
			
		||||
static double _end;
 | 
			
		||||
@@ -49,3 +49,22 @@ void stopProgress()
 | 
			
		||||
    printf("\n");
 | 
			
		||||
    fflush(stdout);
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
FILE* initResidualWriter()
 | 
			
		||||
{
 | 
			
		||||
    FILE* fp;
 | 
			
		||||
    fp = fopen("residual.dat", "w");
 | 
			
		||||
 | 
			
		||||
    if (fp == NULL) {
 | 
			
		||||
        printf("Error!\n");
 | 
			
		||||
        exit(EXIT_FAILURE);
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    return fp;
 | 
			
		||||
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void writeResidual(FILE* fp, double ts, double res)
 | 
			
		||||
{
 | 
			
		||||
    fprintf(fp, "%f, %f\n", ts, res);
 | 
			
		||||
}
 | 
			
		||||
 
 | 
			
		||||
@@ -10,5 +10,6 @@
 | 
			
		||||
extern void initProgress(double);
 | 
			
		||||
extern void printProgress(double);
 | 
			
		||||
extern void stopProgress();
 | 
			
		||||
 | 
			
		||||
extern FILE* initResidualWriter(void);
 | 
			
		||||
extern void writeResidual(FILE*, double, double);
 | 
			
		||||
#endif
 | 
			
		||||
 
 | 
			
		||||
							
								
								
									
										302
									
								
								BasicSolver/2D-mpi/src/solver-mg.c
									
									
									
									
									
										Normal file
									
								
							
							
						
						
									
										302
									
								
								BasicSolver/2D-mpi/src/solver-mg.c
									
									
									
									
									
										Normal file
									
								
							@@ -0,0 +1,302 @@
 | 
			
		||||
/*
 | 
			
		||||
 * Copyright (C) NHR@FAU, University Erlangen-Nuremberg.
 | 
			
		||||
 * All rights reserved. This file is part of nusif-solver.
 | 
			
		||||
 * Use of this source code is governed by a MIT style
 | 
			
		||||
 * license that can be found in the LICENSE file.
 | 
			
		||||
 */
 | 
			
		||||
#include <stdio.h>
 | 
			
		||||
#include <stdlib.h>
 | 
			
		||||
 | 
			
		||||
#include "allocate.h"
 | 
			
		||||
#include "solver.h"
 | 
			
		||||
#include "util.h"
 | 
			
		||||
 | 
			
		||||
#define FINEST_LEVEL   0
 | 
			
		||||
#define COARSEST_LEVEL (s->levels - 1)
 | 
			
		||||
// #define S(i, j)        s[(j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define E(i, j)   e[(j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define R(i, j)   r[(j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define OLD(i, j) old[(j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
 | 
			
		||||
static void restrictMG(Solver* s, int level, Comm* comm)
 | 
			
		||||
{
 | 
			
		||||
    int imaxLocal = comm->imaxLocal;
 | 
			
		||||
    int jmaxLocal = comm->jmaxLocal;
 | 
			
		||||
 | 
			
		||||
    double* r   = s->r[level + 1];
 | 
			
		||||
    double* old = s->r[level];
 | 
			
		||||
 | 
			
		||||
#ifdef _MPI
 | 
			
		||||
    commExchange(comm, old);
 | 
			
		||||
#endif
 | 
			
		||||
 | 
			
		||||
    for (int j = 1; j < (jmaxLocal / 2) + 1; j++) {
 | 
			
		||||
        for (int i = 1; i < (imaxLocal / 2) + 1; i++) {
 | 
			
		||||
            R(i, j) = (OLD(2 * i - 1, 2 * j - 1) + OLD(2 * i, 2 * j - 1) * 2 +
 | 
			
		||||
                          OLD(2 * i + 1, 2 * j - 1) + OLD(2 * i - 1, 2 * j) * 2 +
 | 
			
		||||
                          OLD(2 * i, 2 * j) * 4 + OLD(2 * i + 1, 2 * j) * 2 +
 | 
			
		||||
                          OLD(2 * i - 1, 2 * j + 1) + OLD(2 * i, 2 * j + 1) * 2 +
 | 
			
		||||
                          OLD(2 * i + 1, 2 * j + 1)) /
 | 
			
		||||
                      16.0;
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
static void prolongate(Solver* s, int level, Comm* comm)
 | 
			
		||||
{
 | 
			
		||||
    int imaxLocal = comm->imaxLocal;
 | 
			
		||||
    int jmaxLocal = comm->jmaxLocal;
 | 
			
		||||
 | 
			
		||||
    double* old = s->r[level + 1];
 | 
			
		||||
    double* e   = s->r[level];
 | 
			
		||||
 | 
			
		||||
    for (int j = 2; j < jmaxLocal + 1; j += 2) {
 | 
			
		||||
        for (int i = 2; i < imaxLocal + 1; i += 2) {
 | 
			
		||||
            E(i, j) = OLD(i / 2, j / 2);
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
static void correct(Solver* s, double* p, int level, Comm* comm)
 | 
			
		||||
{
 | 
			
		||||
    double* e     = s->e[level];
 | 
			
		||||
    int imaxLocal = comm->imaxLocal;
 | 
			
		||||
    int jmaxLocal = comm->jmaxLocal;
 | 
			
		||||
 | 
			
		||||
    for (int j = 1; j < jmaxLocal + 1; ++j) {
 | 
			
		||||
        for (int i = 1; i < imaxLocal + 1; ++i) {
 | 
			
		||||
            P(i, j) += E(i, j);
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
static void setBoundaryCondition(Solver* s, double* p, int imaxLocal, int jmaxLocal)
 | 
			
		||||
{
 | 
			
		||||
#ifdef _MPI
 | 
			
		||||
    if (commIsBoundary(s->comm, BOTTOM)) { // set bottom bc
 | 
			
		||||
        for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
            P(i, 0) = P(i, 1);
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    if (commIsBoundary(s->comm, TOP)) { // set top bc
 | 
			
		||||
        for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
            P(i, jmaxLocal + 1) = P(i, jmaxLocal);
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    if (commIsBoundary(s->comm, LEFT)) { // set left bc
 | 
			
		||||
        for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
            P(0, j) = P(1, j);
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    if (commIsBoundary(s->comm, RIGHT)) { // set right bc
 | 
			
		||||
        for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
            P(imaxLocal + 1, j) = P(imaxLocal, j);
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
#else
 | 
			
		||||
    for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
        P(i, 0)             = P(i, 1);
 | 
			
		||||
        P(i, jmaxLocal + 1) = P(i, jmaxLocal);
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
        P(0, j)             = P(1, j);
 | 
			
		||||
        P(imaxLocal + 1, j) = P(imaxLocal, j);
 | 
			
		||||
    }
 | 
			
		||||
#endif
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
static double smooth(Solver* s, double* p, double* rhs, int level, Comm* comm)
 | 
			
		||||
{
 | 
			
		||||
    int imaxLocal = comm->imaxLocal;
 | 
			
		||||
    int jmaxLocal = comm->jmaxLocal;
 | 
			
		||||
 | 
			
		||||
    int imax = s->grid->imax;
 | 
			
		||||
    int jmax = s->grid->jmax;
 | 
			
		||||
 | 
			
		||||
    double dx2  = s->grid->dx * s->grid->dx;
 | 
			
		||||
    double dy2  = s->grid->dy * s->grid->dy;
 | 
			
		||||
    double idx2 = 1.0 / dx2;
 | 
			
		||||
    double idy2 = 1.0 / dy2;
 | 
			
		||||
 | 
			
		||||
    double factor = s->omega * 0.5 * (dx2 * dy2) / (dx2 + dy2);
 | 
			
		||||
    double* r     = s->r[level];
 | 
			
		||||
 | 
			
		||||
    double res = 1.0;
 | 
			
		||||
    int pass, jsw, isw;
 | 
			
		||||
 | 
			
		||||
    jsw = 1;
 | 
			
		||||
 | 
			
		||||
    for (pass = 0; pass < 2; pass++) {
 | 
			
		||||
        isw = jsw;
 | 
			
		||||
 | 
			
		||||
#ifdef _MPI
 | 
			
		||||
        commExchange(comm, p);
 | 
			
		||||
#endif
 | 
			
		||||
 | 
			
		||||
        for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
            for (int i = isw; i < imaxLocal + 1; i += 2) {
 | 
			
		||||
 | 
			
		||||
                P(i, j) -= factor *
 | 
			
		||||
                           (RHS(i, j) -
 | 
			
		||||
                               ((P(i + 1, j) - 2.0 * P(i, j) + P(i - 1, j)) * idx2 +
 | 
			
		||||
                                   (P(i, j + 1) - 2.0 * P(i, j) + P(i, j - 1)) * idy2));
 | 
			
		||||
            }
 | 
			
		||||
            isw = 3 - isw;
 | 
			
		||||
        }
 | 
			
		||||
        jsw = 3 - jsw;
 | 
			
		||||
    }
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
static double calculateResidual(Solver* s, double* p, double* rhs, int level, Comm* comm)
 | 
			
		||||
{
 | 
			
		||||
    int imax      = s->grid->imax;
 | 
			
		||||
    int jmax      = s->grid->jmax;
 | 
			
		||||
    int imaxLocal = comm->imaxLocal;
 | 
			
		||||
    int jmaxLocal = comm->jmaxLocal;
 | 
			
		||||
 | 
			
		||||
    double dx2    = s->grid->dx * s->grid->dx;
 | 
			
		||||
    double dy2    = s->grid->dy * s->grid->dy;
 | 
			
		||||
    double idx2   = 1.0 / dx2;
 | 
			
		||||
    double idy2   = 1.0 / dy2;
 | 
			
		||||
    double factor = s->omega * 0.5 * (dx2 * dy2) / (dx2 + dy2);
 | 
			
		||||
    double* r     = s->r[level];
 | 
			
		||||
    double res    = 1.0;
 | 
			
		||||
    int pass, jsw, isw;
 | 
			
		||||
 | 
			
		||||
    jsw = 1;
 | 
			
		||||
 | 
			
		||||
    for (pass = 0; pass < 2; pass++) {
 | 
			
		||||
        isw = jsw;
 | 
			
		||||
 | 
			
		||||
#ifdef _MPI
 | 
			
		||||
        commExchange(comm, p);
 | 
			
		||||
#endif
 | 
			
		||||
 | 
			
		||||
        for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
            for (int i = isw; i < imaxLocal + 1; i += 2) {
 | 
			
		||||
 | 
			
		||||
                R(i, j) = RHS(i, j) -
 | 
			
		||||
                          ((P(i + 1, j) - 2.0 * P(i, j) + P(i - 1, j)) * idx2 +
 | 
			
		||||
                              (P(i, j + 1) - 2.0 * P(i, j) + P(i, j - 1)) * idy2);
 | 
			
		||||
 | 
			
		||||
                res += (R(i, j) * R(i, j));
 | 
			
		||||
            }
 | 
			
		||||
            isw = 3 - isw;
 | 
			
		||||
        }
 | 
			
		||||
        jsw = 3 - jsw;
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
#ifdef _MPI
 | 
			
		||||
    commReduction(&res, SUM);
 | 
			
		||||
#endif
 | 
			
		||||
 | 
			
		||||
    res = res / (double)(imax * jmax);
 | 
			
		||||
#ifdef DEBUG
 | 
			
		||||
    if (commIsMaster(s->comm)) {
 | 
			
		||||
        printf("%d Residuum: %e\n", it, res);
 | 
			
		||||
    }
 | 
			
		||||
#endif
 | 
			
		||||
    return res;
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
static double multiGrid(Solver* s, double* p, double* rhs, int level, Comm* comm)
 | 
			
		||||
{
 | 
			
		||||
    double res = 0.0;
 | 
			
		||||
 | 
			
		||||
    // coarsest level
 | 
			
		||||
    if (level == COARSEST_LEVEL) {
 | 
			
		||||
        for (int i = 0; i < 5; i++) {
 | 
			
		||||
            smooth(s, p, rhs, level, comm);
 | 
			
		||||
        }
 | 
			
		||||
        return res;
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    // pre-smoothing
 | 
			
		||||
    for (int i = 0; i < s->presmooth; i++) {
 | 
			
		||||
        smooth(s, p, rhs, level, comm);
 | 
			
		||||
        if (level == FINEST_LEVEL)
 | 
			
		||||
            setBoundaryCondition(s, p, comm->imaxLocal, comm->jmaxLocal);
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    // calculate residuals
 | 
			
		||||
    res = calculateResidual(s, p, rhs, level, comm);
 | 
			
		||||
 | 
			
		||||
    // restrict
 | 
			
		||||
    restrictMG(s, level, comm);
 | 
			
		||||
 | 
			
		||||
    Comm newcomm;
 | 
			
		||||
    commUpdateDatatypes(s->comm, &newcomm, comm->imaxLocal, comm->jmaxLocal);
 | 
			
		||||
 | 
			
		||||
 | 
			
		||||
    // MGSolver on residual and error.
 | 
			
		||||
    multiGrid(s, s->e[level + 1], s->r[level + 1], level + 1, &newcomm);
 | 
			
		||||
 | 
			
		||||
    commFreeCommunicator(&newcomm);
 | 
			
		||||
 | 
			
		||||
    // prolongate
 | 
			
		||||
    prolongate(s, level, comm);
 | 
			
		||||
 | 
			
		||||
    // correct p on finer level using residual
 | 
			
		||||
    correct(s, p, level, comm);
 | 
			
		||||
 | 
			
		||||
    if (level == FINEST_LEVEL)
 | 
			
		||||
        setBoundaryCondition(s, p, comm->imaxLocal, comm->jmaxLocal);
 | 
			
		||||
 | 
			
		||||
    // post-smoothing
 | 
			
		||||
    for (int i = 0; i < s->postsmooth; i++) {
 | 
			
		||||
        smooth(s, p, rhs, level, comm);
 | 
			
		||||
        if (level == FINEST_LEVEL)
 | 
			
		||||
            setBoundaryCondition(s, p, comm->imaxLocal, comm->jmaxLocal);
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    return res;
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void initSolver(Solver* s, Discretization* d, Parameter* p)
 | 
			
		||||
{
 | 
			
		||||
    s->eps        = p->eps;
 | 
			
		||||
    s->omega      = p->omg;
 | 
			
		||||
    s->itermax    = p->itermax;
 | 
			
		||||
    s->levels     = p->levels;
 | 
			
		||||
    s->grid       = &d->grid;
 | 
			
		||||
    s->comm       = &d->comm;
 | 
			
		||||
    s->presmooth  = p->presmooth;
 | 
			
		||||
    s->postsmooth = p->postsmooth;
 | 
			
		||||
 | 
			
		||||
    int imax   = s->grid->imax;
 | 
			
		||||
    int jmax   = s->grid->jmax;
 | 
			
		||||
    int levels = s->levels;
 | 
			
		||||
    printf("Using Multigrid solver with %d levels\n", levels);
 | 
			
		||||
 | 
			
		||||
    s->r = malloc(levels * sizeof(double*));
 | 
			
		||||
    s->e = malloc(levels * sizeof(double*));
 | 
			
		||||
 | 
			
		||||
    size_t size = (imax + 2) * (jmax + 2) * sizeof(double);
 | 
			
		||||
 | 
			
		||||
    for (int j = 0; j < levels; j++) {
 | 
			
		||||
        s->r[j] = allocate(64, size);
 | 
			
		||||
        s->e[j] = allocate(64, size);
 | 
			
		||||
 | 
			
		||||
        for (int i = 0; i < (imax + 2) * (jmax + 2); i++) {
 | 
			
		||||
            s->r[j][i] = 0.0;
 | 
			
		||||
            s->e[j][i] = 0.0;
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
double solve(Solver* s, double* p, double* rhs)
 | 
			
		||||
{
 | 
			
		||||
    double res = multiGrid(s, p, rhs, 0, s->comm);
 | 
			
		||||
 | 
			
		||||
#ifdef VERBOSE
 | 
			
		||||
    if (commIsMaster(s->comm)) {
 | 
			
		||||
        printf("Residuum: %.6f\n", res);
 | 
			
		||||
    }
 | 
			
		||||
#endif
 | 
			
		||||
 | 
			
		||||
    return res;
 | 
			
		||||
}
 | 
			
		||||
							
								
								
									
										106
									
								
								BasicSolver/2D-mpi/src/solver-rb.c
									
									
									
									
									
										Normal file
									
								
							
							
						
						
									
										106
									
								
								BasicSolver/2D-mpi/src/solver-rb.c
									
									
									
									
									
										Normal file
									
								
							@@ -0,0 +1,106 @@
 | 
			
		||||
/*
 | 
			
		||||
 * Copyright (C)  NHR@FAU, University Erlangen-Nuremberg.
 | 
			
		||||
 * All rights reserved. This file is part of nusif-solver.
 | 
			
		||||
 * Use of this source code is governed by a MIT style
 | 
			
		||||
 * license that can be found in the LICENSE file.
 | 
			
		||||
 */
 | 
			
		||||
#include <math.h>
 | 
			
		||||
#include <stdio.h>
 | 
			
		||||
 | 
			
		||||
#include "allocate.h"
 | 
			
		||||
#include "comm.h"
 | 
			
		||||
#include "discretization.h"
 | 
			
		||||
#include "parameter.h"
 | 
			
		||||
#include "solver.h"
 | 
			
		||||
#include "util.h"
 | 
			
		||||
 | 
			
		||||
void initSolver(Solver* s, Discretization* d, Parameter* p)
 | 
			
		||||
{
 | 
			
		||||
    s->grid    = &d->grid;
 | 
			
		||||
    s->eps     = p->eps;
 | 
			
		||||
    s->omega   = p->omg;
 | 
			
		||||
    s->itermax = p->itermax;
 | 
			
		||||
    s->comm    = &d->comm;
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
double solve(Solver* s, double* p, double* rhs)
 | 
			
		||||
{
 | 
			
		||||
    int imax      = s->grid->imax;
 | 
			
		||||
    int jmax      = s->grid->jmax;
 | 
			
		||||
    int imaxLocal = s->comm->imaxLocal;
 | 
			
		||||
    int jmaxLocal = s->comm->jmaxLocal;
 | 
			
		||||
    double eps    = s->eps;
 | 
			
		||||
    int itermax   = s->itermax;
 | 
			
		||||
    double dx2    = s->grid->dx * s->grid->dx;
 | 
			
		||||
    double dy2    = s->grid->dy * s->grid->dy;
 | 
			
		||||
    double idx2   = 1.0 / dx2;
 | 
			
		||||
    double idy2   = 1.0 / dy2;
 | 
			
		||||
    double factor = s->omega * 0.5 * (dx2 * dy2) / (dx2 + dy2);
 | 
			
		||||
    double epssq  = eps * eps;
 | 
			
		||||
    int pass, jsw, isw;
 | 
			
		||||
    int it     = 0;
 | 
			
		||||
    double res = 1.0;
 | 
			
		||||
 | 
			
		||||
    while ((res >= epssq) && (it < itermax)) {
 | 
			
		||||
        jsw = 1;
 | 
			
		||||
        for (pass = 0; pass < 2; pass++) {
 | 
			
		||||
            isw = jsw;
 | 
			
		||||
            commExchange(s->comm, p);
 | 
			
		||||
 | 
			
		||||
            for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                for (int i = isw; i < imaxLocal + 1; i += 2) {
 | 
			
		||||
 | 
			
		||||
                    double r = RHS(i, j) -
 | 
			
		||||
                               ((P(i + 1, j) - 2.0 * P(i, j) + P(i - 1, j)) * idx2 +
 | 
			
		||||
                                   (P(i, j + 1) - 2.0 * P(i, j) + P(i, j - 1)) * idy2);
 | 
			
		||||
 | 
			
		||||
                    P(i, j) -= (factor * r);
 | 
			
		||||
                    res += (r * r);
 | 
			
		||||
                }
 | 
			
		||||
                isw = 3 - isw;
 | 
			
		||||
            }
 | 
			
		||||
            jsw = 3 - jsw;
 | 
			
		||||
        }
 | 
			
		||||
 | 
			
		||||
        if (commIsBoundary(s->comm, BOTTOM)) { // set bottom bc
 | 
			
		||||
            for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                P(i, 0) = P(i, 1);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
 | 
			
		||||
        if (commIsBoundary(s->comm, TOP)) { // set top bc
 | 
			
		||||
            for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                P(i, jmaxLocal + 1) = P(i, jmaxLocal);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
 | 
			
		||||
        if (commIsBoundary(s->comm, LEFT)) { // set left bc
 | 
			
		||||
            for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                P(0, j) = P(1, j);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
 | 
			
		||||
        if (commIsBoundary(s->comm, RIGHT)) { // set right bc
 | 
			
		||||
            for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                P(imaxLocal + 1, j) = P(imaxLocal, j);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
 | 
			
		||||
        commReduction(&res, SUM);
 | 
			
		||||
        res = res / (double)(imax * jmax);
 | 
			
		||||
#ifdef DEBUG
 | 
			
		||||
        if (commIsMaster(s->comm)) {
 | 
			
		||||
            printf("%d Residuum: %e\n", it, res);
 | 
			
		||||
        }
 | 
			
		||||
#endif
 | 
			
		||||
        it++;
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
#ifdef VERBOSE
 | 
			
		||||
    if (commIsMaster(s->comm)) {
 | 
			
		||||
        printf("Solver took %d iterations to reach %f\n", it, sqrt(res));
 | 
			
