/*************************************************************************
* This file randomly permutes the contents of an array.
* flag == 0, don't initialize perm
* flag == 1, set p[i] = i
**************************************************************************/
void FastRandomPermute(int n, idxtype *p, int flag)
{
int i, u, v;
idxtype tmp;
/* this is for very small arrays */
if (n < 25) {
RandomPermute(n, p, flag);
return;
}
if (flag == 1) {
for (i=0; i<n; i++)
p[i] = i;
}
for (i=0; i<n; i+=8) {
v = RandomInRange(n-4);
u = RandomInRange(n-4);
SWAP(p[v], p[u], tmp);
SWAP(p[v+1], p[u+1], tmp);
SWAP(p[v+2], p[u+2], tmp);
SWAP(p[v+3], p[u+3], tmp);
}
}
void test_dkvsort()
{
gk_idx_t i;
gk_dkv_t array[N];
/* test the increasing sort */
printf("Testing dkvsorti...\n");
for (i=0; i<N; i++) {
array[i].key = RandomInRange(123432)/(1.0+RandomInRange(645323));
array[i].val = i;
}
gk_dkvsorti(N, array);
for (i=0; i<N-1; i++) {
if (array[i].key > array[i+1].key)
printf("gk_dkvsorti error at index %jd [%lf %lf] [%jd %jd]\n", (intmax_t)i, array[i].key, array[i+1].key, (intmax_t)array[i].val, (intmax_t)array[i+1].val);
}
/* test the decreasing sort */
printf("Testing dkvsortd...\n");
for (i=0; i<N; i++) {
array[i].key = RandomInRange(123432)/(1.0+RandomInRange(645323));
array[i].val = i;
}
gk_dkvsortd(N, array);
for (i=0; i<N-1; i++) {
if (array[i].key < array[i+1].key)
printf("gk_dkvsortd error at index %jd [%lf %lf] [%jd %jd]\n", (intmax_t)i, array[i].key, array[i+1].key, (intmax_t)array[i].val, (intmax_t)array[i+1].val);
}
}
void test_idxsort()
{
gk_idx_t i;
gk_idx_t array[N];
/* test the increasing sort */
printf("Testing idxsorti...\n");
for (i=0; i<N; i++)
array[i] = RandomInRange(123432);
gk_idxsorti(N, array);
for (i=0; i<N-1; i++) {
if (array[i] > array[i+1])
printf("gk_idxsorti error at index %zd [%zd %zd]\n", (ssize_t)i, (ssize_t)array[i], (ssize_t)array[i+1]);
}
/* test the decreasing sort */
printf("Testing idxsortd...\n");
for (i=0; i<N; i++)
array[i] = RandomInRange(123432);
gk_idxsortd(N, array);
for (i=0; i<N-1; i++) {
if (array[i] < array[i+1])
printf("gk_idxsortd error at index %zd [%zd %zd]\n", (ssize_t)i, (ssize_t)array[i], (ssize_t)array[i+1]);
}
}
void test_fsort()
{
gk_idx_t i;
float array[N];
/* test the increasing sort */
printf("Testing ifsort...\n");
for (i=0; i<N; i++)
array[i] = RandomInRange(123432)/(1.0+RandomInRange(645323));
gk_fsorti(N, array);
for (i=0; i<N-1; i++) {
if (array[i] > array[i+1])
printf("gk_fsorti error at index %jd [%f %f]\n", (intmax_t)i, array[i], array[i+1]);
}
/* test the decreasing sort */
printf("Testing dfsort...\n");
for (i=0; i<N; i++)
array[i] = RandomInRange(123432)/(1.0+RandomInRange(645323));
gk_fsortd(N, array);
for (i=0; i<N-1; i++) {
if (array[i] < array[i+1])
printf("gk_fsortd error at index %jd [%f %f]\n", (intmax_t)i, array[i], array[i+1]);
}
}
void test_isort()
{
gk_idx_t i;
int array[N];
/* test the increasing sort */
printf("Testing iisort...\n");
for (i=0; i<N; i++)
array[i] = RandomInRange(123432);
gk_isorti(N, array);
for (i=0; i<N-1; i++) {
if (array[i] > array[i+1])
printf("gk_isorti error at index %jd [%d %d]\n", (intmax_t)i, array[i], array[i+1]);
}
/* test the decreasing sort */
printf("Testing disort...\n");
for (i=0; i<N; i++)
array[i] = RandomInRange(123432);
gk_isortd(N, array);
for (i=0; i<N-1; i++) {
if (array[i] < array[i+1])
printf("gk_isortd error at index %jd [%d %d]\n", (intmax_t)i, array[i], array[i+1]);
}
}
/*************************************************************************
* This function implements a simple graph adaption strategy.
**************************************************************************/
void AdaptGraph(graph_t *graph, idx_t afactor, MPI_Comm comm)
{
idx_t i, nvtxs, nadapt, firstvtx, lastvtx;
idx_t npes, mype, mypwgt, max, min, sum;
idx_t *vwgt, *xadj, *adjncy, *adjwgt, *perm;
gkMPI_Comm_size(comm, &npes);
gkMPI_Comm_rank(comm, &mype);
srand(mype*afactor);
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
if (graph->adjwgt == NULL)
adjwgt = graph->adjwgt = ismalloc(graph->nedges, 1, "AdaptGraph: adjwgt");
else
adjwgt = graph->adjwgt;
vwgt = graph->vwgt;
firstvtx = graph->vtxdist[mype];
lastvtx = graph->vtxdist[mype+1];
perm = imalloc(nvtxs, "AdaptGraph: perm");
FastRandomPermute(nvtxs, perm, 1);
nadapt = RandomInRange(nvtxs);
nadapt = RandomInRange(nvtxs);
nadapt = RandomInRange(nvtxs);
for (i=0; i<nadapt; i++)
vwgt[perm[i]] = afactor*vwgt[perm[i]];
/*
for (i=0; i<nvtxs; i++) {
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = adjncy[j];
if (k >= firstvtx && k < lastvtx) {
adjwgt[j] = (int)pow(1.0*(gk_min(vwgt[i],vwgt[k-firstvtx])), .6667);
if (adjwgt[j] == 0)
adjwgt[j] = 1;
}
}
}
*/
mypwgt = isum(nvtxs, vwgt, 1);
MPI_Allreduce((void *)&mypwgt, (void *)&max, 1, IDX_T, MPI_MAX, comm);
MPI_Allreduce((void *)&mypwgt, (void *)&min, 1, IDX_T, MPI_MIN, comm);
MPI_Allreduce((void *)&mypwgt, (void *)&sum, 1, IDX_T, MPI_SUM, comm);
if (mype == 0)
printf("Initial Load Imbalance: %5.4"PRREAL", [%5"PRIDX" %5"PRIDX" %5"PRIDX"] for afactor: %"PRIDX"\n", (1.0*max*npes)/(1.0*sum), min, max, sum, afactor);
free(perm);
}
void ScatterInRectangle(RNGSeed *seed, F32 x, F32 y, F32 w, F32 h, U32 rows, U32 cols, std::function<void(V2)> procedure) {
F32 xstep = w / cols;
F32 ystep = h / rows;
for (size_t iy = 0; iy < cols + 1; iy++) {
for (size_t ix = 0; ix < rows + 1; ix++) {
F32 p = 0.4f;
V2 point;
point.x = x + (ix * xstep) + (RandomInRange(-p, p, *seed) * xstep);
point.y = y + (iy * ystep) + (RandomInRange(-p, p, *seed) * ystep);
procedure(point);
}
}
}
/*************************************************************************
* This file randomly permutes the contents of an array.
