//--------------------------------------------------------------------------------------
// FUNCTION: RCCE_barrier
//--------------------------------------------------------------------------------------
// very simple, linear barrier
//--------------------------------------------------------------------------------------
int RCCE_barrier(RCCE_COMM *comm) {
int counter, i, error;
int ROOT = 0;
volatile unsigned char cyclechar[RCCE_LINE_SIZE];
volatile unsigned char valchar[RCCE_LINE_SIZE];
volatile int *cycle;
volatile int *val;
counter = 0;
cycle = (volatile int *)cyclechar;
val = (volatile int *)valchar;
if (RCCE_debug_synch)
fprintf(STDERR,"UE %d has checked into barrier\n", RCCE_IAM);
// flip local barrier variable
if (error = RCCE_get(cyclechar, (t_vcharp)(comm->gather), RCCE_LINE_SIZE, RCCE_IAM))
return(RCCE_error_return(RCCE_debug_synch,error));
*cycle = !(*cycle);
if (error = RCCE_put((t_vcharp)(comm->gather), cyclechar, RCCE_LINE_SIZE, RCCE_IAM))
return(RCCE_error_return(RCCE_debug_synch,error));
if (RCCE_IAM==comm->member[ROOT]) {
// read "remote" gather flags; once all equal "cycle" (i.e counter==comm->size),
// we know all UEs have reached the barrier
while (counter != comm->size) {
// skip the first member (#0), because that is the ROOT
for (counter=i=1; i<comm->size; i++) {
/* copy flag values out of comm buffer */
if (error = RCCE_get(valchar, (t_vcharp)(comm->gather), RCCE_LINE_SIZE,
comm->member[i]))
return(RCCE_error_return(RCCE_debug_synch,error));
if (*val == *cycle) counter++;
}
}
// set release flags
for (i=1; i<comm->size; i++) {
if (error = RCCE_flag_write(&(comm->release), *cycle, comm->member[i]))
return(RCCE_error_return(RCCE_debug_synch,error));
}
}
else {
if (error = RCCE_wait_until(comm->release, *cycle))
return(RCCE_error_return(RCCE_debug_synch,error));
}
if (RCCE_debug_synch) fprintf(STDERR,"UE %d has cleared barrier\n", RCCE_IAM);
return(RCCE_SUCCESS);
}
//--------------------------------------------------------------------------------------
// FUNCTION: RCCE_barrier
//--------------------------------------------------------------------------------------
// very simple, linear barrier
//--------------------------------------------------------------------------------------
int RCCE_barrier(RCCE_COMM *comm) {
int counter, i, error;
int ROOT = 0;
t_vchar cyclechar[RCCE_LINE_SIZE];
t_vchar valchar [RCCE_LINE_SIZE];
t_vcharp gatherp, releasep;
RCCE_FLAG_STATUS cycle;
counter = 0;
gatherp = comm->gather.line_address;
if (RCCE_debug_synch)
fprintf(STDERR,"UE %d has checked into barrier\n", RCCE_IAM);
// flip local barrier variable
if (error = RCCE_get(cyclechar, gatherp, RCCE_LINE_SIZE, RCCE_IAM))
return(RCCE_error_return(RCCE_debug_synch,error));
cycle = RCCE_flip_bit_value(cyclechar, comm->gather.location);
if (error = RCCE_put(comm->gather.line_address, cyclechar, RCCE_LINE_SIZE, RCCE_IAM))
return(RCCE_error_return(RCCE_debug_synch,error));
if (RCCE_IAM==comm->member[ROOT]) {
// read "remote" gather flags; once all equal "cycle" (i.e counter==comm->size),
// we know all UEs have reached the barrier
while (counter != comm->size) {
// skip the first member (#0), because that is the ROOT
for (counter=i=1; i<comm->size; i++) {
// copy flag values out of comm buffer
if (error = RCCE_get(valchar, comm->gather.line_address, RCCE_LINE_SIZE,
comm->member[i]))
return(RCCE_error_return(RCCE_debug_synch,error));
if (RCCE_bit_value(valchar, comm->gather.location) == cycle) counter++;
}
}
// set release flags
for (i=1; i<comm->size; i++)
if (error = RCCE_flag_write(&(comm->release), cycle, comm->member[i]))
return(RCCE_error_return(RCCE_debug_synch,error));
}
else {
if (error = RCCE_wait_until(comm->release, cycle))
return(RCCE_error_return(RCCE_debug_synch,error));
}
if (RCCE_debug_synch) fprintf(STDERR,"UE %d has cleared barrier\n", RCCE_IAM);
return(RCCE_SUCCESS);
}
//--------------------------------------------------------------------------------------
// FUNCTION: RCCE_nb_barrier
//--------------------------------------------------------------------------------------
// non-blocking version of the linear barrier
//--------------------------------------------------------------------------------------
int RCCE_nb_barrier(RCCE_COMM *comm) {
volatile unsigned char cyclechar[RCCE_LINE_SIZE] __attribute__ ((aligned (RCCE_LINE_SIZE)));
