static void initialize_border_values( Tile_t** tile_matrix, int n_tiles_width, int n_tiles_height, int tile_width, int tile_height ) {
int i, j;
/* Create a datablock for the bottom right element for tile[0][0] */
ocrGuid_t db_guid_0_0_br;
void* db_guid_0_0_br_data;
ocrDbCreate( &db_guid_0_0_br, &db_guid_0_0_br_data, sizeof(int), FLAGS, NULL, NO_ALLOC );
/* Satisfy the bottom right event for tile[0][0] with the respective datablock */
int* allocated = (int*)db_guid_0_0_br_data;
allocated[0] = 0;
ocrEventSatisfy(tile_matrix[0][0].bottom_right_event_guid, db_guid_0_0_br);
/* Create datablocks for the bottom right elements and bottom rows for tiles[0][j]
* and Satisfy the bottom row event for tile[0][j] with the respective datablock */
for ( j = 1; j < n_tiles_width + 1; ++j ) {
ocrGuid_t db_guid_0_j_brow;
void* db_guid_0_j_brow_data;
ocrDbCreate( &db_guid_0_j_brow, &db_guid_0_j_brow_data, sizeof(int)*tile_width, FLAGS, NULL, NO_ALLOC );
allocated = (int*)db_guid_0_j_brow_data;
for( i = 0; i < tile_width ; ++i ) {
allocated[i] = GAP_PENALTY*((j-1)*tile_width+i+1);
}
ocrEventSatisfy(tile_matrix[0][j].bottom_row_event_guid, db_guid_0_j_brow);
ocrGuid_t db_guid_0_j_br;
void* db_guid_0_j_br_data;
ocrDbCreate( &db_guid_0_j_br, &db_guid_0_j_br_data, sizeof(int), FLAGS, NULL, NO_ALLOC );
allocated = (int*)db_guid_0_j_br_data;
allocated[0] = GAP_PENALTY*(j*tile_width); //sagnak: needed to handle tilesize 2
ocrEventSatisfy(tile_matrix[0][j].bottom_right_event_guid, db_guid_0_j_br);
}
/* Create datablocks for the right columns for tiles[i][0]
* and Satisfy the right column event for tile[i][0] with the respective datablock */
for ( i = 1; i < n_tiles_height + 1; ++i ) {
ocrGuid_t db_guid_i_0_rc;
void* db_guid_i_0_rc_data;
ocrDbCreate( &db_guid_i_0_rc, &db_guid_i_0_rc_data, sizeof(int)*tile_height, FLAGS, NULL, NO_ALLOC );
allocated = (int*)db_guid_i_0_rc_data;
for ( j = 0; j < tile_height ; ++j ) {
allocated[j] = GAP_PENALTY*((i-1)*tile_height+j+1);
}
ocrEventSatisfy(tile_matrix[i][0].right_column_event_guid, db_guid_i_0_rc);
ocrGuid_t db_guid_i_0_br;
void* db_guid_i_0_br_data;
ocrDbCreate( &db_guid_i_0_br, &db_guid_i_0_br_data, sizeof(int), FLAGS, NULL, NO_ALLOC );
allocated = (int*)db_guid_i_0_br_data;
allocated[0] = GAP_PENALTY*(i*tile_height); //sagnak: needed to handle tilesize 2
ocrEventSatisfy(tile_matrix[i][0].bottom_right_event_guid, db_guid_i_0_br);
}
}
ocrGuid_t mainEdt(u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
// Current thread is '0' and goes on with user code.
ocrGuid_t ndepv[1];
ocrGuid_t eventGuid;
ocrEventCreate(&eventGuid, OCR_EVENT_STICKY_T, true);
ndepv[0] = eventGuid;
// Creates the EDT
u32 nparamc = 1;
u64 nparamv = 32;
// 'ndepv' stores dependencies, so no need to call
// ocrAddDependence later on to register events.
ocrGuid_t edtGuid;
ocrGuid_t taskForEdtTemplateGuid;
ocrEdtTemplateCreate(&taskForEdtTemplateGuid, taskForEdt, nparamc, 1 /*depc*/);
ocrEdtCreate(&edtGuid, taskForEdtTemplateGuid, EDT_PARAM_DEF, &nparamv, EDT_PARAM_DEF, /*depv=*/ndepv,
/*properties=*/0, NULL_HINT, /*outEvent=*/NULL);
int *k;
ocrGuid_t dbGuid;
ocrDbCreate(&dbGuid,(void **) &k,
sizeof(int), /*flags=*/DB_PROP_NONE,
/*location=*/NULL_HINT,
NO_ALLOC);
*k = 42;
ocrEventSatisfy(eventGuid, dbGuid);
return NULL_GUID;
}
// This task solves a triangular matrix equation, which is the
// second step in the tiled Cholesky factorization
ocrGuid_t cblas_dtrsm_task ( u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
u64 *func_args = paramv;
u32 k = (u32) func_args[0];
u32 j = (u32) func_args[1];
u32 tileSize = (u32) func_args[2];
ocrGuid_t out_lkji_jkkp1_event_guid = (ocrGuid_t) func_args[3];
// PRINTF("RUNNING trisolve (%d, %d)\n", k, j);
double* aBlock = (double*) (depv[0].ptr);
double* liBlock = (double*) (depv[1].ptr);
ocrGuid_t out_lkji_jkkp1_db_guid;
ocrGuid_t out_lkji_jkkp1_db_affinity = NULL_GUID;
// Solve Ax = alpha B where alpha = 1 and A is liBlock triangular matrix output from
// dpotrf and B is the next block to solve in factorization
cblas_dtrsm(CblasColMajor, CblasLeft, CblasUpper, CblasTrans, CblasNonUnit, tileSize, tileSize, 1.0,
liBlock, tileSize, // A matrix to solve for x in A x = B
aBlock, tileSize); // B matrix, solution put here
ocrEventSatisfy(out_lkji_jkkp1_event_guid, depv[0].guid);
return NULL_GUID;
}
inline static void satisfyInitialTiles(u32 numTiles, u32 tileSize, double** matrix,
ocrGuid_t*** lkji_event_guids) {
u32 i,j,index;
u32 A_i, A_j, T_i, T_j;
for( i = 0 ; i < numTiles ; ++i ) {
for( j = 0 ; j <= i ; ++j ) {
ocrGuid_t db_guid;
ocrGuid_t db_affinity = NULL_GUID;
void* temp_db;
ocrGuid_t tmpdb_guid;
ocrDbCreate(&db_guid, &temp_db, sizeof(double)*tileSize*tileSize,
FLAGS, db_affinity, NO_ALLOC);
double* temp = (double*) temp_db;
double** temp2D;
ocrDbCreate(&tmpdb_guid, (void *)&temp2D, sizeof(double*)*tileSize,
FLAGS, NULL_GUID, NO_ALLOC);
for( index = 0; index < tileSize; ++index )
temp2D [index] = &(temp[index*tileSize]);
// Split the matrix u32o tiles and write it u32o the item space at time 0.