		||||
    }
 | 
			
		||||
#endif
 | 
			
		||||
 | 
			
		||||
    return res;
 | 
			
		||||
}
 | 
			
		||||
@@ -9,6 +9,7 @@
 | 
			
		||||
#include "comm.h"
 | 
			
		||||
#include "discretization.h"
 | 
			
		||||
#include "grid.h"
 | 
			
		||||
#include "mpi.h"
 | 
			
		||||
#include "parameter.h"
 | 
			
		||||
 | 
			
		||||
typedef struct {
 | 
			
		||||
@@ -17,10 +18,12 @@ typedef struct {
 | 
			
		||||
    /* parameters */
 | 
			
		||||
    double eps, omega;
 | 
			
		||||
    int itermax;
 | 
			
		||||
    int levels, presmooth, postsmooth;
 | 
			
		||||
    double **r, **e;
 | 
			
		||||
    /* communication */
 | 
			
		||||
    Comm* comm;
 | 
			
		||||
} Solver;
 | 
			
		||||
 | 
			
		||||
void initSolver(Solver*, Discretization*, Parameter*);
 | 
			
		||||
void solve(Solver*, double*, double*);
 | 
			
		||||
double solve(Solver*, double*, double*);
 | 
			
		||||
#endif
 | 
			
		||||
 
 | 
			
		||||
@@ -38,9 +38,22 @@ $(BUILD_DIR)/%.s:  %.c
 | 
			
		||||
	$(info ===>  GENERATE ASM  $@)
 | 
			
		||||
	$(CC) -S $(CPPFLAGS) $(CFLAGS) $< -o $@
 | 
			
		||||
 | 
			
		||||
.PHONY: clean distclean tags info asm format
 | 
			
		||||
.PHONY: clean distclean vis vis_clean tags info asm format
 | 
			
		||||
 | 
			
		||||
clean:
 | 
			
		||||
vis:
 | 
			
		||||
	$(info ===>  GENERATE VISUALIZATION)
 | 
			
		||||
	@gnuplot -e "filename='pressure.dat'" ./surface.plot
 | 
			
		||||
	@gnuplot -e "filename='velocity.dat'" ./vector.plot
 | 
			
		||||
	@gnuplot -e "filename='residual.dat'" ./residual.plot
 | 
			
		||||
 | 
			
		||||
vis_clean:
 | 
			
		||||
	$(info ===>  CLEAN VISUALIZATION)
 | 
			
		||||
	@rm -f *.dat
 | 
			
		||||
	@rm -f *.png
 | 
			
		||||
	@rm -f ./vis_files/*.dat
 | 
			
		||||
	@rm -f ./vis_files/*.gif
 | 
			
		||||
 | 
			
		||||
clean: vis_clean
 | 
			
		||||
	$(info ===>  CLEAN)
 | 
			
		||||
	@rm -rf $(BUILD_DIR)
 | 
			
		||||
	@rm -f tags
 | 
			
		||||
 
 | 
			
		||||
@@ -36,6 +36,13 @@ te      100.0   # final time
 | 
			
		||||
dt      0.02    # time stepsize
 | 
			
		||||
tau     0.5     # safety factor for time stepsize control (<0 constant delt)
 | 
			
		||||
 | 
			
		||||
# Multigrid data:
 | 
			
		||||
# ---------
 | 
			
		||||
 | 
			
		||||
levels        3         # Multigrid levels
 | 
			
		||||
presmooth     5         # Pre-smoothning iterations
 | 
			
		||||
postsmooth    5         # Post-smoothning iterations
 | 
			
		||||
 | 
			
		||||
# Pressure Iteration Data:
 | 
			
		||||
# -----------------------
 | 
			
		||||
 | 
			
		||||
 
 | 
			
		||||
@@ -1,12 +1,12 @@
 | 
			
		||||
# Supported: GCC, CLANG, ICC
 | 
			
		||||
TAG ?= CLANG
 | 
			
		||||
TAG ?= ICC
 | 
			
		||||
ENABLE_OPENMP ?= false
 | 
			
		||||
# Supported: sor, rb, mg
 | 
			
		||||
SOLVER ?= rb
 | 
			
		||||
SOLVER ?= mg
 | 
			
		||||
# Run in debug settings
 | 
			
		||||
DEBUG ?= false
 | 
			
		||||
 | 
			
		||||
#Feature options
 | 
			
		||||
OPTIONS +=  -DARRAY_ALIGNMENT=64
 | 
			
		||||
#OPTIONS +=  -DVERBOSE
 | 
			
		||||
OPTIONS +=  -DVERBOSE
 | 
			
		||||
#OPTIONS +=  -DDEBUG
 | 
			
		||||
 
 | 
			
		||||
@@ -36,6 +36,13 @@ te      10.0		# final time
 | 
			
		||||
dt      0.02	    # time stepsize
 | 
			
		||||
tau     0.5	    	# safety factor for time stepsize control (<0 constant delt)
 | 
			
		||||
 | 
			
		||||
# Multigrid data:
 | 
			
		||||
# ---------
 | 
			
		||||
 | 
			
		||||
levels        2         # Multigrid levels
 | 
			
		||||
presmooth     20         # Pre-smoothning iterations
 | 
			
		||||
postsmooth    5         # Post-smoothning iterations
 | 
			
		||||
 | 
			
		||||
# Solver Data:
 | 
			
		||||
# -----------------------
 | 
			
		||||
 | 
			
		||||
@@ -44,5 +51,4 @@ eps      0.001		# stopping tolerance for pressure iteration
 | 
			
		||||
rho      0.5
 | 
			
		||||
omg      1.7		# relaxation parameter for SOR iteration
 | 
			
		||||
gamma    0.9		# upwind differencing factor gamma
 | 
			
		||||
levels   5         # Multigrid levels
 | 
			
		||||
#===============================================================================
 | 
			
		||||
 
 | 
			
		||||
							
								
								
									
										9
									
								
								BasicSolver/2D-seq/residual.plot
									
									
									
									
									
										Normal file
									
								
							
							
						
						
									
										9
									
								
								BasicSolver/2D-seq/residual.plot
									
									
									
									
									
										Normal file
									
								
							@@ -0,0 +1,9 @@
 | 
			
		||||
set terminal png size 1800,768 enhanced font ,12
 | 
			
		||||
set output 'residual.png'
 | 
			
		||||
set datafile separator whitespace
 | 
			
		||||
set xlabel "Timestep"
 | 
			
		||||
set ylabel "Residual"
 | 
			
		||||
 | 
			
		||||
set logscale y 2
 | 
			
		||||
 | 
			
		||||
plot 'residual.dat' using 1:2 title "Residual"
 | 
			
		||||
@@ -20,8 +20,10 @@ int main(int argc, char** argv)
 | 
			
		||||
    Parameter p;
 | 
			
		||||
    Discretization d;
 | 
			
		||||
    Solver s;
 | 
			
		||||
    
 | 
			
		||||
 | 
			
		||||
    initParameter(&p);
 | 
			
		||||
    FILE* fp;
 | 
			
		||||
    fp = initResidualWriter();
 | 
			
		||||
 | 
			
		||||
    if (argc != 2) {
 | 
			
		||||
        printf("Usage: %s <configFile>\n", argv[0]);
 | 
			
		||||
@@ -32,7 +34,7 @@ int main(int argc, char** argv)
 | 
			
		||||
    printParameter(&p);
 | 
			
		||||
    initDiscretization(&d, &p);
 | 
			
		||||
    initSolver(&s, &d, &p);
 | 
			
		||||
    
 | 
			
		||||
 | 
			
		||||
#ifndef VERBOSE
 | 
			
		||||
    initProgress(d.te);
 | 
			
		||||
#endif
 | 
			
		||||
@@ -41,8 +43,9 @@ int main(int argc, char** argv)
 | 
			
		||||
    double te  = d.te;
 | 
			
		||||
    double t   = 0.0;
 | 
			
		||||
    int nt     = 0;
 | 
			
		||||
    double res = 0.0;
 | 
			
		||||
 | 
			
		||||
    timeStart = getTimeStamp();
 | 
			
		||||
    timeStart  = getTimeStamp();
 | 
			
		||||
    while (t <= te) {
 | 
			
		||||
        if (tau > 0.0) computeTimestep(&d);
 | 
			
		||||
        setBoundaryConditions(&d);
 | 
			
		||||
@@ -50,17 +53,21 @@ int main(int argc, char** argv)
 | 
			
		||||
        computeFG(&d);
 | 
			
		||||
        computeRHS(&d);
 | 
			
		||||
        if (nt % 100 == 0) normalizePressure(&d);
 | 
			
		||||
        solve(&s, d.p, d.rhs);
 | 
			
		||||
        res = solve(&s, d.p, d.rhs);
 | 
			
		||||
        adaptUV(&d);
 | 
			
		||||
 | 
			
		||||
        writeResidual(fp, t, res);
 | 
			
		||||
 | 
			
		||||
        t += d.dt;
 | 
			
		||||
        nt++;
 | 
			
		||||
 | 
			
		||||
#ifdef VERBOSE
 | 
			
		||||
        printf("TIME %f , TIMESTEP %f\n", t, solver.dt);
 | 
			
		||||
        printf("TIME %f , TIMESTEP %f\n", t, d.dt);
 | 
			
		||||
#else
 | 
			
		||||
        printProgress(t);
 | 
			
		||||
#endif
 | 
			
		||||
    }
 | 
			
		||||
    fclose(fp);
 | 
			
		||||
    timeStop = getTimeStamp();
 | 
			
		||||
    stopProgress();
 | 
			
		||||
    printf("Solution took %.2fs\n", timeStop - timeStart);
 | 
			
		||||
 
 | 
			
		||||
@@ -24,6 +24,8 @@ void initParameter(Parameter* param)
 | 
			
		||||
    param->gamma   = 0.9;
 | 
			
		||||
    param->tau     = 0.5;
 | 
			
		||||
    param->levels  = 5;
 | 
			
		||||
    param->presmooth = 5;
 | 
			
		||||
    param->postsmooth = 5;
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void readParameter(Parameter* param, const char* filename)
 | 
			
		||||
@@ -79,6 +81,8 @@ void readParameter(Parameter* param, const char* filename)
 | 
			
		||||
            PARSE_REAL(u_init);
 | 
			
		||||
            PARSE_REAL(v_init);
 | 
			
		||||
            PARSE_REAL(p_init);
 | 
			
		||||
            PARSE_INT(presmooth);
 | 
			
		||||
            PARSE_INT(postsmooth);
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
 
 | 
			
		||||
@@ -18,6 +18,7 @@ typedef struct {
 | 
			
		||||
    char* name;
 | 
			
		||||
    int bcLeft, bcRight, bcBottom, bcTop;
 | 
			
		||||
    double u_init, v_init, p_init;
 | 
			
		||||
    int presmooth, postsmooth;
 | 
			
		||||
} Parameter;
 | 
			
		||||
 | 
			
		||||
void initParameter(Parameter*);
 | 
			
		||||
 
 | 
			
		||||
@@ -49,3 +49,22 @@ void stopProgress()
 | 
			
		||||
    printf("\n");
 | 
			
		||||
    fflush(stdout);
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
FILE* initResidualWriter()
 | 
			
		||||
{
 | 
			
		||||
    FILE* fp;
 | 
			
		||||
    fp = fopen("residual.dat", "w");
 | 
			
		||||
 | 
			
		||||
    if (fp == NULL) {
 | 
			
		||||
        printf("Error!\n");
 | 
			
		||||
        exit(EXIT_FAILURE);
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    return fp;
 | 
			
		||||
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void writeResidual(FILE* fp, double ts, double res)
 | 
			
		||||
{
 | 
			
		||||
    fprintf(fp, "%f, %f\n", ts, res);
 | 
			
		||||
}
 | 
			
		||||
 
 | 
			
		||||
@@ -10,5 +10,7 @@
 | 
			
		||||
extern void initProgress(double);
 | 
			
		||||
extern void printProgress(double);
 | 
			
		||||
extern void stopProgress(void);
 | 
			
		||||
extern FILE* initResidualWriter(void);
 | 
			
		||||
extern void writeResidual(FILE*, double, double);
 | 
			
		||||
 | 
			
		||||
#endif
 | 
			
		||||
 
 | 
			
		||||
@@ -84,6 +84,37 @@ static double smooth(Solver* s, double* p, double* rhs, int level, int imax, int
 | 
			
		||||
 | 
			
		||||
    jsw = 1;
 | 
			
		||||
 | 
			
		||||
    for (pass = 0; pass < 2; pass++) {
 | 
			
		||||
        isw = jsw;
 | 
			
		||||
 | 
			
		||||
        for (int j = 1; j < jmax + 1; j++) {
 | 
			
		||||
            for (int i = isw; i < imax + 1; i += 2) {
 | 
			
		||||
 | 
			
		||||
                P(i, j) -= factor * (RHS(i, j) -
 | 
			
		||||
                          ((P(i + 1, j) - 2.0 * P(i, j) + P(i - 1, j)) * idx2 +
 | 
			
		||||
                              (P(i, j + 1) - 2.0 * P(i, j) + P(i, j - 1)) * idy2));
 | 
			
		||||
 | 
			
		||||
            }
 | 
			
		||||
            isw = 3 - isw;
 | 
			
		||||
        }
 | 
			
		||||
        jsw = 3 - jsw;
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
static double calculateResidual(Solver* s, double* p, double* rhs, int level, int imax, int jmax)
 | 
			
		||||
{
 | 
			
		||||
    double dx2    = s->grid->dx * s->grid->dx;
 | 
			
		||||
    double dy2    = s->grid->dy * s->grid->dy;
 | 
			
		||||
    double idx2   = 1.0 / dx2;
 | 
			
		||||
    double idy2   = 1.0 / dy2;
 | 
			
		||||
    double factor = s->omega * 0.5 * (dx2 * dy2) / (dx2 + dy2);
 | 
			
		||||
    double* r     = s->r[level];
 | 
			
		||||
    double res    = 1.0;
 | 
			
		||||
    int pass, jsw, isw;
 | 
			
		||||
 | 
			
		||||
    jsw = 1;
 | 
			
		||||
 | 
			
		||||
    for (pass = 0; pass < 2; pass++) {
 | 
			
		||||
        isw = jsw;
 | 
			
		||||
 | 
			
		||||
@@ -94,7 +125,6 @@ static double smooth(Solver* s, double* p, double* rhs, int level, int imax, int
 | 
			
		||||
                          ((P(i + 1, j) - 2.0 * P(i, j) + P(i - 1, j)) * idx2 +
 | 
			
		||||
                              (P(i, j + 1) - 2.0 * P(i, j) + P(i, j - 1)) * idy2);
 | 
			
		||||
 | 
			
		||||
                P(i, j) -= (factor * R(i, j));
 | 
			
		||||
                res += (R(i, j) * R(i, j));
 | 
			
		||||
            }
 | 
			
		||||
            isw = 3 - isw;
 | 
			
		||||
@@ -110,7 +140,7 @@ static double multiGrid(Solver* s, double* p, double* rhs, int level, int imax,
 | 
			
		||||
{
 | 
			
		||||
    double res = 0.0;
 | 
			
		||||
 | 
			
		||||
    // coarsest level TODO: Use direct solver?
 | 
			
		||||
    // coarsest level
 | 
			
		||||
    if (level == COARSEST_LEVEL) {
 | 
			
		||||
        for (int i = 0; i < 5; i++) {
 | 
			
		||||
            smooth(s, p, rhs, level, imax, jmax);
 | 
			
		||||
@@ -118,17 +148,18 @@ static double multiGrid(Solver* s, double* p, double* rhs, int level, int imax,
 | 
			
		||||
        return res;
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    // pre-smoothing TODO: Make smoothing steps configurable?
 | 
			
		||||
    for (int i = 0; i < 5; i++) {
 | 
			
		||||
    // pre-smoothing
 | 
			
		||||
    for (int i = 0; i < s->presmooth; i++) {
 | 
			
		||||
        smooth(s, p, rhs, level, imax, jmax);
 | 
			
		||||
        if (level == FINEST_LEVEL) setBoundaryCondition(p, imax, jmax);
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    res = calculateResidual(s, p, rhs, level, imax, jmax);
 | 
			
		||||
 | 
			
		||||
    // restrict
 | 
			
		||||
    restrictMG(s, level, imax, jmax);
 | 
			
		||||
 | 
			
		||||
    // MGSolver on residual and error.
 | 
			
		||||
    // TODO: What if there is a rest?
 | 
			
		||||
    multiGrid(s, s->e[level + 1], s->r[level + 1], level + 1, imax / 2, jmax / 2);
 | 
			
		||||
 | 
			
		||||
    // prolongate
 | 
			
		||||
@@ -139,8 +170,8 @@ static double multiGrid(Solver* s, double* p, double* rhs, int level, int imax,
 | 
			
		||||
    if (level == FINEST_LEVEL) setBoundaryCondition(p, imax, jmax);
 | 
			
		||||
 | 
			
		||||
    // post-smoothing
 | 
			
		||||
    for (int i = 0; i < 5; i++) {
 | 
			
		||||
        res = smooth(s, p, rhs, level, imax, jmax);
 | 
			
		||||
    for (int i = 0; i < s->postsmooth; i++) {
 | 
			
		||||
        smooth(s, p, rhs, level, imax, jmax);
 | 
			
		||||
        if (level == FINEST_LEVEL) setBoundaryCondition(p, imax, jmax);
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
@@ -154,6 +185,8 @@ void initSolver(Solver* s, Discretization* d, Parameter* p)
 | 
			
		||||
    s->itermax = p->itermax;
 | 
			
		||||
    s->levels  = p->levels;
 | 
			
		||||
    s->grid    = &d->grid;
 | 
			
		||||
    s->presmooth = p->presmooth;
 | 
			
		||||
    s->postsmooth = p->postsmooth;
 | 
			
		||||
 | 
			
		||||
    int imax   = s->grid->imax;
 | 
			
		||||
    int jmax   = s->grid->jmax;
 | 
			
		||||
@@ -176,11 +209,13 @@ void initSolver(Solver* s, Discretization* d, Parameter* p)
 | 
			
		||||
    }
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void solve(Solver* s, double* p, double* rhs)
 | 
			
		||||
double solve(Solver* s, double* p, double* rhs)
 | 
			
		||||
{
 | 
			
		||||
    double res = multiGrid(s, p, rhs, 0, s->grid->imax, s->grid->jmax);
 | 
			
		||||
 | 
			
		||||
#ifdef VERBOSE
 | 
			
		||||
    printf("Residuum: %.6f\n", res);
 | 
			
		||||
#endif
 | 
			
		||||
 | 
			
		||||
return res;
 | 
			
		||||
}
 | 
			
		||||
 
 | 
			
		||||
@@ -15,7 +15,7 @@ void initSolver(Solver* s, Discretization* d, Parameter* p)
 | 
			
		||||
    s->omega   = p->omg;
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void solve(Solver* solver, double* p, double* rhs)
 | 
			
		||||
double solve(Solver* solver, double* p, double* rhs)
 | 
			
		||||
{
 | 
			
		||||
    int imax      = solver->grid->imax;
 | 
			
		||||
    int jmax      = solver->grid->jmax;
 | 
			
		||||
@@ -73,4 +73,6 @@ void solve(Solver* solver, double* p, double* rhs)
 | 
			
		||||
#ifdef VERBOSE
 | 
			
		||||
    printf("Solver took %d iterations to reach %f\n", it, sqrt(res));
 | 
			
		||||
#endif
 | 
			
		||||
 | 
			
		||||
    return res;
 | 
			
		||||
}
 | 
			
		||||
 
 | 
			
		||||
@@ -15,7 +15,7 @@ void initSolver(Solver* s, Discretization* d, Parameter* p)
 | 
			
		||||
    s->omega   = p->omg;
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void solve(Solver* solver, double* p, double* rhs)
 | 
			
		||||
double solve(Solver* solver, double* p, double* rhs)
 | 
			
		||||
{
 | 
			
		||||
    int imax      = solver->grid->imax;
 | 
			
		||||
    int jmax      = solver->grid->jmax;
 | 
			
		||||
@@ -65,4 +65,6 @@ void solve(Solver* solver, double* p, double* rhs)
 | 
			
		||||
#ifdef VERBOSE
 | 
			
		||||
    printf("Solver took %d iterations to reach %f\n", it, sqrt(res));
 | 
			
		||||
#endif
 | 
			
		||||
 | 
			
		||||
    return res;
 | 
			
		||||
}
 | 
			
		||||
 
 | 
			
		||||
@@ -18,9 +18,10 @@ typedef struct {
 | 
			
		||||
    int itermax;
 | 
			
		||||
    int levels;
 | 
			
		||||
    double **r, **e;
 | 
			
		||||
        int presmooth, postsmooth;
 | 
			
		||||
} Solver;
 | 
			
		||||
 | 
			
		||||
extern void initSolver(Solver*, Discretization*, Parameter*);
 | 
			
		||||
extern void solve(Solver*, double*, double*);
 | 
			
		||||
extern double solve(Solver*, double*, double*);
 | 
			
		||||
 | 
			
		||||
#endif
 | 
			
		||||
 
 | 
			
		||||
@@ -22,6 +22,7 @@ SRC       = $(filter-out $(wildcard $(SRC_DIR)/*-*.c),$(wildcard $(SRC_DIR)/*.c)
 | 
			
		||||
ASM       = $(patsubst $(SRC_DIR)/%.c, $(BUILD_DIR)/%.s, $(SRC))
 | 
			
		||||
OBJ       = $(patsubst $(SRC_DIR)/%.c, $(BUILD_DIR)/%.o, $(SRC))
 | 
			
		||||
OBJ      += $(BUILD_DIR)/vtkWriter-$(VTK_OUTPUT_FMT).o
 | 
			
		||||
OBJ      += $(BUILD_DIR)/solver-$(SOLVER).o
 | 
			
		||||
SOURCES   = $(SRC) $(wildcard $(SRC_DIR)/*.h)
 | 
			
		||||
ifeq ($(VTK_OUTPUT_FMT),mpi)
 | 
			
		||||
DEFINES  += -D_VTK_WRITER_MPI
 | 
			
		||||
@@ -42,9 +43,19 @@ $(BUILD_DIR)/%.s:  %.c
 | 
			
		||||
	$(info ===>  GENERATE ASM  $@)
 | 
			
		||||
	$(CC) -S $(CPPFLAGS) $(CFLAGS) $< -o $@
 | 
			
		||||
 | 
			
		||||
.PHONY: clean distclean tags info asm format
 | 
			
		||||
.PHONY: clean distclean vis vis_clean tags info asm format
 | 
			
		||||
 | 
			
		||||
clean:
 | 
			
		||||
vis:
 | 
			
		||||
	$(info ===>  GENERATE VISUALIZATION)
 | 
			
		||||
	@gnuplot -e "filename='residual.dat'" ./residual.plot
 | 
			
		||||
 | 
			
		||||
vis_clean:
 | 
			
		||||
	$(info ===>  CLEAN VISUALIZATION)
 | 
			
		||||
	@rm -f *.dat
 | 
			
		||||
	@rm -f *.vtk
 | 
			
		||||
	@rm -f *.png
 | 
			
		||||
 | 
			
		||||
clean: vis_clean
 | 
			
		||||
	$(info ===>  CLEAN)
 | 
			
		||||
	@rm -rf $(BUILD_DIR)
 | 
			
		||||
	@rm -f tags
 | 
			
		||||
 
 | 
			
		||||
@@ -32,21 +32,28 @@ xlength       30.0     # domain size in x-direction
 | 
			
		||||
ylength       4.0	   # domain size in y-direction
 | 
			
		||||
zlength       4.0	   # domain size in z-direction
 | 
			
		||||
imax          200      # number of interior cells in x-direction
 | 
			
		||||
jmax          50	   # number of interior cells in y-direction
 | 
			
		||||
kmax          50	   # number of interior cells in z-direction
 | 
			
		||||
jmax          40	   # number of interior cells in y-direction
 | 
			
		||||
kmax          40	   # number of interior cells in z-direction
 | 
			
		||||
 | 
			
		||||
# Time Data:
 | 
			
		||||
# ---------
 | 
			
		||||
 | 
			
		||||
te       100.0   # final time
 | 
			
		||||
te       60.0   # final time
 | 
			
		||||
dt       0.02    # time stepsize
 | 
			
		||||
tau      0.5     # safety factor for time stepsize control (<0 constant delt)
 | 
			
		||||
 | 
			
		||||
# Multigrid data:
 | 
			
		||||
# ---------
 | 
			
		||||
 | 
			
		||||
levels        2         # Multigrid levels
 | 
			
		||||
presmooth     20         # Pre-smoothning iterations
 | 
			
		||||
postsmooth    5         # Post-smoothning iterations
 | 
			
		||||
 | 
			
		||||
# Pressure Iteration Data:
 | 
			
		||||
# -----------------------
 | 
			
		||||
 | 
			
		||||
itermax       500       # maximal number of pressure iteration in one time step
 | 
			
		||||
eps           0.0001   # stopping tolerance for pressure iteration
 | 
			
		||||
omg           1.3       # relaxation parameter for SOR iteration
 | 
			
		||||
omg           1.7       # relaxation parameter for SOR iteration
 | 
			
		||||
gamma         0.9       # upwind differencing factor gamma
 | 
			
		||||
#===============================================================================
 | 
			
		||||
 
 | 
			
		||||
@@ -1,7 +1,10 @@
 | 
			
		||||
# Supported: GCC, CLANG, ICC
 | 
			
		||||
TAG ?= CLANG
 | 
			
		||||
TAG ?= ICC
 | 
			
		||||
# Supported: true, false
 | 
			
		||||
ENABLE_MPI ?= true
 | 
			
		||||
ENABLE_OPENMP ?= false
 | 
			
		||||
# Supported: rb, mg
 | 
			
		||||
SOLVER ?= mg
 | 
			
		||||
# Supported: seq, mpi
 | 
			
		||||
VTK_OUTPUT_FMT ?= seq
 | 
			
		||||
 | 
			
		||||
 
 | 
			
		||||
@@ -18,7 +18,7 @@ gx    0.0			# Body forces (e.g. gravity)
 | 
			
		||||
gy    0.0			#
 | 
			
		||||
gz    0.0			#
 | 
			
		||||
 | 
			
		||||
re    1000.0		    # Reynolds number
 | 
			
		||||
re    1000.0		# Reynolds number
 | 
			
		||||
 | 
			
		||||
u_init    0.0		# initial value for velocity in x-direction
 | 
			
		||||
v_init    0.0		# initial value for velocity in y-direction
 | 
			
		||||
@@ -42,6 +42,13 @@ te       10.0		# final time
 | 
			
		||||
dt       0.02	    # time stepsize
 | 
			
		||||
tau      0.5		# safety factor for time stepsize control (<0 constant delt)
 | 
			
		||||
 | 
			
		||||
# Multigrid data:
 | 
			
		||||
# ---------
 | 
			
		||||
 | 
			
		||||
levels        3         # Multigrid levels
 | 
			
		||||
presmooth     20        # Pre-smoothning iterations
 | 
			
		||||
postsmooth    5         # Post-smoothning iterations
 | 
			
		||||
 | 
			
		||||
# Pressure Iteration Data:
 | 
			
		||||
# -----------------------
 | 
			
		||||
 | 
			
		||||
 
 | 
			
		||||
							
								
								