* flag == 0, don't initialize perm
* flag == 1, set p[i] = i
**************************************************************************/
void gk_RandomPermute(size_t n, int *p, int flag)
{
size_t i, u, v;
int tmp;
if (flag == 1) {
for (i=0; i<n; i++)
p[i] = i;
}
for (i=0; i<n/2; i++) {
v = RandomInRange(n);
u = RandomInRange(n);
gk_SWAP(p[v], p[u], tmp);
}
}
/*************************************************************************
* This file randomly permutes the contents of an array.
* flag == 0, don't initialize perm
* flag == 1, set p[i] = i
**************************************************************************/
void RandomPermute(int n, idxtype *p, int flag)
{
int i, u, v;
idxtype tmp;
if (flag == 1) {
for (i=0; i<n; i++)
p[i] = i;
}
for (i=0; i<n; i++) {
v = RandomInRange(n);
u = RandomInRange(n);
SWAP(p[v], p[u], tmp);
}
}
/*************************************************************************
* This function implements a simple graph adaption strategy.
**************************************************************************/
void Mc_AdaptGraph(graph_t *graph, idx_t *part, idx_t ncon, idx_t nparts, MPI_Comm comm)
{
idx_t h, i;
idx_t nvtxs;
idx_t npes, mype;
idx_t *vwgt, *pwgts;
gkMPI_Comm_size(comm, &npes);
gkMPI_Comm_rank(comm, &mype);
nvtxs = graph->nvtxs;
vwgt = graph->vwgt;
pwgts = ismalloc(nparts*ncon, 1, "pwgts");
if (mype == 0) {
for (i=0; i<nparts; i++)
for (h=0; h<ncon; h++)
pwgts[i*ncon+h] = RandomInRange(20)+1;
}
MPI_Bcast((void *)pwgts, nparts*ncon, IDX_T, 0, comm);
for (i=0; i<nvtxs; i++)
for (h=0; h<ncon; h++)
vwgt[i*ncon+h] = pwgts[part[i]*ncon+h];
free(pwgts);
return;
}
bool GameController::Initialize(HWND aWindow)
{
if (XACore::GetStatus() != XACore::OK) // Check if something went horribly wrong with XACore.
return false;
// Figure out the starting position for the monster.
// NOTE: THE FUNCTION WAS DESIGNED TO BE USED WITH FLOATS FOR ANOTHER COURSEWORK PROJECT SO CAST THE RETURN AS AN INT.
int monsterPositionX = static_cast<int>(RandomInRange(MIN_MONSTER_SPAWN_RANGE, MAX_MONSTER_SPAWN_RANGE));
int monsterPositionY = static_cast<int>(RandomInRange(MIN_MONSTER_SPAWN_RANGE, MAX_MONSTER_SPAWN_RANGE));
// The variables above need to be randomized more so that the monster doesn't start in front and to the right of the player all the time.
RandomSign(monsterPositionX);
RandomSign(monsterPositionY);
// Figure out the position of the music box, as with above.
int musicBoxPositionX = static_cast<int>(RandomInRange(MIN_MUSIC_BOX_SPAWN_RANGE, MAX_MUSIC_BOX_SPAWN_RANGE));
int musicBoxPositionY = static_cast<int>(RandomInRange(MIN_MUSIC_BOX_SPAWN_RANGE, MAX_MUSIC_BOX_SPAWN_RANGE));
// The variables above need to be randomized more so that the music box doesn't always appear to the front and right of the player.
RandomSign(musicBoxPositionX);
RandomSign(musicBoxPositionY);
// Figure out the delay timer for the music box to start playing.
float musicBoxStartDelay = RandomInRange(MAX_MUSIC_BOX_START_DELAY, MIN_MUSIC_BOX_START_DELAY);
// Initialize the game objects.
world.reset(new World(musicBoxPositionX, musicBoxPositionY, musicBoxStartDelay));
player.reset(new Player);
monster.reset(new Monster(monsterPositionX, monsterPositionY));
// Load Ambient Audio
world->LoadAmbientSound("Audio/AmbientAudio.wav");
world->LoadMusicBoxSound("Audio/MusicBox.wav");
// Load Player Audio
player->LoadBreathingAudio("Audio/PlayerBreathing.wav");
player->LoadWalkingAudio("Audio/PlayerWalking.wav");
player->LoadDeathAudio("Audio/PlayerDeath.wav");
// Load Monster Audio
monster->LoadBreathingAudio("Audio/MonsterBreathing.wav");
monster->LoadWalkingAudio("Audio/MonsterWalking.wav");
monster->LoadDeathAudio("Audio/MonsterDeath.wav");
return true;
} // End of Initialize function
/*************************************************************************
* This function randomly permutes the locally stored adjacency lists
**************************************************************************/
void GraphRandomPermute(GraphType *graph)
{
int i, j, k, tmp;
for (i=0; i<graph->nvtxs; i++) {
for (j=graph->xadj[i]; j<graph->xadj[i+1]; j++) {
k = graph->xadj[i] + RandomInRange(graph->xadj[i+1]-graph->xadj[i]);
SWAP(graph->adjncy[j], graph->adjncy[k], tmp);
SWAP(graph->adjwgt[j], graph->adjwgt[k], tmp);
}
}
}
inline void SSAOInit(SSAO* ssao, U32 viewportWidth, U32 viewportHeight) {
glGenFramebuffers(1, &ssao->framebuffer);
glBindFramebuffer(GL_FRAMEBUFFER, ssao->framebuffer);
glGenTextures(1, &ssao->bufferTexture);
glBindTexture(GL_TEXTURE_2D, ssao->bufferTexture);
glTexStorage2D(GL_TEXTURE_2D, 1, GL_R32F, viewportWidth, viewportHeight);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
glFramebufferTexture2D(GL_FRAMEBUFFER, GL_COLOR_ATTACHMENT0, GL_TEXTURE_2D, ssao->bufferTexture, 0);
GLuint attachments[] = { GL_COLOR_ATTACHMENT0 };
glDrawBuffers(1, attachments);
glBindFramebuffer(GL_FRAMEBUFFER, 0);
RNGSeed seed(7);
V3 ssaoNoise[SSAO_NOISE_SIZE * SSAO_NOISE_SIZE];
for (size_t i = 0; i < SSAO_NOISE_SIZE * SSAO_NOISE_SIZE; i++) {
ssaoNoise[i].x = RandomInRange(-1.0f, 1.0f, seed);
ssaoNoise[i].y = RandomInRange(-1.0f, 1.0f, seed);
ssaoNoise[i].z = 0.0f;
}
glGenTextures(1, &ssao->noiseTexture);
glBindTexture(GL_TEXTURE_2D, ssao->noiseTexture);
glTexImage2D(GL_TEXTURE_2D, 0, GL_RGB32F, SSAO_NOISE_SIZE, SSAO_NOISE_SIZE, 0, GL_RGB, GL_FLOAT, ssaoNoise);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MIN_FILTER, GL_NEAREST);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_MAG_FILTER, GL_NEAREST);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_S, GL_REPEAT);
glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_WRAP_T, GL_REPEAT);
V3 sample;
for (size_t i = 0; i < SSAO_SAMPLE_COUNT; i++) {
sample.x = RandomInRange(-1.0, 1.0, seed);
sample.y = RandomInRange(-1.0, 1.0, seed);
sample.z = RandomInRange(-1.0, 1.0, seed);
sample = Normalize(sample);
ssao->kernelSamples[i] = sample;
}
}
/*************************************************************************
* This function implements a simple graph adaption strategy.
