/* =============================================================================
* TMheap_insert
* -- Returns false on failure
* =============================================================================
*/
bool
TMheap_insert (TM_ARGDECL heap_t* heapPtr, void* dataPtr)
{
long size = (long)TM_SHARED_READ_L(heapPtr->size);
long capacity = (long)TM_SHARED_READ_L(heapPtr->capacity);
if ((size + 1) >= capacity) {
long newCapacity = capacity * 2;
void** newElements = (void**)TM_MALLOC(newCapacity * sizeof(void*));
if (newElements == NULL) {
return false;
}
TM_SHARED_WRITE_L(heapPtr->capacity, newCapacity);
long i;
void** elements = TM_SHARED_READ_P(heapPtr->elements);
for (i = 0; i <= size; i++) {
newElements[i] = (void*)TM_SHARED_READ_P(elements[i]);
}
TM_FREE(elements);
TM_SHARED_WRITE_P(heapPtr->elements, newElements);
}
size++;
TM_SHARED_WRITE_L(heapPtr->size, size);
void** elements = (void**)TM_SHARED_READ_P(heapPtr->elements);
TM_SHARED_WRITE_P(elements[size], dataPtr);
TMsiftUp(TM_ARG heapPtr, size);
return true;
}
/* =============================================================================
* TMlist_insert
* -- Return TRUE on success, else FALSE
* =============================================================================
*/
bool_t
TMlist_insert (TM_ARGDECL list_t* listPtr, void* dataPtr)
{
list_node_t* prevPtr;
list_node_t* nodePtr;
list_node_t* currPtr;
prevPtr = TMfindPrevious(TM_ARG listPtr, dataPtr);
currPtr = (list_node_t*)TM_SHARED_READ_P(prevPtr->nextPtr);
#ifdef LIST_NO_DUPLICATES
if ((currPtr != NULL) &&
listPtr->comparator->compare_tm(TM_ARG TM_SHARED_READ_P(currPtr->dataPtr), dataPtr) == 0) {
return FALSE;
}
#endif
nodePtr = TMallocNode(TM_ARG dataPtr);
if (nodePtr == NULL) {
return FALSE;
}
TM_SHARED_WRITE_P(nodePtr->nextPtr, currPtr);
TM_SHARED_WRITE_P(prevPtr->nextPtr, nodePtr);
TM_SHARED_WRITE_L(listPtr->size, (TM_SHARED_READ_L(listPtr->size) + 1));
return TRUE;
}
/* =============================================================================
* TMgrid_addPath
* =============================================================================
*/
void
TMgrid_addPath (TM_ARGDECL grid_t* gridPtr, vector_t* pointVectorPtr)
{
long i;
long n = vector_getSize(pointVectorPtr);
for (i = 1; i < (n-1); i++) {
long* gridPointPtr = (long*)vector_at(pointVectorPtr, i);
long value = (long)TM_SHARED_READ_L(*gridPointPtr);
if (value != GRID_POINT_EMPTY) {
TM_RESTART();
}
TM_SHARED_WRITE_L(*gridPointPtr, (long)GRID_POINT_FULL);
}
}
/* =============================================================================
* TMheap_remove
* -- Returns NULL if empty
* =============================================================================
*/
void*
TMheap_remove (TM_ARGDECL heap_t* heapPtr)
{
long size = (long)TM_SHARED_READ_L(heapPtr->size);
if (size < 1) {
return NULL;
}
void** elements = (void**)TM_SHARED_READ_P(heapPtr->elements);
void* dataPtr = (void*)TM_SHARED_READ_P(elements[1]);
TM_SHARED_WRITE_P(elements[1], TM_SHARED_READ_P(elements[size]));
TM_SHARED_WRITE_L(heapPtr->size, (size - 1));
TMheapify(TM_ARG heapPtr, 1);
return dataPtr;
}
/* =============================================================================
* TMlist_remove
* -- Returns TRUE if successful, else FALSE
* =============================================================================
*/
bool_t
TMlist_remove (TM_ARGDECL list_t* listPtr, void* dataPtr)
{
list_node_t* prevPtr;
list_node_t* nodePtr;
prevPtr = TMfindPrevious(TM_ARG listPtr, dataPtr);
nodePtr = (list_node_t*)TM_SHARED_READ_P(prevPtr->nextPtr);
if ((nodePtr != NULL) &&
(listPtr->comparator->compare_tm(TM_ARG TM_SHARED_READ_P(nodePtr->dataPtr), dataPtr) == 0))
{
TM_SHARED_WRITE_P(prevPtr->nextPtr, TM_SHARED_READ_P(nodePtr->nextPtr));
TM_SHARED_WRITE_P(nodePtr->nextPtr, (struct list_node*)NULL);
TMfreeNode(TM_ARG nodePtr);
TM_SHARED_WRITE_L(listPtr->size, (TM_SHARED_READ_L(listPtr->size) - 1));
assert(listPtr->size >= 0);
return TRUE;
}
return FALSE;
}
/* =============================================================================
* work
* =============================================================================
*/
static void
work (void* argPtr)
{
TM_THREAD_ENTER();
args_t* args = (args_t*)argPtr;
float** feature = args->feature;
int nfeatures = args->nfeatures;
int npoints = args->npoints;
int nclusters = args->nclusters;
int* membership = args->membership;
float** clusters = args->clusters;
int** new_centers_len = args->new_centers_len;
float** new_centers = args->new_centers;
float delta = 0.0;
int index;
long i;
int j;
int start;
int stop;
int myId;
myId = thread_getId();
start = myId * CHUNK;
while (start < npoints) {
stop = (((start + CHUNK) < npoints) ? (start + CHUNK) : npoints);
for (i = start; i < stop; TMHT_LOCAL_WRITE(i, i+1)) {
index = common_findNearestPoint(feature[i],
nfeatures,
clusters,
nclusters);
/*
* If membership changes, increase delta by 1.