									
										9
									
								
								BasicSolver/3D-mpi/residual.plot
									
									
									
									
									
										Normal file
									
								
							
							
						
						
									
										9
									
								
								BasicSolver/3D-mpi/residual.plot
									
									
									
									
									
										Normal file
									
								
							@@ -0,0 +1,9 @@
 | 
			
		||||
set terminal png size 1800,768 enhanced font ,12
 | 
			
		||||
set output 'residual.png'
 | 
			
		||||
set datafile separator whitespace
 | 
			
		||||
set xlabel "Timestep"
 | 
			
		||||
set ylabel "Residual"
 | 
			
		||||
 | 
			
		||||
set logscale y 2
 | 
			
		||||
 | 
			
		||||
plot 'residual.dat' using 1:2 title "Residual"
 | 
			
		||||
@@ -167,11 +167,12 @@ static void assembleResult(Comm* c,
 | 
			
		||||
    MPI_Waitall(numRequests, requests, MPI_STATUSES_IGNORE);
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
static int sum(int* sizes, int position)
 | 
			
		||||
// subroutines local to this module
 | 
			
		||||
static int sum(int* sizes, int init, int offset, int coord)
 | 
			
		||||
{
 | 
			
		||||
    int sum = 0;
 | 
			
		||||
 | 
			
		||||
    for (int i = 0; i < position; i++) {
 | 
			
		||||
    for (int i = init - offset; coord > 0; i -= offset, --coord) {
 | 
			
		||||
        sum += sizes[i];
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
@@ -336,7 +337,6 @@ void commCollectResult(Comm* c,
 | 
			
		||||
    int imaxLocal = c->imaxLocal;
 | 
			
		||||
    int jmaxLocal = c->jmaxLocal;
 | 
			
		||||
    int kmaxLocal = c->kmaxLocal;
 | 
			
		||||
 | 
			
		||||
#if defined(_MPI)
 | 
			
		||||
    int offset[c->size * NDIMS];
 | 
			
		||||
    int imaxLocalAll[c->size];
 | 
			
		||||
@@ -351,9 +351,13 @@ void commCollectResult(Comm* c,
 | 
			
		||||
        for (int i = 0; i < c->size; i++) {
 | 
			
		||||
            int coords[NCORDS];
 | 
			
		||||
            MPI_Cart_coords(c->comm, i, NDIMS, coords);
 | 
			
		||||
            offset[i * NDIMS + IDIM] = sum(imaxLocalAll, coords[ICORD]);
 | 
			
		||||
            offset[i * NDIMS + JDIM] = sum(jmaxLocalAll, coords[JCORD]);
 | 
			
		||||
            offset[i * NDIMS + KDIM] = sum(kmaxLocalAll, coords[KCORD]);
 | 
			
		||||
            offset[i * NDIMS + IDIM] = sum(imaxLocalAll,
 | 
			
		||||
                i,
 | 
			
		||||
                c->dims[IDIM] * c->dims[JDIM],
 | 
			
		||||
                coords[ICORD]);
 | 
			
		||||
            offset[i * NDIMS + JDIM] = sum(jmaxLocalAll, i, c->dims[IDIM], coords[JCORD]);
 | 
			
		||||
            offset[i * NDIMS + KDIM] = sum(kmaxLocalAll, i, 1, coords[KCORD]);
 | 
			
		||||
 | 
			
		||||
            printf("Rank: %d, Coords(k,j,i): %d %d %d, Size(k,j,i): %d %d %d, "
 | 
			
		||||
                   "Offset(k,j,i): %d %d %d\n",
 | 
			
		||||
                i,
 | 
			
		||||
@@ -562,9 +566,9 @@ void commPartition(Comm* c, int kmax, int jmax, int imax)
 | 
			
		||||
    MPI_Cart_shift(c->comm, KCORD, 1, &c->neighbours[FRONT], &c->neighbours[BACK]);
 | 
			
		||||
    MPI_Cart_get(c->comm, NCORDS, c->dims, periods, c->coords);
 | 
			
		||||
 | 
			
		||||
    c->imaxLocal = sizeOfRank(c->rank, dims[ICORD], imax);
 | 
			
		||||
    c->jmaxLocal = sizeOfRank(c->rank, dims[JCORD], jmax);
 | 
			
		||||
    c->kmaxLocal = sizeOfRank(c->rank, dims[KCORD], kmax);
 | 
			
		||||
    c->imaxLocal = sizeOfRank(c->coords[KDIM], dims[ICORD], imax);
 | 
			
		||||
    c->jmaxLocal = sizeOfRank(c->coords[JDIM], dims[JCORD], jmax);
 | 
			
		||||
    c->kmaxLocal = sizeOfRank(c->coords[IDIM], dims[KCORD], kmax);
 | 
			
		||||
 | 
			
		||||
    // setup buffer types for communication
 | 
			
		||||
    setupCommunication(c, LEFT, BULK);
 | 
			
		||||
@@ -597,3 +601,47 @@ void commFinalize(Comm* c)
 | 
			
		||||
    MPI_Finalize();
 | 
			
		||||
#endif
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void commUpdateDatatypes(
 | 
			
		||||
    Comm* oldcomm, Comm* newcomm, int imaxLocal, int jmaxLocal, int kmaxLocal)
 | 
			
		||||
{
 | 
			
		||||
#if defined _MPI
 | 
			
		||||
 | 
			
		||||
    int result = MPI_Comm_dup(oldcomm->comm, &newcomm->comm);
 | 
			
		||||
 | 
			
		||||
    if (result == MPI_ERR_COMM) {
 | 
			
		||||
        printf("\nNull communicator. Duplication failed !!\n");
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    newcomm->rank = oldcomm->rank;
 | 
			
		||||
    newcomm->size = oldcomm->size;
 | 
			
		||||
 | 
			
		||||
    newcomm->imaxLocal = imaxLocal / 2;
 | 
			
		||||
    newcomm->jmaxLocal = jmaxLocal / 2;
 | 
			
		||||
    newcomm->kmaxLocal = kmaxLocal / 2;
 | 
			
		||||
 | 
			
		||||
    setupCommunication(newcomm, LEFT, BULK);
 | 
			
		||||
    setupCommunication(newcomm, LEFT, HALO);
 | 
			
		||||
    setupCommunication(newcomm, RIGHT, BULK);
 | 
			
		||||
    setupCommunication(newcomm, RIGHT, HALO);
 | 
			
		||||
    setupCommunication(newcomm, BOTTOM, BULK);
 | 
			
		||||
    setupCommunication(newcomm, BOTTOM, HALO);
 | 
			
		||||
    setupCommunication(newcomm, TOP, BULK);
 | 
			
		||||
    setupCommunication(newcomm, TOP, HALO);
 | 
			
		||||
    setupCommunication(newcomm, FRONT, BULK);
 | 
			
		||||
    setupCommunication(newcomm, FRONT, HALO);
 | 
			
		||||
    setupCommunication(newcomm, BACK, BULK);
 | 
			
		||||
    setupCommunication(newcomm, BACK, HALO);
 | 
			
		||||
#else
 | 
			
		||||
    newcomm->imaxLocal = imaxLocal;
 | 
			
		||||
    newcomm->jmaxLocal = jmaxLocal;
 | 
			
		||||
    newcomm->kmaxLocal = kmaxLocal;
 | 
			
		||||
#endif
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void commFreeCommunicator(Comm* comm)
 | 
			
		||||
{
 | 
			
		||||
#ifdef _MPI
 | 
			
		||||
    MPI_Comm_free(&comm->comm);
 | 
			
		||||
#endif
 | 
			
		||||
}
 | 
			
		||||
@@ -23,6 +23,7 @@ typedef enum dimension { KDIM = 0, JDIM, IDIM, NDIMS } Dimension;
 | 
			
		||||
enum layer { HALO = 0, BULK };
 | 
			
		||||
enum op { MAX = 0, SUM };
 | 
			
		||||
 | 
			
		||||
 | 
			
		||||
typedef struct {
 | 
			
		||||
    int rank;
 | 
			
		||||
    int size;
 | 
			
		||||
@@ -45,6 +46,9 @@ extern void commShift(Comm* c, double* f, double* g, double* h);
 | 
			
		||||
extern void commReduction(double* v, int op);
 | 
			
		||||
extern int commIsBoundary(Comm* c, Direction direction);
 | 
			
		||||
extern void commGetOffsets(Comm* c, int offsets[], int kmax, int jmax, int imax);
 | 
			
		||||
extern void commFreeCommunicator(Comm* comm);
 | 
			
		||||
extern void commUpdateDatatypes(
 | 
			
		||||
    Comm* oldcomm, Comm* newcomm, int imaxLocal, int jmaxLocal, int kmaxLocal);
 | 
			
		||||
extern void commCollectResult(Comm* c,
 | 
			
		||||
    double* ug,
 | 
			
		||||
    double* vg,
 | 
			
		||||
 
 | 
			
		||||
							
								
								
									
										729
									
								
								BasicSolver/3D-mpi/src/discretization.c
									
									
									
									
									
										Normal file
									
								
							
							
						
						
									
										729
									
								
								BasicSolver/3D-mpi/src/discretization.c
									
									
									
									
									
										Normal file
									
								
							@@ -0,0 +1,729 @@
 | 
			
		||||
/*
 | 
			
		||||
 * Copyright (C) NHR@FAU, University Erlangen-Nuremberg.
 | 
			
		||||
 * All rights reserved. This file is part of nusif-solver.
 | 
			
		||||
 * Use of this source code is governed by a MIT style
 | 
			
		||||
 * license that can be found in the LICENSE file.
 | 
			
		||||
 */
 | 
			
		||||
#include <float.h>
 | 
			
		||||
#include <math.h>
 | 
			
		||||
#include <stdio.h>
 | 
			
		||||
#include <string.h>
 | 
			
		||||
 | 
			
		||||
#include "allocate.h"
 | 
			
		||||
#include "discretization.h"
 | 
			
		||||
#include "parameter.h"
 | 
			
		||||
#include "util.h"
 | 
			
		||||
 | 
			
		||||
#define P(i, j, k)                                                                       \
 | 
			
		||||
    p[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define F(i, j, k)                                                                       \
 | 
			
		||||
    f[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define G(i, j, k)                                                                       \
 | 
			
		||||
    g[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define H(i, j, k)                                                                       \
 | 
			
		||||
    h[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define U(i, j, k)                                                                       \
 | 
			
		||||
    u[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define V(i, j, k)                                                                       \
 | 
			
		||||
    v[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define W(i, j, k)                                                                       \
 | 
			
		||||
    w[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define RHS(i, j, k)                                                                     \
 | 
			
		||||
    rhs[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
 | 
			
		||||
static void printConfig(Discretization* s)
 | 
			
		||||
{
 | 
			
		||||
    if (commIsMaster(&s->comm)) {
 | 
			
		||||
        printf("Parameters for #%s#\n", s->problem);
 | 
			
		||||
        printf("BC Left:%d Right:%d Bottom:%d Top:%d Front:%d Back:%d\n",
 | 
			
		||||
            s->bcLeft,
 | 
			
		||||
            s->bcRight,
 | 
			
		||||
            s->bcBottom,
 | 
			
		||||
            s->bcTop,
 | 
			
		||||
            s->bcFront,
 | 
			
		||||
            s->bcBack);
 | 
			
		||||
        printf("\tReynolds number: %.2f\n", s->re);
 | 
			
		||||
        printf("\tGx Gy: %.2f %.2f %.2f\n", s->gx, s->gy, s->gz);
 | 
			
		||||
        printf("Geometry data:\n");
 | 
			
		||||
        printf("\tDomain box size (x, y, z): %.2f, %.2f, %.2f\n",
 | 
			
		||||
            s->grid.xlength,
 | 
			
		||||
            s->grid.ylength,
 | 
			
		||||
            s->grid.zlength);
 | 
			
		||||
        printf("\tCells (x, y, z): %d, %d, %d\n",
 | 
			
		||||
            s->grid.imax,
 | 
			
		||||
            s->grid.jmax,
 | 
			
		||||
            s->grid.kmax);
 | 
			
		||||
        printf("\tCell size (dx, dy, dz): %f, %f, %f\n",
 | 
			
		||||
            s->grid.dx,
 | 
			
		||||
            s->grid.dy,
 | 
			
		||||
            s->grid.dz);
 | 
			
		||||
        printf("Timestep parameters:\n");
 | 
			
		||||
        printf("\tDefault stepsize: %.2f, Final time %.2f\n", s->dt, s->te);
 | 
			
		||||
        printf("\tdt bound: %.6f\n", s->dtBound);
 | 
			
		||||
        printf("\tTau factor: %.2f\n", s->tau);
 | 
			
		||||
        printf("Iterative parameters:\n");
 | 
			
		||||
        printf("\tMax iterations: %d\n", s->itermax);
 | 
			
		||||
        printf("\tepsilon (stopping tolerance) : %f\n", s->eps);
 | 
			
		||||
        printf("\tgamma factor: %f\n", s->gamma);
 | 
			
		||||
        printf("\tomega (SOR relaxation): %f\n", s->omega);
 | 
			
		||||
    }
 | 
			
		||||
    commPrintConfig(&s->comm);
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void initDiscretization(Discretization* s, Parameter* params)
 | 
			
		||||
{
 | 
			
		||||
    s->problem  = params->name;
 | 
			
		||||
    s->bcLeft   = params->bcLeft;
 | 
			
		||||
    s->bcRight  = params->bcRight;
 | 
			
		||||
    s->bcBottom = params->bcBottom;
 | 
			
		||||
    s->bcTop    = params->bcTop;
 | 
			
		||||
    s->bcFront  = params->bcFront;
 | 
			
		||||
    s->bcBack   = params->bcBack;
 | 
			
		||||
 | 
			
		||||
    s->grid.imax    = params->imax;
 | 
			
		||||
    s->grid.jmax    = params->jmax;
 | 
			
		||||
    s->grid.kmax    = params->kmax;
 | 
			
		||||
    s->grid.xlength = params->xlength;
 | 
			
		||||
    s->grid.ylength = params->ylength;
 | 
			
		||||
    s->grid.zlength = params->zlength;
 | 
			
		||||
    s->grid.dx      = params->xlength / params->imax;
 | 
			
		||||
    s->grid.dy      = params->ylength / params->jmax;
 | 
			
		||||
    s->grid.dz      = params->zlength / params->kmax;
 | 
			
		||||
 | 
			
		||||
    s->eps     = params->eps;
 | 
			
		||||
    s->omega   = params->omg;
 | 
			
		||||
    s->itermax = params->itermax;
 | 
			
		||||
    s->re      = params->re;
 | 
			
		||||
    s->gx      = params->gx;
 | 
			
		||||
    s->gy      = params->gy;
 | 
			
		||||
    s->gz      = params->gz;
 | 
			
		||||
    s->dt      = params->dt;
 | 
			
		||||
    s->te      = params->te;
 | 
			
		||||
    s->tau     = params->tau;
 | 
			
		||||
    s->gamma   = params->gamma;
 | 
			
		||||
 | 
			
		||||
    /* allocate arrays */
 | 
			
		||||
    int imaxLocal = s->comm.imaxLocal;
 | 
			
		||||
    int jmaxLocal = s->comm.jmaxLocal;
 | 
			
		||||
    int kmaxLocal = s->comm.kmaxLocal;
 | 
			
		||||
    size_t size   = (imaxLocal + 2) * (jmaxLocal + 2) * (kmaxLocal + 2);
 | 
			
		||||
 | 
			
		||||
    s->u   = allocate(64, size * sizeof(double));
 | 
			
		||||
    s->v   = allocate(64, size * sizeof(double));
 | 
			
		||||
    s->w   = allocate(64, size * sizeof(double));
 | 
			
		||||
    s->p   = allocate(64, size * sizeof(double));
 | 
			
		||||
    s->rhs = allocate(64, size * sizeof(double));
 | 
			
		||||
    s->f   = allocate(64, size * sizeof(double));
 | 
			
		||||
    s->g   = allocate(64, size * sizeof(double));
 | 
			
		||||
    s->h   = allocate(64, size * sizeof(double));
 | 
			
		||||
 | 
			
		||||
    for (int i = 0; i < size; i++) {
 | 
			
		||||
        s->u[i]   = params->u_init;
 | 
			
		||||
        s->v[i]   = params->v_init;
 | 
			
		||||
        s->w[i]   = params->w_init;
 | 
			
		||||
        s->p[i]   = params->p_init;
 | 
			
		||||
        s->rhs[i] = 0.0;
 | 
			
		||||
        s->f[i]   = 0.0;
 | 
			
		||||
        s->g[i]   = 0.0;
 | 
			
		||||
        s->h[i]   = 0.0;
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    double dx = s->grid.dx;
 | 
			
		||||
    double dy = s->grid.dy;
 | 
			
		||||
    double dz = s->grid.dz;
 | 
			
		||||
 | 
			
		||||
    double invSqrSum = 1.0 / (dx * dx) + 1.0 / (dy * dy) + 1.0 / (dz * dz);
 | 
			
		||||
    s->dtBound       = 0.5 * s->re * 1.0 / invSqrSum;
 | 
			
		||||
 | 
			
		||||
#ifdef VERBOSE
 | 
			
		||||
    printConfig(s);
 | 
			
		||||
#endif /* VERBOSE */
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void setBoundaryConditions(Discretization* s)
 | 
			
		||||
{
 | 
			
		||||
    int imaxLocal = s->comm.imaxLocal;
 | 
			
		||||
    int jmaxLocal = s->comm.jmaxLocal;
 | 
			
		||||
    int kmaxLocal = s->comm.kmaxLocal;
 | 
			
		||||
 | 
			
		||||
    double* u = s->u;
 | 
			
		||||
    double* v = s->v;
 | 
			
		||||
    double* w = s->w;
 | 
			
		||||
 | 
			
		||||
    if (commIsBoundary(&s->comm, TOP)) {
 | 
			
		||||
        switch (s->bcTop) {
 | 
			
		||||
        case NOSLIP:
 | 
			
		||||
            for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
                for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                    U(i, jmaxLocal + 1, k) = -U(i, jmaxLocal, k);
 | 
			
		||||
                    V(i, jmaxLocal, k)     = 0.0;
 | 
			
		||||
                    W(i, jmaxLocal + 1, k) = -W(i, jmaxLocal, k);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
            break;
 | 
			
		||||
        case SLIP:
 | 
			
		||||
            for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
                for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                    U(i, jmaxLocal + 1, k) = U(i, jmaxLocal, k);
 | 
			
		||||
                    V(i, jmaxLocal, k)     = 0.0;
 | 
			
		||||
                    W(i, jmaxLocal + 1, k) = W(i, jmaxLocal, k);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
            break;
 | 
			
		||||
        case OUTFLOW:
 | 
			
		||||
            for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
                for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                    U(i, jmaxLocal + 1, k) = U(i, jmaxLocal, k);
 | 
			
		||||
                    V(i, jmaxLocal, k)     = V(i, jmaxLocal - 1, k);
 | 
			
		||||
                    W(i, jmaxLocal + 1, k) = W(i, jmaxLocal, k);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
            break;
 | 
			
		||||
        case PERIODIC:
 | 
			
		||||
            break;
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    if (commIsBoundary(&s->comm, BOTTOM)) {
 | 
			
		||||
        switch (s->bcBottom) {
 | 
			
		||||
        case NOSLIP:
 | 
			
		||||
            for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
                for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                    U(i, 0, k) = -U(i, 1, k);
 | 
			
		||||
                    V(i, 0, k) = 0.0;
 | 
			
		||||
                    W(i, 0, k) = -W(i, 1, k);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
            break;
 | 
			
		||||
        case SLIP:
 | 
			
		||||
            for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
                for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                    U(i, 0, k) = U(i, 1, k);
 | 
			