**************************************************************************/
void AdaptGraph2(graph_t *graph, idx_t afactor, MPI_Comm comm)
{
idx_t i, j, k, nvtxs, firstvtx, lastvtx;
idx_t npes, mype, mypwgt, max, min, sum;
idx_t *vwgt, *xadj, *adjncy, *adjwgt;
gkMPI_Comm_size(comm, &npes);
gkMPI_Comm_rank(comm, &mype);
srand(mype*afactor);
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
if (graph->adjwgt == NULL)
adjwgt = graph->adjwgt = ismalloc(graph->nedges, 1, "AdaptGraph: adjwgt");
else
adjwgt = graph->adjwgt;
vwgt = graph->vwgt;
firstvtx = graph->vtxdist[mype];
lastvtx = graph->vtxdist[mype+1];
/* if (RandomInRange(npes+1) < .05*npes) { */
if (RandomInRange(npes+1) < 2) {
printf("[%"PRIDX"] is adapting\n", mype);
for (i=0; i<nvtxs; i++)
vwgt[i] = afactor*vwgt[i];
}
for (i=0; i<nvtxs; i++) {
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = adjncy[j];
if (k >= firstvtx && k < lastvtx) {
adjwgt[j] = (int)pow(1.0*(gk_min(vwgt[i],vwgt[k-firstvtx])), .6667);
if (adjwgt[j] == 0)
adjwgt[j] = 1;
}
}
}
mypwgt = isum(nvtxs, vwgt, 1);
gkMPI_Allreduce((void *)&mypwgt, (void *)&max, 1, IDX_T, MPI_MAX, comm);
gkMPI_Allreduce((void *)&mypwgt, (void *)&min, 1, IDX_T, MPI_MIN, comm);
gkMPI_Allreduce((void *)&mypwgt, (void *)&sum, 1, IDX_T, MPI_SUM, comm);
if (mype == 0)
printf("Initial Load Imbalance: %5.4"PRREAL", [%5"PRIDX" %5"PRIDX" %5"PRIDX"]\n", (1.0*max*npes)/(1.0*sum), min, max, sum);
}
void test_idxkvsort()
{
gk_idx_t i;
gk_idxkv_t array[N];
/* test the increasing sort */
printf("Testing idxkvsorti...\n");
for (i=0; i<N; i++) {
array[i].key = RandomInRange(123432);
array[i].val = i;
}
gk_idxkvsorti(N, array);
for (i=0; i<N-1; i++) {
if (array[i].key > array[i+1].key)
printf("gk_idxkvsorti error at index %zd [%zd %zd] [%zd %zd]\n",
(ssize_t)i, (ssize_t)array[i].key, (ssize_t)array[i+1].key,
(ssize_t)array[i].val, (ssize_t)array[i+1].val);
}
/* test the decreasing sort */
printf("Testing idxkvsortd...\n");
for (i=0; i<N; i++) {
array[i].key = RandomInRange(123432);
array[i].val = i;
}
gk_idxkvsortd(N, array);
for (i=0; i<N-1; i++) {
if (array[i].key < array[i+1].key)
printf("gk_idxkvsortd error at index %zd [%zd %zd] [%zd %zd]\n",
(ssize_t)i, (ssize_t)array[i].key, (ssize_t)array[i+1].key,
(ssize_t)array[i].val, (ssize_t)array[i+1].val);
}
}
void test_skvsort()
{
gk_idx_t i;
gk_skv_t array[N];
char line[256];
/* test the increasing sort */
printf("Testing skvsorti...\n");
for (i=0; i<N; i++) {
sprintf(line, "%d", RandomInRange(123432));
array[i].key = gk_strdup(line);
array[i].val = i;
}
gk_skvsorti(N, array);
for (i=0; i<N-1; i++) {
if (strcmp(array[i].key, array[i+1].key) > 0)
printf("gk_skvsorti error at index %jd [%s %s] [%jd %jd]\n", (intmax_t)i, array[i].key, array[i+1].key, (intmax_t)array[i].val, (intmax_t)array[i+1].val);
}
/* test the decreasing sort */
printf("Testing skvsortd...\n");
for (i=0; i<N; i++) {
sprintf(line, "%d", RandomInRange(123432));
array[i].key = gk_strdup(line);
array[i].val = i;
}
gk_skvsortd(N, array);
for (i=0; i<N-1; i++) {
/*printf("%s\n", array[i].key);*/
if (strcmp(array[i].key, array[i+1].key) < 0)
printf("gk_skvsortd error at index %jd [%s %s] [%jd %jd]\n", (intmax_t)i, array[i].key, array[i+1].key, (intmax_t)array[i].val, (intmax_t)array[i+1].val);
}
}
/*************************************************************************
* This file randomly permutes the contents of an array.
* flag == 0, don't initialize perm
* flag == 1, set p[i] = i
**************************************************************************/
void RandomPermute(int n, idxtype *p, int flag)
{
int i, u, v;
idxtype tmp;
if (flag == 1) {
for (i=0; i<n; i++)
p[i] = i;
}
if (n <= 4)
return;
for (i=0; i<n; i+=16) {
u = RandomInRange(n-4);
v = RandomInRange(n-4);
SWAP(p[v], p[u], tmp);
SWAP(p[v+1], p[u+1], tmp);
SWAP(p[v+2], p[u+2], tmp);
SWAP(p[v+3], p[u+3], tmp);
}
}
/*************************************************************************
* This function randomly permutes the adjacency lists of a graph
**************************************************************************/
void RandomizeGraph(GraphType *graph)
{
int i, j, k, l, tmp, nvtxs;
idxtype *xadj, *adjncy, *adjwgt;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
for (i=0; i<nvtxs; i++) {
l = xadj[i+1]-xadj[i];
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = xadj[i] + RandomInRange(l);
SWAP(adjncy[j], adjncy[k], tmp);
SWAP(adjwgt[j], adjwgt[k], tmp);
}
}
}
/*************************************************************************
* This function takes a graph and produces a bisection by using a region
* growing algorithm. The resulting partition is returned in
* graph->where
**************************************************************************/
void MocGrowBisection(CtrlType *ctrl, GraphType *graph, float *tpwgts, float ubfactor)
{
int i, j, k, nvtxs, ncon, from, bestcut, mincut, nbfs;
idxtype *bestwhere, *where;
nvtxs = graph->nvtxs;
MocAllocate2WayPartitionMemory(ctrl, graph);
where = graph->where;
bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere");
nbfs = 2*(nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS);
bestcut = idxsum(graph->nedges, graph->adjwgt);
for (; nbfs>0; nbfs--) {
idxset(nvtxs, 1, where);
where[RandomInRange(nvtxs)] = 0;
MocCompute2WayPartitionParams(ctrl, graph);
MocInit2WayBalance(ctrl, graph, tpwgts);
MocFM_2WayEdgeRefine(ctrl, graph, tpwgts, 4);
MocBalance2Way(ctrl, graph, tpwgts, 1.02);
MocFM_2WayEdgeRefine(ctrl, graph, tpwgts, 4);
if (bestcut >= graph->mincut) {
bestcut = graph->mincut;
idxcopy(nvtxs, where, bestwhere);
if (bestcut == 0)
break;
}
}
graph->mincut = bestcut;
idxcopy(nvtxs, bestwhere, where);
/*GKfree(&bestwhere, LTERM);*/
GKfree1((void**)&bestwhere);
}
/*************************************************************************
* This function takes a graph and produces a bisection by using a region
* growing algorithm. The resulting partition is returned in
* graph->where
**************************************************************************/
void MocGrowBisection2(CtrlType *ctrl, GraphType *graph, float *tpwgts, float *ubvec)
{
int /*i, j, k,*/ nvtxs, /*ncon, from,*/ bestcut, /*mincut,*/ nbfs;
idxtype *bestwhere, *where;
nvtxs = graph->nvtxs;
MocAllocate2WayPartitionMemory(ctrl, graph);
where = graph->where;
bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere");