* membership[i] cannot be changed by other threads
*/
if (membership[i] != index) {
delta += 1.0;
}
/* Assign the membership to object i */
/* membership[i] can't be changed by other thread */
membership[i] = index;
/* Update new cluster centers : sum of objects located within */
TM_BEGIN();
TM_SHARED_WRITE_I(*new_centers_len[index],
TM_SHARED_READ_I(*new_centers_len[index]) + 1);
for (j = 0; j < nfeatures; j++) {
TM_SHARED_WRITE_F(
new_centers[index][j],
(TM_SHARED_READ_F(new_centers[index][j]) + feature[i][j])
);
}
TM_END();
}
/* Update task queue */
if (start + CHUNK < npoints) {
TM_BEGIN();
start = (int)TM_SHARED_READ_L(global_i);
TM_SHARED_WRITE_L(global_i, (long)(start + CHUNK));
TM_END();
} else {
break;
}
}
TM_BEGIN();
TM_SHARED_WRITE_F(global_delta, TM_SHARED_READ_F(global_delta) + delta);
TM_END();
TM_THREAD_EXIT();
}
/* =============================================================================
* sequencer_run
* =============================================================================
*/
void
sequencer_run (void* argPtr)
{
TM_THREAD_ENTER();
long threadId = thread_getId();
sequencer_t* sequencerPtr = (sequencer_t*)argPtr;
hashtable_t* uniqueSegmentsPtr;
endInfoEntry_t* endInfoEntries;
table_t** startHashToConstructEntryTables;
constructEntry_t* constructEntries;
table_t* hashToConstructEntryTable;
uniqueSegmentsPtr = sequencerPtr->uniqueSegmentsPtr;
endInfoEntries = sequencerPtr->endInfoEntries;
startHashToConstructEntryTables = sequencerPtr->startHashToConstructEntryTables;
constructEntries = sequencerPtr->constructEntries;
hashToConstructEntryTable = sequencerPtr->hashToConstructEntryTable;
segments_t* segmentsPtr = sequencerPtr->segmentsPtr;
assert(segmentsPtr);
vector_t* segmentsContentsPtr = segmentsPtr->contentsPtr;
long numSegment = vector_getSize(segmentsContentsPtr);
long segmentLength = segmentsPtr->length;
long i;
long j;
long i_start;
long i_stop;
long numUniqueSegment;
long substringLength;
long entryIndex;
/*
* Step 1: Remove duplicate segments
*/
#if defined(HTM) || defined(STM)
long numThread = thread_getNumThread();
{
/* Choose disjoint segments [i_start,i_stop) for each thread */
long partitionSize = (numSegment + numThread/2) / numThread; /* with rounding */
i_start = threadId * partitionSize;
if (threadId == (numThread - 1)) {
i_stop = numSegment;
} else {
i_stop = i_start + partitionSize;
}
}
#else /* !(HTM || STM) */
i_start = 0;
i_stop = numSegment;
#endif /* !(HTM || STM) */
for (i = i_start; i < i_stop; i+=CHUNK_STEP1) {
TM_BEGIN();
{
long ii;
long ii_stop = MIN(i_stop, (i+CHUNK_STEP1));
for (ii = i; ii < ii_stop; ii++) {
void* segment = vector_at(segmentsContentsPtr, ii);
TMHASHTABLE_INSERT(uniqueSegmentsPtr,
segment,
segment);
} /* ii */
}
TM_END();
}
thread_barrier_wait();
/*
* Step 2a: Iterate over unique segments and compute hashes.
*
* For the gene "atcg", the hashes for the end would be:
*
* "t", "tc", and "tcg"
*
* And for the gene "tcgg", the hashes for the start would be:
*
* "t", "tc", and "tcg"
*
* The names are "end" and "start" because if a matching pair is found,
* they are the substring of the end part of the pair and the start
* part of the pair respectively. In the above example, "tcg" is the
* matching substring so:
*
* (end) (start)
* a[tcg] + [tcg]g = a[tcg]g (overlap = "tcg")
*/
/* uniqueSegmentsPtr is constant now */
numUniqueSegment = hashtable_getSize(uniqueSegmentsPtr);
entryIndex = 0;
#if defined(HTM) || defined(STM)
{
/* Choose disjoint segments [i_start,i_stop) for each thread */
long num = uniqueSegmentsPtr->numBucket;
long partitionSize = (num + numThread/2) / numThread; /* with rounding */
i_start = threadId * partitionSize;
if (threadId == (numThread - 1)) {
i_stop = num;
} else {
i_stop = i_start + partitionSize;
}
}
{
/* Approximate disjoint segments of element allocation in constructEntries */
long partitionSize = (numUniqueSegment + numThread/2) / numThread; /* with rounding */
entryIndex = threadId * partitionSize;
}
#else /* !(HTM || STM) */
i_start = 0;
i_stop = uniqueSegmentsPtr->numBucket;
entryIndex = 0;
#endif /* !(HTM || STM) */
for (i = i_start; i < i_stop; i++) {
list_t* chainPtr = uniqueSegmentsPtr->buckets[i];
list_iter_t it;
list_iter_reset(&it, chainPtr);
while (list_iter_hasNext(&it, chainPtr)) {
char* segment =
(char*)((pair_t*)list_iter_next(&it, chainPtr))->firstPtr;
constructEntry_t* constructEntryPtr;
long j;
unsigned long startHash;
bool status;
/* Find an empty constructEntries entry */
TM_BEGIN();
while (((void*)TM_SHARED_READ_P(constructEntries[entryIndex].segment)) != NULL) {
entryIndex = (entryIndex + 1) % numUniqueSegment; /* look for empty */
}
constructEntryPtr = &constructEntries[entryIndex];
TM_SHARED_WRITE_P(constructEntryPtr->segment, segment);
TM_END();
entryIndex = (entryIndex + 1) % numUniqueSegment;
/*
* Save hashes (sdbm algorithm) of segment substrings
*
* endHashes will be computed for shorter substrings after matches
* have been made (in the next phase of the code). This will reduce
* the number of substrings for which hashes need to be computed.
*
* Since we can compute startHashes incrementally, we go ahead
* and compute all of them here.
*/
/* constructEntryPtr is local now */
constructEntryPtr->endHash = hashString(&segment[1]);
startHash = 0;
for (j = 1; j < segmentLength; j++) {
startHash = (unsigned long)segment[j-1] +
(startHash << 6) + (startHash << 16) - startHash;
TM_BEGIN();
status = TMTABLE_INSERT(startHashToConstructEntryTables[j],
(unsigned long)startHash,
(void*)constructEntryPtr );
TM_END();
assert(status);
}
/*
* For looking up construct entries quickly
*/
startHash = (unsigned long)segment[j-1] +
(startHash << 6) + (startHash << 16) - startHash;
TM_BEGIN();
status = TMTABLE_INSERT(hashToConstructEntryTable,
(unsigned long)startHash,
(void*)constructEntryPtr);
TM_END();
assert(status);
}
}
thread_barrier_wait();
/*
* Step 2b: Match ends to starts by using hash-based string comparison.