		||||
                    V(i, 0, k) = 0.0;
 | 
			
		||||
                    W(i, 0, k) = W(i, 1, k);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
            break;
 | 
			
		||||
        case OUTFLOW:
 | 
			
		||||
            for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
                for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                    U(i, 0, k) = U(i, 1, k);
 | 
			
		||||
                    V(i, 0, k) = V(i, 1, k);
 | 
			
		||||
                    W(i, 0, k) = W(i, 1, k);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
            break;
 | 
			
		||||
        case PERIODIC:
 | 
			
		||||
            break;
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    if (commIsBoundary(&s->comm, LEFT)) {
 | 
			
		||||
        switch (s->bcLeft) {
 | 
			
		||||
        case NOSLIP:
 | 
			
		||||
            for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
                for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                    U(0, j, k) = 0.0;
 | 
			
		||||
                    V(0, j, k) = -V(1, j, k);
 | 
			
		||||
                    W(0, j, k) = -W(1, j, k);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
            break;
 | 
			
		||||
        case SLIP:
 | 
			
		||||
            for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
                for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                    U(0, j, k) = 0.0;
 | 
			
		||||
                    V(0, j, k) = V(1, j, k);
 | 
			
		||||
                    W(0, j, k) = W(1, j, k);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
            break;
 | 
			
		||||
        case OUTFLOW:
 | 
			
		||||
            for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
                for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                    U(0, j, k) = U(1, j, k);
 | 
			
		||||
                    V(0, j, k) = V(1, j, k);
 | 
			
		||||
                    W(0, j, k) = W(1, j, k);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
            break;
 | 
			
		||||
        case PERIODIC:
 | 
			
		||||
            break;
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    if (commIsBoundary(&s->comm, RIGHT)) {
 | 
			
		||||
        switch (s->bcRight) {
 | 
			
		||||
        case NOSLIP:
 | 
			
		||||
            for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
                for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                    U(imaxLocal, j, k)     = 0.0;
 | 
			
		||||
                    V(imaxLocal + 1, j, k) = -V(imaxLocal, j, k);
 | 
			
		||||
                    W(imaxLocal + 1, j, k) = -W(imaxLocal, j, k);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
            break;
 | 
			
		||||
        case SLIP:
 | 
			
		||||
            for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
                for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                    U(imaxLocal, j, k)     = 0.0;
 | 
			
		||||
                    V(imaxLocal + 1, j, k) = V(imaxLocal, j, k);
 | 
			
		||||
                    W(imaxLocal + 1, j, k) = W(imaxLocal, j, k);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
            break;
 | 
			
		||||
        case OUTFLOW:
 | 
			
		||||
            for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
                for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                    U(imaxLocal, j, k)     = U(imaxLocal - 1, j, k);
 | 
			
		||||
                    V(imaxLocal + 1, j, k) = V(imaxLocal, j, k);
 | 
			
		||||
                    W(imaxLocal + 1, j, k) = W(imaxLocal, j, k);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
            break;
 | 
			
		||||
        case PERIODIC:
 | 
			
		||||
            break;
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    if (commIsBoundary(&s->comm, FRONT)) {
 | 
			
		||||
        switch (s->bcFront) {
 | 
			
		||||
        case NOSLIP:
 | 
			
		||||
            for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                    U(i, j, 0) = -U(i, j, 1);
 | 
			
		||||
                    V(i, j, 0) = -V(i, j, 1);
 | 
			
		||||
                    W(i, j, 0) = 0.0;
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
            break;
 | 
			
		||||
        case SLIP:
 | 
			
		||||
            for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                    U(i, j, 0) = U(i, j, 1);
 | 
			
		||||
                    V(i, j, 0) = V(i, j, 1);
 | 
			
		||||
                    W(i, j, 0) = 0.0;
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
            break;
 | 
			
		||||
        case OUTFLOW:
 | 
			
		||||
            for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                    U(i, j, 0) = U(i, j, 1);
 | 
			
		||||
                    V(i, j, 0) = V(i, j, 1);
 | 
			
		||||
                    W(i, j, 0) = W(i, j, 1);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
            break;
 | 
			
		||||
        case PERIODIC:
 | 
			
		||||
            break;
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    if (commIsBoundary(&s->comm, BACK)) {
 | 
			
		||||
        switch (s->bcBack) {
 | 
			
		||||
        case NOSLIP:
 | 
			
		||||
            for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                    U(i, j, kmaxLocal + 1) = -U(i, j, kmaxLocal);
 | 
			
		||||
                    V(i, j, kmaxLocal + 1) = -V(i, j, kmaxLocal);
 | 
			
		||||
                    W(i, j, kmaxLocal)     = 0.0;
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
            break;
 | 
			
		||||
        case SLIP:
 | 
			
		||||
            for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                    U(i, j, kmaxLocal + 1) = U(i, j, kmaxLocal);
 | 
			
		||||
                    V(i, j, kmaxLocal + 1) = V(i, j, kmaxLocal);
 | 
			
		||||
                    W(i, j, kmaxLocal)     = 0.0;
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
            break;
 | 
			
		||||
        case OUTFLOW:
 | 
			
		||||
            for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                    U(i, j, kmaxLocal + 1) = U(i, j, kmaxLocal);
 | 
			
		||||
                    V(i, j, kmaxLocal + 1) = V(i, j, kmaxLocal);
 | 
			
		||||
                    W(i, j, kmaxLocal)     = W(i, j, kmaxLocal - 1);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
            break;
 | 
			
		||||
        case PERIODIC:
 | 
			
		||||
            break;
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void computeRHS(Discretization* s)
 | 
			
		||||
{
 | 
			
		||||
    int imaxLocal = s->comm.imaxLocal;
 | 
			
		||||
    int jmaxLocal = s->comm.jmaxLocal;
 | 
			
		||||
    int kmaxLocal = s->comm.kmaxLocal;
 | 
			
		||||
 | 
			
		||||
    double idx = 1.0 / s->grid.dx;
 | 
			
		||||
    double idy = 1.0 / s->grid.dy;
 | 
			
		||||
    double idz = 1.0 / s->grid.dz;
 | 
			
		||||
    double idt = 1.0 / s->dt;
 | 
			
		||||
 | 
			
		||||
    double* rhs = s->rhs;
 | 
			
		||||
    double* f   = s->f;
 | 
			
		||||
    double* g   = s->g;
 | 
			
		||||
    double* h   = s->h;
 | 
			
		||||
 | 
			
		||||
    commShift(&s->comm, f, g, h);
 | 
			
		||||
 | 
			
		||||
    for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
        for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
            for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                RHS(i, j, k) = ((F(i, j, k) - F(i - 1, j, k)) * idx +
 | 
			
		||||
                                   (G(i, j, k) - G(i, j - 1, k)) * idy +
 | 
			
		||||
                                   (H(i, j, k) - H(i, j, k - 1)) * idz) *
 | 
			
		||||
                               idt;
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void setSpecialBoundaryCondition(Discretization* s)
 | 
			
		||||
{
 | 
			
		||||
    int imaxLocal = s->comm.imaxLocal;
 | 
			
		||||
    int jmaxLocal = s->comm.jmaxLocal;
 | 
			
		||||
    int kmaxLocal = s->comm.kmaxLocal;
 | 
			
		||||
 | 
			
		||||
    double* u = s->u;
 | 
			
		||||
 | 
			
		||||
    if (strcmp(s->problem, "dcavity") == 0) {
 | 
			
		||||
        if (commIsBoundary(&s->comm, TOP)) {
 | 
			
		||||
            for (int k = 1; k < kmaxLocal; k++) {
 | 
			
		||||
                for (int i = 1; i < imaxLocal; i++) {
 | 
			
		||||
                    U(i, jmaxLocal + 1, k) = 2.0 - U(i, jmaxLocal, k);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    } else if (strcmp(s->problem, "canal") == 0) {
 | 
			
		||||
        if (commIsBoundary(&s->comm, LEFT)) {
 | 
			
		||||
            for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
                for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                    U(0, j, k) = 2.0;
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
static double maxElement(Discretization* s, double* m)
 | 
			
		||||
{
 | 
			
		||||
    int size = (s->comm.imaxLocal + 2) * (s->comm.jmaxLocal + 2) *
 | 
			
		||||
               (s->comm.kmaxLocal + 2);
 | 
			
		||||
    double maxval = DBL_MIN;
 | 
			
		||||
 | 
			
		||||
    for (int i = 0; i < size; i++) {
 | 
			
		||||
        maxval = MAX(maxval, fabs(m[i]));
 | 
			
		||||
    }
 | 
			
		||||
    commReduction(&maxval, MAX);
 | 
			
		||||
    return maxval;
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void normalizePressure(Discretization* s)
 | 
			
		||||
{
 | 
			
		||||
    int imaxLocal = s->comm.imaxLocal;
 | 
			
		||||
    int jmaxLocal = s->comm.jmaxLocal;
 | 
			
		||||
    int kmaxLocal = s->comm.kmaxLocal;
 | 
			
		||||
 | 
			
		||||
    double* p   = s->p;
 | 
			
		||||
    double avgP = 0.0;
 | 
			
		||||
 | 
			
		||||
    for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
        for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
            for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                avgP += P(i, j, k);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
    commReduction(&avgP, SUM);
 | 
			
		||||
    avgP /= (s->grid.imax * s->grid.jmax * s->grid.kmax);
 | 
			
		||||
 | 
			
		||||
    for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
        for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
            for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                P(i, j, k) = P(i, j, k) - avgP;
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void computeTimestep(Discretization* s)
 | 
			
		||||
{
 | 
			
		||||
    double dt = s->dtBound;
 | 
			
		||||
    double dx = s->grid.dx;
 | 
			
		||||
    double dy = s->grid.dy;
 | 
			
		||||
    double dz = s->grid.dz;
 | 
			
		||||
 | 
			
		||||
    double umax = maxElement(s, s->u);
 | 
			
		||||
    double vmax = maxElement(s, s->v);
 | 
			
		||||
    double wmax = maxElement(s, s->w);
 | 
			
		||||
 | 
			
		||||
    if (umax > 0) {
 | 
			
		||||
        dt = (dt > dx / umax) ? dx / umax : dt;
 | 
			
		||||
    }
 | 
			
		||||
    if (vmax > 0) {
 | 
			
		||||
        dt = (dt > dy / vmax) ? dy / vmax : dt;
 | 
			
		||||
    }
 | 
			
		||||
    if (wmax > 0) {
 | 
			
		||||
        dt = (dt > dz / wmax) ? dz / wmax : dt;
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    s->dt = dt * s->tau;
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
 | 
			
		||||
 | 
			
		||||
void computeFG(Discretization* s)
 | 
			
		||||
{
 | 
			
		||||
    int imaxLocal = s->comm.imaxLocal;
 | 
			
		||||
    int jmaxLocal = s->comm.jmaxLocal;
 | 
			
		||||
    int kmaxLocal = s->comm.kmaxLocal;
 | 
			
		||||
 | 
			
		||||
    double* u = s->u;
 | 
			
		||||
    double* v = s->v;
 | 
			
		||||
    double* w = s->w;
 | 
			
		||||
    double* f = s->f;
 | 
			
		||||
    double* g = s->g;
 | 
			
		||||
    double* h = s->h;
 | 
			
		||||
 | 
			
		||||
    double gx = s->gx;
 | 
			
		||||
    double gy = s->gy;
 | 
			
		||||
    double gz = s->gz;
 | 
			
		||||
    double dt = s->dt;
 | 
			
		||||
 | 
			
		||||
    double gamma     = s->gamma;
 | 
			
		||||
    double inverseRe = 1.0 / s->re;
 | 
			
		||||
    double inverseDx = 1.0 / s->grid.dx;
 | 
			
		||||
    double inverseDy = 1.0 / s->grid.dy;
 | 
			
		||||
    double inverseDz = 1.0 / s->grid.dz;
 | 
			
		||||
    double du2dx, dv2dy, dw2dz;
 | 
			
		||||
    double duvdx, duwdx, duvdy, dvwdy, duwdz, dvwdz;
 | 
			
		||||
    double du2dx2, du2dy2, du2dz2;
 | 
			
		||||
    double dv2dx2, dv2dy2, dv2dz2;
 | 
			
		||||
    double dw2dx2, dw2dy2, dw2dz2;
 | 
			
		||||
 | 
			
		||||
    commExchange(&s->comm, u);
 | 
			
		||||
    commExchange(&s->comm, v);
 | 
			
		||||
    commExchange(&s->comm, w);
 | 
			
		||||
 | 
			
		||||
    for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
        for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
            for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                du2dx = inverseDx * 0.25 *
 | 
			
		||||
                            ((U(i, j, k) + U(i + 1, j, k)) *
 | 
			
		||||
                                    (U(i, j, k) + U(i + 1, j, k)) -
 | 
			
		||||
                                (U(i, j, k) + U(i - 1, j, k)) *
 | 
			
		||||
                                    (U(i, j, k) + U(i - 1, j, k))) +
 | 
			
		||||
                        gamma * inverseDx * 0.25 *
 | 
			
		||||
                            (fabs(U(i, j, k) + U(i + 1, j, k)) *
 | 
			
		||||
                                    (U(i, j, k) - U(i + 1, j, k)) +
 | 
			
		||||
                                fabs(U(i, j, k) + U(i - 1, j, k)) *
 | 
			
		||||
                                    (U(i, j, k) - U(i - 1, j, k)));
 | 
			
		||||
 | 
			
		||||
                duvdy = inverseDy * 0.25 *
 | 
			
		||||
                            ((V(i, j, k) + V(i + 1, j, k)) *
 | 
			
		||||
                                    (U(i, j, k) + U(i, j + 1, k)) -
 | 
			
		||||
                                (V(i, j - 1, k) + V(i + 1, j - 1, k)) *
 | 
			
		||||
                                    (U(i, j, k) + U(i, j - 1, k))) +
 | 
			
		||||
                        gamma * inverseDy * 0.25 *
 | 
			
		||||
                            (fabs(V(i, j, k) + V(i + 1, j, k)) *
 | 
			
		||||
                                    (U(i, j, k) - U(i, j + 1, k)) +
 | 
			
		||||
                                fabs(V(i, j - 1, k) + V(i + 1, j - 1, k)) *
 | 
			
		||||
                                    (U(i, j, k) - U(i, j - 1, k)));
 | 
			
		||||
 | 
			
		||||
                duwdz = inverseDz * 0.25 *
 | 
			
		||||
                            ((W(i, j, k) + W(i + 1, j, k)) *
 | 
			
		||||
                                    (U(i, j, k) + U(i, j, k + 1)) -
 | 
			
		||||
                                (W(i, j, k - 1) + W(i + 1, j, k - 1)) *
 | 
			
		||||
                                    (U(i, j, k) + U(i, j, k - 1))) +
 | 
			
		||||
                        gamma * inverseDz * 0.25 *
 | 
			
		||||
                            (fabs(W(i, j, k) + W(i + 1, j, k)) *
 | 
			
		||||
                                    (U(i, j, k) - U(i, j, k + 1)) +
 | 
			
		||||
                                fabs(W(i, j, k - 1) + W(i + 1, j, k - 1)) *
 | 
			
		||||
                                    (U(i, j, k) - U(i, j, k - 1)));
 | 
			
		||||
 | 
			
		||||
                du2dx2 = inverseDx * inverseDx *
 | 
			
		||||
                         (U(i + 1, j, k) - 2.0 * U(i, j, k) + U(i - 1, j, k));
 | 
			
		||||
                du2dy2 = inverseDy * inverseDy *
 | 
			
		||||
                         (U(i, j + 1, k) - 2.0 * U(i, j, k) + U(i, j - 1, k));
 | 
			
		||||
                du2dz2 = inverseDz * inverseDz *
 | 
			
		||||
                         (U(i, j, k + 1) - 2.0 * U(i, j, k) + U(i, j, k - 1));
 | 
			
		||||
                F(i, j, k) = U(i, j, k) + dt * (inverseRe * (du2dx2 + du2dy2 + du2dz2) -
 | 
			
		||||
                                                   du2dx - duvdy - duwdz + gx);
 | 
			
		||||
 | 
			
		||||
                duvdx = inverseDx * 0.25 *
 | 
			
		||||
                            ((U(i, j, k) + U(i, j + 1, k)) *
 | 
			
		||||
                                    (V(i, j, k) + V(i + 1, j, k)) -
 | 
			
		||||
                                (U(i - 1, j, k) + U(i - 1, j + 1, k)) *
 | 
			
		||||
                                    (V(i, j, k) + V(i - 1, j, k))) +
 | 
			
		||||
                        gamma * inverseDx * 0.25 *
 | 
			
		||||
                            (fabs(U(i, j, k) + U(i, j + 1, k)) *
 | 
			
		||||
                                    (V(i, j, k) - V(i + 1, j, k)) +
 | 
			
		||||
                                fabs(U(i - 1, j, k) + U(i - 1, j + 1, k)) *
 | 
			
		||||
                                    (V(i, j, k) - V(i - 1, j, k)));
 | 
			
		||||
 | 
			
		||||
                dv2dy = inverseDy * 0.25 *
 | 
			
		||||
                            ((V(i, j, k) + V(i, j + 1, k)) *
 | 
			
		||||
                                    (V(i, j, k) + V(i, j + 1, k)) -
 | 
			
		||||
                                (V(i, j, k) + V(i, j - 1, k)) *
 | 
			
		||||
                                    (V(i, j, k) + V(i, j - 1, k))) +
 | 
			
		||||
                        gamma * inverseDy * 0.25 *
 | 
			
		||||
                            (fabs(V(i, j, k) + V(i, j + 1, k)) *
 | 
			
		||||
                                    (V(i, j, k) - V(i, j + 1, k)) +
 | 
			
		||||
                                fabs(V(i, j, k) + V(i, j - 1, k)) *
 | 
			
		||||
                                    (V(i, j, k) - V(i, j - 1, k)));
 | 
			
		||||
 | 
			
		||||
                dvwdz = inverseDz * 0.25 *
 | 
			
		||||
                            ((W(i, j, k) + W(i, j + 1, k)) *
 | 
			
		||||
                                    (V(i, j, k) + V(i, j, k + 1)) -
 | 
			
		||||
                                (W(i, j, k - 1) + W(i, j + 1, k - 1)) *
 | 
			
		||||
                                    (V(i, j, k) + V(i, j, k + 1))) +
 | 
			
		||||
                        gamma * inverseDz * 0.25 *
 | 
			
		||||
                            (fabs(W(i, j, k) + W(i, j + 1, k)) *
 | 
			
		||||
                                    (V(i, j, k) - V(i, j, k + 1)) +
 | 
			
		||||
                                fabs(W(i, j, k - 1) + W(i, j + 1, k - 1)) *
 | 
			
		||||
                                    (V(i, j, k) - V(i, j, k + 1)));
 | 
			
		||||
 | 
			
		||||
                dv2dx2 = inverseDx * inverseDx *
 | 
			
		||||
                         (V(i + 1, j, k) - 2.0 * V(i, j, k) + V(i - 1, j, k));
 | 
			
		||||
                dv2dy2 = inverseDy * inverseDy *
 | 
			
		||||
                         (V(i, j + 1, k) - 2.0 * V(i, j, k) + V(i, j - 1, k));
 | 
			
		||||
                dv2dz2 = inverseDz * inverseDz *
 | 
			
		||||
                         (V(i, j, k + 1) - 2.0 * V(i, j, k) + V(i, j, k - 1));
 | 
			
		||||
                G(i, j, k) = V(i, j, k) + dt * (inverseRe * (dv2dx2 + dv2dy2 + dv2dz2) -
 | 
			
		||||
                                                   duvdx - dv2dy - dvwdz + gy);
 | 
			
		||||
 | 
			
		||||
                duwdx = inverseDx * 0.25 *
 | 
			
		||||
                            ((U(i, j, k) + U(i, j, k + 1)) *
 | 
			
		||||
                                    (W(i, j, k) + W(i + 1, j, k)) -
 | 
			
		||||
                                (U(i - 1, j, k) + U(i - 1, j, k + 1)) *
 | 
			
		||||
                                    (W(i, j, k) + W(i - 1, j, k))) +
 | 
			
		||||
                        gamma * inverseDx * 0.25 *
 | 
			
		||||
                            (fabs(U(i, j, k) + U(i, j, k + 1)) *
 | 
			
		||||
                                    (W(i, j, k) - W(i + 1, j, k)) +
 | 
			
		||||
                                fabs(U(i - 1, j, k) + U(i - 1, j, k + 1)) *
 | 
			
		||||
                                    (W(i, j, k) - W(i - 1, j, k)));
 | 
			
		||||
 | 
			
		||||
                dvwdy = inverseDy * 0.25 *
 | 
			
		||||
                            ((V(i, j, k) + V(i, j, k + 1)) *
 | 
			
		||||
                                    (W(i, j, k) + W(i, j + 1, k)) -
 | 
			
		||||
                                (V(i, j - 1, k + 1) + V(i, j - 1, k)) *
 | 
			
		||||
                                    (W(i, j, k) + W(i, j - 1, k))) +
 | 
			
		||||
                        gamma * inverseDy * 0.25 *
 | 
			
		||||
                            (fabs(V(i, j, k) + V(i, j, k + 1)) *
 | 
			
		||||
                                    (W(i, j, k) - W(i, j + 1, k)) +
 | 
			
		||||
                                fabs(V(i, j - 1, k + 1) + V(i, j - 1, k)) *
 | 
			
		||||
                                    (W(i, j, k) - W(i, j - 1, k)));
 | 
			
		||||
 | 
			
		||||
                dw2dz = inverseDz * 0.25 *
 | 
			
		||||
                            ((W(i, j, k) + W(i, j, k + 1)) *
 | 
			
		||||
                                    (W(i, j, k) + W(i, j, k + 1)) -
 | 
			
		||||
                                (W(i, j, k) + W(i, j, k - 1)) *
 | 
			
		||||
                                    (W(i, j, k) + W(i, j, k - 1))) +
 | 
			
		||||
                        gamma * inverseDz * 0.25 *
 | 
			
		||||
                            (fabs(W(i, j, k) + W(i, j, k + 1)) *
 | 
			
		||||
                                    (W(i, j, k) - W(i, j, k + 1)) +
 | 
			
		||||
                                fabs(W(i, j, k) + W(i, j, k - 1)) *
 | 
			
		||||
                                    (W(i, j, k) - W(i, j, k - 1)));
 | 
			
		||||
 | 
			
		||||
                dw2dx2 = inverseDx * inverseDx *
 | 
			
		||||
                         (W(i + 1, j, k) - 2.0 * W(i, j, k) + W(i - 1, j, k));
 | 
			
		||||
                dw2dy2 = inverseDy * inverseDy *
 | 
			
		||||
                         (W(i, j + 1, k) - 2.0 * W(i, j, k) + W(i, j - 1, k));
 | 
			
		||||
                dw2dz2 = inverseDz * inverseDz *
 | 
			
		||||
                         (W(i, j, k + 1) - 2.0 * W(i, j, k) + W(i, j, k - 1));
 | 
			
		||||
                H(i, j, k) = W(i, j, k) + dt * (inverseRe * (dw2dx2 + dw2dy2 + dw2dz2) -
 | 
			
		||||
                                                   duwdx - dvwdy - dw2dz + gz);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    /* ----------------------------- boundary of F ---------------------------
 | 
			
		||||
     */
 | 
			
		||||
    if (commIsBoundary(&s->comm, LEFT)) {
 | 
			
		||||
        for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
            for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                F(0, j, k) = U(0, j, k);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    if (commIsBoundary(&s->comm, RIGHT)) {
 | 
			
		||||
        for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
            for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                F(imaxLocal, j, k) = U(imaxLocal, j, k);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    /* ----------------------------- boundary of G ---------------------------
 | 
			
		||||
     */
 | 
			
		||||
    if (commIsBoundary(&s->comm, BOTTOM)) {
 | 
			
		||||
        for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
            for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                G(i, 0, k) = V(i, 0, k);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    if (commIsBoundary(&s->comm, TOP)) {
 | 
			
		||||
        for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
            for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                G(i, jmaxLocal, k) = V(i, jmaxLocal, k);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    /* ----------------------------- boundary of H ---------------------------
 | 
			
		||||
     */
 | 
			
		||||
    if (commIsBoundary(&s->comm, FRONT)) {
 | 
			
		||||
        for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
            for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                H(i, j, 0) = W(i, j, 0);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    if (commIsBoundary(&s->comm, BACK)) {
 | 
			
		||||
        for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
            for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                H(i, j, kmaxLocal) = W(i, j, kmaxLocal);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void adaptUV(Discretization* s)
 | 
			
		||||
{
 | 
			
		||||
    int imaxLocal = s->comm.imaxLocal;
 | 
			
		||||
    int jmaxLocal = s->comm.jmaxLocal;
 | 
			
		||||
    int kmaxLocal = s->comm.kmaxLocal;
 | 
			
		||||
 | 
			
		||||
    double* p = s->p;
 | 
			
		||||
    double* u = s->u;
 | 
			
		||||
    double* v = s->v;
 | 
			
		||||
    double* w = s->w;
 | 
			
		||||
    double* f = s->f;
 | 
			
		||||
    double* g = s->g;
 | 
			
		||||
    double* h = s->h;
 | 
			
		||||
 | 
			
		||||
    double factorX = s->dt / s->grid.dx;
 | 
			
		||||
    double factorY = s->dt / s->grid.dy;
 | 
			
		||||
    double factorZ = s->dt / s->grid.dz;
 | 
			
		||||
 | 
			
		||||
    for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
        for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
            for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                U(i, j, k) = F(i, j, k) - (P(i + 1, j, k) - P(i, j, k)) * factorX;
 | 
			
		||||
                V(i, j, k) = G(i, j, k) - (P(i, j + 1, k) - P(i, j, k)) * factorY;
 | 
			
		||||
                W(i, j, k) = H(i, j, k) - (P(i, j, k + 1) - P(i, j, k)) * factorZ;
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
}
 | 
			
		||||
							
								
								