// The tiles are indexed by tile indices (which are tag values).
for( A_i = i*tileSize, T_i = 0 ; T_i < tileSize; ++A_i, ++T_i ) {
for( A_j = j*tileSize, T_j = 0 ; T_j < tileSize; ++A_j, ++T_j ) {
temp2D[ T_i ][ T_j ] = matrix[ A_i ][ A_j ];
}
}
ocrEventSatisfy(lkji_event_guids[i][j][0], db_guid);
ocrDbDestroy(tmpdb_guid);
}
}
}
inline static void satisfyInitialTiles(u32 numTiles, u32 tileSize,
ocrGuid_t*** lkji_event_guids) {
u32 i,j;
u32 T_i, T_j;
ocrGuid_t db_guid;
ocrGuid_t db_affinity;
void* temp_db;
FILE *fin;
fin = fopen("inputfile", "r");
if(fin == NULL) PRINTF("Error opening input file\n");
for( i = 0 ; i < numTiles ; ++i ) {
for( j = 0 ; j <= i ; ++j ) {
ocrDbCreate(&db_guid, &temp_db, sizeof(double)*tileSize*tileSize,
FLAGS, db_affinity, NO_ALLOC);
fread(temp_db, sizeof(double)*tileSize*tileSize, 1, fin);
ocrEventSatisfy(lkji_event_guids[i][j][0], db_guid);
ocrDbRelease(db_guid);
}
}
hal_fence();
fclose(fin);
}
// This task computes the Cholesky factorization of a symmetric positive definite matrix...
// This is the first step in the tile Cholesky factorization.
ocrGuid_t lapacke_dpotrf_task ( u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
u32 info;
u64 *func_args = paramv;
u32 k = (u32) func_args[0];
u32 tileSize = (u32) func_args[1];
ocrGuid_t out_lkji_kkkp1_event_guid = (ocrGuid_t) func_args[2];
double* aBlock = (double*) (depv[0].ptr);
// PRINTF("RUNNING sequential_cholesky %d with 0x%llx to satisfy\n", k, (u64)(out_lkji_kkkp1_event_guid));
ocrGuid_t out_lkji_kkkp1_db_guid;
ocrGuid_t out_lkji_kkkp1_db_affinity = NULL_GUID;
info = LAPACKE_dpotrf(LAPACK_ROW_MAJOR, 'L', tileSize, aBlock, tileSize );
if (info != 0)
{
if (info > 0) PRINTF("Matrix A is not Symmetric Positive Definite (SPD)");
else PRINTF("i-th parameter had an illegal value.");
ASSERT(0);
ocrShutdown();
return NULL_GUID;
}
ocrEventSatisfy(out_lkji_kkkp1_event_guid, depv[0].guid);
return NULL_GUID;
}
ocrGuid_t mainEdt(u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
// Current thread is '0' and goes on with user code.
ocrGuid_t e0;
ocrEventCreate(&e0, OCR_EVENT_ONCE_T, EVT_PROP_TAKES_ARG);
// Creates the EDTa
ocrGuid_t edtGuid;
ocrGuid_t taskForEdtTemplateGuid;
ocrEdtTemplateCreate(&taskForEdtTemplateGuid, taskForEdt, 0 /*paramc*/, 2/*depc*/);
ocrEdtCreate(&edtGuid, taskForEdtTemplateGuid, EDT_PARAM_DEF, /*paramv=*/NULL, EDT_PARAM_DEF, /*depv=*/NULL,
/*properties=*/0, NULL_HINT, /*outEvent=*/NULL);
// Register a dependence between an event and an edt
ocrAddDependence(e0, edtGuid, 0, DB_MODE_CONST);
ocrAddDependence(e0, edtGuid, 1, DB_MODE_CONST);
int *k;
ocrGuid_t dbGuid;
ocrDbCreate(&dbGuid,(void **) &k,
sizeof(int), /*flags=*/DB_PROP_NONE,
/*location=*/NULL_HINT,
NO_ALLOC);
*k = 42;
// Satisfy event's chain head
ocrEventSatisfy(e0, dbGuid);
return NULL_GUID;
}
// Takes two events and satisfy a result event
//depv[0] db for domainSetup_t
//depv[1] db for domainKernel_t
ocrGuid_t combineKernelEdt(u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
// timestamp_t stopTimer;
// get_time(&stopTimer);
ocrGuid_t kernelEdtDoneEvt; // Result event
kernelEdtDoneEvt.guid = paramv[0];
// Impl-specific here, we know to expect two DBs
ocrGuid_t setupDb = depv[0].guid;
domainSetup_t * setupDbPtr = depv[0].ptr;
ocrGuid_t kernelDb = depv[1].guid;
domainKernel_t * kernelDbPtr = depv[1].ptr;
// kernelDbPtr->stopTimer=stopTimer;
ocrGuid_t timeDbGuid;
long * timeDbPtr;
ocrDbCreate(&timeDbGuid, (void**) &timeDbPtr, sizeof(long), 0, NULL_HINT, NO_ALLOC);
domainKernelCombine(setupDbPtr, kernelDbPtr, timeDbPtr);
// PRINTF("combineEdt timeDbGuid=0x%lx\n", timeDbGuid);
ocrDbRelease(timeDbGuid);
ocrDbRelease(kernelDb);
ocrDbRelease(setupDb);
ocrDbDestroy(kernelDb);
ocrDbDestroy(setupDb);
ocrEventSatisfy(kernelEdtDoneEvt, timeDbGuid);
return NULL_GUID;
}
void test () {
// Current thread is '0' and goes on with user code.