*/
for (substringLength = segmentLength-1; substringLength > 0; substringLength--) {
table_t* startHashToConstructEntryTablePtr =
startHashToConstructEntryTables[substringLength];
list_t** buckets = startHashToConstructEntryTablePtr->buckets;
long numBucket = startHashToConstructEntryTablePtr->numBucket;
long index_start;
long index_stop;
#if defined(HTM) || defined(STM)
{
/* Choose disjoint segments [index_start,index_stop) for each thread */
long partitionSize = (numUniqueSegment + numThread/2) / numThread; /* with rounding */
index_start = threadId * partitionSize;
if (threadId == (numThread - 1)) {
index_stop = numUniqueSegment;
} else {
index_stop = index_start + partitionSize;
}
}
#else /* !(HTM || STM) */
index_start = 0;
index_stop = numUniqueSegment;
#endif /* !(HTM || STM) */
/* Iterating over disjoint itervals in the range [0, numUniqueSegment) */
for (entryIndex = index_start;
entryIndex < index_stop;
entryIndex += endInfoEntries[entryIndex].jumpToNext)
{
if (!endInfoEntries[entryIndex].isEnd) {
continue;
}
/* ConstructEntries[entryIndex] is local data */
constructEntry_t* endConstructEntryPtr =
&constructEntries[entryIndex];
char* endSegment = endConstructEntryPtr->segment;
unsigned long endHash = endConstructEntryPtr->endHash;
list_t* chainPtr = buckets[endHash % numBucket]; /* buckets: constant data */
list_iter_t it;
list_iter_reset(&it, chainPtr);
/* Linked list at chainPtr is constant */
while (list_iter_hasNext(&it, chainPtr)) {
constructEntry_t* startConstructEntryPtr =
(constructEntry_t*)list_iter_next(&it, chainPtr);
char* startSegment = startConstructEntryPtr->segment;
long newLength = 0;
/* endConstructEntryPtr is local except for properties startPtr/endPtr/length */
TM_BEGIN();
/* Check if matches */
if (TM_SHARED_READ_L(startConstructEntryPtr->isStart) &&
(TM_SHARED_READ_P(endConstructEntryPtr->startPtr) != startConstructEntryPtr) &&
(strncmp(startSegment,
&endSegment[segmentLength - substringLength],
substringLength) == 0))
{
TM_SHARED_WRITE_L(startConstructEntryPtr->isStart, false);
constructEntry_t* startConstructEntry_endPtr;
constructEntry_t* endConstructEntry_startPtr;
/* Update endInfo (appended something so no longer end) */
TM_LOCAL_WRITE_L(endInfoEntries[entryIndex].isEnd, false);
/* Update segment chain construct info */
startConstructEntry_endPtr =
(constructEntry_t*)TM_SHARED_READ_P(startConstructEntryPtr->endPtr);
endConstructEntry_startPtr =
(constructEntry_t*)TM_SHARED_READ_P(endConstructEntryPtr->startPtr);
assert(startConstructEntry_endPtr);
assert(endConstructEntry_startPtr);
TM_SHARED_WRITE_P(startConstructEntry_endPtr->startPtr,
endConstructEntry_startPtr);
TM_LOCAL_WRITE_P(endConstructEntryPtr->nextPtr,
startConstructEntryPtr);
TM_SHARED_WRITE_P(endConstructEntry_startPtr->endPtr,
startConstructEntry_endPtr);
TM_SHARED_WRITE_L(endConstructEntryPtr->overlap, substringLength);
newLength = (long)TM_SHARED_READ_L(endConstructEntry_startPtr->length) +
(long)TM_SHARED_READ_L(startConstructEntryPtr->length) -
substringLength;
TM_SHARED_WRITE_L(endConstructEntry_startPtr->length, newLength);
} /* if (matched) */
TM_END();
if (!endInfoEntries[entryIndex].isEnd) { /* if there was a match */
break;
}
} /* iterate over chain */
} /* for (endIndex < numUniqueSegment) */
thread_barrier_wait();
/*
* Step 2c: Update jump values and hashes
*
* endHash entries of all remaining ends are updated to the next
* substringLength. Additionally jumpToNext entries are updated such
* that they allow to skip non-end entries. Currently this is sequential
* because parallelization did not perform better.