									
										44
									
								
								BasicSolver/3D-mpi/src/discretization.h
									
									
									
									
									
										Normal file
									
								
							
							
						
						
									
										44
									
								
								BasicSolver/3D-mpi/src/discretization.h
									
									
									
									
									
										Normal file
									
								
							@@ -0,0 +1,44 @@
 | 
			
		||||
/*
 | 
			
		||||
 * Copyright (C)  NHR@FAU, University Erlangen-Nuremberg.
 | 
			
		||||
 * All rights reserved. This file is part of nusif-solver.
 | 
			
		||||
 * Use of this source code is governed by a MIT style
 | 
			
		||||
 * license that can be found in the LICENSE file.
 | 
			
		||||
 */
 | 
			
		||||
#ifndef __DISCRETIZATION_H_
 | 
			
		||||
#define __DISCRETIZATION_H_
 | 
			
		||||
 | 
			
		||||
#include "grid.h"
 | 
			
		||||
#include "parameter.h"
 | 
			
		||||
#include "comm.h"
 | 
			
		||||
 | 
			
		||||
enum BC { NOSLIP = 1, SLIP, OUTFLOW, PERIODIC };
 | 
			
		||||
 | 
			
		||||
typedef struct {
 | 
			
		||||
    /* geometry and grid information */
 | 
			
		||||
    Grid grid;
 | 
			
		||||
    /* arrays */
 | 
			
		||||
    double *p, *rhs;
 | 
			
		||||
    double *f, *g, *h;
 | 
			
		||||
    double *u, *v, *w;
 | 
			
		||||
    /* parameters */
 | 
			
		||||
    double eps, omega;
 | 
			
		||||
    double re, tau, gamma;
 | 
			
		||||
    double gx, gy, gz;
 | 
			
		||||
    /* time stepping */
 | 
			
		||||
    int itermax;
 | 
			
		||||
    double dt, te;
 | 
			
		||||
    double dtBound;
 | 
			
		||||
    char* problem;
 | 
			
		||||
    int bcLeft, bcRight, bcBottom, bcTop, bcFront, bcBack;
 | 
			
		||||
    Comm comm;
 | 
			
		||||
} Discretization;
 | 
			
		||||
 | 
			
		||||
extern void initDiscretization(Discretization*, Parameter*);
 | 
			
		||||
extern void computeRHS(Discretization*);
 | 
			
		||||
extern void normalizePressure(Discretization*);
 | 
			
		||||
extern void computeTimestep(Discretization*);
 | 
			
		||||
extern void setBoundaryConditions(Discretization*);
 | 
			
		||||
extern void setSpecialBoundaryCondition(Discretization*);
 | 
			
		||||
extern void computeFG(Discretization*);
 | 
			
		||||
extern void adaptUV(Discretization*);
 | 
			
		||||
#endif
 | 
			
		||||
@@ -9,7 +9,7 @@
 | 
			
		||||
#include <unistd.h>
 | 
			
		||||
 | 
			
		||||
#include "allocate.h"
 | 
			
		||||
#include "comm.h"
 | 
			
		||||
#include "discretization.h"
 | 
			
		||||
#include "parameter.h"
 | 
			
		||||
#include "progress.h"
 | 
			
		||||
#include "solver.h"
 | 
			
		||||
@@ -21,9 +21,12 @@ int main(int argc, char** argv)
 | 
			
		||||
    double timeStart, timeStop;
 | 
			
		||||
    Parameter p;
 | 
			
		||||
    Solver s;
 | 
			
		||||
    Discretization d;
 | 
			
		||||
 | 
			
		||||
    commInit(&s.comm, argc, argv);
 | 
			
		||||
    commInit(&d.comm, argc, argv);
 | 
			
		||||
    initParameter(&p);
 | 
			
		||||
    FILE* fp;
 | 
			
		||||
    if (commIsMaster(&d.comm)) fp = initResidualWriter();
 | 
			
		||||
 | 
			
		||||
    if (argc != 2) {
 | 
			
		||||
        printf("Usage: %s <configFile>\n", argv[0]);
 | 
			
		||||
@@ -31,33 +34,44 @@ int main(int argc, char** argv)
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    readParameter(&p, argv[1]);
 | 
			
		||||
    commPartition(&s.comm, p.kmax, p.jmax, p.imax);
 | 
			
		||||
    if (commIsMaster(&s.comm)) {
 | 
			
		||||
    commPartition(&d.comm, p.kmax, p.jmax, p.imax);
 | 
			
		||||
 | 
			
		||||
    if (commIsMaster(&d.comm)) {
 | 
			
		||||
        printParameter(&p);
 | 
			
		||||
    }
 | 
			
		||||
    initSolver(&s, &p);
 | 
			
		||||
    
 | 
			
		||||
    initDiscretization(&d, &p);
 | 
			
		||||
    initSolver(&s, &d, &p);
 | 
			
		||||
 | 
			
		||||
#ifndef VERBOSE
 | 
			
		||||
    initProgress(s.te);
 | 
			
		||||
    initProgress(d.te);
 | 
			
		||||
#endif
 | 
			
		||||
 | 
			
		||||
    double tau = s.tau;
 | 
			
		||||
    double te  = s.te;
 | 
			
		||||
    double tau = d.tau;
 | 
			
		||||
    double te  = d.te;
 | 
			
		||||
    double t   = 0.0;
 | 
			
		||||
    int nt     = 0;
 | 
			
		||||
    double res = 0.0;
 | 
			
		||||
 | 
			
		||||
    timeStart = getTimeStamp();
 | 
			
		||||
    while (t <= te) {
 | 
			
		||||
        if (tau > 0.0) computeTimestep(&s);
 | 
			
		||||
        setBoundaryConditions(&s);
 | 
			
		||||
        setSpecialBoundaryCondition(&s);
 | 
			
		||||
        computeFG(&s);
 | 
			
		||||
        computeRHS(&s);
 | 
			
		||||
        solve(&s);
 | 
			
		||||
        adaptUV(&s);
 | 
			
		||||
        t += s.dt;
 | 
			
		||||
        if (tau > 0.0) computeTimestep(&d);
 | 
			
		||||
        setBoundaryConditions(&d);
 | 
			
		||||
        setSpecialBoundaryCondition(&d);
 | 
			
		||||
        computeFG(&d);
 | 
			
		||||
        computeRHS(&d);
 | 
			
		||||
        if (nt % 100 == 0) normalizePressure(&d);
 | 
			
		||||
        res = solve(&s, d.p, d.rhs);
 | 
			
		||||
        adaptUV(&d);
 | 
			
		||||
 | 
			
		||||
        if (commIsMaster(&d.comm)) writeResidual(fp, t, res);
 | 
			
		||||
 | 
			
		||||
        t += d.dt;
 | 
			
		||||
        nt++;
 | 
			
		||||
 | 
			
		||||
#ifdef VERBOSE
 | 
			
		||||
        if (commIsMaster(&s.comm)) {
 | 
			
		||||
            printf("TIME %f , TIMESTEP %f\n", t, s.dt);
 | 
			
		||||
        if (commIsMaster(s.comm)) {
 | 
			
		||||
            printf("TIME %f , TIMESTEP %f\n", t, d.dt);
 | 
			
		||||
        }
 | 
			
		||||
#else
 | 
			
		||||
        printProgress(t);
 | 
			
		||||
@@ -67,7 +81,7 @@ int main(int argc, char** argv)
 | 
			
		||||
#ifndef VERBOSE
 | 
			
		||||
    stopProgress();
 | 
			
		||||
#endif
 | 
			
		||||
    if (commIsMaster(&s.comm)) {
 | 
			
		||||
    if (commIsMaster(s.comm)) {
 | 
			
		||||
        printf("Solution took %.2fs\n", timeStop - timeStart);
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
@@ -75,14 +89,16 @@ int main(int argc, char** argv)
 | 
			
		||||
#ifdef _VTK_WRITER_MPI
 | 
			
		||||
    VtkOptions opts = { .grid = s.grid, .comm = s.comm };
 | 
			
		||||
    vtkOpen(&opts, s.problem);
 | 
			
		||||
    vtkScalar(&opts, "pressure", s.p);
 | 
			
		||||
    vtkVector(&opts, "velocity", (VtkVector) { s.u, s.v, s.w });
 | 
			
		||||
    vtkScalar(&opts, "pressure", d.p);
 | 
			
		||||
    vtkVector(&opts, "velocity", (VtkVector) { d.u, d.v, d.w });
 | 
			
		||||
    vtkClose(&opts);
 | 
			
		||||
#else
 | 
			
		||||
    if (commIsMaster(&d.comm)) fclose(fp);
 | 
			
		||||
 | 
			
		||||
    double *pg, *ug, *vg, *wg;
 | 
			
		||||
 | 
			
		||||
    if (commIsMaster(&s.comm)) {
 | 
			
		||||
        size_t bytesize = s.grid.imax * s.grid.jmax * s.grid.kmax * sizeof(double);
 | 
			
		||||
    if (commIsMaster(s.comm)) {
 | 
			
		||||
        size_t bytesize = s.grid->imax * s.grid->jmax * s.grid->kmax * sizeof(double);
 | 
			
		||||
 | 
			
		||||
        pg = allocate(64, bytesize);
 | 
			
		||||
        ug = allocate(64, bytesize);
 | 
			
		||||
@@ -90,34 +106,35 @@ int main(int argc, char** argv)
 | 
			
		||||
        wg = allocate(64, bytesize);
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    commCollectResult(&s.comm,
 | 
			
		||||
    commCollectResult(s.comm,
 | 
			
		||||
        ug,
 | 
			
		||||
        vg,
 | 
			
		||||
        wg,
 | 
			
		||||
        pg,
 | 
			
		||||
        s.u,
 | 
			
		||||
        s.v,
 | 
			
		||||
        s.w,
 | 
			
		||||
        s.p,
 | 
			
		||||
        s.grid.kmax,
 | 
			
		||||
        s.grid.jmax,
 | 
			
		||||
        s.grid.imax);
 | 
			
		||||
        d.u,
 | 
			
		||||
        d.v,
 | 
			
		||||
        d.w,
 | 
			
		||||
        d.p,
 | 
			
		||||
        s.grid->kmax,
 | 
			
		||||
        s.grid->jmax,
 | 
			
		||||
        s.grid->imax);
 | 
			
		||||
 | 
			
		||||
    if (commIsMaster(&s.comm)) {
 | 
			
		||||
    if (commIsMaster(s.comm)) {
 | 
			
		||||
        VtkOptions opts = { .grid = s.grid };
 | 
			
		||||
        vtkOpen(&opts, s.problem);
 | 
			
		||||
        vtkScalar(&opts, "pressure", pg);
 | 
			
		||||
        vtkVector(&opts, "velocity", (VtkVector) { ug, vg, wg });
 | 
			
		||||
        vtkClose(&opts);
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
#endif
 | 
			
		||||
 | 
			
		||||
    timeStop = getTimeStamp();
 | 
			
		||||
 | 
			
		||||
    if (commIsMaster(&s.comm)) {
 | 
			
		||||
    if (commIsMaster(s.comm)) {
 | 
			
		||||
        printf("Result output took %.2fs\n", timeStop - timeStart);
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    commFinalize(&s.comm);
 | 
			
		||||
    commFinalize(s.comm);
 | 
			
		||||
    return EXIT_SUCCESS;
 | 
			
		||||
}
 | 
			
		||||
 
 | 
			
		||||
@@ -14,18 +14,22 @@
 | 
			
		||||
 | 
			
		||||
void initParameter(Parameter* param)
 | 
			
		||||
{
 | 
			
		||||
    param->xlength = 1.0;
 | 
			
		||||
    param->ylength = 1.0;
 | 
			
		||||
    param->zlength = 1.0;
 | 
			
		||||
    param->imax    = 100;
 | 
			
		||||
    param->jmax    = 100;
 | 
			
		||||
    param->kmax    = 100;
 | 
			
		||||
    param->itermax = 1000;
 | 
			
		||||
    param->eps     = 0.0001;
 | 
			
		||||
    param->omg     = 1.7;
 | 
			
		||||
    param->re      = 100.0;
 | 
			
		||||
    param->gamma   = 0.9;
 | 
			
		||||
    param->tau     = 0.5;
 | 
			
		||||
    param->xlength     = 1.0;
 | 
			
		||||
    param->ylength     = 1.0;
 | 
			
		||||
    param->zlength     = 1.0;
 | 
			
		||||
    param->imax        = 100;
 | 
			
		||||
    param->jmax        = 100;
 | 
			
		||||
    param->kmax        = 100;
 | 
			
		||||
    param->itermax     = 1000;
 | 
			
		||||
    param->eps         = 0.0001;
 | 
			
		||||
    param->omg         = 1.7;
 | 
			
		||||
    param->re          = 100.0;
 | 
			
		||||
    param->gamma       = 0.9;
 | 
			
		||||
    param->tau         = 0.5;
 | 
			
		||||
    param->levels     = 5;
 | 
			
		||||
    param->presmooth  = 5;
 | 
			
		||||
    param->postsmooth = 5;
 | 
			
		||||
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void readParameter(Parameter* param, const char* filename)
 | 
			
		||||
@@ -65,6 +69,9 @@ void readParameter(Parameter* param, const char* filename)
 | 
			
		||||
            PARSE_INT(jmax);
 | 
			
		||||
            PARSE_INT(kmax);
 | 
			
		||||
            PARSE_INT(itermax);
 | 
			
		||||
            PARSE_INT(levels);
 | 
			
		||||
            PARSE_INT(presmooth);
 | 
			
		||||
            PARSE_INT(postsmooth);
 | 
			
		||||
            PARSE_REAL(eps);
 | 
			
		||||
            PARSE_REAL(omg);
 | 
			
		||||
            PARSE_REAL(re);
 | 
			
		||||
 
 | 
			
		||||
@@ -18,6 +18,7 @@ typedef struct {
 | 
			
		||||
    char* name;
 | 
			
		||||
    int bcLeft, bcRight, bcBottom, bcTop, bcFront, bcBack;
 | 
			
		||||
    double u_init, v_init, w_init, p_init;
 | 
			
		||||
    int levels, presmooth, postsmooth;
 | 
			
		||||
} Parameter;
 | 
			
		||||
 | 
			
		||||
void initParameter(Parameter*);
 | 
			
		||||
 
 | 
			
		||||
@@ -4,12 +4,12 @@
 | 
			
		||||
 * Use of this source code is governed by a MIT style
 | 
			
		||||
 * license that can be found in the LICENSE file.
 | 
			
		||||
 */
 | 
			
		||||
#include "progress.h"
 | 
			
		||||
#include <math.h>
 | 
			
		||||
#include <stdio.h>
 | 
			
		||||
#include <stdlib.h>
 | 
			
		||||
#include <string.h>
 | 
			
		||||
 | 
			
		||||
#include "progress.h"
 | 
			
		||||
 | 
			
		||||
static double _end;
 | 
			
		||||
static int _current;
 | 
			
		||||
 | 
			
		||||
@@ -48,3 +48,18 @@ void stopProgress()
 | 
			
		||||
    printf("\n");
 | 
			
		||||
    fflush(stdout);
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
FILE* initResidualWriter()
 | 
			
		||||
{
 | 
			
		||||
    FILE* fp;
 | 
			
		||||
    fp = fopen("residual.dat", "w");
 | 
			
		||||
 | 
			
		||||
    if (fp == NULL) {
 | 
			
		||||
        printf("Error!\n");
 | 
			
		||||
        exit(EXIT_FAILURE);
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    return fp;
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void writeResidual(FILE* fp, double ts, double res) { fprintf(fp, "%f, %f\n", ts, res); }
 | 
			
		||||
@@ -4,11 +4,14 @@
 | 
			
		||||
 * Use of this source code is governed by a MIT-style
 | 
			
		||||
 * license that can be found in the LICENSE file.
 | 
			
		||||
 */
 | 
			
		||||
 | 
			
		||||
#include <stdio.h>
 | 
			
		||||
 | 
			
		||||
#ifndef __PROGRESS_H_
 | 
			
		||||
#define __PROGRESS_H_
 | 
			
		||||
 | 
			
		||||
extern void initProgress(double);
 | 
			
		||||
extern void printProgress(double);
 | 
			
		||||
extern void stopProgress(void);
 | 
			
		||||
 | 
			
		||||
extern FILE* initResidualWriter(void);
 | 
			
		||||
extern void writeResidual(FILE*, double, double);
 | 
			
		||||
#endif
 | 
			
		||||
 
 | 
			
		||||
							
								
								
									
										403
									
								
								BasicSolver/3D-mpi/src/solver-mg.c
									
									
									
									
									
										Normal file
									
								
							
							
						
						
									
										403
									
								
								BasicSolver/3D-mpi/src/solver-mg.c
									
									
									
									
									