ocrGuid_t eventGuid;
ocrEventCreate(&eventGuid, OCR_EVENT_STICKY_T, true);
// Creates the EDT
ocrGuid_t edtGuid;
ocrGuid_t taskForEdtTemplateGuid;
ocrEdtTemplateCreate(&taskForEdtTemplateGuid, taskForEdt, 0 /*paramc*/, 1 /*depc*/);
ocrEdtCreate(&edtGuid, taskForEdtTemplateGuid, EDT_PARAM_DEF, /*paramv=*/NULL,
EDT_PARAM_DEF, /*depv=*/NULL, 0, NULL_GUID, NULL);
// Register a dependence between an event and an edt
ocrAddDependence(eventGuid, edtGuid, 0, DB_MODE_RO);
int *k;
ocrGuid_t db_guid;
ocrDbCreate(&db_guid,(void **) &k,
sizeof(int), /*flags=*/FLAGS,
/*location=*/NULL_GUID,
NO_ALLOC);
*k = 42;
ocrEventSatisfy(eventGuid, db_guid);
}
// paramv[0]: event to satisfy when kernel is done
// depv[0]: setupEdt completed (may carry a DB)
ocrGuid_t kernelEdt(u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
// PRINTF("kernelEdt\n");
ocrGuid_t kernelEdtDoneEvt;
kernelEdtDoneEvt.guid = paramv[0];
ocrGuid_t setupDb = depv[0].guid;
domainSetup_t * setupDbPtr = depv[0].ptr;
// The sub kernel done event
ocrGuid_t subKernelDoneEvt;
ocrEventCreate(&subKernelDoneEvt, OCR_EVENT_ONCE_T, true);
// This EDT done event
ocrGuid_t selfDoneEvt;
ocrEventCreate(&selfDoneEvt, OCR_EVENT_ONCE_T, true);
// Combine those in a combine EDT that satisfies kernelEdtDoneEvt
//TODO same issue of allocating tpl every iteration
ocrGuid_t curAffGuid;
ocrAffinityGetCurrent(&curAffGuid);
ocrGuid_t combEdtTplGuid;
ocrEdtTemplateCreate(&combEdtTplGuid, combineKernelEdt, 1, 2);
ocrGuid_t combineEdtGuid;
combine(&combineEdtGuid, combEdtTplGuid, curAffGuid,
selfDoneEvt, subKernelDoneEvt, kernelEdtDoneEvt);
ocrEdtTemplateDestroy(combEdtTplGuid);
timestamp_t timer;
domainKernel(subKernelDoneEvt, setupDbPtr, &timer);
// Satisfy self event with the timer information
ocrEventSatisfy(selfDoneEvt, setupDb);
return NULL_GUID;
}
/* Do the addition */
u8 summer(u32 paramc, u64 * params, void* paramv[], u32 depc, ocrEdtDep_t depv[]) {
int *result;
ocrGuid_t resultGuid;
/* Get both numbers */
int *n1 = (int*)depv[0].ptr, *n2 = (int*)depv[1].ptr;
/* Get event to satisfy */
ocrGuid_t *evt = (ocrGuid_t*)depv[2].ptr;
/* Create data-block to put result */
ocrDbCreate(&resultGuid, (void**)&result, sizeof(int), /*flags=*/0, /*location=*/NULL, NO_ALLOC);
*result = *n1 + *n2;
/* Say hello */
printf("I am summing %d (GUID: 0x%lx) and %d (GUID: 0x%lx) and passing along %d (GUID: 0x%lx)\n",
*n1, (u64)depv[0].guid, *n2, (u64)depv[1].guid,
*result, (u64)resultGuid);
/* Satisfy whomever is waiting on me */
ocrEventSatisfy(*evt, resultGuid);
/* Free inputs */
ocrDbDestroy(depv[0].guid);
ocrDbDestroy(depv[1].guid);
ocrDbDestroy(depv[2].guid);
return 0;
}
ocrGuid_t complete(u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
u64 arg = (u64)paramv[0];
int my_ID = paramv[1];
ocrGuid_t inDep;
u32 in1, in2;
u32 out;
inDep = (ocrGuid_t)arg;
/* When we run, we got our inputs from fib(n-1) and fib(n-2) */
in1 = *(u32*)depv[0].ptr;
in2 = *(u32*)depv[1].ptr;
out = *(u32*)depv[2].ptr;
PRINTF("r%d Done with %d (%d + %d)\n", my_ID, out, in1, in2);
/* we return our answer in the 3rd db passed in as an argument */
*((u32*)(depv[2].ptr)) = in1 + in2;
/* The app is done with the answers from fib(n-1) and fib(n-2) */
ocrDbDestroy(depv[0].guid);
ocrDbDestroy(depv[1].guid);
/* and let our parent's completion know we're done with fib(n) */
ocrEventSatisfy(inDep, depv[2].guid);
return NULL_GUID;
}
int main (int argc, char ** argv) {
ocrEdt_t fctPtrArray [1];
fctPtrArray[0] = &task_for_edt;
ocrInit(&argc, argv, 1, fctPtrArray);
// Current thread is '0' and goes on with user code.