nbfs = 2*(nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS);
bestcut = idxsum(graph->nedges, graph->adjwgt);
for (; nbfs>0; nbfs--) {
idxset(nvtxs, 1, where);
where[RandomInRange(nvtxs)] = 0;
MocCompute2WayPartitionParams(ctrl, graph);
MocBalance2Way2(ctrl, graph, tpwgts, ubvec);
MocFM_2WayEdgeRefine2(ctrl, graph, tpwgts, ubvec, 4);
MocBalance2Way2(ctrl, graph, tpwgts, ubvec);
MocFM_2WayEdgeRefine2(ctrl, graph, tpwgts, ubvec, 4);
if (bestcut > graph->mincut) {
bestcut = graph->mincut;
idxcopy(nvtxs, where, bestwhere);
if (bestcut == 0)
break;
}
}
graph->mincut = bestcut;
idxcopy(nvtxs, bestwhere, where);
GKfree((void**)&bestwhere, LTERM);
}
int Partition(int data[], int length, int start, int end) {
if (data == NULL || length <= 0 || start < 0 || end >= length)
return -1;
int index = RandomInRange(start, end);
Swap(data[index], data[end]);
int small = start - 1;
for (index = start; index < end; index++) {
if (data[index] < data[end]) {
small++;
if (small != index)
Swap(data[index], data[small]);
}
}
small++;
Swap(data[small], data[end]);
return small;
}
/*************************************************************************
* This function takes a graph and produces a bisection by using a region
* growing algorithm. The resulting partition is returned in
* graph->where
**************************************************************************/
void MocGrowBisectionNew2(CtrlType *ctrl, GraphType *graph, float *tpwgts, float *ubvec)
{
idxtype i, j, k, nvtxs, ncon, from, bestcut, mincut, nbfs, inbfs;
idxtype *bestwhere, *where;
nvtxs = graph->nvtxs;
MocAllocate2WayPartitionMemory(ctrl, graph);
where = graph->where;
bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere");
nbfs = 2*(nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS);
for (inbfs=0; inbfs<nbfs; inbfs++) {
idxset(nvtxs, 1, where);
where[RandomInRange(nvtxs)] = 0;
MocCompute2WayPartitionParams(ctrl, graph);
MocInit2WayBalance2(ctrl, graph, tpwgts, ubvec);
MocFM_2WayEdgeRefine2(ctrl, graph, tpwgts, ubvec, 4);
if (inbfs == 0 || bestcut > graph->mincut) {
bestcut = graph->mincut;
idxcopy(nvtxs, where, bestwhere);
if (bestcut == 0)
break;
}
}
graph->mincut = bestcut;
idxcopy(nvtxs, bestwhere, where);
gk_free((void **)&bestwhere, LTERM);
}
int main(int argc, char *argv[])
{
ssize_t i, j, niter;
params_t *params;
gk_csr_t *mat;
FILE *fpout;
/* get command-line options */
params = parse_cmdline(argc, argv);
/* read the data */
mat = gk_csr_Read(params->infile, GK_CSR_FMT_METIS, 1, 1);
/* display some basic stats */
print_init_info(params, mat);
if (params->ntvs != -1) {
/* compute the pr for different randomly generated restart-distribution vectors */
float **prs;
prs = gk_fAllocMatrix(params->ntvs, mat->nrows, 0.0, "main: prs");
/* generate the random restart vectors */
for (j=0; j<params->ntvs; j++) {
for (i=0; i<mat->nrows; i++)
prs[j][i] = RandomInRange(931);
gk_fscale(mat->nrows, 1.0/gk_fsum(mat->nrows, prs[j], 1), prs[j], 1);
niter = gk_rw_PageRank(mat, params->lamda, params->eps, params->niter, prs[j]);
printf("tvs#: %zd; niters: %zd\n", j, niter);
}
/* output the computed pr scores */
fpout = gk_fopen(params->outfile, "w", "main: outfile");
for (i=0; i<mat->nrows; i++) {
for (j=0; j<params->ntvs; j++)
fprintf(fpout, "%.4e ", prs[j][i]);
fprintf(fpout, "\n");
}
gk_fclose(fpout);
gk_fFreeMatrix(&prs, params->ntvs, mat->nrows);
}
else if (params->ppr != -1) {
/* compute the personalized pr from the specified vertex */
float *pr;
pr = gk_fsmalloc(mat->nrows, 0.0, "main: pr");
pr[params->ppr-1] = 1.0;
niter = gk_rw_PageRank(mat, params->lamda, params->eps, params->niter, pr);
printf("ppr: %d; niters: %zd\n", params->ppr, niter);
/* output the computed pr scores */
fpout = gk_fopen(params->outfile, "w", "main: outfile");
for (i=0; i<mat->nrows; i++)
fprintf(fpout, "%.4e\n", pr[i]);
gk_fclose(fpout);
gk_free((void **)&pr, LTERM);
}
else {
/* compute the standard pr */
int jmax;
float diff, maxdiff;
float *pr;
pr = gk_fsmalloc(mat->nrows, 1.0/mat->nrows, "main: pr");
niter = gk_rw_PageRank(mat, params->lamda, params->eps, params->niter, pr);
printf("pr; niters: %zd\n", niter);
/* output the computed pr scores */
fpout = gk_fopen(params->outfile, "w", "main: outfile");
for (i=0; i<mat->nrows; i++) {
for (jmax=i, maxdiff=0.0, j=mat->rowptr[i]; j<mat->rowptr[i+1]; j++) {
if ((diff = fabs(pr[i]-pr[mat->rowind[j]])) > maxdiff) {
maxdiff = diff;
jmax = mat->rowind[j];
}
}
fprintf(fpout, "%.4e %10zd %.4e %10d\n", pr[i],
mat->rowptr[i+1]-mat->rowptr[i], maxdiff, jmax+1);
}
gk_fclose(fpout);
gk_free((void **)&pr, LTERM);
}
gk_csr_Free(&mat);
/* display some final stats */
print_final_info(params);
}
/*************************************************************************
* This function takes a graph and produces a bisection by using a region
* growing algorithm. The resulting partition is returned in
* graph->where
**************************************************************************/
void GrowBisection(CtrlType *ctrl, GraphType *graph, int *tpwgts, float ubfactor)
{
int i, j, k, nvtxs, drain, nleft, first, last, pwgts[2], minpwgt[2], maxpwgt[2], from, bestcut, icut, mincut, me, pass, nbfs;
idxtype *xadj, *vwgt, *adjncy, *adjwgt, *where;
idxtype *queue, *touched, *gain, *bestwhere;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
vwgt = graph->vwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
Allocate2WayPartitionMemory(ctrl, graph);
where = graph->where;
bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere");
queue = idxmalloc(nvtxs, "BisectGraph: queue");
touched = idxmalloc(nvtxs, "BisectGraph: touched");
ASSERTP(tpwgts[0]+tpwgts[1] == idxsum(nvtxs, vwgt), ("%d %d\n", tpwgts[0]+tpwgts[1], idxsum(nvtxs, vwgt)));
maxpwgt[0] = ubfactor*tpwgts[0];
maxpwgt[1] = ubfactor*tpwgts[1];
minpwgt[0] = (1.0/ubfactor)*tpwgts[0];
minpwgt[1] = (1.0/ubfactor)*tpwgts[1];
nbfs = (nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS);
bestcut = idxsum(nvtxs, graph->adjwgtsum)+1; /* The +1 is for the 0 edges case */
for (; nbfs>0; nbfs--) {
idxset(nvtxs, 0, touched);
pwgts[1] = tpwgts[0]+tpwgts[1];
pwgts[0] = 0;
idxset(nvtxs, 1, where);
queue[0] = RandomInRange(nvtxs);
touched[queue[0]] = 1;
first = 0; last = 1;
nleft = nvtxs-1;
drain = 0;
/* Start the BFS from queue to get a partition */
for (;;) {