										Normal file
									
								
							@@ -0,0 +1,403 @@
 | 
			
		||||
/*
 | 
			
		||||
 * Copyright (C) NHR@FAU, University Erlangen-Nuremberg.
 | 
			
		||||
 * All rights reserved. This file is part of nusif-solver.
 | 
			
		||||
 * Use of this source code is governed by a MIT style
 | 
			
		||||
 * license that can be found in the LICENSE file.
 | 
			
		||||
 */
 | 
			
		||||
#include <stdio.h>
 | 
			
		||||
#include <stdlib.h>
 | 
			
		||||
 | 
			
		||||
#include "allocate.h"
 | 
			
		||||
#include "solver.h"
 | 
			
		||||
#include "util.h"
 | 
			
		||||
 | 
			
		||||
#define FINEST_LEVEL   0
 | 
			
		||||
#define COARSEST_LEVEL (s->levels - 1)
 | 
			
		||||
#define S(i, j, k)                                                                       \
 | 
			
		||||
    s[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define E(i, j, k)                                                                       \
 | 
			
		||||
    e[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define R(i, j, k)                                                                       \
 | 
			
		||||
    r[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define OLD(i, j, k)                                                                     \
 | 
			
		||||
    old[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
 | 
			
		||||
static void restrictMG(Solver* s, int level, Comm* comm)
 | 
			
		||||
{
 | 
			
		||||
    double* r   = s->r[level + 1];
 | 
			
		||||
    double* old = s->r[level];
 | 
			
		||||
 | 
			
		||||
    int imaxLocal = comm->imaxLocal;
 | 
			
		||||
    int jmaxLocal = comm->jmaxLocal;
 | 
			
		||||
    int kmaxLocal = comm->kmaxLocal;
 | 
			
		||||
 | 
			
		||||
    commExchange(comm, old);
 | 
			
		||||
 | 
			
		||||
    for (int k = 1; k < comm->kmaxLocal + 1; k++) {
 | 
			
		||||
        for (int j = 1; j < comm->jmaxLocal + 1; j++) {
 | 
			
		||||
            for (int i = 1; i < comm->imaxLocal + 1; ++i) {
 | 
			
		||||
                R(i, j, k) = (OLD(2 * i - 1, 2 * j - 1, 2 * k) +
 | 
			
		||||
                                 OLD(2 * i, 2 * j - 1, 2 * k) * 2 +
 | 
			
		||||
                                 OLD(2 * i + 1, 2 * j - 1, 2 * k) +
 | 
			
		||||
                                 OLD(2 * i - 1, 2 * j, 2 * k) * 2 +
 | 
			
		||||
                                 OLD(2 * i, 2 * j, 2 * k) * 8 +
 | 
			
		||||
                                 OLD(2 * i + 1, 2 * j, 2 * k) * 2 +
 | 
			
		||||
                                 OLD(2 * i - 1, 2 * j + 1, 2 * k) +
 | 
			
		||||
                                 OLD(2 * i, 2 * j + 1, 2 * k) * 2 +
 | 
			
		||||
                                 OLD(2 * i + 1, 2 * j + 1, 2 * k) +
 | 
			
		||||
 | 
			
		||||
                                 OLD(2 * i - 1, 2 * j - 1, 2 * k - 1) +
 | 
			
		||||
                                 OLD(2 * i, 2 * j - 1, 2 * k - 1) * 2 +
 | 
			
		||||
                                 OLD(2 * i + 1, 2 * j - 1, 2 * k - 1) +
 | 
			
		||||
                                 OLD(2 * i - 1, 2 * j, 2 * k - 1) * 2 +
 | 
			
		||||
                                 OLD(2 * i, 2 * j, 2 * k - 1) * 4 +
 | 
			
		||||
                                 OLD(2 * i + 1, 2 * j, 2 * k - 1) * 2 +
 | 
			
		||||
                                 OLD(2 * i - 1, 2 * j + 1, 2 * k - 1) +
 | 
			
		||||
                                 OLD(2 * i, 2 * j + 1, 2 * k - 1) * 2 +
 | 
			
		||||
                                 OLD(2 * i + 1, 2 * j + 1, 2 * k - 1) +
 | 
			
		||||
 | 
			
		||||
                                 OLD(2 * i - 1, 2 * j - 1, 2 * k + 1) +
 | 
			
		||||
                                 OLD(2 * i, 2 * j - 1, 2 * k + 1) * 2 +
 | 
			
		||||
                                 OLD(2 * i + 1, 2 * j - 1, 2 * k + 1) +
 | 
			
		||||
                                 OLD(2 * i - 1, 2 * j, 2 * k + 1) * 2 +
 | 
			
		||||
                                 OLD(2 * i, 2 * j, 2 * k + 1) * 4 +
 | 
			
		||||
                                 OLD(2 * i + 1, 2 * j, 2 * k + 1) * 2 +
 | 
			
		||||
                                 OLD(2 * i - 1, 2 * j + 1, 2 * k + 1) +
 | 
			
		||||
                                 OLD(2 * i, 2 * j + 1, 2 * k + 1) * 2 +
 | 
			
		||||
                                 OLD(2 * i + 1, 2 * j + 1, 2 * k + 1)) /
 | 
			
		||||
                             64.0;
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
static void prolongate(Solver* s, int level, Comm* comm)
 | 
			
		||||
{
 | 
			
		||||
    double* old = s->r[level + 1];
 | 
			
		||||
    double* e   = s->r[level];
 | 
			
		||||
 | 
			
		||||
    int imaxLocal = comm->imaxLocal;
 | 
			
		||||
    int jmaxLocal = comm->jmaxLocal;
 | 
			
		||||
    int kmaxLocal = comm->kmaxLocal;
 | 
			
		||||
 | 
			
		||||
    for (int k = 2; k < kmaxLocal + 1; k += 2) {
 | 
			
		||||
        for (int j = 2; j < jmaxLocal + 1; j += 2) {
 | 
			
		||||
            for (int i = 2; i < imaxLocal + 1; i += 2) {
 | 
			
		||||
                E(i, j, k) = OLD(i / 2, j / 2, k / 2);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
static void correct(Solver* s, double* p, int level, Comm* comm)
 | 
			
		||||
{
 | 
			
		||||
    double* e = s->e[level];
 | 
			
		||||
 | 
			
		||||
    int imaxLocal = comm->imaxLocal;
 | 
			
		||||
    int jmaxLocal = comm->jmaxLocal;
 | 
			
		||||
    int kmaxLocal = comm->kmaxLocal;
 | 
			
		||||
 | 
			
		||||
    for (int k = 1; k < kmaxLocal + 1; ++k) {
 | 
			
		||||
        for (int j = 1; j < jmaxLocal + 1; ++j) {
 | 
			
		||||
            for (int i = 1; i < imaxLocal + 1; ++i) {
 | 
			
		||||
                P(i, j, k) += E(i, j, k);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
static void setBoundaryCondition(
 | 
			
		||||
    Solver* s, double* p, int imaxLocal, int jmaxLocal, int kmaxLocal)
 | 
			
		||||
{
 | 
			
		||||
#ifdef _MPI
 | 
			
		||||
    if (commIsBoundary(s->comm, FRONT)) {
 | 
			
		||||
        for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
            for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                P(i, j, 0) = P(i, j, 1);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    if (commIsBoundary(s->comm, BACK)) {
 | 
			
		||||
        for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
            for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                P(i, j, kmaxLocal + 1) = P(i, j, kmaxLocal);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    if (commIsBoundary(s->comm, BOTTOM)) {
 | 
			
		||||
        for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
            for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                P(i, 0, k) = P(i, 1, k);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    if (commIsBoundary(s->comm, TOP)) {
 | 
			
		||||
        for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
            for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                P(i, jmaxLocal + 1, k) = P(i, jmaxLocal, k);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    if (commIsBoundary(s->comm, LEFT)) {
 | 
			
		||||
        for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
            for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                P(0, j, k) = P(1, j, k);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    if (commIsBoundary(s->comm, RIGHT)) {
 | 
			
		||||
        for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
            for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                P(imaxLocal + 1, j, k) = P(imaxLocal, j, k);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
#else
 | 
			
		||||
    for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
        for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
            P(i, j, 0)             = P(i, j, 1);
 | 
			
		||||
            P(i, j, kmaxLocal + 1) = P(i, j, kmaxLocal);
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
        for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
            P(i, 0, k)             = P(i, 1, k);
 | 
			
		||||
            P(i, jmaxLocal + 1, k) = P(i, jmaxLocal, k);
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
        for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
            P(0, j, k)             = P(1, j, k);
 | 
			
		||||
            P(imaxLocal + 1, j, k) = P(imaxLocal, j, k);
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
#endif
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
static double smooth(Solver* s, double* p, double* rhs, int level, Comm* comm)
 | 
			
		||||
{
 | 
			
		||||
    int imaxLocal = comm->imaxLocal;
 | 
			
		||||
    int jmaxLocal = comm->jmaxLocal;
 | 
			
		||||
    int kmaxLocal = comm->kmaxLocal;
 | 
			
		||||
 | 
			
		||||
    double eps    = s->eps;
 | 
			
		||||
    int itermax   = s->itermax;
 | 
			
		||||
    double dx2    = s->grid->dx * s->grid->dx;
 | 
			
		||||
    double dy2    = s->grid->dy * s->grid->dy;
 | 
			
		||||
    double dz2    = s->grid->dz * s->grid->dz;
 | 
			
		||||
    double idx2   = 1.0 / dx2;
 | 
			
		||||
    double idy2   = 1.0 / dy2;
 | 
			
		||||
    double idz2   = 1.0 / dz2;
 | 
			
		||||
    double factor = s->omega * 0.5 * (dx2 * dy2 * dz2) /
 | 
			
		||||
                    (dy2 * dz2 + dx2 * dz2 + dx2 * dy2);
 | 
			
		||||
    double* r    = s->r[level];
 | 
			
		||||
    double epssq = eps * eps;
 | 
			
		||||
    int it       = 0;
 | 
			
		||||
    int pass, ksw, jsw, isw;
 | 
			
		||||
    double res = 1.0;
 | 
			
		||||
 | 
			
		||||
    ksw = 1;
 | 
			
		||||
 | 
			
		||||
    for (pass = 0; pass < 2; pass++) {
 | 
			
		||||
        jsw = ksw;
 | 
			
		||||
 | 
			
		||||
        commExchange(comm, p);
 | 
			
		||||
 | 
			
		||||
        for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
            isw = jsw;
 | 
			
		||||
            for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                for (int i = isw; i < imaxLocal + 1; i += 2) {
 | 
			
		||||
 | 
			
		||||
                    P(i, j, k) -=
 | 
			
		||||
                        factor *
 | 
			
		||||
                        (RHS(i, j, k) -
 | 
			
		||||
                            ((P(i + 1, j, k) - 2.0 * P(i, j, k) + P(i - 1, j, k)) * idx2 +
 | 
			
		||||
                                (P(i, j + 1, k) - 2.0 * P(i, j, k) + P(i, j - 1, k)) *
 | 
			
		||||
                                    idy2 +
 | 
			
		||||
                                (P(i, j, k + 1) - 2.0 * P(i, j, k) + P(i, j, k - 1)) *
 | 
			
		||||
                                    idz2));
 | 
			
		||||
                }
 | 
			
		||||
                isw = 3 - isw;
 | 
			
		||||
            }
 | 
			
		||||
            jsw = 3 - jsw;
 | 
			
		||||
        }
 | 
			
		||||
        ksw = 3 - ksw;
 | 
			
		||||
    }
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
static double calculateResidual(Solver* s, double* p, double* rhs, int level, Comm* comm)
 | 
			
		||||
{
 | 
			
		||||
 | 
			
		||||
    int imaxLocal = comm->imaxLocal;
 | 
			
		||||
    int jmaxLocal = comm->jmaxLocal;
 | 
			
		||||
    int kmaxLocal = comm->kmaxLocal;
 | 
			
		||||
 | 
			
		||||
    double eps    = s->eps;
 | 
			
		||||
    int itermax   = s->itermax;
 | 
			
		||||
    double dx2    = s->grid->dx * s->grid->dx;
 | 
			
		||||
    double dy2    = s->grid->dy * s->grid->dy;
 | 
			
		||||
    double dz2    = s->grid->dz * s->grid->dz;
 | 
			
		||||
    double idx2   = 1.0 / dx2;
 | 
			
		||||
    double idy2   = 1.0 / dy2;
 | 
			
		||||
    double idz2   = 1.0 / dz2;
 | 
			
		||||
    double factor = s->omega * 0.5 * (dx2 * dy2 * dz2) /
 | 
			
		||||
                    (dy2 * dz2 + dx2 * dz2 + dx2 * dy2);
 | 
			
		||||
    double* r    = s->r[level];
 | 
			
		||||
    double epssq = eps * eps;
 | 
			
		||||
    int it       = 0;
 | 
			
		||||
    int pass, ksw, jsw, isw;
 | 
			
		||||
    double res = 1.0;
 | 
			
		||||
 | 
			
		||||
    ksw = 1;
 | 
			
		||||
 | 
			
		||||
    for (pass = 0; pass < 2; pass++) {
 | 
			
		||||
        jsw = ksw;
 | 
			
		||||
 | 
			
		||||
        commExchange(comm, p);
 | 
			
		||||
 | 
			
		||||
        for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
            isw = jsw;
 | 
			
		||||
            for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                for (int i = isw; i < imaxLocal + 1; i += 2) {
 | 
			
		||||
 | 
			
		||||
                    R(i,
 | 
			
		||||
                        j,
 | 
			
		||||
                        k) = (RHS(i, j, k) -
 | 
			
		||||
                              ((P(i + 1, j, k) - 2.0 * P(i, j, k) + P(i - 1, j, k)) *
 | 
			
		||||
                                      idx2 +
 | 
			
		||||
                                  (P(i, j + 1, k) - 2.0 * P(i, j, k) + P(i, j - 1, k)) *
 | 
			
		||||
                                      idy2 +
 | 
			
		||||
                                  (P(i, j, k + 1) - 2.0 * P(i, j, k) + P(i, j, k - 1)) *
 | 
			
		||||
                                      idz2));
 | 
			
		||||
 | 
			
		||||
                    res += (R(i, j, k) * R(i, j, k));
 | 
			
		||||
                }
 | 
			
		||||
                isw = 3 - isw;
 | 
			
		||||
            }
 | 
			
		||||
            jsw = 3 - jsw;
 | 
			
		||||
        }
 | 
			
		||||
        ksw = 3 - ksw;
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    commReduction(&res, SUM);
 | 
			
		||||
 | 
			
		||||
    res = res / (double)(imaxLocal * jmaxLocal * kmaxLocal);
 | 
			
		||||
#ifdef DEBUG
 | 
			
		||||
    if (commIsMaster(s->comm)) {
 | 
			
		||||
        printf("%d Residuum: %e\n", it, res);
 | 
			
		||||
    }
 | 
			
		||||
#endif
 | 
			
		||||
    return res;
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
static double multiGrid(Solver* s, double* p, double* rhs, int level, Comm* comm)
 | 
			
		||||
{
 | 
			
		||||
    int imaxLocal = comm->imaxLocal;
 | 
			
		||||
    int jmaxLocal = comm->jmaxLocal;
 | 
			
		||||
    int kmaxLocal = comm->kmaxLocal;
 | 
			
		||||
 | 
			
		||||
    double res = 0.0;
 | 
			
		||||
 | 
			
		||||
    // coarsest level
 | 
			
		||||
    if (level == COARSEST_LEVEL) {
 | 
			
		||||
        for (int i = 0; i < 5; i++) {
 | 
			
		||||
            smooth(s, p, rhs, level, comm);
 | 
			
		||||
        }
 | 
			
		||||
        return res;
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    // pre-smoothing
 | 
			
		||||
    for (int i = 0; i < s->presmooth; i++) {
 | 
			
		||||
        smooth(s, p, rhs, level, comm);
 | 
			
		||||
        if (level == FINEST_LEVEL)
 | 
			
		||||
            setBoundaryCondition(s, p, imaxLocal, jmaxLocal, kmaxLocal);
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    res = calculateResidual(s, p, rhs, level, comm);
 | 
			
		||||
 | 
			
		||||
    // restrict
 | 
			
		||||
    restrictMG(s, level, comm);
 | 
			
		||||
 | 
			
		||||
    Comm newcomm;
 | 
			
		||||
    commUpdateDatatypes(s->comm,
 | 
			
		||||
        &newcomm,
 | 
			
		||||
        imaxLocal,
 | 
			
		||||
        jmaxLocal,
 | 
			
		||||
        kmaxLocal);
 | 
			
		||||
 | 
			
		||||
    // MGSolver on residual and error.
 | 
			
		||||
    multiGrid(s, s->e[level + 1], s->r[level + 1], level + 1, &newcomm);
 | 
			
		||||
 | 
			
		||||
    commFreeCommunicator(&newcomm);
 | 
			
		||||
 | 
			
		||||
    // prolongate
 | 
			
		||||
    prolongate(s, level, comm);
 | 
			
		||||
 | 
			
		||||
    // correct p on finer level using residual
 | 
			
		||||
    correct(s, p, level, comm);
 | 
			
		||||
    if (level == FINEST_LEVEL)
 | 
			
		||||
        setBoundaryCondition(s, p, imaxLocal, jmaxLocal, kmaxLocal);
 | 
			
		||||
 | 
			
		||||
    // post-smoothing
 | 
			
		||||
    for (int i = 0; i < s->postsmooth; i++) {
 | 
			
		||||
        smooth(s, p, rhs, level, comm);
 | 
			
		||||
        if (level == FINEST_LEVEL)
 | 
			
		||||
            setBoundaryCondition(s, p, imaxLocal, jmaxLocal, kmaxLocal);
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    return res;
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void initSolver(Solver* s, Discretization* d, Parameter* p)
 | 
			
		||||
{
 | 
			
		||||
    s->eps        = p->eps;
 | 
			
		||||
    s->omega      = p->omg;
 | 
			
		||||
    s->itermax    = p->itermax;
 | 
			
		||||
    s->levels     = p->levels;
 | 
			
		||||
    s->grid       = &d->grid;
 | 
			
		||||
    s->presmooth  = p->presmooth;
 | 
			
		||||
    s->postsmooth = p->postsmooth;
 | 
			
		||||
    s->comm       = &d->comm;
 | 
			
		||||
    s->problem    = p->name;
 | 
			
		||||
 | 
			
		||||
    int imax   = s->grid->imax;
 | 
			
		||||
    int jmax   = s->grid->jmax;
 | 
			
		||||
    int kmax   = s->grid->kmax;
 | 
			
		||||
    int levels = s->levels;
 | 
			
		||||
    if (commIsMaster(s->comm)) printf("Using Multigrid solver with %d levels\n", levels);
 | 
			
		||||
 | 
			
		||||
    s->r = malloc(levels * sizeof(double*));
 | 
			
		||||
    s->e = malloc(levels * sizeof(double*));
 | 
			
		||||
 | 
			
		||||
    size_t size = (imax + 2) * (jmax + 2) * (kmax + 2);
 | 
			
		||||
 | 
			
		||||
    for (int j = 0; j < levels; j++) {
 | 
			
		||||
        s->r[j] = allocate(64, size * sizeof(double));
 | 
			
		||||
        s->e[j] = allocate(64, size * sizeof(double));
 | 
			
		||||
 | 
			
		||||
        for (size_t i = 0; i < size; i++) {
 | 
			
		||||
            s->r[j][i] = 0.0;
 | 
			
		||||
            s->e[j][i] = 0.0;
 | 
			
		||||
        }
 | 
			
		||||
    }
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
double solve(Solver* s, double* p, double* rhs)
 | 
			
		||||
{
 | 
			
		||||
    double res = multiGrid(s, p, rhs, 0, s->comm);
 | 
			
		||||
 | 
			
		||||
#ifdef VERBOSE
 | 
			
		||||
    if (commIsMaster(s->comm)) {
 | 
			
		||||
        printf("Residuum: %.6f\n", res);
 | 
			
		||||
    }
 | 
			
		||||
#endif
 | 
			
		||||
 | 
			
		||||
    return res;
 | 
			
		||||
}
 | 
			
		||||
							
								
								
									
										175
									
								
								BasicSolver/3D-mpi/src/solver-rb.c
									
									
									
									
									
										Normal file
									
								
							
							
						
						
									
										175
									
								
								BasicSolver/3D-mpi/src/solver-rb.c
									
									
									
									
									
										Normal file
									
								
							@@ -0,0 +1,175 @@
 | 
			
		||||
/*
 | 
			
		||||
 * Copyright (C)  NHR@FAU, University Erlangen-Nuremberg.
 | 
			
		||||
 * All rights reserved. This file is part of nusif-solver.
 | 
			
		||||
 * Use of this source code is governed by a MIT style
 | 
			
		||||
 * license that can be found in the LICENSE file.
 | 
			
		||||
 */
 | 
			
		||||
#include <float.h>
 | 
			
		||||
#include <math.h>
 | 
			
		||||
#include <stdio.h>
 | 
			
		||||
#include <stdlib.h>
 | 
			
		||||
#include <string.h>
 | 
			
		||||
 | 
			
		||||
#include "allocate.h"
 | 
			
		||||
#include "comm.h"
 | 
			
		||||
#include "parameter.h"
 | 
			
		||||
#include "solver.h"
 | 
			
		||||
#include "util.h"
 | 
			
		||||
 | 
			
		||||
void initSolver(Solver* s, Discretization* d, Parameter* p)
 | 
			
		||||
{
 | 
			
		||||
    s->eps        = p->eps;
 | 
			
		||||
    s->omega      = p->omg;
 | 
			
		||||
    s->itermax    = p->itermax;
 | 
			
		||||
    s->levels     = p->levels;
 | 
			
		||||
    s->grid       = &d->grid;
 | 
			
		||||
    s->presmooth  = p->presmooth;
 | 
			
		||||
    s->postsmooth = p->postsmooth;
 | 
			
		||||
    s->comm       = &d->comm;
 | 
			
		||||
    s->problem    = p->name;
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
double solve(Solver* s, double* p, double* rhs)
 | 
			
		||||
{
 | 
			
		||||
    int imaxLocal = s->comm->imaxLocal;
 | 
			
		||||
    int jmaxLocal = s->comm->jmaxLocal;
 | 
			
		||||
    int kmaxLocal = s->comm->kmaxLocal;
 | 
			
		||||
 | 
			
		||||
    int imax = s->grid->imax;
 | 
			
		||||
    int jmax = s->grid->jmax;
 | 
			
		||||
    int kmax = s->grid->kmax;
 | 
			
		||||
 | 
			
		||||
    double eps  = s->eps;
 | 
			
		||||
    int itermax = s->itermax;
 | 
			
		||||
    double dx2  = s->grid->dx * s->grid->dx;
 | 
			
		||||
    double dy2  = s->grid->dy * s->grid->dy;
 | 
			
		||||
    double dz2  = s->grid->dz * s->grid->dz;
 | 
			
		||||
    double idx2 = 1.0 / dx2;
 | 
			
		||||
    double idy2 = 1.0 / dy2;
 | 
			
		||||
    double idz2 = 1.0 / dz2;
 | 
			
		||||
 | 
			
		||||
    double factor = s->omega * 0.5 * (dx2 * dy2 * dz2) /
 | 
			
		||||
                    (dy2 * dz2 + dx2 * dz2 + dx2 * dy2);
 | 
			
		||||
    double epssq = eps * eps;
 | 
			
		||||
    int it       = 0;
 | 
			
		||||
    double res   = 1.0;
 | 
			
		||||
    int pass, ksw, jsw, isw;
 | 
			
		||||
 | 
			
		||||
    while ((res >= epssq) && (it < itermax)) {
 | 
			
		||||
        ksw = 1;
 | 
			
		||||
 | 
			
		||||
        for (pass = 0; pass < 2; pass++) {
 | 
			
		||||
            jsw = ksw;
 | 
			
		||||
            commExchange(s->comm, p);
 | 
			
		||||
 | 
			
		||||
            for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
                isw = jsw;
 | 
			
		||||
                for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                    for (int i = isw; i < imaxLocal + 1; i += 2) {
 | 
			
		||||
 | 
			
		||||
                        double r =
 | 
			
		||||
                            RHS(i, j, k) -
 | 
			
		||||
                            ((P(i + 1, j, k) - 2.0 * P(i, j, k) + P(i - 1, j, k)) * idx2 +
 | 
			
		||||
                                (P(i, j + 1, k) - 2.0 * P(i, j, k) + P(i, j - 1, k)) *
 | 
			
		||||
                                    idy2 +
 | 
			
		||||
                                (P(i, j, k + 1) - 2.0 * P(i, j, k) + P(i, j, k - 1)) *
 | 
			
		||||
                                    idz2);
 | 
			
		||||
 | 
			
		||||
                        P(i, j, k) -= (factor * r);
 | 
			
		||||
                        res += (r * r);
 | 
			
		||||
                    }
 | 
			
		||||
                    isw = 3 - isw;
 | 
			
		||||
                }
 | 
			
		||||
                jsw = 3 - jsw;
 | 
			
		||||
            }
 | 
			
		||||
            ksw = 3 - ksw;
 | 
			
		||||
        }
 | 
			
		||||
#ifdef _MPI
 | 
			
		||||
        if (commIsBoundary(s->comm, FRONT)) {
 | 
			
		||||
            for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                    P(i, j, 0) = P(i, j, 1);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
 | 
			
		||||
        if (commIsBoundary(s->comm, BACK)) {
 | 
			
		||||
            for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                    P(i, j, kmaxLocal + 1) = P(i, j, kmaxLocal);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
 | 
			