ocrGuid_t event_guid;
ocrEventCreate(&event_guid, OCR_EVENT_STICKY_T, true);
// Creates the EDT
ocrGuid_t edt_guid;
ocrEdtCreate(&edt_guid, task_for_edt, /*paramc=*/0, /*params=*/ NULL,
/*paramv=*/NULL, /*properties=*/0,
/*depc=*/1, /*depv=*/NULL);
// Register a dependence between an event and an edt
ocrAddDependence(event_guid, edt_guid, 0);
// Schedule the EDT (will run when dependences satisfied)
ocrEdtSchedule(edt_guid);
int *k;
ocrGuid_t db_guid;
ocrDbCreate(&db_guid, (void **) &k, sizeof(int), /*flags=*/FLAGS,
/*location=*/NULL, NO_ALLOC);
*k = 42;
ocrEventSatisfy(event_guid, db_guid);
ocrCleanup();
return 0;
}
// paramv[0]: continuation after iterations
// depv[0]: info
// depv[1]: done event for the work spawned by the iteration, carries the time DB
ocrGuid_t iterationEdt(u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
ocrGuid_t iterationsDoneEvt;
iterationsDoneEvt.guid = paramv[0];
ocrGuid_t infoGuid = depv[0].guid;
info_t * info = (info_t *) depv[0].ptr;
ocrGuid_t timeDbGuid = depv[1].guid;
long * timePrevIt = (long *) depv[1].ptr;
// PRINTF("iteration %d\n", info->i);
if (timePrevIt != NULL) {
info->timer += (*timePrevIt);
ocrDbRelease(timeDbGuid);
ocrDbDestroy(timeDbGuid);
}
if (info->i < info->max) {
//TODO do a pipeline EDT
u64 affCount = 0;
ocrAffinityCount(AFFINITY_PD, &affCount);
ocrGuid_t affinities[affCount];
ocrAffinityGet(AFFINITY_PD, &affCount, affinities);
ocrGuid_t stageInit;
ocrEventCreate(&stageInit, OCR_EVENT_ONCE_T, false);
ocrGuid_t stagePrev = stageInit;
u32 i = PIPE_START;
while(i < (PIPE_START+PIPE_SZ)) {
ocrGuid_t stageEdtDoneEvt;
ocrEventCreate(&stageEdtDoneEvt, OCR_EVENT_ONCE_T, false);
ocrGuid_t stageEdtGuid;
//TODO I wonder if we shouldn't give the whole info to a functor and invoke that
chain(&stageEdtGuid, info->edtTemplGuids[i], affinities[info->edtAffinities[i]], stagePrev, stageEdtDoneEvt);
stagePrev = stageEdtDoneEvt;
i++;
}
info->i+=1;
ocrDbRelease(infoGuid);
ocrGuid_t nextItEdtGuid;
// 'iterationsDoneEvt' is passed on and on til the last iteration
iterate(&nextItEdtGuid, info->edtTemplGuids[ITER_IDX],
/*prev*/stagePrev, /*data*/infoGuid, /*next*/iterationsDoneEvt);
// Start the pipeline
ocrEventSatisfy(stageInit, NULL_GUID);
} else {
ocrEventSatisfy(iterationsDoneEvt, infoGuid);
}
return NULL_GUID;
}
// For control dependence to sync up framework setup and user setup
ocrGuid_t combineSetupEdt(u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
ocrGuid_t doneEvt;
doneEvt.guid = paramv[0];
ocrGuid_t setupDbGuid = depv[0].guid;
ocrDbRelease(setupDbGuid);
ocrEventSatisfy(doneEvt, setupDbGuid);
return NULL_GUID;
}
ocrGuid_t remoteEdt(u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
PRINTF("[remote] RemoteEdt: executing\n");
ocrGuid_t eventGuid = ((ocrGuid_t *) depv[0].ptr)[0];
// see comment at top of file
// ocrGuid_t expectedAffinityGuid = ((ocrGuid_t *) depv[0].ptr)[1];
// ocrGuid_t currentAffinity;
// ocrAffinityGetCurrent(¤tAffinity);
// ASSERT(expectedAffinityGuid == currentAffinity);
ocrEventSatisfy(eventGuid, NULL_GUID);
return NULL_GUID;
}
ocrGuid_t mainEdt ( u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
u64 size = -1;
u64 offset = -1;
u32 i = 0;
u64 argc;
void *programArg = depv[0].ptr;
argc = getArgc(programArg);
if ( argc != 4 ) {
PRINTF("Usage: ./basicIO offset size fileName \n");
ocrShutdown();
return 1;
}
offset = atoi(getArgv(programArg, 1));
size = atoi(getArgv(programArg, 2));
u64 nparamv[2];
nparamv[0] = offset;
nparamv[1] = size;
FILE *in;
in = fopen(getArgv(programArg, 3), "r");
if( !in ) {
PRINTF("Cannot find file: %s\n", getArgv(programArg, 3));
ocrShutdown();
return NULL_GUID;
}
ocrGuid_t dataGuid;
// Data can be passed as parameter also , there was no
// necessary need of creating data block in this example.