if (first == last) { /* Empty. Disconnected graph! */
if (nleft == 0 || drain)
break;
k = RandomInRange(nleft);
for (i=0; i<nvtxs; i++) {
if (touched[i] == 0) {
if (k == 0)
break;
else
k--;
}
}
queue[0] = i;
touched[i] = 1;
first = 0; last = 1;;
nleft--;
}
i = queue[first++];
if (pwgts[0] > 0 && pwgts[1]-vwgt[i] < minpwgt[1]) {
drain = 1;
continue;
}
where[i] = 0;
INC_DEC(pwgts[0], pwgts[1], vwgt[i]);
if (pwgts[1] <= maxpwgt[1])
break;
drain = 0;
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = adjncy[j];
if (touched[k] == 0) {
queue[last++] = k;
touched[k] = 1;
nleft--;
}
}
}
/* Check to see if we hit any bad limiting cases */
if (pwgts[1] == 0) {
i = RandomInRange(nvtxs);
where[i] = 1;
INC_DEC(pwgts[1], pwgts[0], vwgt[i]);
}
/*************************************************************
* Do some partition refinement
**************************************************************/
Compute2WayPartitionParams(ctrl, graph);
/*printf("IPART: %3d [%5d %5d] [%5d %5d] %5d\n", graph->nvtxs, pwgts[0], pwgts[1], graph->pwgts[0], graph->pwgts[1], graph->mincut); */
Balance2Way(ctrl, graph, tpwgts, ubfactor);
/*printf("BPART: [%5d %5d] %5d\n", graph->pwgts[0], graph->pwgts[1], graph->mincut);*/
FM_2WayEdgeRefine(ctrl, graph, tpwgts, 4);
/*printf("RPART: [%5d %5d] %5d\n", graph->pwgts[0], graph->pwgts[1], graph->mincut);*/
if (bestcut > graph->mincut) {
bestcut = graph->mincut;
idxcopy(nvtxs, where, bestwhere);
if (bestcut == 0)
break;
}
}
graph->mincut = bestcut;
idxcopy(nvtxs, bestwhere, where);
GKfree(&bestwhere, &queue, &touched, LTERM);
}
/*************************************************************************
* This function takes a graph and produces a bisection by using a region
* growing algorithm. The resulting partition is returned in
* graph->where
**************************************************************************/
void GrowBisectionNode(CtrlType *ctrl, GraphType *graph, float ubfactor)
{
int i, j, k, nvtxs, drain, nleft, first, last, pwgts[2], tpwgts[2], minpwgt[2], maxpwgt[2], from, bestcut, icut, mincut, me, pass, nbfs;
idxtype *xadj, *vwgt, *adjncy, *adjwgt, *where, *bndind;
idxtype *queue, *touched, *gain, *bestwhere;
nvtxs = graph->nvtxs;
xadj = graph->xadj;
vwgt = graph->vwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
bestwhere = idxmalloc(nvtxs, "BisectGraph: bestwhere");
queue = idxmalloc(nvtxs, "BisectGraph: queue");
touched = idxmalloc(nvtxs, "BisectGraph: touched");
tpwgts[0] = idxsum(nvtxs, vwgt);
tpwgts[1] = tpwgts[0]/2;
tpwgts[0] -= tpwgts[1];
maxpwgt[0] = ubfactor*tpwgts[0];
maxpwgt[1] = ubfactor*tpwgts[1];
minpwgt[0] = (1.0/ubfactor)*tpwgts[0];
minpwgt[1] = (1.0/ubfactor)*tpwgts[1];
/* Allocate memory for graph->rdata. Allocate sufficient memory for both edge and node */
graph->rdata = idxmalloc(5*nvtxs+3, "GrowBisectionNode: graph->rdata");
graph->pwgts = graph->rdata;
graph->where = graph->rdata + 3;
graph->bndptr = graph->rdata + nvtxs + 3;
graph->bndind = graph->rdata + 2*nvtxs + 3;
graph->nrinfo = (NRInfoType *)(graph->rdata + 3*nvtxs + 3);
graph->id = graph->rdata + 3*nvtxs + 3;
graph->ed = graph->rdata + 4*nvtxs + 3;
where = graph->where;
bndind = graph->bndind;
nbfs = (nvtxs <= ctrl->CoarsenTo ? SMALLNIPARTS : LARGENIPARTS);
bestcut = tpwgts[0]+tpwgts[1];
for (nbfs++; nbfs>0; nbfs--) {
idxset(nvtxs, 0, touched);
pwgts[1] = tpwgts[0]+tpwgts[1];
pwgts[0] = 0;
idxset(nvtxs, 1, where);
queue[0] = RandomInRange(nvtxs);
touched[queue[0]] = 1;
first = 0; last = 1;
nleft = nvtxs-1;
drain = 0;
/* Start the BFS from queue to get a partition */
if (nbfs >= 1) {
for (;;) {
if (first == last) { /* Empty. Disconnected graph! */
if (nleft == 0 || drain)
break;
k = RandomInRange(nleft);
for (i=0; i<nvtxs; i++) {
if (touched[i] == 0) {
if (k == 0)
break;
else
k--;
}
}
queue[0] = i;
touched[i] = 1;
first = 0; last = 1;;
nleft--;
}
i = queue[first++];
if (pwgts[1]-vwgt[i] < minpwgt[1]) {
drain = 1;
continue;
}
where[i] = 0;
INC_DEC(pwgts[0], pwgts[1], vwgt[i]);
if (pwgts[1] <= maxpwgt[1])
break;
drain = 0;
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = adjncy[j];
if (touched[k] == 0) {
queue[last++] = k;
touched[k] = 1;
nleft--;
}
}
}
}
/*************************************************************
* Do some partition refinement
**************************************************************/
Compute2WayPartitionParams(ctrl, graph);
Balance2Way(ctrl, graph, tpwgts, ubfactor);
FM_2WayEdgeRefine(ctrl, graph, tpwgts, 4);
/* Construct and refine the vertex separator */
for (i=0; i<graph->nbnd; i++)
where[bndind[i]] = 2;
Compute2WayNodePartitionParams(ctrl, graph);
FM_2WayNodeRefine(ctrl, graph, ubfactor, 6);
/* printf("ISep: [%d %d %d] %d\n", graph->pwgts[0], graph->pwgts[1], graph->pwgts[2], bestcut); */
if (bestcut > graph->mincut) {
bestcut = graph->mincut;
idxcopy(nvtxs, where, bestwhere);
}
}
graph->mincut = bestcut;
idxcopy(nvtxs, bestwhere, where);
Compute2WayNodePartitionParams(ctrl, graph);
GKfree(&bestwhere, &queue, &touched, LTERM);
}
/*************************************************************************
* This function performs a k-way directed diffusion
**************************************************************************/
real_t WavefrontDiffusion(ctrl_t *ctrl, graph_t *graph, idx_t *home)
{
idx_t ii, i, j, k, l, nvtxs, nedges, nparts;
idx_t from, to, edge, done, nswaps, noswaps, totalv, wsize;
idx_t npasses, first, second, third, mind, maxd;
idx_t *xadj, *adjncy, *adjwgt, *where, *perm;
idx_t *rowptr, *colind, *ed, *psize;
real_t *transfer, *tmpvec;
real_t balance = -1.0, *load, *solution, *workspace;
real_t *nvwgt, *npwgts, flowFactor, cost, ubfactor;
matrix_t matrix;
ikv_t *cand;
idx_t ndirty, nclean, dptr, clean;
nvtxs = graph->nvtxs;
nedges = graph->nedges;
xadj = graph->xadj;
nvwgt = graph->nvwgt;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
where = graph->where;
nparts = ctrl->nparts;
ubfactor = ctrl->ubvec[0];
matrix.nrows = nparts;
flowFactor = 0.35;
flowFactor = (ctrl->mype == 2) ? 0.50 : flowFactor;
flowFactor = (ctrl->mype == 3) ? 0.75 : flowFactor;
flowFactor = (ctrl->mype == 4) ? 1.00 : flowFactor;
/* allocate memory */
solution = rmalloc(4*nparts+2*nedges, "WavefrontDiffusion: solution");
tmpvec = solution + nparts;
npwgts = solution + 2*nparts;
load = solution + 3*nparts;
matrix.values = solution + 4*nparts;