		||||
        if (commIsBoundary(s->comm, BOTTOM)) {
 | 
			
		||||
            for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
                for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                    P(i, 0, k) = P(i, 1, k);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
 | 
			
		||||
        if (commIsBoundary(s->comm, TOP)) {
 | 
			
		||||
            for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
                for (int i = 1; i < imaxLocal + 1; i++) {
 | 
			
		||||
                    P(i, jmaxLocal + 1, k) = P(i, jmaxLocal, k);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
 | 
			
		||||
        if (commIsBoundary(s->comm, LEFT)) {
 | 
			
		||||
            for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
                for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                    P(0, j, k) = P(1, j, k);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
 | 
			
		||||
        if (commIsBoundary(s->comm, RIGHT)) {
 | 
			
		||||
            for (int k = 1; k < kmaxLocal + 1; k++) {
 | 
			
		||||
                for (int j = 1; j < jmaxLocal + 1; j++) {
 | 
			
		||||
                    P(imaxLocal + 1, j, k) = P(imaxLocal, j, k);
 | 
			
		||||
                }
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
#else
 | 
			
		||||
        for (int j = 1; j < jmax + 1; j++) {
 | 
			
		||||
            for (int i = 1; i < imax + 1; i++) {
 | 
			
		||||
                P(i, j, 0)        = P(i, j, 1);
 | 
			
		||||
                P(i, j, kmax + 1) = P(i, j, kmax);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
 | 
			
		||||
        for (int k = 1; k < kmax + 1; k++) {
 | 
			
		||||
            for (int i = 1; i < imax + 1; i++) {
 | 
			
		||||
                P(i, 0, k)        = P(i, 1, k);
 | 
			
		||||
                P(i, jmax + 1, k) = P(i, jmax, k);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
 | 
			
		||||
        for (int k = 1; k < kmax + 1; k++) {
 | 
			
		||||
            for (int j = 1; j < jmax + 1; j++) {
 | 
			
		||||
                P(0, j, k)        = P(1, j, k);
 | 
			
		||||
                P(imax + 1, j, k) = P(imax, j, k);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
#endif
 | 
			
		||||
        commReduction(&res, SUM);
 | 
			
		||||
        res = res / (double)(imax * jmax * kmax);
 | 
			
		||||
#ifdef DEBUG
 | 
			
		||||
        if (commIsMaster(&s->comm)) {
 | 
			
		||||
            printf("%d Residuum: %e\n", it, res);
 | 
			
		||||
        }
 | 
			
		||||
#endif
 | 
			
		||||
        commExchange(s->comm, p);
 | 
			
		||||
        it++;
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
#ifdef VERBOSE
 | 
			
		||||
    if (commIsMaster(s->comm)) {
 | 
			
		||||
        printf("Solver took %d iterations to reach %f\n", it, sqrt(res));
 | 
			
		||||
    }
 | 
			
		||||
#endif
 | 
			
		||||
 | 
			
		||||
return res;
 | 
			
		||||
}
 | 
			
		||||
@@ -9,12 +9,11 @@
 | 
			
		||||
#include "comm.h"
 | 
			
		||||
#include "grid.h"
 | 
			
		||||
#include "parameter.h"
 | 
			
		||||
 | 
			
		||||
enum BC { NOSLIP = 1, SLIP, OUTFLOW, PERIODIC };
 | 
			
		||||
#include "discretization.h"
 | 
			
		||||
 | 
			
		||||
typedef struct {
 | 
			
		||||
    /* geometry and grid information */
 | 
			
		||||
    Grid grid;
 | 
			
		||||
    Grid* grid;
 | 
			
		||||
    /* arrays */
 | 
			
		||||
    double *p, *rhs;
 | 
			
		||||
    double *f, *g, *h;
 | 
			
		||||
@@ -30,16 +29,11 @@ typedef struct {
 | 
			
		||||
    char* problem;
 | 
			
		||||
    int bcLeft, bcRight, bcBottom, bcTop, bcFront, bcBack;
 | 
			
		||||
    /* communication */
 | 
			
		||||
    Comm comm;
 | 
			
		||||
    double **r, **e;
 | 
			
		||||
    int levels, presmooth, postsmooth;
 | 
			
		||||
    Comm* comm;
 | 
			
		||||
} Solver;
 | 
			
		||||
 | 
			
		||||
extern void initSolver(Solver*, Parameter*);
 | 
			
		||||
extern void computeRHS(Solver*);
 | 
			
		||||
extern void solve(Solver*);
 | 
			
		||||
extern void normalizePressure(Solver*);
 | 
			
		||||
extern void computeTimestep(Solver*);
 | 
			
		||||
extern void setBoundaryConditions(Solver*);
 | 
			
		||||
extern void setSpecialBoundaryCondition(Solver*);
 | 
			
		||||
extern void computeFG(Solver*);
 | 
			
		||||
extern void adaptUV(Solver*);
 | 
			
		||||
extern double solve(Solver* , double* , double* );
 | 
			
		||||
extern void initSolver(Solver*, Discretization*, Parameter*);
 | 
			
		||||
#endif
 | 
			
		||||
 
 | 
			
		||||
@@ -14,10 +14,10 @@
 | 
			
		||||
 | 
			
		||||
void testInit(Solver* s)
 | 
			
		||||
{
 | 
			
		||||
    int imaxLocal = s->comm.imaxLocal;
 | 
			
		||||
    int jmaxLocal = s->comm.jmaxLocal;
 | 
			
		||||
    int kmaxLocal = s->comm.kmaxLocal;
 | 
			
		||||
    int myrank    = s->comm.rank;
 | 
			
		||||
    int imaxLocal = s->comm->imaxLocal;
 | 
			
		||||
    int jmaxLocal = s->comm->jmaxLocal;
 | 
			
		||||
    int kmaxLocal = s->comm->kmaxLocal;
 | 
			
		||||
    int myrank    = s->comm->rank;
 | 
			
		||||
    double* p     = s->p;
 | 
			
		||||
    double* f     = s->f;
 | 
			
		||||
    double* g     = s->g;
 | 
			
		||||
@@ -76,11 +76,11 @@ static char* direction2String(Direction dir)
 | 
			
		||||
 | 
			
		||||
static void printPlane(Solver* s, double* a, int ymax, int xmax, Direction dir)
 | 
			
		||||
{
 | 
			
		||||
    int imaxLocal = s->comm.imaxLocal;
 | 
			
		||||
    int jmaxLocal = s->comm.jmaxLocal;
 | 
			
		||||
    int kmaxLocal = s->comm.kmaxLocal;
 | 
			
		||||
    int imaxLocal = s->comm->imaxLocal;
 | 
			
		||||
    int jmaxLocal = s->comm->jmaxLocal;
 | 
			
		||||
    int kmaxLocal = s->comm->kmaxLocal;
 | 
			
		||||
    char filename[50];
 | 
			
		||||
    snprintf(filename, 50, "halo-%s-r%d.txt", direction2String(dir), s->comm.rank);
 | 
			
		||||
    snprintf(filename, 50, "halo-%s-r%d.txt", direction2String(dir), s->comm->rank);
 | 
			
		||||
    FILE* fh = fopen(filename, "w");
 | 
			
		||||
 | 
			
		||||
    for (int y = 0; y < ymax; y++) {
 | 
			
		||||
@@ -116,9 +116,9 @@ static void printPlane(Solver* s, double* a, int ymax, int xmax, Direction dir)
 | 
			
		||||
 | 
			
		||||
void testPrintHalo(Solver* s, double* a)
 | 
			
		||||
{
 | 
			
		||||
    int imaxLocal = s->comm.imaxLocal;
 | 
			
		||||
    int jmaxLocal = s->comm.jmaxLocal;
 | 
			
		||||
    int kmaxLocal = s->comm.kmaxLocal;
 | 
			
		||||
    int imaxLocal = s->comm->imaxLocal;
 | 
			
		||||
    int jmaxLocal = s->comm->jmaxLocal;
 | 
			
		||||
    int kmaxLocal = s->comm->kmaxLocal;
 | 
			
		||||
 | 
			
		||||
    printPlane(s, a, kmaxLocal + 2, imaxLocal + 2, BOTTOM);
 | 
			
		||||
    printPlane(s, a, kmaxLocal + 2, imaxLocal + 2, TOP);
 | 
			
		||||
 
 | 
			
		||||
@@ -19,4 +19,13 @@
 | 
			
		||||
#define ABS(a) ((a) >= 0 ? (a) : -(a))
 | 
			
		||||
#endif
 | 
			
		||||
 | 
			
		||||
#define P(i, j, k)   p[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define F(i, j, k)   f[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define G(i, j, k)   g[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define H(i, j, k)   h[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define U(i, j, k)   u[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define V(i, j, k)   v[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define W(i, j, k)   w[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
#define RHS(i, j, k) rhs[(k) * (imaxLocal + 2) * (jmaxLocal + 2) + (j) * (imaxLocal + 2) + (i)]
 | 
			
		||||
 | 
			
		||||
#endif // __UTIL_H_
 | 
			
		||||
 
 | 
			
		||||
@@ -46,18 +46,18 @@ static void writeHeader(VtkOptions* o)
 | 
			
		||||
    cursor += sprintf(cursor, "DATASET STRUCTURED_POINTS\n");
 | 
			
		||||
    cursor += sprintf(cursor,
 | 
			
		||||
        "DIMENSIONS %d %d %d\n",
 | 
			
		||||
        o->grid.imax,
 | 
			
		||||
        o->grid.jmax,
 | 
			
		||||
        o->grid.kmax);
 | 
			
		||||
        o->grid->imax,
 | 
			
		||||
        o->grid->jmax,
 | 
			
		||||
        o->grid->kmax);
 | 
			
		||||
    cursor += sprintf(cursor,
 | 
			
		||||
        "ORIGIN %f %f %f\n",
 | 
			
		||||
        o->grid.dx * 0.5,
 | 
			
		||||
        o->grid.dy * 0.5,
 | 
			
		||||
        o->grid.dz * 0.5);
 | 
			
		||||
    cursor += sprintf(cursor, "SPACING %f %f %f\n", o->grid.dx, o->grid.dy, o->grid.dz);
 | 
			
		||||
        o->grid->dx * 0.5,
 | 
			
		||||
        o->grid->dy * 0.5,
 | 
			
		||||
        o->grid->dz * 0.5);
 | 
			
		||||
    cursor += sprintf(cursor, "SPACING %f %f %f\n", o->grid->dx, o->grid->dy, o->grid->dz);
 | 
			
		||||
    cursor += sprintf(cursor,
 | 
			
		||||
        "POINT_DATA %d\n",
 | 
			
		||||
        o->grid.imax * o->grid.jmax * o->grid.kmax);
 | 
			
		||||
        o->grid->imax * o->grid->jmax * o->grid->kmax);
 | 
			
		||||
 | 
			
		||||
    if (commIsMaster(&o->comm)) {
 | 
			
		||||
        MPI_File_write(o->fh, header, (int)strlen(header), MPI_CHAR, MPI_STATUS_IGNORE);
 | 
			
		||||
@@ -67,7 +67,6 @@ static void writeHeader(VtkOptions* o)
 | 
			
		||||
void vtkOpen(VtkOptions* o, char* problem)
 | 
			
		||||
{
 | 
			
		||||
    char filename[50];
 | 
			
		||||
 | 
			
		||||
    snprintf(filename, 50, "%s-p%d.vtk", problem, o->comm.size);
 | 
			
		||||
    MPI_File_open(o->comm.comm,
 | 
			
		||||
        filename,
 | 
			
		||||
@@ -98,7 +97,7 @@ void vtkScalar(VtkOptions* o, char* name, double* s)
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    int offsets[NDIMS];
 | 
			
		||||
    commGetOffsets(&o->comm, offsets, o->grid.kmax, o->grid.jmax, o->grid.imax);
 | 
			
		||||
    commGetOffsets(&o->comm, offsets, o->grid->kmax, o->grid->jmax, o->grid->imax);
 | 
			
		||||
 | 
			
		||||
    // set global view in file
 | 
			
		||||
    MPI_Offset disp;
 | 
			
		||||
@@ -108,7 +107,7 @@ void vtkScalar(VtkOptions* o, char* name, double* s)
 | 
			
		||||
    MPI_File_get_size(o->fh, &disp);
 | 
			
		||||
 | 
			
		||||
    MPI_Type_create_subarray(NDIMS,
 | 
			
		||||
        (int[NDIMS]) { o->grid.kmax, o->grid.jmax, o->grid.imax },
 | 
			
		||||
        (int[NDIMS]) { o->grid->kmax, o->grid->jmax, o->grid->imax },
 | 
			
		||||
        (int[NDIMS]) { o->comm.kmaxLocal, o->comm.jmaxLocal, o->comm.imaxLocal },
 | 
			
		||||
        offsets,
 | 
			
		||||
        MPI_ORDER_C,
 | 
			
		||||
@@ -177,7 +176,7 @@ void vtkVector(VtkOptions* o, char* name, VtkVector vec)
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    int offsets[NDIMS];
 | 
			
		||||
    commGetOffsets(&o->comm, offsets, o->grid.kmax, o->grid.jmax, o->grid.imax);
 | 
			
		||||
    commGetOffsets(&o->comm, offsets, o->grid->kmax, o->grid->jmax, o->grid->imax);
 | 
			
		||||
 | 
			
		||||
    // set global view in file
 | 
			
		||||
    MPI_Offset disp;
 | 
			
		||||
@@ -190,7 +189,7 @@ void vtkVector(VtkOptions* o, char* name, VtkVector vec)
 | 
			
		||||
    MPI_Type_commit(&vectorType);
 | 
			
		||||
 | 
			
		||||
    MPI_Type_create_subarray(NDIMS,
 | 
			
		||||
        (int[NDIMS]) { o->grid.kmax, o->grid.jmax, o->grid.imax },
 | 
			
		||||
        (int[NDIMS]) { o->grid->kmax, o->grid->jmax, o->grid->imax },
 | 
			
		||||
        (int[NDIMS]) { kmaxLocal, jmaxLocal, imaxLocal },
 | 
			
		||||
        offsets,
 | 
			
		||||
        MPI_ORDER_C,
 | 
			
		||||
 
 | 
			
		||||
@@ -41,14 +41,14 @@ static void writeHeader(VtkOptions* o)
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    fprintf(o->fh, "DATASET STRUCTURED_POINTS\n");
 | 
			
		||||
    fprintf(o->fh, "DIMENSIONS %d %d %d\n", o->grid.imax, o->grid.jmax, o->grid.kmax);
 | 
			
		||||
    fprintf(o->fh, "DIMENSIONS %d %d %d\n", o->grid->imax, o->grid->jmax, o->grid->kmax);
 | 
			
		||||
    fprintf(o->fh,
 | 
			
		||||
        "ORIGIN %f %f %f\n",
 | 
			
		||||
        o->grid.dx * 0.5,
 | 
			
		||||
        o->grid.dy * 0.5,
 | 
			
		||||
        o->grid.dz * 0.5);
 | 
			
		||||
    fprintf(o->fh, "SPACING %f %f %f\n", o->grid.dx, o->grid.dy, o->grid.dz);
 | 
			
		||||
    fprintf(o->fh, "POINT_DATA %d\n", o->grid.imax * o->grid.jmax * o->grid.kmax);
 | 
			
		||||
        o->grid->dx * 0.5,
 | 
			
		||||
        o->grid->dy * 0.5,
 | 
			
		||||
        o->grid->dz * 0.5);
 | 
			
		||||
    fprintf(o->fh, "SPACING %f %f %f\n", o->grid->dx, o->grid->dy, o->grid->dz);
 | 
			
		||||
    fprintf(o->fh, "POINT_DATA %d\n", o->grid->imax * o->grid->jmax * o->grid->kmax);
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void vtkOpen(VtkOptions* o, char* problem)
 | 
			
		||||
@@ -64,9 +64,9 @@ void vtkOpen(VtkOptions* o, char* problem)
 | 
			
		||||
 | 
			
		||||
static void writeScalar(VtkOptions* o, double* s)
 | 
			
		||||
{
 | 
			
		||||
    int imax = o->grid.imax;
 | 
			
		||||
    int jmax = o->grid.jmax;
 | 
			
		||||
    int kmax = o->grid.kmax;
 | 
			
		||||
    int imax = o->grid->imax;
 | 
			
		||||
    int jmax = o->grid->jmax;
 | 
			
		||||
    int kmax = o->grid->kmax;
 | 
			
		||||
 | 
			
		||||
    for (int k = 0; k < kmax; k++) {
 | 
			
		||||
        for (int j = 0; j < jmax; j++) {
 | 
			
		||||
@@ -105,9 +105,9 @@ void vtkScalar(VtkOptions* o, char* name, double* s)
 | 
			
		||||
 | 
			
		||||
static void writeVector(VtkOptions* o, VtkVector vec)
 | 
			
		||||
{
 | 
			
		||||
    int imax = o->grid.imax;
 | 
			
		||||
    int jmax = o->grid.jmax;
 | 
			
		||||
    int kmax = o->grid.kmax;
 | 
			
		||||
    int imax = o->grid->imax;
 | 
			
		||||
    int jmax = o->grid->jmax;
 | 
			
		||||
    int kmax = o->grid->kmax;
 | 
			
		||||
 | 
			
		||||
    for (int k = 0; k < kmax; k++) {
 | 
			
		||||
        for (int j = 0; j < jmax; j++) {
 | 
			
		||||
 
 | 
			
		||||
@@ -14,7 +14,7 @@
 | 
			
		||||
typedef enum VtkFormat { ASCII = 0, BINARY } VtkFormat;
 | 
			
		||||
 | 
			
		||||
typedef struct VtkOptions {
 | 
			
		||||
    Grid grid;
 | 
			
		||||
    Grid* grid;
 | 
			
		||||
#ifdef _VTK_WRITER_MPI
 | 
			
		||||
    MPI_File fh;
 | 
			
		||||
#else
 | 
			
		||||
 
 | 
			
		||||
@@ -40,6 +40,18 @@ $(BUILD_DIR)/%.s:  %.c
 | 
			
		||||
 | 
			
		||||
.PHONY: clean distclean tags info asm format
 | 
			
		||||
 | 
			
		||||
vis:
 | 
			
		||||
	$(info ===>  GENERATE VISUALIZATION)
 | 
			
		||||
	@gnuplot -e "filename='residual.dat'" ./residual.plot
 | 
			
		||||
 | 
			
		||||
vis_clean:
 | 
			
		||||
	$(info ===>  CLEAN VISUALIZATION)
 | 
			
		||||
	@rm -f *.dat
 | 
			
		||||
	@rm -f *.vtk
 | 
			
		||||
	@rm -f *.png
 | 
			
		||||
 | 
			
		||||
clean: vis_clean
 | 
			
		||||
 | 
			
		||||
clean:
 | 
			
		||||
	$(info ===>  CLEAN)
 | 
			
		||||
	@rm -rf $(BUILD_DIR)
 | 
			
		||||
 
 | 
			
		||||
@@ -38,10 +38,17 @@ kmax          50	   # number of interior cells in z-direction
 | 
			
		||||
# Time Data:
 | 
			
		||||
# ---------
 | 
			
		||||
 | 
			
		||||
te       100.0   # final time
 | 
			
		||||
te       60.0   # final time
 | 
			
		||||
dt       0.02    # time stepsize
 | 
			
		||||
tau      0.5     # safety factor for time stepsize control (<0 constant delt)
 | 
			
		||||
 | 
			
		||||
# Multigrid data:
 | 
			
		||||
# ---------
 | 
			
		||||
 | 
			
		||||
levels        3         # Multigrid levels
 | 
			
		||||
presmooth     5         # Pre-smoothning iterations
 | 
			
		||||
postsmooth    5         # Post-smoothning iterations
 | 
			
		||||
 | 
			
		||||
# Pressure Iteration Data:
 | 
			
		||||
# -----------------------
 | 
			
		||||
 | 
			
		||||
 
 | 
			
		||||
@@ -1,12 +1,12 @@
 | 
			
		||||
# Supported: GCC, CLANG, ICC
 | 
			
		||||
TAG ?= CLANG
 | 
			
		||||
TAG ?= ICC
 | 
			
		||||
ENABLE_OPENMP ?= false
 | 
			
		||||
# Supported: sor, mg
 | 
			
		||||
# Supported: rb, mg
 | 
			
		||||
SOLVER ?= mg
 | 
			
		||||
# Run in debug settings
 | 
			
		||||
DEBUG ?= false
 | 
			
		||||
 | 
			
		||||
#Feature options
 | 
			
		||||
OPTIONS +=  -DARRAY_ALIGNMENT=64
 | 
			
		||||
#OPTIONS +=  -DVERBOSE
 | 
			
		||||
OPTIONS +=  -DVERBOSE
 | 
			
		||||
#OPTIONS +=  -DDEBUG
 | 
			
		||||
 
 | 
			
		||||
@@ -38,10 +38,17 @@ kmax       128		# number of interior cells in z-direction
 | 
			
		||||
# Time Data:
 | 
			
		||||
# ---------
 | 
			
		||||
 | 
			
		||||
te      2.0		# final time
 | 
			
		||||
te       10.0		# final time
 | 
			
		||||
dt       0.02	    # time stepsize
 | 
			
		||||
tau      0.5		# safety factor for time stepsize control (<0 constant delt)
 | 
			
		||||
 | 
			
		||||
# Multigrid data:
 | 
			
		||||
# ---------
 | 
			
		||||
 | 
			
		||||
levels        3         # Multigrid levels
 | 
			
		||||
presmooth     20         # Pre-smoothning iterations
 | 
			
		||||
postsmooth    5         # Post-smoothning iterations
 | 
			
		||||
 | 
			
		||||
# Pressure Iteration Data:
 | 
			
		||||
# -----------------------
 | 
			
		||||
 | 
			
		||||
@@ -50,5 +57,4 @@ eps      0.001		# stopping tolerance for pressure iteration
 | 
			
		||||
rho      0.5
 | 
			
		||||
omg      1.7		# relaxation parameter for SOR iteration
 | 
			
		||||
gamma    0.9		# upwind differencing factor gamma
 | 
			
		||||
levels   5         # Multigrid levels
 | 
			
		||||
#===============================================================================
 | 
			
		||||
 
 | 
			
		||||
							
								
								
									
										9
									
								
								BasicSolver/3D-seq/residual.plot
									
									
									
									
									
										Normal file
									
								
							
							
						
						
									
										9
									
								
								BasicSolver/3D-seq/residual.plot
									
									
									
									
									