// Its has been created for demo purpose
//Create datablock to hold a block of 'size' elements
u64 *inputarr;
ocrDbCreate(&dataGuid, (void**)&inputarr, sizeof(u64)*size,0,NULL_GUID, NO_ALLOC);
#ifndef TG_ARCH
while(fscanf(in,"%llu\n",&inputarr[i++])!=EOF);
#else
fread(inputarr, sizeof(u64),size , in);
#endif
fclose(in);
ocrGuid_t addEdtTemplateGuid;
ocrEdtTemplateCreate(&addEdtTemplateGuid, add_edt, 2 /*paramc*/, 1 /*depc*/);
ocrGuid_t add_edt_guid;
// Create the EDT not specifying the dependence vector at creation
ocrEdtCreate(&add_edt_guid, addEdtTemplateGuid, EDT_PARAM_DEF, nparamv,1,NULL ,EDT_PROP_FINISH , NULL_GUID, NULL);
ocrGuid_t triggerEventGuid;
ocrEventCreate(&triggerEventGuid, OCR_EVENT_STICKY_T, true);
ocrAddDependence(triggerEventGuid, add_edt_guid, 0, DB_MODE_EW);
ocrEventSatisfy(triggerEventGuid, dataGuid);
return NULL_GUID;
}
ocrGuid_t mainEdt(u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
ocrGuid_t e0;
ocrEventCreate(&e0, OCR_EVENT_STICKY_T, false);
ocrGuid_t terminateEdtGuid;
ocrGuid_t terminateEdtTemplateGuid;
ocrEdtTemplateCreate(&terminateEdtTemplateGuid, terminateEdt, 0 /*paramc*/, EDT_PARAM_UNK /*depc*/);
ocrEdtCreate(&terminateEdtGuid, terminateEdtTemplateGuid, EDT_PARAM_DEF, NULL, 1, NULL_GUID,
/*properties=*/EDT_PROP_FINISH, NULL_GUID, /*outEvent=*/ NULL);
ocrAddDependence(e0, terminateEdtGuid, 0, DB_MODE_RO);
ocrEventSatisfy(e0, NULL_GUID);
return NULL_GUID;
}
ocrGuid_t postEdt(u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
// PRINTF("postEdt timeDbGuid=0x%lx\n", depv[0].guid);
ocrGuid_t postEdtDoneEvt;
postEdtDoneEvt.guid = paramv[0];
ocrGuid_t timeDbGuid = depv[0].guid;
long * timeDbPtr = depv[0].ptr;
//TODO I think there's an issue here with this
//edt doing an implicit release at the end
//concurrently with the iterationEdt destroying it.
ocrDbRelease(depv[0].guid);
// domainPost(kernelDbPtr);
ocrEventSatisfy(postEdtDoneEvt, timeDbGuid);
return NULL_GUID;
}
ocrGuid_t remoteSetupUserEdt(u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
ocrGuid_t dsetupGuid = depv[0].guid;
domainSetup_t * dsetup = (domainSetup_t *) depv[0].ptr;
// Create 'remote' (local to here) latch event
ocrGuid_t evtGuid;
ocrEventParams_t params;
params.EVENT_LATCH.counter = NB_SATISFY;
ocrEventCreateParams(&evtGuid, OCR_EVENT_LATCH_T, false, ¶ms);
dsetup->remoteLatchEvent = evtGuid;
// Setup callback for when the latch event fires
ocrAddDependence(evtGuid, dsetup->stopTimerEvt, 0, DB_MODE_NULL);
ocrGuid_t userSetupDoneEvt = dsetup->userSetupDoneEvt;
ocrDbRelease(dsetupGuid);
// Global setup is done
ocrEventSatisfy(userSetupDoneEvt, NULL_GUID);
return NULL_GUID;
}
// This task performs a symmetric rank-k update, which is the
// third step in the tiled Cholesky factorization
ocrGuid_t cblas_dsyrk_task ( u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
u64 *func_args = paramv;
u32 k = (u32) func_args[0];
u32 j = (u32) func_args[1];
u32 i = (u32) func_args[2];
u32 tileSize = (u32) func_args[3];
ocrGuid_t out_lkji_jjkp1_event_guid = (ocrGuid_t) func_args[4];
// PRINTF("RUNNING update_diagonal (%d, %d, %d)\n", k, j, i);
double* aBlock = (double*) (depv[0].ptr);
double* l2Block = (double*) (depv[1].ptr);
// Symmetric Rank-k Update C = alpha AA' + beta C, where alpha = -1, beta = 1
cblas_dsyrk(CblasRowMajor, CblasLower, CblasNoTrans, tileSize, tileSize, -1.0,
l2Block, tileSize, 1.0, // A matrix
aBlock, tileSize); // C matrix, solution put here in lower triangle
ocrEventSatisfy(out_lkji_jjkp1_event_guid, depv[0].guid);
return NULL_GUID;
}
int main (int argc, char ** argv) {
ocrEdt_t fctPtrArray [1];
fctPtrArray[0] = &task_for_edt;
ocrInit(&argc, argv, 1, fctPtrArray);
// Current thread is '0' and goes on with user code.
ocrGuid_t event_guid;
ocrEventCreate(&event_guid, OCR_EVENT_STICKY_T, true);
// Creates the EDT
u32 paramc = 1;
u64 params[1];
params[0] = sizeof(int);
int * paramv = (int *) malloc(sizeof(int));
paramv[0] = 32;
ocrGuid_t edt_guid;
ocrEdtCreate(&edt_guid, task_for_edt, paramc, params, (void**) ¶mv, 0, 1, NULL);
// Register a dependence between an event and an edt
ocrAddDependence(event_guid, edt_guid, 0);
int *k;
ocrGuid_t db_guid;
ocrDbCreate(&db_guid,(void **) &k,
sizeof(int), /*flags=*/FLAGS,
/*location=*/NULL,
NO_ALLOC);
*k = 42;
ocrEventSatisfy(event_guid, db_guid);
ocrEdtSchedule(edt_guid);
ocrCleanup();
return 0;
}
// This computes a matrix-matrix product to complete the last step in
// the Cholesky factorization prior to the next iteration.