. */
if (threadId == 0) {
if (substringLength > 1) {
long index = segmentLength - substringLength + 1;
/* initialization if j and i: with i being the next end after j=0 */
for (i = 1; !endInfoEntries[i].isEnd; i+=endInfoEntries[i].jumpToNext) {
/* find first non-null */
}
/* entry 0 is handled seperately from the loop below */
endInfoEntries[0].jumpToNext = i;
if (endInfoEntries[0].isEnd) {
constructEntry_t* constructEntryPtr = &constructEntries[0];
char* segment = constructEntryPtr->segment;
constructEntryPtr->endHash = hashString(&segment[index]);
}
/* Continue scanning (do not reset i) */
for (j = 0; i < numUniqueSegment; i+=endInfoEntries[i].jumpToNext) {
if (endInfoEntries[i].isEnd) {
constructEntry_t* constructEntryPtr = &constructEntries[i];
char* segment = constructEntryPtr->segment;
constructEntryPtr->endHash = hashString(&segment[index]);
endInfoEntries[j].jumpToNext = MAX(1, (i - j));
j = i;
}
}
endInfoEntries[j].jumpToNext = i - j;
}
}
thread_barrier_wait();
} /* for (substringLength > 0) */
thread_barrier_wait();
/*
* Step 3: Build sequence string
*/
if (threadId == 0) {
long totalLength = 0;
for (i = 0; i < numUniqueSegment; i++) {
constructEntry_t* constructEntryPtr = &constructEntries[i];
if (constructEntryPtr->isStart) {
totalLength += constructEntryPtr->length;
}
}
sequencerPtr->sequence = (char*)P_MALLOC((totalLength+1) * sizeof(char));
char* sequence = sequencerPtr->sequence;
assert(sequence);
char* copyPtr = sequence;
long sequenceLength = 0;
for (i = 0; i < numUniqueSegment; i++) {
constructEntry_t* constructEntryPtr = &constructEntries[i];
/* If there are several start segments, we append in arbitrary order */
if (constructEntryPtr->isStart) {
long newSequenceLength = sequenceLength + constructEntryPtr->length;
assert( newSequenceLength <= totalLength );
copyPtr = sequence + sequenceLength;
sequenceLength = newSequenceLength;
do {
long numChar = segmentLength - constructEntryPtr->overlap;
if ((copyPtr + numChar) > (sequence + newSequenceLength)) {
TM_PRINT0("ERROR: sequence length != actual length\n");
break;
}
memcpy(copyPtr,
constructEntryPtr->segment,
(numChar * sizeof(char)));
copyPtr += numChar;
} while ((constructEntryPtr = constructEntryPtr->nextPtr) != NULL);
assert(copyPtr <= (sequence + sequenceLength));
}
}
assert(sequence != NULL);
sequence[sequenceLength] = '\0';
}
TM_THREAD_EXIT();
}
/* =============================================================================
* computeGraph
* =============================================================================
*/
void
computeGraph (void* argPtr)
{
TM_THREAD_ENTER();
graph* GPtr = ((computeGraph_arg_t*)argPtr)->GPtr;
graphSDG* SDGdataPtr = ((computeGraph_arg_t*)argPtr)->SDGdataPtr;
long myId = thread_getId();
long numThread = thread_getNumThread();
ULONGINT_T j;
ULONGINT_T maxNumVertices = 0;
ULONGINT_T numEdgesPlaced = SDGdataPtr->numEdgesPlaced;
/*
* First determine the number of vertices by scanning the tuple
* startVertex list
*/
long i;
long i_start;
long i_stop;
createPartition(0, numEdgesPlaced, myId, numThread, &i_start, &i_stop);
for (i = i_start; i < i_stop; i++) {
if (SDGdataPtr->startVertex[i] > maxNumVertices) {
maxNumVertices = SDGdataPtr->startVertex[i];
}
}
TM_BEGIN();
long tmp_maxNumVertices = (long)TM_SHARED_READ_L(global_maxNumVertices);
long new_maxNumVertices = MAX(tmp_maxNumVertices, maxNumVertices) + 1;
TM_SHARED_WRITE_L(global_maxNumVertices, new_maxNumVertices);
TM_END();
thread_barrier_wait();
maxNumVertices = global_maxNumVertices;
if (myId == 0) {
GPtr->numVertices = maxNumVertices;
GPtr->numEdges = numEdgesPlaced;
GPtr->intWeight = SDGdataPtr->intWeight;
GPtr->strWeight = SDGdataPtr->strWeight;
for (i = 0; i < numEdgesPlaced; i++) {
if (GPtr->intWeight[numEdgesPlaced-i-1] < 0) {
GPtr->numStrEdges = -(GPtr->intWeight[numEdgesPlaced-i-1]) + 1;
GPtr->numIntEdges = numEdgesPlaced - GPtr->numStrEdges;
break;
}
}
GPtr->outDegree =
(LONGINT_T*)P_MALLOC((GPtr->numVertices) * sizeof(LONGINT_T));
assert(GPtr->outDegree);
GPtr->outVertexIndex =
(ULONGINT_T*)P_MALLOC((GPtr->numVertices) * sizeof(ULONGINT_T));
assert(GPtr->outVertexIndex);
}
thread_barrier_wait();
createPartition(0, GPtr->numVertices, myId, numThread, &i_start, &i_stop);
for (i = i_start; i < i_stop; i++) {
GPtr->outDegree[i] = 0;
GPtr->outVertexIndex[i] = 0;
}
ULONGINT_T outVertexListSize = 0;
thread_barrier_wait();
ULONGINT_T i0 = -1UL;
for (i = i_start; i < i_stop; i++) {
ULONGINT_T k = i;
if ((outVertexListSize == 0) && (k != 0)) {
while (i0 == -1UL) {
for (j = 0; j < numEdgesPlaced; j++) {
if (k == SDGdataPtr->startVertex[j]) {
i0 = j;
break;
}
}
k--;
}
}
if ((outVertexListSize == 0) && (k == 0)) {
i0 = 0;
}
for (j = i0; j < numEdgesPlaced; j++) {
if (i == GPtr->numVertices-1) {
break;