										Normal file
									
								
							@@ -0,0 +1,9 @@
 | 
			
		||||
set terminal png size 1800,768 enhanced font ,12
 | 
			
		||||
set output 'residual.png'
 | 
			
		||||
set datafile separator whitespace
 | 
			
		||||
set xlabel "Timestep"
 | 
			
		||||
set ylabel "Residual"
 | 
			
		||||
 | 
			
		||||
set logscale y 2
 | 
			
		||||
 | 
			
		||||
plot 'residual.dat' using 1:2 title "Residual"
 | 
			
		||||
@@ -106,7 +106,7 @@ void initDiscretization(Discretization* d, Parameter* p)
 | 
			
		||||
    d->dtBound       = 0.5 * d->re * 1.0 / invSqrSum;
 | 
			
		||||
 | 
			
		||||
#ifdef VERBOSE
 | 
			
		||||
    printConfig(s);
 | 
			
		||||
    printConfig(d);
 | 
			
		||||
#endif /* VERBOSE */
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
 
 | 
			
		||||
@@ -73,6 +73,9 @@ int main(int argc, char** argv)
 | 
			
		||||
    Solver s;
 | 
			
		||||
    initParameter(&p);
 | 
			
		||||
 | 
			
		||||
    FILE* fp;
 | 
			
		||||
    fp = initResidualWriter();
 | 
			
		||||
 | 
			
		||||
    if (argc != 2) {
 | 
			
		||||
        printf("Usage: %s <configFile>\n", argv[0]);
 | 
			
		||||
        exit(EXIT_SUCCESS);
 | 
			
		||||
@@ -90,6 +93,7 @@ int main(int argc, char** argv)
 | 
			
		||||
    double te  = d.te;
 | 
			
		||||
    double t   = 0.0;
 | 
			
		||||
    int nt     = 0;
 | 
			
		||||
    double res = 0.0;
 | 
			
		||||
 | 
			
		||||
    timeStart = getTimeStamp();
 | 
			
		||||
    while (t <= te) {
 | 
			
		||||
@@ -99,13 +103,16 @@ int main(int argc, char** argv)
 | 
			
		||||
        computeFG(&d);
 | 
			
		||||
        computeRHS(&d);
 | 
			
		||||
        if (nt % 100 == 0) normalizePressure(&d);
 | 
			
		||||
        solve(&s, d.p, d.rhs);
 | 
			
		||||
        res = solve(&s, d.p, d.rhs);
 | 
			
		||||
        adaptUV(&d);
 | 
			
		||||
 | 
			
		||||
        writeResidual(fp, t, res);
 | 
			
		||||
 | 
			
		||||
        t += d.dt;
 | 
			
		||||
        nt++;
 | 
			
		||||
 | 
			
		||||
#ifdef VERBOSE
 | 
			
		||||
        printf("TIME %f , TIMESTEP %f\n", t, solver.dt);
 | 
			
		||||
        printf("TIME %f , TIMESTEP %f\n", t, d.dt);
 | 
			
		||||
#else
 | 
			
		||||
        printProgress(t);
 | 
			
		||||
#endif
 | 
			
		||||
@@ -123,7 +130,8 @@ int main(int argc, char** argv)
 | 
			
		||||
    ug = allocate(64, bytesize);
 | 
			
		||||
    vg = allocate(64, bytesize);
 | 
			
		||||
    wg = allocate(64, bytesize);
 | 
			
		||||
 | 
			
		||||
    
 | 
			
		||||
    fclose(fp);
 | 
			
		||||
    createBulkArrays(&d, pg, ug, vg, wg);
 | 
			
		||||
    VtkOptions opts = { .grid = d.grid };
 | 
			
		||||
    vtkOpen(&opts, d.problem);
 | 
			
		||||
 
 | 
			
		||||
@@ -14,19 +14,21 @@
 | 
			
		||||
 | 
			
		||||
void initParameter(Parameter* param)
 | 
			
		||||
{
 | 
			
		||||
    param->xlength = 1.0;
 | 
			
		||||
    param->ylength = 1.0;
 | 
			
		||||
    param->zlength = 1.0;
 | 
			
		||||
    param->imax    = 100;
 | 
			
		||||
    param->jmax    = 100;
 | 
			
		||||
    param->kmax    = 100;
 | 
			
		||||
    param->itermax = 1000;
 | 
			
		||||
    param->eps     = 0.0001;
 | 
			
		||||
    param->omg     = 1.7;
 | 
			
		||||
    param->re      = 100.0;
 | 
			
		||||
    param->gamma   = 0.9;
 | 
			
		||||
    param->tau     = 0.5;
 | 
			
		||||
    param->levels  = 5;
 | 
			
		||||
    param->xlength    = 1.0;
 | 
			
		||||
    param->ylength    = 1.0;
 | 
			
		||||
    param->zlength    = 1.0;
 | 
			
		||||
    param->imax       = 100;
 | 
			
		||||
    param->jmax       = 100;
 | 
			
		||||
    param->kmax       = 100;
 | 
			
		||||
    param->itermax    = 1000;
 | 
			
		||||
    param->eps        = 0.0001;
 | 
			
		||||
    param->omg        = 1.7;
 | 
			
		||||
    param->re         = 100.0;
 | 
			
		||||
    param->gamma      = 0.9;
 | 
			
		||||
    param->tau        = 0.5;
 | 
			
		||||
    param->levels     = 5;
 | 
			
		||||
    param->presmooth  = 5;
 | 
			
		||||
    param->postsmooth = 5;
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void readParameter(Parameter* param, const char* filename)
 | 
			
		||||
 
 | 
			
		||||
@@ -18,6 +18,7 @@ typedef struct {
 | 
			
		||||
    char* name;
 | 
			
		||||
    int bcLeft, bcRight, bcBottom, bcTop, bcFront, bcBack;
 | 
			
		||||
    double u_init, v_init, w_init, p_init;
 | 
			
		||||
    int presmooth, postsmooth;
 | 
			
		||||
} Parameter;
 | 
			
		||||
 | 
			
		||||
void initParameter(Parameter*);
 | 
			
		||||
 
 | 
			
		||||
@@ -49,3 +49,22 @@ void stopProgress()
 | 
			
		||||
    printf("\n");
 | 
			
		||||
    fflush(stdout);
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
FILE* initResidualWriter()
 | 
			
		||||
{
 | 
			
		||||
    FILE* fp;
 | 
			
		||||
    fp = fopen("residual.dat", "w");
 | 
			
		||||
 | 
			
		||||
    if (fp == NULL) {
 | 
			
		||||
        printf("Error!\n");
 | 
			
		||||
        exit(EXIT_FAILURE);
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    return fp;
 | 
			
		||||
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void writeResidual(FILE* fp, double ts, double res)
 | 
			
		||||
{
 | 
			
		||||
    fprintf(fp, "%f, %f\n", ts, res);
 | 
			
		||||
}
 | 
			
		||||
@@ -10,5 +10,6 @@
 | 
			
		||||
extern void initProgress(double);
 | 
			
		||||
extern void printProgress(double);
 | 
			
		||||
extern void stopProgress(void);
 | 
			
		||||
 | 
			
		||||
extern FILE* initResidualWriter(void);
 | 
			
		||||
extern void writeResidual(FILE*, double, double);
 | 
			
		||||
#endif
 | 
			
		||||
 
 | 
			
		||||
@@ -141,13 +141,62 @@ static double smooth(
 | 
			
		||||
            for (int j = 1; j < jmax + 1; j++) {
 | 
			
		||||
                for (int i = isw; i < imax + 1; i += 2) {
 | 
			
		||||
 | 
			
		||||
                    R(i, j, k) =
 | 
			
		||||
                        RHS(i, j, k) -
 | 
			
		||||
                        ((P(i + 1, j, k) - 2.0 * P(i, j, k) + P(i - 1, j, k)) * idx2 +
 | 
			
		||||
                            (P(i, j + 1, k) - 2.0 * P(i, j, k) + P(i, j - 1, k)) * idy2 +
 | 
			
		||||
                            (P(i, j, k + 1) - 2.0 * P(i, j, k) + P(i, j, k - 1)) * idz2);
 | 
			
		||||
                    P(i, j, k) -=
 | 
			
		||||
                        factor *
 | 
			
		||||
                        (RHS(i, j, k) -
 | 
			
		||||
                            ((P(i + 1, j, k) - 2.0 * P(i, j, k) + P(i - 1, j, k)) * idx2 +
 | 
			
		||||
                                (P(i, j + 1, k) - 2.0 * P(i, j, k) + P(i, j - 1, k)) *
 | 
			
		||||
                                    idy2 +
 | 
			
		||||
                                (P(i, j, k + 1) - 2.0 * P(i, j, k) + P(i, j, k - 1)) *
 | 
			
		||||
                                    idz2));
 | 
			
		||||
                }
 | 
			
		||||
                isw = 3 - isw;
 | 
			
		||||
            }
 | 
			
		||||
            jsw = 3 - jsw;
 | 
			
		||||
        }
 | 
			
		||||
        ksw = 3 - ksw;
 | 
			
		||||
    }
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
static double calculateResidual(
 | 
			
		||||
    Solver* s, double* p, double* rhs, int level, int imax, int jmax, int kmax)
 | 
			
		||||
{
 | 
			
		||||
    double eps    = s->eps;
 | 
			
		||||
    int itermax   = s->itermax;
 | 
			
		||||
    double dx2    = s->grid->dx * s->grid->dx;
 | 
			
		||||
    double dy2    = s->grid->dy * s->grid->dy;
 | 
			
		||||
    double dz2    = s->grid->dz * s->grid->dz;
 | 
			
		||||
    double idx2   = 1.0 / dx2;
 | 
			
		||||
    double idy2   = 1.0 / dy2;
 | 
			
		||||
    double idz2   = 1.0 / dz2;
 | 
			
		||||
    double factor = s->omega * 0.5 * (dx2 * dy2 * dz2) /
 | 
			
		||||
                    (dy2 * dz2 + dx2 * dz2 + dx2 * dy2);
 | 
			
		||||
    double* r    = s->r[level];
 | 
			
		||||
    double epssq = eps * eps;
 | 
			
		||||
    int it       = 0;
 | 
			
		||||
    int pass, ksw, jsw, isw;
 | 
			
		||||
    double res = 1.0;
 | 
			
		||||
 | 
			
		||||
    ksw = 1;
 | 
			
		||||
 | 
			
		||||
    for (pass = 0; pass < 2; pass++) {
 | 
			
		||||
        jsw = ksw;
 | 
			
		||||
 | 
			
		||||
        for (int k = 1; k < kmax + 1; k++) {
 | 
			
		||||
            isw = jsw;
 | 
			
		||||
            for (int j = 1; j < jmax + 1; j++) {
 | 
			
		||||
                for (int i = isw; i < imax + 1; i += 2) {
 | 
			
		||||
 | 
			
		||||
                    R(i,
 | 
			
		||||
                        j,
 | 
			
		||||
                        k) = (RHS(i, j, k) -
 | 
			
		||||
                              ((P(i + 1, j, k) - 2.0 * P(i, j, k) + P(i - 1, j, k)) *
 | 
			
		||||
                                      idx2 +
 | 
			
		||||
                                  (P(i, j + 1, k) - 2.0 * P(i, j, k) + P(i, j - 1, k)) *
 | 
			
		||||
                                      idy2 +
 | 
			
		||||
                                  (P(i, j, k + 1) - 2.0 * P(i, j, k) + P(i, j, k - 1)) *
 | 
			
		||||
                                      idz2));
 | 
			
		||||
 | 
			
		||||
                    P(i, j, k) -= (factor * R(i, j, k));
 | 
			
		||||
                    res += (R(i, j, k) * R(i, j, k));
 | 
			
		||||
                }
 | 
			
		||||
                isw = 3 - isw;
 | 
			
		||||
@@ -167,7 +216,7 @@ static double multiGrid(
 | 
			
		||||
{
 | 
			
		||||
    double res = 0.0;
 | 
			
		||||
 | 
			
		||||
    // coarsest level TODO: Use direct solver?
 | 
			
		||||
    // coarsest level
 | 
			
		||||
    if (level == COARSEST_LEVEL) {
 | 
			
		||||
        for (int i = 0; i < 5; i++) {
 | 
			
		||||
            smooth(s, p, rhs, level, imax, jmax, kmax);
 | 
			
		||||
@@ -175,17 +224,18 @@ static double multiGrid(
 | 
			
		||||
        return res;
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    // pre-smoothing TODO: Make smoothing steps configurable?
 | 
			
		||||
    for (int i = 0; i < 5; i++) {
 | 
			
		||||
    // pre-smoothing
 | 
			
		||||
    for (int i = 0; i < s->presmooth; i++) {
 | 
			
		||||
        smooth(s, p, rhs, level, imax, jmax, kmax);
 | 
			
		||||
        if (level == FINEST_LEVEL) setBoundaryCondition(p, imax, jmax, kmax);
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
    res = calculateResidual(s, p, rhs, level, imax, jmax, kmax);
 | 
			
		||||
 | 
			
		||||
    // restrict
 | 
			
		||||
    restrictMG(s, level, imax, jmax, kmax);
 | 
			
		||||
 | 
			
		||||
    // MGSolver on residual and error.
 | 
			
		||||
    // TODO: What if there is a rest?
 | 
			
		||||
    multiGrid(s,
 | 
			
		||||
        s->e[level + 1],
 | 
			
		||||
        s->r[level + 1],
 | 
			
		||||
@@ -202,8 +252,8 @@ static double multiGrid(
 | 
			
		||||
    if (level == FINEST_LEVEL) setBoundaryCondition(p, imax, jmax, kmax);
 | 
			
		||||
 | 
			
		||||
    // post-smoothing
 | 
			
		||||
    for (int i = 0; i < 5; i++) {
 | 
			
		||||
        res = smooth(s, p, rhs, level, imax, jmax, kmax);
 | 
			
		||||
    for (int i = 0; i < s->postsmooth; i++) {
 | 
			
		||||
        smooth(s, p, rhs, level, imax, jmax, kmax);
 | 
			
		||||
        if (level == FINEST_LEVEL) setBoundaryCondition(p, imax, jmax, kmax);
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
@@ -212,11 +262,13 @@ static double multiGrid(
 | 
			
		||||
 | 
			
		||||
void initSolver(Solver* s, Discretization* d, Parameter* p)
 | 
			
		||||
{
 | 
			
		||||
    s->eps     = p->eps;
 | 
			
		||||
    s->omega   = p->omg;
 | 
			
		||||
    s->itermax = p->itermax;
 | 
			
		||||
    s->levels  = p->levels;
 | 
			
		||||
    s->grid    = &d->grid;
 | 
			
		||||
    s->eps        = p->eps;
 | 
			
		||||
    s->omega      = p->omg;
 | 
			
		||||
    s->itermax    = p->itermax;
 | 
			
		||||
    s->levels     = p->levels;
 | 
			
		||||
    s->grid       = &d->grid;
 | 
			
		||||
    s->presmooth  = p->presmooth;
 | 
			
		||||
    s->postsmooth = p->postsmooth;
 | 
			
		||||
 | 
			
		||||
    int imax   = s->grid->imax;
 | 
			
		||||
    int jmax   = s->grid->jmax;
 | 
			
		||||
@@ -240,11 +292,13 @@ void initSolver(Solver* s, Discretization* d, Parameter* p)
 | 
			
		||||
    }
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
void solve(Solver* s, double* p, double* rhs)
 | 
			
		||||
double solve(Solver* s, double* p, double* rhs)
 | 
			
		||||
{
 | 
			
		||||
    double res = multiGrid(s, p, rhs, 0, s->grid->imax, s->grid->jmax, s->grid->kmax);
 | 
			
		||||
 | 
			
		||||
#ifdef VERBOSE
 | 
			
		||||
    printf("Residuum: %.6f\n", res);
 | 
			
		||||
#endif
 | 
			
		||||
 | 
			
		||||
return res;
 | 
			
		||||
}
 | 
			
		||||
 
 | 
			
		||||
							
								
								
									
										101
									
								
								BasicSolver/3D-seq/src/solver-rb.c
									
									
									
									
									
										Normal file
									
								
							
							
						
						
									
										101
									
								
								BasicSolver/3D-seq/src/solver-rb.c
									
									
									
									
									
										Normal file
									
								
							@@ -0,0 +1,101 @@
 | 
			
		||||
/*
 | 
			
		||||
 * Copyright (C) NHR@FAU, University Erlangen-Nuremberg.
 | 
			
		||||
 * All rights reserved. This file is part of nusif-solver.
 | 
			
		||||
 * Use of this source code is governed by a MIT style
 | 
			
		||||
 * license that can be found in the LICENSE file.
 | 
			
		||||
 */
 | 
			
		||||
#include "solver.h"
 | 
			
		||||
#include "util.h"
 | 
			
		||||
 | 
			
		||||
void initSolver(Solver* s, Discretization* d, Parameter* p)
 | 
			
		||||
{
 | 
			
		||||
    s->grid    = &d->grid;
 | 
			
		||||
    s->itermax = p->itermax;
 | 
			
		||||
    s->eps     = p->eps;
 | 
			
		||||
    s->omega   = p->omg;
 | 
			
		||||
}
 | 
			
		||||
 | 
			
		||||
double solve(Solver* s, double* p, double* rhs)
 | 
			
		||||
{
 | 
			
		||||
    int imax      = s->grid->imax;
 | 
			
		||||
    int jmax      = s->grid->jmax;
 | 
			
		||||
    int kmax      = s->grid->kmax;
 | 
			
		||||
    double eps    = s->eps;
 | 
			
		||||
    int itermax   = s->itermax;
 | 
			
		||||
    double dx2    = s->grid->dx * s->grid->dx;
 | 
			
		||||
    double dy2    = s->grid->dy * s->grid->dy;
 | 
			
		||||
    double dz2    = s->grid->dz * s->grid->dz;
 | 
			
		||||
    double idx2   = 1.0 / dx2;
 | 
			
		||||
    double idy2   = 1.0 / dy2;
 | 
			
		||||
    double idz2   = 1.0 / dz2;
 | 
			
		||||
    double factor = s->omega * 0.5 * (dx2 * dy2 * dz2) /
 | 
			
		||||
                    (dy2 * dz2 + dx2 * dz2 + dx2 * dy2);
 | 
			
		||||
    double epssq = eps * eps;
 | 
			
		||||
    int it       = 0;
 | 
			
		||||
    double res   = 1.0;
 | 
			
		||||
    int pass, ksw, jsw, isw;
 | 
			
		||||
 | 
			
		||||
    while ((res >= epssq) && (it < itermax)) {
 | 
			
		||||
        res = 0.0;
 | 
			
		||||
        ksw = 1;
 | 
			
		||||
 | 
			
		||||
        for (pass = 0; pass < 2; pass++) {
 | 
			
		||||
            jsw = ksw;
 | 
			
		||||
 | 
			
		||||
            for (int k = 1; k < kmax + 1; k++) {
 | 
			
		||||
                isw = jsw;
 | 
			
		||||
                for (int j = 1; j < jmax + 1; j++) {
 | 
			
		||||
                    for (int i = isw; i < imax + 1; i += 2) {
 | 
			
		||||
 | 
			
		||||
                        double r =
 | 
			
		||||
                            RHS(i, j, k) -
 | 
			
		||||
                            ((P(i + 1, j, k) - 2.0 * P(i, j, k) + P(i - 1, j, k)) * idx2 +
 | 
			
		||||
                                (P(i, j + 1, k) - 2.0 * P(i, j, k) + P(i, j - 1, k)) *
 | 
			
		||||
                                    idy2 +
 | 
			
		||||
                                (P(i, j, k + 1) - 2.0 * P(i, j, k) + P(i, j, k - 1)) *
 | 
			
		||||
                                    idz2);
 | 
			
		||||
 | 
			
		||||
                        P(i, j, k) -= (factor * r);
 | 
			
		||||
                        res += (r * r);
 | 
			
		||||
                    }
 | 
			
		||||
                    isw = 3 - isw;
 | 
			
		||||
                }
 | 
			
		||||
                jsw = 3 - jsw;
 | 
			
		||||
            }
 | 
			
		||||
            ksw = 3 - ksw;
 | 
			
		||||
        }
 | 
			
		||||
 | 
			
		||||
        for (int j = 1; j < jmax + 1; j++) {
 | 
			
		||||
            for (int i = 1; i < imax + 1; i++) {
 | 
			
		||||
                P(i, j, 0)        = P(i, j, 1);
 | 
			
		||||
                P(i, j, kmax + 1) = P(i, j, kmax);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
 | 
			
		||||
        for (int k = 1; k < kmax + 1; k++) {
 | 
			
		||||
            for (int i = 1; i < imax + 1; i++) {
 | 
			
		||||
                P(i, 0, k)        = P(i, 1, k);
 | 
			
		||||
                P(i, jmax + 1, k) = P(i, jmax, k);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
 | 
			
		||||
        for (int k = 1; k < kmax + 1; k++) {
 | 
			
		||||
            for (int j = 1; j < jmax + 1; j++) {
 | 
			
		||||
                P(0, j, k)        = P(1, j, k);
 | 
			
		||||
                P(imax + 1, j, k) = P(imax, j, k);
 | 
			
		||||
            }
 | 
			
		||||
        }
 | 
			
		||||
 | 
			
		||||
        res = res / (double)(imax * jmax * kmax);
 | 
			
		||||
#ifdef DEBUG
 | 
			
		||||
        printf("%d Residuum: %e\n", it, res);
 | 
			
		||||
#endif
 | 
			
		||||
        it++;
 | 
			
		||||
    }
 | 
			
		||||
 | 
			
		||||
#ifdef VERBOSE
 | 
			
		||||
    printf("Solver took %d iterations to reach %f\n", it, sqrt(res));
 | 
			
		||||
#endif
 | 
			
		||||
 | 
			
		||||
return res;
 | 
			
		||||
}
 | 
			
		||||
@@ -18,9 +18,10 @@ typedef struct {
 | 
			
		||||
    int itermax;
 | 
			
		||||
    int levels;
 | 
			
		||||
    double **r, **e;
 | 
			
		||||
    int presmooth, postsmooth;
 | 
			
		||||
} Solver;
 | 
			
		||||
 | 
			
		||||
extern void initSolver(Solver*, Discretization*, Parameter*);
 | 
			
		||||
extern void solve(Solver*, double*, double*);
 | 
			
		||||
extern double solve(Solver*, double*, double*);
 | 
			
		||||
 | 
			
		||||
#endif
 | 
			
		||||
 
 | 
			
		||||
@@ -70,6 +70,7 @@ void vtkScalar(VtkOptions* o, char* name, double* s)
 | 
			
		||||
        exit(EXIT_FAILURE);
 | 
			
		||||
    }
 | 
			
		||||
    fprintf(o->fh, "SCALARS %s float\n", name);
 | 
			
		||||
    fprintf(o->fh, "LOOKUP_TABLE default\n");
 | 
			
		||||
 | 
			
		||||
    for (int k = 0; k < kmax; k++) {
 | 
			
		||||
        for (int j = 0; j < jmax; j++) {
 | 
			
		||||
 
 | 
			
		||||
		Reference in New Issue
	
	Block a user