ocrGuid_t cblas_dgemm_task ( u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
u64 *func_args = paramv;
u32 k = (u32) func_args[0];
u32 j = (u32) func_args[1];
u32 i = (u32) func_args[2];
u32 tileSize = (u32) func_args[3];
ocrGuid_t out_lkji_jikp1_event_guid = (ocrGuid_t) func_args[4];
// PRINTF("RUNNING update_nondiagonal (%d, %d, %d)\n", k, j, i);
double* aBlock = (double*) (depv[0].ptr);
double* l1Block = (double*) (depv[1].ptr);
double* l2Block = (double*) (depv[2].ptr);
cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasTrans, tileSize, tileSize, tileSize, -1.0,
l1Block, tileSize, // A matrix
l2Block, tileSize, 1.0, // B matrix
aBlock, tileSize); // C matrix, solution put here
ocrEventSatisfy(out_lkji_jikp1_event_guid, depv[0].guid);
return NULL_GUID;
}
ocrGuid_t mainEdt(u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
u64 arraySize = 100;
//Create argument vector to hold array size
u64 nparamv[1];
nparamv[0] = arraySize;
u64 * dataArray;
ocrGuid_t dataGuid;
//Create datablock to hold a block of 'array size' u64 elements
ocrDbCreate(&dataGuid, (void **) &dataArray, sizeof(u64)*arraySize,
/*flags=*/0, /*loc=*/NULL_GUID, NO_ALLOC);
ocrGuid_t stepAEdtTemplateGuid;
ocrEdtTemplateCreate(&stepAEdtTemplateGuid, stepA_edt, 1 /*paramc*/, 1 /*depc*/);
// Create the EDT not specifying the dependence vector at creation
ocrGuid_t stepAEdtGuid;
ocrEdtCreate(&stepAEdtGuid, stepAEdtTemplateGuid, EDT_PARAM_DEF, nparamv, 1, NULL,
/*prop=*/EDT_PROP_NONE, NULL_GUID, NULL);
ocrGuid_t triggerEventGuid;
//TODO Setup event used to trigger stepA
ocrEventCreate(&triggerEventGuid, OCR_EVENT_STICKY_T, EVT_PROP_TAKES_ARG);
//END-TODO
//TODO Setup dependence between event and stepA's EDTs slot 0
ocrAddDependence(triggerEventGuid, stepAEdtGuid, 0, DB_MODE_EW);
//END-TODO
//TODO Satisfy the event with the datablock
ocrEventSatisfy(triggerEventGuid, dataGuid);
//END-TODO
return NULL_GUID;
}
// Input
// - Completion event to be satisfied when setup is done (paramv[0])
ocrGuid_t setupEdt(u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
// Setup all done event
ocrGuid_t setupEdtDoneEvt;
setupEdtDoneEvt.guid = paramv[0];
ocrGuid_t curAffGuid;
ocrAffinityGetCurrent(&curAffGuid);
ocrGuid_t setupDbGuid;
domainSetup_t * setupDbPtr;
ocrHint_t dbHint;
ocrHintInit(&dbHint, OCR_HINT_DB_T);
ocrSetHintValue(&dbHint, OCR_HINT_DB_AFFINITY, ocrAffinityToHintValue(curAffGuid));
ocrDbCreate(&setupDbGuid, (void**) &setupDbPtr, sizeof(domainSetup_t), 0, &dbHint, NO_ALLOC);
setupDbPtr->self = setupDbGuid;
// This EDT done event
ocrGuid_t selfDoneEvt;
ocrEventCreate(&selfDoneEvt, OCR_EVENT_ONCE_T, true);
// Create a done event for the user code
ocrGuid_t subSetupDoneEvt;
ocrEventCreate(&subSetupDoneEvt, OCR_EVENT_ONCE_T, true);
ocrGuid_t combEdtTplGuid;
ocrEdtTemplateCreate(&combEdtTplGuid, combineSetupEdt, 1, 2);
ocrGuid_t combineEdtGuid;
combine(&combineEdtGuid, combEdtTplGuid, curAffGuid,
selfDoneEvt, subSetupDoneEvt, setupEdtDoneEvt);
ocrEdtTemplateDestroy(combEdtTplGuid);
domainSetup(subSetupDoneEvt, setupDbPtr);
ocrDbRelease(setupDbGuid);
ocrEventSatisfy(selfDoneEvt, setupDbGuid);
return NULL_GUID;
}
u8 smith_waterman_task ( u32 paramc, u64 * params, void* paramv[], u32 depc, ocrEdtDep_t depv[]) {
int index, ii, jj;
/* Unbox parameters */
intptr_t* typed_paramv = *paramv;
int i = (int) typed_paramv[0];
int j = (int) typed_paramv[1];
int tile_width = (int) typed_paramv[2];
int tile_height = (int) typed_paramv[3];
Tile_t** tile_matrix = (Tile_t**) typed_paramv[4];
signed char* string_1 = (signed char* ) typed_paramv[5];
signed char* string_2 = (signed char* ) typed_paramv[6];
int n_tiles_height = (int) typed_paramv[7];
int n_tiles_width = (int)typed_paramv[8];
/* Get the input datablock data pointers acquired from dependences */
int* left_tile_right_column = (int *) depv[0].ptr;
int* above_tile_bottom_row = (int *) depv[1].ptr;
int* diagonal_tile_bottom_right = (int *) depv[2].ptr;
/* Allocate a haloed local matrix for calculating 'this' tile*/
int * curr_tile_tmp = (int*)malloc(sizeof(int)*(1+tile_width)*(1+tile_height));