transfer = matrix.transfer = solution + 4*nparts + nedges;
perm = imalloc(2*nvtxs+2*nparts+nedges+1, "WavefrontDiffusion: perm");
ed = perm + nvtxs;
psize = perm + 2*nvtxs;
rowptr = matrix.rowptr = perm + 2*nvtxs + nparts;
colind = matrix.colind = perm + 2*nvtxs + 2*nparts + 1;
/*GKTODO - Potential problem with this malloc */
wsize = gk_max(sizeof(real_t)*nparts*6, sizeof(idx_t)*(nvtxs+nparts*2+1));
workspace = (real_t *)gk_malloc(wsize, "WavefrontDiffusion: workspace");
cand = ikvmalloc(nvtxs, "WavefrontDiffusion: cand");
/*****************************/
/* Populate empty subdomains */
/*****************************/
iset(nparts, 0, psize);
for (i=0; i<nvtxs; i++)
psize[where[i]]++;
mind = iargmin(nparts, psize);
maxd = iargmax(nparts, psize);
if (psize[mind] == 0) {
for (i=0; i<nvtxs; i++) {
k = (RandomInRange(nvtxs)+i)%nvtxs;
if (where[k] == maxd) {
where[k] = mind;
psize[mind]++;
psize[maxd]--;
break;
}
}
}
iset(nvtxs, 0, ed);
rset(nparts, 0.0, npwgts);
for (i=0; i<nvtxs; i++) {
npwgts[where[i]] += nvwgt[i];
for (j=xadj[i]; j<xadj[i+1]; j++)
ed[i] += (where[i] != where[adjncy[j]] ? adjwgt[j] : 0);
}
ComputeLoad(graph, nparts, load, ctrl->tpwgts, 0);
done = 0;
/* zero out the tmpvec array */
rset(nparts, 0.0, tmpvec);
npasses = gk_min(nparts/2, NGD_PASSES);
for (l=0; l<npasses; l++) {
/* Set-up and solve the diffusion equation */
nswaps = 0;
/************************/
/* Solve flow equations */
/************************/
SetUpConnectGraph(graph, &matrix, (idx_t *)workspace);
/* check for disconnected subdomains */
for(i=0; i<matrix.nrows; i++) {
if (matrix.rowptr[i]+1 == matrix.rowptr[i+1]) {
cost = (real_t)(ctrl->mype);
goto CleanUpAndExit;
}
}
ConjGrad2(&matrix, load, solution, 0.001, workspace);
ComputeTransferVector(1, &matrix, solution, transfer, 0);
GetThreeMax(nparts, load, &first, &second, &third);
if (l%3 == 0) {
FastRandomPermute(nvtxs, perm, 1);
}
else {
/*****************************/
/* move dirty vertices first */
/*****************************/
ndirty = 0;
for (i=0; i<nvtxs; i++) {
if (where[i] != home[i])
ndirty++;
}
dptr = 0;
for (i=0; i<nvtxs; i++) {
if (where[i] != home[i])
perm[dptr++] = i;
else
perm[ndirty++] = i;
}
PASSERT(ctrl, ndirty == nvtxs);
ndirty = dptr;
nclean = nvtxs-dptr;
FastRandomPermute(ndirty, perm, 0);
FastRandomPermute(nclean, perm+ndirty, 0);
}
if (ctrl->mype == 0) {
for (j=nvtxs, k=0, ii=0; ii<nvtxs; ii++) {
i = perm[ii];
if (ed[i] != 0) {
cand[k].key = -ed[i];
cand[k++].val = i;
}
else {
cand[--j].key = 0;
cand[j].val = i;
}
}
ikvsorti(k, cand);
}
for (ii=0; ii<nvtxs/3; ii++) {
i = (ctrl->mype == 0) ? cand[ii].val : perm[ii];
from = where[i];
/* don't move out the last vertex in a subdomain */
if (psize[from] == 1)
continue;
clean = (from == home[i]) ? 1 : 0;
/* only move from top three or dirty vertices */
if (from != first && from != second && from != third && clean)
continue;
/* Scatter the sparse transfer row into the dense tmpvec row */
for (j=rowptr[from]+1; j<rowptr[from+1]; j++)
tmpvec[colind[j]] = transfer[j];
for (j=xadj[i]; j<xadj[i+1]; j++) {
to = where[adjncy[j]];
if (from != to) {
if (tmpvec[to] > (flowFactor * nvwgt[i])) {
tmpvec[to] -= nvwgt[i];
INC_DEC(psize[to], psize[from], 1);
INC_DEC(npwgts[to], npwgts[from], nvwgt[i]);
INC_DEC(load[to], load[from], nvwgt[i]);
where[i] = to;
nswaps++;
/* Update external degrees */
ed[i] = 0;
for (k=xadj[i]; k<xadj[i+1]; k++) {
edge = adjncy[k];
ed[i] += (to != where[edge] ? adjwgt[k] : 0);
if (where[edge] == from)
ed[edge] += adjwgt[k];
if (where[edge] == to)
ed[edge] -= adjwgt[k];
}
break;
}
}
}
/* Gather the dense tmpvec row into the sparse transfer row */
for (j=rowptr[from]+1; j<rowptr[from+1]; j++) {
transfer[j] = tmpvec[colind[j]];
tmpvec[colind[j]] = 0.0;
}
ASSERT(fabs(rsum(nparts, tmpvec, 1)) < .0001)
}
if (l % 2 == 1) {
balance = rmax(nparts, npwgts)*nparts;
if (balance < ubfactor + 0.035)
done = 1;
if (GlobalSESum(ctrl, done) > 0)
break;
noswaps = (nswaps > 0) ? 0 : 1;
if (GlobalSESum(ctrl, noswaps) > ctrl->npes/2)
break;
}
}
graph->mincut = ComputeSerialEdgeCut(graph);
totalv = Mc_ComputeSerialTotalV(graph, home);
cost = ctrl->ipc_factor * (real_t)graph->mincut + ctrl->redist_factor * (real_t)totalv;
CleanUpAndExit:
gk_free((void **)&solution, (void **)&perm, (void **)&workspace, (void **)&cand, LTERM);
return cost;
}
void HeightMap::GenerationDiamondSquareFractal(void)
{
// Declare some internal variable
int w, x, z;
float dH, dHFactor;
int nP1, nP2, nP3, nP4, nPmid;
float fMinH, fMaxH;
dH = m_nSize * 0.5f; // dH is length of height map in one dimension
dHFactor = pow(2.0f, -ROUGHNESS);
fMinH = fMaxH = 0.0f;
// Fill initial value for height map
for (w = 0 ; w < m_nSize*m_nSize ; ++w)
m_pHeightMap[w] = SEED;
for (w = m_nSize ; w > 0 ; w /= 2, dH *= dHFactor)
{
// Diamond stage
for (z = 0; z < m_nSize; z += w)
{
for (x = 0; x < m_nSize; x += w)
{
nP1 = GetHeightIndexAt(x, z);
nP2 = GetHeightIndexAt(x + w, z);
nP3 = GetHeightIndexAt(x, z + w);
nP4 = GetHeightIndexAt(x + w, z + w);
nPmid = GetHeightIndexAt(x + w/2, z + w/2);
// Value of midpoint is calculated by averagint the four corner values, plus a random
// amount
m_pHeightMap[nPmid] = RandomInRange(-dH, dH)
+ ((m_pHeightMap[nP1] + m_pHeightMap[nP2] + m_pHeightMap[nP3] + m_pHeightMap[nP4]) * 0.25f);
fMinH = std::min(fMinH, m_pHeightMap[nPmid]);
fMaxH = std::max(fMaxH, m_pHeightMap[nPmid]);
}
}
// Square stage
for (z = 0; z < m_nSize; z += w)
{
for (x = 0; x < m_nSize; x += w)
{
nP1 = GetHeightIndexAt(x, z);
nP2 = GetHeightIndexAt(x + w, z);
nP3 = GetHeightIndexAt(x + w/2, z - w/2);
nP4 = GetHeightIndexAt(x + w/2, z + w/2);
nPmid = GetHeightIndexAt(x + w/2, z);
m_pHeightMap[nPmid] = RandomInRange(-dH, dH)
+ ((m_pHeightMap[nP1] + m_pHeightMap[nP2] + m_pHeightMap[nP3] + m_pHeightMap[nP4]) * 0.25f);
fMinH = std::min(fMinH, m_pHeightMap[nPmid]);
fMaxH = std::max(fMaxH, m_pHeightMap[nPmid]);
nP1 = GetHeightIndexAt(x, z);
nP2 = GetHeightIndexAt(x, z + w);
nP3 = GetHeightIndexAt(x + w/2, z + w/2);
nP4 = GetHeightIndexAt(x - w/2, z + w/2);
nPmid = GetHeightIndexAt(x, z + w/2);
m_pHeightMap[nPmid] = RandomInRange(-dH, dH)
+ ((m_pHeightMap[nP1] + m_pHeightMap[nP2] + m_pHeightMap[nP3] + m_pHeightMap[nP4]) * 0.25f);
fMinH = std::min(fMinH, m_pHeightMap[nPmid]);
fMaxH = std::max(fMaxH, m_pHeightMap[nPmid]);
}
}
}
// Normalize height field so altitudes fall into range [0,255].