}
if ((i != SDGdataPtr->startVertex[j])) {
if ((j > 0) && (i == SDGdataPtr->startVertex[j-1])) {
if (j-i0 >= 1) {
outVertexListSize++;
GPtr->outDegree[i]++;
ULONGINT_T t;
for (t = i0+1; t < j; t++) {
if (SDGdataPtr->endVertex[t] !=
SDGdataPtr->endVertex[t-1])
{
outVertexListSize++;
GPtr->outDegree[i] = GPtr->outDegree[i]+1;
}
}
}
}
i0 = j;
break;
}
}
if (i == GPtr->numVertices-1) {
if (numEdgesPlaced-i0 >= 0) {
outVertexListSize++;
GPtr->outDegree[i]++;
ULONGINT_T t;
for (t = i0+1; t < numEdgesPlaced; t++) {
if (SDGdataPtr->endVertex[t] != SDGdataPtr->endVertex[t-1]) {
outVertexListSize++;
GPtr->outDegree[i]++;
}
}
}
}
} /* for i */
thread_barrier_wait();
prefix_sums(GPtr->outVertexIndex, GPtr->outDegree, GPtr->numVertices);
thread_barrier_wait();
TM_BEGIN();
TM_SHARED_WRITE_L(
global_outVertexListSize,
((long)TM_SHARED_READ_L(global_outVertexListSize) + outVertexListSize)
);
TM_END();
thread_barrier_wait();
outVertexListSize = global_outVertexListSize;
if (myId == 0) {
GPtr->numDirectedEdges = outVertexListSize;
GPtr->outVertexList =
(ULONGINT_T*)P_MALLOC(outVertexListSize * sizeof(ULONGINT_T));
assert(GPtr->outVertexList);
GPtr->paralEdgeIndex =
(ULONGINT_T*)P_MALLOC(outVertexListSize * sizeof(ULONGINT_T));
assert(GPtr->paralEdgeIndex);
GPtr->outVertexList[0] = SDGdataPtr->endVertex[0];
}
thread_barrier_wait();
/*
* Evaluate outVertexList
*/
i0 = -1UL;
for (i = i_start; i < i_stop; i++) {
ULONGINT_T k = i;
while ((i0 == -1UL) && (k != 0)) {
for (j = 0; j < numEdgesPlaced; j++) {
if (k == SDGdataPtr->startVertex[j]) {
i0 = j;
break;
}
}
k--;
}
if ((i0 == -1) && (k == 0)) {
i0 = 0;
}
for (j = i0; j < numEdgesPlaced; j++) {
if (i == GPtr->numVertices-1) {
break;
}
if (i != SDGdataPtr->startVertex[j]) {
if ((j > 0) && (i == SDGdataPtr->startVertex[j-1])) {
if (j-i0 >= 1) {
long ii = GPtr->outVertexIndex[i];
ULONGINT_T r = 0;
GPtr->paralEdgeIndex[ii] = i0;
GPtr->outVertexList[ii] = SDGdataPtr->endVertex[i0];
r++;
ULONGINT_T t;
for (t = i0+1; t < j; t++) {
if (SDGdataPtr->endVertex[t] !=
SDGdataPtr->endVertex[t-1])
{
GPtr->paralEdgeIndex[ii+r] = t;
GPtr->outVertexList[ii+r] = SDGdataPtr->endVertex[t];
r++;
}
}
}
}
i0 = j;
break;
}
} /* for j */
if (i == GPtr->numVertices-1) {
ULONGINT_T r = 0;
if (numEdgesPlaced-i0 >= 0) {
long ii = GPtr->outVertexIndex[i];
GPtr->paralEdgeIndex[ii+r] = i0;
GPtr->outVertexList[ii+r] = SDGdataPtr->endVertex[i0];
r++;
ULONGINT_T t;
for (t = i0+1; t < numEdgesPlaced; t++) {
if (SDGdataPtr->endVertex[t] != SDGdataPtr->endVertex[t-1]) {
GPtr->paralEdgeIndex[ii+r] = t;
GPtr->outVertexList[ii+r] = SDGdataPtr->endVertex[t];
r++;
}
}
}
}
} /* for i */
thread_barrier_wait();
if (myId == 0) {
P_FREE(SDGdataPtr->startVertex);
P_FREE(SDGdataPtr->endVertex);
GPtr->inDegree =
(LONGINT_T*)P_MALLOC(GPtr->numVertices * sizeof(LONGINT_T));
assert(GPtr->inDegree);
GPtr->inVertexIndex =
(ULONGINT_T*)P_MALLOC(GPtr->numVertices * sizeof(ULONGINT_T));
assert(GPtr->inVertexIndex);
}
thread_barrier_wait();
for (i = i_start; i < i_stop; i++) {
GPtr->inDegree[i] = 0;
GPtr->inVertexIndex[i] = 0;
}
/* A temp. array to store the inplied edges */
ULONGINT_T* impliedEdgeList;
if (myId == 0) {
impliedEdgeList = (ULONGINT_T*)P_MALLOC(GPtr->numVertices
* MAX_CLUSTER_SIZE
* sizeof(ULONGINT_T));
global_impliedEdgeList = impliedEdgeList;
}
thread_barrier_wait();
impliedEdgeList = global_impliedEdgeList;
createPartition(0,
(GPtr->numVertices * MAX_CLUSTER_SIZE),
myId,
numThread,
&i_start,
&i_stop);
for (i = i_start; i < i_stop; i++) {
impliedEdgeList[i] = 0;
}
/*
* An auxiliary array to store implied edges, in case we overshoot
* MAX_CLUSTER_SIZE
*/
ULONGINT_T** auxArr;
if (myId == 0) {
auxArr = (ULONGINT_T**)P_MALLOC(GPtr->numVertices * sizeof(ULONGINT_T*));
assert(auxArr);
global_auxArr = auxArr;
}
thread_barrier_wait();
auxArr = global_auxArr;
createPartition(0, GPtr->numVertices, myId, numThread, &i_start, &i_stop);
for (i = i_start; i < i_stop; i++) {
/* Inspect adjacency list of vertex i */
for (j = GPtr->outVertexIndex[i];
j < (GPtr->outVertexIndex[i] + GPtr->outDegree[i]);
j++)
{
ULONGINT_T v = GPtr->outVertexList[j];
ULONGINT_T k;
for (k = GPtr->outVertexIndex[v];
k < (GPtr->outVertexIndex[v] + GPtr->outDegree[v]);
k++)
{
if (GPtr->outVertexList[k] == i) {
break;
}
}
if (k == GPtr->outVertexIndex[v]+GPtr->outDegree[v]) {
TM_BEGIN();
/* Add i to the impliedEdgeList of v */
long inDegree = (long)TM_SHARED_READ_L(GPtr->inDegree[v]);
TM_SHARED_WRITE_L(GPtr->inDegree[v], (inDegree + 1));
if (inDegree < MAX_CLUSTER_SIZE) {
TM_SHARED_WRITE_L(impliedEdgeList[v*MAX_CLUSTER_SIZE+inDegree],
i);
} else {
/* Use auxiliary array to store the implied edge */
/* Create an array if it's not present already */
ULONGINT_T* a = NULL;
if ((inDegree % MAX_CLUSTER_SIZE) == 0) {