/* 2D-ify it for readability */
int ** curr_tile = (int**)malloc(sizeof(int*)*(1+tile_height));
for (index = 0; index < tile_height+1; ++index) {
curr_tile[index] = &curr_tile_tmp[index*(1+tile_width)];
}
/* Initialize halo from neighbouring tiles */
/* Set local_tile[0][0] (top left) from the bottom right of the northwest tile */
curr_tile[0][0] = diagonal_tile_bottom_right[0];
/* Set local_tile[i+1][0] (left column) from the right column of the left tile */
for ( index = 1; index < tile_height+1; ++index ) {
curr_tile[index][0] = left_tile_right_column[index-1];
}
/* Set local_tile[0][j+1] (top row) from the bottom row of the above tile */
for ( index = 1; index < tile_width+1; ++index ) {
curr_tile[0][index] = above_tile_bottom_row[index-1];
}
/* Run a smith-waterman on the local tile */
for ( ii = 1; ii < tile_height+1; ++ii ) {
for ( jj = 1; jj < tile_width+1; ++jj ) {
signed char char_from_1 = string_1[(j-1)*tile_width+(jj-1)];
signed char char_from_2 = string_2[(i-1)*tile_height+(ii-1)];
/* Get score from northwest, north and west */
int diag_score = curr_tile[ii-1][jj-1] + alignment_score_matrix[char_from_2][char_from_1];
int left_score = curr_tile[ii ][jj-1] + alignment_score_matrix[char_from_1][GAP];
int top_score = curr_tile[ii-1][jj ] + alignment_score_matrix[GAP][char_from_2];
int bigger_of_left_top = (left_score > top_score) ? left_score : top_score;
/* Set the local tile[i][j] to the maximum value of northwest, north and west */
curr_tile[ii][jj] = (bigger_of_left_top > diag_score) ? bigger_of_left_top : diag_score;
}
}
/* Allocate datablock for bottom right of the local tile */
ocrGuid_t db_guid_i_j_br;
void* db_guid_i_j_br_data;
ocrDbCreate( &db_guid_i_j_br, &db_guid_i_j_br_data, sizeof(int), FLAGS, NULL, NO_ALLOC );
/* Satisfy the bottom right event of local tile with the data block allocated above */
int* curr_bottom_right = (int*)db_guid_i_j_br_data;
curr_bottom_right[0] = curr_tile[tile_height][tile_width];
ocrEventSatisfy(tile_matrix[i][j].bottom_right_event_guid, db_guid_i_j_br);
/* Allocate datablock for right column of the local tile */
ocrGuid_t db_guid_i_j_rc;
void* db_guid_i_j_rc_data;
ocrDbCreate( &db_guid_i_j_rc, &db_guid_i_j_rc_data, sizeof(int)*tile_height, FLAGS, NULL, NO_ALLOC );
/* Satisfy the right column event of local tile with the data block allocated above */
int* curr_right_column = (int*)db_guid_i_j_rc_data;
for ( index = 0; index < tile_height; ++index ) {
curr_right_column[index] = curr_tile[index+1][tile_width];
}
ocrEventSatisfy(tile_matrix[i][j].right_column_event_guid, db_guid_i_j_rc);
/* Allocate datablock for bottom row of the local tile */
ocrGuid_t db_guid_i_j_brow;
void* db_guid_i_j_brow_data;
ocrDbCreate( &db_guid_i_j_brow, &db_guid_i_j_brow_data, sizeof(int)*tile_width, FLAGS, NULL, NO_ALLOC );
/* Satisfy the bottom row event of local tile with the data block allocated above */
int* curr_bottom_row = (int*)db_guid_i_j_brow_data;
for ( index = 0; index < tile_width; ++index ) {
curr_bottom_row[index] = curr_tile[tile_height][index+1];
}
ocrEventSatisfy(tile_matrix[i][j].bottom_row_event_guid, db_guid_i_j_brow);
free(curr_tile);
free(curr_tile_tmp);
/* If this is the last tile (bottom right most tile), finish */
if ( i == n_tiles_height && j == n_tiles_width ) {
fprintf(stdout, "score: %d\n", curr_bottom_row[tile_width-1]);
ocrFinish();
}
}
ocrGuid_t fibEdt(u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
void* ptr;
ocrGuid_t inDep;
ocrGuid_t fib0, fib1, comp;
ocrGuid_t fibDone[2];
ocrGuid_t fibArg[2];
inDep = (ocrGuid_t)paramv[0];
int my_ID = paramv[1];
u32 n = *(u32*)(depv[0].ptr);
PRINTF("r%d Starting fibEdt(%u)\n", my_ID, n);
if (n < 2) {
PRINTF("r%d In fibEdt(%d) -- done (sat %lx)\n", my_ID, n, inDep);
ocrEventSatisfy(inDep, depv[0].guid);
return NULL_GUID;
}
PRINTF("r%d In fibEdt(%d) -- spawning children\n", my_ID, n);
/* create the completion EDT and pass it the in/out argument as a dependency */
/* create the EDT with the done_event as the argument */
{
u64 paramv[] = {(u64)inDep, my_ID};
ocrGuid_t templateGuid;