for (int i = 0; i < m_nSize * m_nSize; ++i)
m_pHeightMap[i] = 255.0f * (m_pHeightMap[i] - fMinH) / (fMaxH - fMinH);
}
void Match_Global(ctrl_t *ctrl, graph_t *graph)
{
idx_t h, i, ii, j, k;
idx_t nnbrs, nvtxs, ncon, cnvtxs, firstvtx, lastvtx, maxi, maxidx, nkept;
idx_t otherlastvtx, nrequests, nchanged, pass, nmatched, wside;
idx_t *xadj, *adjncy, *adjwgt, *vtxdist, *home, *myhome;
idx_t *match;
idx_t *peind, *sendptr, *recvptr;
idx_t *perm, *iperm, *nperm, *changed;
real_t *nvwgt, maxnvwgt;
idx_t *nreqs_pe;
ikv_t *match_requests, *match_granted, *pe_requests;
idx_t last_unmatched;
WCOREPUSH;
maxnvwgt = 0.75/((real_t)(ctrl->CoarsenTo));
graph->match_type = PARMETIS_MTYPE_GLOBAL;
IFSET(ctrl->dbglvl, DBG_TIME, gkMPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->MatchTmr));
nvtxs = graph->nvtxs;
ncon = graph->ncon;
xadj = graph->xadj;
adjncy = graph->adjncy;
adjwgt = graph->adjwgt;
home = graph->home;
nvwgt = graph->nvwgt;
vtxdist = graph->vtxdist;
firstvtx = vtxdist[ctrl->mype];
lastvtx = vtxdist[ctrl->mype+1];
nnbrs = graph->nnbrs;
peind = graph->peind;
sendptr = graph->sendptr;
recvptr = graph->recvptr;
match = graph->match = ismalloc(nvtxs+graph->nrecv, UNMATCHED, "GlobalMatch: match");
/* wspacemalloc'ed arrays */
myhome = iset(nvtxs+graph->nrecv, UNMATCHED, iwspacemalloc(ctrl, nvtxs+graph->nrecv));
nreqs_pe = iset(nnbrs, 0, iwspacemalloc(ctrl, nnbrs));
perm = iwspacemalloc(ctrl, nvtxs);
iperm = iwspacemalloc(ctrl, nvtxs);
nperm = iwspacemalloc(ctrl, nnbrs);
changed = iwspacemalloc(ctrl, nvtxs);
match_requests = ikvwspacemalloc(ctrl, graph->nsend);
match_granted = ikvwspacemalloc(ctrl, graph->nrecv);
/* create the traversal order */
FastRandomPermute(nvtxs, perm, 1);
for (i=0; i<nvtxs; i++)
iperm[perm[i]] = i;
iincset(nnbrs, 0, nperm);
/* if coasening for adaptive/repartition, exchange home information */
if (ctrl->partType == ADAPTIVE_PARTITION || ctrl->partType == REFINE_PARTITION) {
PASSERT(ctrl, home != NULL);
icopy(nvtxs, home, myhome);
CommInterfaceData(ctrl, graph, myhome, myhome+nvtxs);
}
/* if coarsening for ordering, replace home with where information */
if (ctrl->partType == ORDER_PARTITION) {
PASSERT(ctrl, graph->where != NULL);
icopy(nvtxs, graph->where, myhome);
CommInterfaceData(ctrl, graph, myhome, myhome+nvtxs);
}
/* mark all heavy vertices as TOO_HEAVY so they will not be matched */
for (nchanged=i=0; i<nvtxs; i++) {
for (h=0; h<ncon; h++) {
if (nvwgt[i*ncon+h] > maxnvwgt) {
match[i] = TOO_HEAVY;
nchanged++;
break;
}
}
}
/* If no heavy vertices, pick one at random and treat it as such so that
at the end of the matching each partition will still have one vertex.