a = (ULONGINT_T*)TM_MALLOC(MAX_CLUSTER_SIZE
* sizeof(ULONGINT_T));
assert(a);
TM_SHARED_WRITE_P(auxArr[v], a);
} else {
a = auxArr[v];
}
TM_SHARED_WRITE_L(a[inDegree % MAX_CLUSTER_SIZE], i);
}
TM_END();
}
}
} /* for i */
thread_barrier_wait();
prefix_sums(GPtr->inVertexIndex, GPtr->inDegree, GPtr->numVertices);
if (myId == 0) {
GPtr->numUndirectedEdges = GPtr->inVertexIndex[GPtr->numVertices-1]
+ GPtr->inDegree[GPtr->numVertices-1];
GPtr->inVertexList =
(ULONGINT_T *)P_MALLOC(GPtr->numUndirectedEdges * sizeof(ULONGINT_T));
}
thread_barrier_wait();
/*
* Create the inVertex List
*/
for (i = i_start; i < i_stop; i++) {
for (j = GPtr->inVertexIndex[i];
j < (GPtr->inVertexIndex[i] + GPtr->inDegree[i]);
j++)
{
if ((j - GPtr->inVertexIndex[i]) < MAX_CLUSTER_SIZE) {
GPtr->inVertexList[j] =
impliedEdgeList[i*MAX_CLUSTER_SIZE+j-GPtr->inVertexIndex[i]];
} else {
GPtr->inVertexList[j] =
auxArr[i][(j-GPtr->inVertexIndex[i]) % MAX_CLUSTER_SIZE];
}
}
}
thread_barrier_wait();
if (myId == 0) {
P_FREE(impliedEdgeList);
}
for (i = i_start; i < i_stop; i++) {
if (GPtr->inDegree[i] > MAX_CLUSTER_SIZE) {
P_FREE(auxArr[i]);
}
}
thread_barrier_wait();
if (myId == 0) {
P_FREE(auxArr);
}
TM_THREAD_EXIT();
}
void client_run (void* argPtr) {
TM_THREAD_ENTER();
/*long id = thread_getId();
volatile long* ptr1 = &(global_array[0].value);
volatile long* ptr2 = &(global_array[100].value);
long tt = 0;
if (id == 0) {
while (1) {
long v1 = 0;
long v2 = 0;
acquire_write(&(local_th_data[phys_id]), &the_lock);
*ptr1 = (*ptr1) + 1;
int f = 1;
int ii;
for(ii = 1; ii <= 100000000; ii++)
{
f *= ii;
}
tt += f;
*ptr2 = (*ptr2) + 1;
v1 = global_array[0].value;
v2 = global_array[100].value;
release_write(cluster_id, &(local_th_data[phys_id]), &the_lock); \
if (v1 != v2) {
printf("different2! %ld %ld\n", v1, v2);
exit(1);
}
}
} else {
while (1) {
int i = 0;
long sum = 0;
for (; i < 100000; i++) {
int status = _xbegin();
if (status == _XBEGIN_STARTED) {
sum += *ptr1;
sum += *ptr2;
_xend();
}
}
while(1) {
long v1 = 0;
long v2 = 0;
int status = _xbegin();
if (status == _XBEGIN_STARTED) {
v1 = *ptr1;
v2 = *ptr2;
_xend();
if (v1 != v2) {
printf("different! %ld %ld\n", v1, v2);
exit(1);
}
}
}
}
}
printf("%ld", tt);*/
random_t* randomPtr = random_alloc();
random_seed(randomPtr, time(0));
// unsigned long myId = thread_getId();
// long numThread = *((long*)argPtr);
long operations = (long)global_params[PARAM_OPERATIONS] / (long)global_params[PARAM_THREADS];
long interval = (long)global_params[PARAM_INTERVAL];
printf("operations: %ld \tinterval: %ld\n", operations, interval);
long total = 0;
long total2 = 0;
long i = 0;
for (; i < operations; i++) {
long random_number = ((long) random_generate(randomPtr)) % ((long)global_params[PARAM_SIZE]);
long random_number2 = ((long) random_generate(randomPtr)) % ((long)global_params[PARAM_SIZE]);
if (random_number == random_number2) {
random_number2 = (random_number2 + 1) % ((long)global_params[PARAM_SIZE]);
}
TM_BEGIN();
long r1 = (long)TM_SHARED_READ_L(global_array[random_number].value);
long r2 = (long)TM_SHARED_READ_L(global_array[random_number2].value);
int repeat = 0;
for (; repeat < (long) global_params[PARAM_CONTENTION]; repeat++) {
total2 += (long) TM_SHARED_READ_L(global_array[((long) random_generate(randomPtr)) % ((long)global_params[PARAM_SIZE])].value);
}
r1 = r1 + 1;
r2 = r2 - 1;
int f = 1;
int ii;
for(ii = 1; ii <= ((unsigned int) global_params[PARAM_WORK]); ii++)
{
f *= ii;
}
total += f / 1000000;
TM_SHARED_WRITE_L(global_array[random_number].value, r1);
TM_SHARED_WRITE_L(global_array[random_number2].value, r2);
TM_END();
long k = 0;
for (;k < (long)global_params[PARAM_INTERVAL]; k++) {
long ru = ((long) random_generate(randomPtr)) % 2;
total += ru;
}
}
TM_THREAD_EXIT();
printf("ru ignore %ld - %ld\n", total, total2);
}
/* =============================================================================
* work
* =============================================================================
*/
static void
work (void* argPtr)
{
TM_THREAD_ENTER();
args_t* args = (args_t*)argPtr;
float** feature = args->feature;
int nfeatures = args->nfeatures;
int npoints = args->npoints;
int nclusters = args->nclusters;
int* membership = args->membership;
float** clusters = args->clusters;
long long int** new_centers_len = args->new_centers_len;
float** new_centers = args->new_centers;
float delta = 0.0;
int index;
int i;
int j;
int start;
int stop;
int myId;
bool indexx[1000];
myId = thread_getId();
start = myId * CHUNK;
int cnt=0;
while (start < npoints) {
stop = (((start + CHUNK) < npoints) ? (start + CHUNK) : npoints);
for (i = start; i < stop; i++) {
index = common_findNearestPoint(feature[i],
nfeatures,
clusters,
nclusters);
/*
* If membership changes, increase delta by 1.