ocrEdtTemplateCreate(&templateGuid, complete, 2, 3);
ocrEdtCreate(&comp, templateGuid, 2, paramv, 3, NULL, EDT_PROP_NONE,
NULL_GUID, NULL);
ocrEdtTemplateDestroy(templateGuid);
}
PRINTF("r%d In fibEdt(%u) -- spawned complete EDT GUID 0x%llx\n", my_ID, n, (u64)comp);
ocrAddDependence(depv[0].guid, comp, 2, DB_DEFAULT_MODE);
/* create the events that the completion EDT will "wait" on */
ocrEventCreate(&fibDone[0], OCR_EVENT_ONCE_T, EVT_PROP_TAKES_ARG);
ocrEventCreate(&fibDone[1], OCR_EVENT_ONCE_T, EVT_PROP_TAKES_ARG);
ocrAddDependence(fibDone[0], comp, 0, DB_DEFAULT_MODE);
ocrAddDependence(fibDone[1], comp, 1, DB_DEFAULT_MODE);
/* allocate the argument to pass to fib(n-1) */
ocrDbCreate(&fibArg[0], (void**)&ptr, sizeof(u32), DB_PROP_NONE, NULL_GUID, NO_ALLOC);
PRINTF("r%d In fibEdt(%u) -- created arg DB GUID 0x%llx\n", my_ID, n, fibArg[0]);
*((u32*)ptr) = n-1;
/* sched the EDT, passing the fibDone event as it's argument */
{
u64 paramv[] = {(u64)fibDone[0], my_ID};
ocrGuid_t depv = fibArg[0];
ocrGuid_t templateGuid;
ocrEdtTemplateCreate(&templateGuid, fibEdt, 2, 1);
ocrEdtCreate(&fib0, templateGuid, 2, paramv, 1, &depv, EDT_PROP_NONE,
NULL_GUID, NULL);
ocrEdtTemplateDestroy(templateGuid);
}
PRINTF("r%d In fibEdt(%u) -- spawned first sub-part EDT GUID 0x%llx\n", my_ID, n, fib0);
/* then do the exact same thing for n-2 */
ocrDbCreate(&fibArg[1], (void**)&ptr, sizeof(u32), DB_PROP_NONE, NULL_GUID, NO_ALLOC);
PRINTF("r%d In fibEdt(%u) -- created arg DB GUID 0x%llx\n", my_ID, n, fibArg[1]);
*((u32*)ptr) = n-2;
{
u64 paramv[] = {(u64)fibDone[1], my_ID};
ocrGuid_t depv = fibArg[1];
ocrGuid_t templateGuid;
ocrEdtTemplateCreate(&templateGuid, fibEdt, 2, 1);
ocrEdtCreate(&fib1, templateGuid, 2, paramv, 1, &depv, EDT_PROP_NONE,
NULL_GUID, NULL);
ocrEdtTemplateDestroy(templateGuid);
}
PRINTF("r%d In fibEdt(%u) -- spawned first sub-part EDT GUID 0x%llx\n", my_ID, n, fib1);
PRINTF("r%d Returning from fibEdt(%u)\n", my_ID, n);
return NULL_GUID;
}
ocrGuid_t remoteEdt(u32 paramc, u64* paramv, u32 depc, ocrEdtDep_t depv[]) {
PRINTF("RemoteEdt: executing\n");
ocrGuid_t eventGuid = ((ocrGuid_t *) depv[0].ptr)[0];
ocrEventSatisfy(eventGuid, NULL_GUID);
return NULL_GUID;
}
int main (int argc, char ** argv) {
ocrEdt_t fctPtrArray[2] = {summer, autumn};
ocrInit(&argc, argv, 2, fctPtrArray); /* No machine description */
if(argc != 3) {
printf("Usage %s <num1> <num2>\n", argv[0]);
return -1;
}
/* Create 2 data-blocks */
ocrGuid_t dbs[2];
int *data[2];
int i;
for(i = 0; i < 2; ++i) {
ocrDbCreate(&dbs[i], (void**)&data[i], sizeof(int), /*flags=*/0,
/*location=*/NULL, NO_ALLOC);
*(data[i]) = atoi(argv[i+1]);
printf("Created a data-block with value %d (GUID: 0x%lx)\n", i, (u64)dbs[i]);
}
ocrGuid_t summerEdt, autumnEdt;
ocrGuid_t autumnEvt, summerEvt[3];
ocrGuid_t summerEvtDbGuid;
ocrGuid_t *summerEvtDb;
/* Create final EDT (autumn) */
ocrEdtCreate(&autumnEdt, autumn, /*paramc=*/0, /*params=*/NULL,
/*paramv=*/NULL, /*properties=*/0, /*depc=*/1, /*depv=*/NULL);
/* Create event */
ocrEventCreate(&autumnEvt, OCR_EVENT_STICKY_T, true);
/* Create summer */
ocrEdtCreate(&summerEdt, summer, /*paramc=*/0, /*params=*/NULL,
/*paramv=*/NULL, /*properties=*/0, /*depc=*/3, /*depv=*/NULL);
/* Create events for summer */
for(i = 0; i < 3; ++i) {
ocrEventCreate(&summerEvt[i], OCR_EVENT_STICKY_T, true);
}
/* Create data-block containing event */
ocrDbCreate(&summerEvtDbGuid, (void**)&summerEvtDb, sizeof(ocrGuid_t), /*flags=*/0,
/*location=*/NULL, NO_ALLOC);
*summerEvtDb = autumnEvt;
/* Link up dependencees */
for(i = 0; i < 3; ++i) {
ocrAddDependence(summerEvt[i], summerEdt, i);
}
ocrAddDependence(autumnEvt, autumnEdt, 0);
/* "Schedule" EDTs (order does not matter) */
ocrEdtSchedule(autumnEdt);
ocrEdtSchedule(summerEdt);
printf("Done all scheduling, now going to satisfy\n");
/* Satisfy dependences passing data */
ocrEventSatisfy(summerEvt[0], dbs[0]);
ocrEventSatisfy(summerEvt[1], dbs[1]);
ocrEventSatisfy(summerEvt[2], summerEvtDbGuid);
/* Finalize */
ocrCleanup();
return 0;
}
void SPMD_Finalize(info_t info)
{
ocrEventSatisfy(info.evt_finalize, NULL_GUID);
}