This is to eliminate the cases in which once a matching has been
computed, a processor ends up having no vertices */
if (nchanged == 0)
match[RandomInRange(nvtxs)] = TOO_HEAVY;
CommInterfaceData(ctrl, graph, match, match+nvtxs);
/* set initial value of nkept based on how over/under weight the
partition is to begin with */
nkept = graph->gnvtxs/ctrl->npes - nvtxs;
/* Find a matching by doing multiple iterations */
for (nmatched=pass=0; pass<NMATCH_PASSES; pass++) {
wside = (graph->level+pass)%2;
nchanged = nrequests = 0;
for (last_unmatched=ii=nmatched; ii<nvtxs; ii++) {
i = perm[ii];
if (match[i] == UNMATCHED) { /* Unmatched */
maxidx = i;
maxi = -1;
/* Deal with islands. Find another vertex and match it with */
if (xadj[i] == xadj[i+1]) {
last_unmatched = gk_max(ii, last_unmatched)+1;
for (; last_unmatched<nvtxs; last_unmatched++) {
k = perm[last_unmatched];
if (match[k] == UNMATCHED && myhome[i] == myhome[k]) {
match[i] = firstvtx+k + (i <= k ? KEEP_BIT : 0);
match[k] = firstvtx+i + (i > k ? KEEP_BIT : 0);
changed[nchanged++] = i;
changed[nchanged++] = k;
break;
}
}
continue;
}
/* Find a heavy-edge matching. */
for (j=xadj[i]; j<xadj[i+1]; j++) {
k = adjncy[j];
if (match[k] == UNMATCHED && myhome[k] == myhome[i]) {
if (ncon == 1) {
if (maxi == -1 || adjwgt[maxi] < adjwgt[j] ||
(adjwgt[maxi] == adjwgt[j] && RandomInRange(xadj[i+1]-xadj[i]) == 0)) {
maxi = j;
maxidx = k;
}
}
else {
if (maxi == -1 || adjwgt[maxi] < adjwgt[j] ||
(adjwgt[maxi] == adjwgt[j] && maxidx < nvtxs && k < nvtxs &&
BetterVBalance(ncon,nvwgt+i*ncon,nvwgt+maxidx*ncon,nvwgt+k*ncon) >= 0)) {
maxi = j;
maxidx = k;
}
}
}
}
if (maxi != -1) {
k = adjncy[maxi];
if (k < nvtxs) { /* Take care the local vertices first */
/* Here we give preference the local matching by granting it right away */
match[i] = firstvtx+k + (i <= k ? KEEP_BIT : 0);
match[k] = firstvtx+i + (i > k ? KEEP_BIT : 0);
changed[nchanged++] = i;
changed[nchanged++] = k;
}
else { /* Take care any remote boundary vertices */
match[k] = MAYBE_MATCHED;
/* Alternate among which vertices will issue the requests */
if ((wside == 0 && firstvtx+i < graph->imap[k]) ||
(wside == 1 && firstvtx+i > graph->imap[k])) {
match[i] = MAYBE_MATCHED;
match_requests[nrequests].key = graph->imap[k];
match_requests[nrequests].val = firstvtx+i;
nrequests++;
}
}
}
}
}
#ifdef DEBUG_MATCH
PrintVector2(ctrl, nvtxs, firstvtx, match, "Match1");
myprintf(ctrl, "[c: %2"PRIDX"] Nlocal: %"PRIDX", Nrequests: %"PRIDX"\n", c, nlocal, nrequests);
#endif
/***********************************************************
* Exchange the match_requests, requests for me are stored in
* match_granted
************************************************************/
/* Issue the receives first. Note that from each PE can receive a maximum
of the interface node that it needs to send it in the case of a mat-vec */
for (i=0; i<nnbrs; i++) {
gkMPI_Irecv((void *)(match_granted+recvptr[i]), 2*(recvptr[i+1]-recvptr[i]), IDX_T,
peind[i], 1, ctrl->comm, ctrl->rreq+i);
}
/* Issue the sends next. This needs some work */
ikvsorti(nrequests, match_requests);
for (j=i=0; i<nnbrs; i++) {
otherlastvtx = vtxdist[peind[i]+1];
for (k=j; k<nrequests && match_requests[k].key < otherlastvtx; k++);
gkMPI_Isend((void *)(match_requests+j), 2*(k-j), IDX_T, peind[i], 1,
ctrl->comm, ctrl->sreq+i);
j = k;
}
/* OK, now get into the loop waiting for the operations to finish */
gkMPI_Waitall(nnbrs, ctrl->rreq, ctrl->statuses);
for (i=0; i<nnbrs; i++) {
gkMPI_Get_count(ctrl->statuses+i, IDX_T, nreqs_pe+i);
nreqs_pe[i] = nreqs_pe[i]/2; /* Adjust for pairs of IDX_T */
}
gkMPI_Waitall(nnbrs, ctrl->sreq, ctrl->statuses);
/***********************************************************
* Now, go and service the requests that you received in
* match_granted
************************************************************/
RandomPermute(nnbrs, nperm, 0);
for (ii=0; ii<nnbrs; ii++) {
i = nperm[ii];
pe_requests = match_granted+recvptr[i];
for (j=0; j<nreqs_pe[i]; j++) {
k = pe_requests[j].key;
PASSERTP(ctrl, k >= firstvtx && k < lastvtx, (ctrl, "%"PRIDX" %"PRIDX" %"PRIDX" %"PRIDX" %"PRIDX"\n", firstvtx, lastvtx, k, j, peind[i]));
/* myprintf(ctrl, "Requesting a match %"PRIDX" %"PRIDX"\n", pe_requests[j].key, pe_requests[j].val); */
if (match[k-firstvtx] == UNMATCHED) { /* Bingo, lets grant this request */
changed[nchanged++] = k-firstvtx;
if (nkept >= 0) { /* decide who to keep it based on local balance */
match[k-firstvtx] = pe_requests[j].val + KEEP_BIT;
nkept--;
}
else {
match[k-firstvtx] = pe_requests[j].val;
pe_requests[j].key += KEEP_BIT;
nkept++;
}
/* myprintf(ctrl, "Request from pe:%"PRIDX" (%"PRIDX" %"PRIDX") granted!\n", peind[i], pe_requests[j].val, pe_requests[j].key); */
}
else { /* We are not granting the request */
/* myprintf(ctrl, "Request from pe:%"PRIDX" (%"PRIDX" %"PRIDX") not granted!\n", peind[i], pe_requests[j].val, pe_requests[j].key); */
pe_requests[j].key = UNMATCHED;
}
}
}
/***********************************************************
* Exchange the match_granted information. It is stored in
* match_requests
************************************************************/
/* Issue the receives first. Note that from each PE can receive a maximum
of the interface node that it needs to send during the case of a mat-vec */
for (i=0; i<nnbrs; i++) {
gkMPI_Irecv((void *)(match_requests+sendptr[i]), 2*(sendptr[i+1]-sendptr[i]), IDX_T,
peind[i], 1, ctrl->comm, ctrl->rreq+i);
}
/* Issue the sends next. */
for (i=0; i<nnbrs; i++) {
gkMPI_Isend((void *)(match_granted+recvptr[i]), 2*nreqs_pe[i], IDX_T,
peind[i], 1, ctrl->comm, ctrl->sreq+i);
}
/* OK, now get into the loop waiting for the operations to finish */
gkMPI_Waitall(nnbrs, ctrl->rreq, ctrl->statuses);
for (i=0; i<nnbrs; i++) {
gkMPI_Get_count(ctrl->statuses+i, IDX_T, nreqs_pe+i);
nreqs_pe[i] = nreqs_pe[i]/2; /* Adjust for pairs of IDX_T */
}
gkMPI_Waitall(nnbrs, ctrl->sreq, ctrl->statuses);
/***********************************************************
* Now, go and through the match_requests and update local
* match information for the matchings that were granted.
************************************************************/
for (i=0; i<nnbrs; i++) {
pe_requests = match_requests+sendptr[i];
for (j=0; j<nreqs_pe[i]; j++) {
match[pe_requests[j].val-firstvtx] = pe_requests[j].key;
if (pe_requests[j].key != UNMATCHED)
changed[nchanged++] = pe_requests[j].val-firstvtx;
}
}
for (i=0; i<nchanged; i++) {
ii = iperm[changed[i]];
perm[ii] = perm[nmatched];
iperm[perm[nmatched]] = ii;
nmatched++;
}
CommChangedInterfaceData(ctrl, graph, nchanged, changed, match,
match_requests, match_granted);
}
/* Traverse the vertices and those that were unmatched, match them with themselves */
cnvtxs = 0;
for (i=0; i<nvtxs; i++) {
if (match[i] == UNMATCHED || match[i] == TOO_HEAVY) {
match[i] = (firstvtx+i) + KEEP_BIT;
cnvtxs++;
}
else if (match[i] >= KEEP_BIT) { /* A matched vertex which I get to keep */
cnvtxs++;
}
}
if (ctrl->dbglvl&DBG_MATCHINFO) {
PrintVector2(ctrl, nvtxs, firstvtx, match, "Match");
myprintf(ctrl, "Cnvtxs: %"PRIDX"\n", cnvtxs);
rprintf(ctrl, "Done with matching...\n");
}
WCOREPOP;
IFSET(ctrl->dbglvl, DBG_TIME, gkMPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->MatchTmr));
IFSET(ctrl->dbglvl, DBG_TIME, starttimer(ctrl->ContractTmr));
CreateCoarseGraph_Global(ctrl, graph, cnvtxs);
IFSET(ctrl->dbglvl, DBG_TIME, gkMPI_Barrier(ctrl->comm));
IFSET(ctrl->dbglvl, DBG_TIME, stoptimer(ctrl->ContractTmr));
}