* membership[i] cannot be changed by other threads
*/
if (membership[i] != index) {
delta += 1.0;
}
/* Assign the membership to object i */
/* membership[i] can't be changed by other thread */
membership[i] = index;
/* Update new cluster centers : sum of objects located within */
TM_BEGIN();
//printf("shared write to begin: \n");
// int write = *new_centers_len[index];
//int* pt = new_centers_len[i];
//int dat = TM_SHARED_READ_I(*new_centers_len[i]);
//printf("in loop write centers lendata: %i %i\n", dat, *new_centers_len[i]);
TM_SHARED_WRITE_I(*new_centers_len[index],
TM_SHARED_READ_I(*new_centers_len[index]) + 1);
//printf("befor loop len P: %p data: %i\n", new_centers_len[index], *new_centers_len[index]);
//new *new_centers_len[index] = *new_centers_len[index] + 1;
//printf("INDEX %i \n" , index);
indexx[index] = true;
/*if(*new_centers_len[index]==0)*/
//printf("in lloop len P: %p data: %i\n", new_centers_len[index], *new_centers_len[index]);
//*new_centers_len[i] = *new_centers_len[i]+1;
// pt = new_centers_len[i];
//dat = TM_SHARED_READ_I(*new_centers_len[i]);
//printf("in loop write centers len data: %i\n", dat);
for (j = 0; j < nfeatures; j++) {
//printf("featurs\n");
//int feat = feature[i][j];
//printf("write\n");
//float read = TM_SHARED_READ_F(new_centers[index][j]);
//printf("write %p " ,write);
//printf("shared write to:\n");
//printf("feature %f", feature[i][j]);
//float feat = feature[i][j];
//float fl = (TM_SHARED_READ_F(new_centers[index][j])+ feat);//feature[i][j]);
//int len = *new_centers_len[index];
TM_SHARED_WRITE_F(
//write,
new_centers[index][j],
//(read + feat)
//fl
(TM_SHARED_READ_F(new_centers[index][j])+ feature[i][j])
//printf("index %p %p\n", (void*)*new_centers[index][j], (void*)(*new_centers[index][j] +1));
//printf("indexnon p %p %p\n", (void*)new_centers[index][j], (void*)(new_centers[index][j] +1));
);
//new new_centers[index][j] = new_centers[index][j] + feature[i][j];
//if(0==*new_centers_len[index]) printf("ISNAN %i\n", len);
//if(isnanf(new_centers[index][j])) printf("ISNAN2\n\n");
// if(isinf(*new_centers_len[index])) printf("ISINF\n\n");
//if(isinf(*new_centers_len[index])) printf("ISINF2\n\n");
}
TM_END();
}
//printf("update \n");
/* Update task queue */
if (start + CHUNK < npoints) {
TM_BEGIN();
start = (int)TM_SHARED_READ_L(*global_i);
TM_SHARED_WRITE_L(*global_i, (long)(start + CHUNK));
TM_END();
} else {
break;
}
}
TM_BEGIN();
//printf("shared write to: %p", *global_delta);
TM_SHARED_WRITE_F(*global_delta, TM_SHARED_READ_F(*global_delta) + delta);
//new *global_delta = *global_delta + delta;
TM_END();
int u1 =0;
/* for(int i1=0; i1<1000; i1++){
if(indexx[i1]) printf("INDEX %i %i\n", i1, u1++);
}*/
TM_THREAD_EXIT();
}
/* =============================================================================
* process
* =============================================================================
*/
void
process (void*)
{
TM_THREAD_ENTER();
heap_t* workHeapPtr = global_workHeapPtr;
mesh_t* meshPtr = global_meshPtr;
region_t* regionPtr;
long totalNumAdded = 0;
long numProcess = 0;
regionPtr = PREGION_ALLOC();
assert(regionPtr);
while (1) {
element_t* elementPtr;
TM_BEGIN();
elementPtr = (element_t*)TMHEAP_REMOVE(workHeapPtr);
TM_END();
if (elementPtr == NULL) {
break;
}
bool isGarbage;
TM_BEGIN();
isGarbage = TMELEMENT_ISGARBAGE(elementPtr);
TM_END();
if (isGarbage) {
/*
* Handle delayed deallocation
*/
PELEMENT_FREE(elementPtr);
continue;
}
long numAdded;
TM_BEGIN();
PREGION_CLEARBAD(regionPtr);
numAdded = TMREGION_REFINE(regionPtr, elementPtr, meshPtr);
TM_END();
TM_BEGIN();
TMELEMENT_SETISREFERENCED(elementPtr, false);
isGarbage = TMELEMENT_ISGARBAGE(elementPtr);
TM_END();
if (isGarbage) {
/*
* Handle delayed deallocation
*/
PELEMENT_FREE(elementPtr);
}
totalNumAdded += numAdded;
TM_BEGIN();
TMREGION_TRANSFERBAD(regionPtr, workHeapPtr);
TM_END();
numProcess++;
}
TM_BEGIN();
TM_SHARED_WRITE_L(global_totalNumAdded,
TM_SHARED_READ_L(global_totalNumAdded) + totalNumAdded);
TM_SHARED_WRITE_L(global_numProcess,
TM_SHARED_READ_L(global_numProcess) + numProcess);
TM_END();
PREGION_FREE(regionPtr);
TM_THREAD_EXIT();
}
/* =============================================================================
* getStartLists
* =============================================================================
*/
void
getStartLists (void* argPtr)
{
TM_THREAD_ENTER();
graph* GPtr = ((getStartLists_arg_t*)argPtr)->GPtr;
edge** maxIntWtListPtr = ((getStartLists_arg_t*)argPtr)->maxIntWtListPtr;
long* maxIntWtListSize = ((getStartLists_arg_t*)argPtr)->maxIntWtListSize;
edge** soughtStrWtListPtr = ((getStartLists_arg_t*)argPtr)->soughtStrWtListPtr;
long* soughtStrWtListSize = ((getStartLists_arg_t*)argPtr)->soughtStrWtListSize;
long myId = thread_getId();
long numThread = thread_getNumThread();
/*
* Find Max Wt on each thread
*/
LONGINT_T maxWeight = 0;
long i;
long i_start;
long i_stop;
createPartition(0, GPtr->numEdges, myId, numThread, &i_start, &i_stop);
for (i = i_start; i < i_stop; i++) {
if (GPtr->intWeight[i] > maxWeight) {
maxWeight = GPtr->intWeight[i];
}
}
TM_BEGIN();
long tmp_maxWeight = (long)TM_SHARED_READ_L(*global_maxWeight);
if (maxWeight > tmp_maxWeight) {
TM_SHARED_WRITE_L(*global_maxWeight, maxWeight);
}
TM_END();
thread_barrier_wait();
maxWeight = *global_maxWeight;
/*
* Create partial lists
*/
/*
* Allocate mem. for temp edge list for each thread
*/
long numTmpEdge = (5+ceil(1.5*(GPtr->numIntEdges)/MAX_INT_WEIGHT));
edge* tmpEdgeList = (edge*)P_MALLOC(numTmpEdge * sizeof(edge));
long i_edgeCounter = 0;
for (i = i_start; i < i_stop; i++) {
if (GPtr->intWeight[i] == maxWeight) {
/* Find the corresponding endVertex */
unsigned long j;
for (j = 0; j < GPtr->numDirectedEdges; j++) {
if (GPtr->paralEdgeIndex[j] > (unsigned long)i) {
break;
}
}
tmpEdgeList[i_edgeCounter].endVertex = GPtr->outVertexList[j-1];
tmpEdgeList[i_edgeCounter].edgeNum = j-1;
unsigned long t;
for (t = 0; t < GPtr->numVertices; t++) {
if (GPtr->outVertexIndex[t] > j-1) {
break;
}
}
tmpEdgeList[i_edgeCounter].startVertex = t-1;
i_edgeCounter++;
}
}
/*
* Merge partial edge lists
*/
long* i_edgeStartCounter;
long* i_edgeEndCounter;
if (myId == 0) {
i_edgeStartCounter = (long*)P_MALLOC(numThread * sizeof(long));
assert(i_edgeStartCounter);
global_i_edgeStartCounter = i_edgeStartCounter;
i_edgeEndCounter = (long*)P_MALLOC(numThread * sizeof(long));
assert(i_edgeEndCounter);
global_i_edgeEndCounter = i_edgeEndCounter;
*maxIntWtListSize = 0;
}
thread_barrier_wait();
i_edgeStartCounter = global_i_edgeStartCounter;
i_edgeEndCounter = global_i_edgeEndCounter;
i_edgeEndCounter[myId] = i_edgeCounter;
i_edgeStartCounter[myId] = 0;
thread_barrier_wait();
if (myId == 0) {
for (i = 1; i < numThread; i++) {
i_edgeEndCounter[i] = i_edgeEndCounter[i-1] + i_edgeEndCounter[i];
i_edgeStartCounter[i] = i_edgeEndCounter[i-1];
}
}
*maxIntWtListSize += i_edgeCounter;
thread_barrier_wait();
edge* maxIntWtList;
if (myId == 0) {
P_FREE(*maxIntWtListPtr);
maxIntWtList = (edge*)P_MALLOC((*maxIntWtListSize) * sizeof(edge));
assert(maxIntWtList);
global_maxIntWtList = maxIntWtList;
}
thread_barrier_wait();
maxIntWtList = global_maxIntWtList;
for (i = i_edgeStartCounter[myId]; i<i_edgeEndCounter[myId]; i++) {
(maxIntWtList[i]).startVertex = tmpEdgeList[i-i_edgeStartCounter[myId]].startVertex;
(maxIntWtList[i]).endVertex = tmpEdgeList[i-i_edgeStartCounter[myId]].endVertex;
(maxIntWtList[i]).edgeNum = tmpEdgeList[i-i_edgeStartCounter[myId]].edgeNum;
}
if (myId == 0) {
*maxIntWtListPtr = maxIntWtList;
}
i_edgeCounter = 0;
createPartition(0, GPtr->numStrEdges, myId, numThread, &i_start, &i_stop);
for (i = i_start; i < i_stop; i++) {
if (strncmp(GPtr->strWeight+i*MAX_STRLEN,
SOUGHT_STRING,
MAX_STRLEN) == 0)
{
/*
* Find the corresponding endVertex
*/
unsigned long t;
for (t = 0; t < GPtr->numEdges; t++) {
if (GPtr->intWeight[t] == -i) {
break;
}
}
unsigned long j;
for (j = 0; j < GPtr->numDirectedEdges; j++) {
if (GPtr->paralEdgeIndex[j] > t) {
break;
}
}
tmpEdgeList[i_edgeCounter].endVertex = GPtr->outVertexList[j-1];
tmpEdgeList[i_edgeCounter].edgeNum = j-1;
for (t = 0; t < GPtr->numVertices; t++) {
if (GPtr->outVertexIndex[t] > j-1) {
break;
}
}
tmpEdgeList[i_edgeCounter].startVertex = t-1;
i_edgeCounter++;
}
}
thread_barrier_wait();
i_edgeEndCounter[myId] = i_edgeCounter;
i_edgeStartCounter[myId] = 0;
if (myId == 0) {
*soughtStrWtListSize = 0;
}
thread_barrier_wait();
if (myId == 0) {
for (i = 1; i < numThread; i++) {
i_edgeEndCounter[i] = i_edgeEndCounter[i-1] + i_edgeEndCounter[i];
i_edgeStartCounter[i] = i_edgeEndCounter[i-1];
}
}
*soughtStrWtListSize += i_edgeCounter;
thread_barrier_wait();
edge* soughtStrWtList;
if (myId == 0) {
P_FREE(*soughtStrWtListPtr);
soughtStrWtList = (edge*)P_MALLOC((*soughtStrWtListSize) * sizeof(edge));
assert(soughtStrWtList);
global_soughtStrWtList = soughtStrWtList;
}
thread_barrier_wait();
soughtStrWtList = global_soughtStrWtList;
for (i = i_edgeStartCounter[myId]; i < i_edgeEndCounter[myId]; i++) {
(soughtStrWtList[i]).startVertex =
tmpEdgeList[i-i_edgeStartCounter[myId]].startVertex;
(soughtStrWtList[i]).endVertex =
tmpEdgeList[i-i_edgeStartCounter[myId]].endVertex;
(soughtStrWtList[i]).edgeNum =
tmpEdgeList[i-i_edgeStartCounter[myId]].edgeNum;
}
thread_barrier_wait();
if (myId == 0) {
*soughtStrWtListPtr = soughtStrWtList;
P_FREE(i_edgeStartCounter);
P_FREE(i_edgeEndCounter);
}
P_FREE(tmpEdgeList);
TM_THREAD_EXIT();
}
/* =============================================================================
* TMelement_setIsGarbage
* =============================================================================
*/
void
TMelement_setIsGarbage (TM_ARGDECL element_t* elementPtr, bool status)
{
TM_SHARED_WRITE_L(elementPtr->isGarbage, (long)status);
}
/* =============================================================================
* TMelement_setIsReferenced
* =============================================================================
*/
void
TMelement_setIsReferenced (TM_ARGDECL element_t* elementPtr, bool status)
{
TM_SHARED_WRITE_L(elementPtr->isReferenced, (long)status);
}