volatile unsigned char valchar[RCCE_LINE_SIZE] __attribute__ ((aligned (RCCE_LINE_SIZE)));
int i, error;
int ROOT = 0;
#ifdef USE_FLAG_EXPERIMENTAL
volatile char *cycle;
volatile char *val;
cycle = (volatile char *)cyclechar;
val = (volatile char *)valchar;
#else
volatile int *cycle;
volatile int *val;
cycle = (volatile int *)cyclechar;
val = (volatile int *)valchar;
#endif
if(comm->label == 1) goto label1;
if(comm->label == 2) goto label2;
comm->count = 0;
if (RCCE_debug_synch)
fprintf(STDERR,"UE %d has checked into barrier\n", RCCE_IAM);
#ifdef USE_FAT_BARRIER
// flip local barrier variable
#ifndef USE_FLAG_EXPERIMENTAL
if ((error = RCCE_get(cyclechar, (t_vcharp)(comm->gather[RCCE_IAM]), RCCE_LINE_SIZE, RCCE_IAM)))
#else
if ((error = RCCE_get_flag(cyclechar, (t_vcharp)(comm->gather[RCCE_IAM]), RCCE_LINE_SIZE, RCCE_IAM)))
#endif
return(RCCE_error_return(RCCE_debug_synch,error));
*cycle = !(*cycle);
#ifndef USE_FLAG_EXPERIMENTAL
if ((error = RCCE_put((t_vcharp)(comm->gather[RCCE_IAM]), cyclechar, RCCE_LINE_SIZE, RCCE_IAM)))
#else
if ((error = RCCE_put_flag((t_vcharp)(comm->gather[RCCE_IAM]), cyclechar, RCCE_LINE_SIZE, RCCE_IAM)))
#endif
return(RCCE_error_return(RCCE_debug_synch,error));
if ((error = RCCE_put((t_vcharp)(comm->gather[RCCE_IAM]), cyclechar, RCCE_LINE_SIZE, comm->member[ROOT])))
return(RCCE_error_return(RCCE_debug_synch,error));
if (RCCE_IAM==comm->member[ROOT]) {
// read "remote" gather flags; once all equal "cycle" (i.e counter==comm->size),
// we know all UEs have reached the barrier
comm->cycle = *cycle;
label1:
while (comm->count != comm->size) {
// skip the first member (#0), because that is the ROOT
for (comm->count=i=1; i<comm->size; i++) {
/* copy flag values out of comm buffer */
#ifndef USE_FLAG_EXPERIMENTAL
if ((error = RCCE_get(valchar, (t_vcharp)(comm->gather[i]), RCCE_LINE_SIZE, RCCE_IAM)))
#else
if ((error = RCCE_get_flag(valchar, (t_vcharp)(comm->gather[i]), RCCE_LINE_SIZE, RCCE_IAM)))
#endif
return(RCCE_error_return(RCCE_debug_synch,error));
if (*val == comm->cycle) comm->count++;
}
if(comm->count != comm->size) {
comm->label = 1;
return(RCCE_PENDING);
}
}
// set release flags
for (i=1; i<comm->size; i++) {
if ((error = RCCE_flag_write(&(comm->release), comm->cycle, comm->member[i])))
return(RCCE_error_return(RCCE_debug_synch,error));
}
}
else {
int test;
comm->cycle = *cycle;
label2:
RCCE_test_flag(comm->release, comm->cycle, &test);
if(!test) {
comm->label = 2;
return(RCCE_PENDING);
}
}
comm->label = 0;
#else // !USE_FAT_BARRIER
// flip local barrier variable
#ifndef USE_FLAG_EXPERIMENTAL
if ((error = RCCE_get(cyclechar, (t_vcharp)(comm->gather[0]), RCCE_LINE_SIZE, RCCE_IAM)))
#else
if ((error = RCCE_get_flag(cyclechar, (t_vcharp)(comm->gather[0]), RCCE_LINE_SIZE, RCCE_IAM)))
#endif
return(RCCE_error_return(RCCE_debug_synch,error));
*cycle = !(*cycle);
#ifndef USE_FLAG_EXPERIMENTAL
if ((error = RCCE_put((t_vcharp)(comm->gather[0]), cyclechar, RCCE_LINE_SIZE, RCCE_IAM)))
#else
if ((error = RCCE_put_flag((t_vcharp)(comm->gather[0]), cyclechar, RCCE_LINE_SIZE, RCCE_IAM)))
#endif
return(RCCE_error_return(RCCE_debug_synch,error));
if (RCCE_IAM==comm->member[ROOT]) {
// read "remote" gather flags; once all equal "cycle" (i.e counter==comm->size),
// we know all UEs have reached the barrier
comm->cycle = *cycle;
label1:
while (comm->count != comm->size) {
// skip the first member (#0), because that is the ROOT
for (comm->count=i=1; i<comm->size; i++) {
/* copy flag values out of comm buffer */
#ifndef USE_FLAG_EXPERIMENTAL
if ((error = RCCE_get(valchar, (t_vcharp)(comm->gather[0]), RCCE_LINE_SIZE,
comm->member[i])))
#else
if ((error = RCCE_get_flag(valchar, (t_vcharp)(comm->gather[0]), RCCE_LINE_SIZE,
comm->member[i])))
#endif
return(RCCE_error_return(RCCE_debug_synch,error));
if (*val == comm->cycle) comm->count++;
}
if(comm->count != comm->size) {
comm->label = 1;
return(RCCE_PENDING);
}
}
// set release flags
for (i=1; i<comm->size; i++) {
if ((error = RCCE_flag_write(&(comm->release), comm->cycle, comm->member[i])))
return(RCCE_error_return(RCCE_debug_synch,error));
}
}
else {
int test;
comm->cycle = *cycle;
label2:
RCCE_test_flag(comm->release, comm->cycle, &test);
if(!test) {
comm->label = 2;
return(RCCE_PENDING);
}
}
comm->label = 0;
#endif // !USE_FAT_BARRIER
if (RCCE_debug_synch) fprintf(STDERR,"UE %d has cleared barrier\n", RCCE_IAM);
return(RCCE_SUCCESS);
}
int RCCE_APP(int argc, char **argv){
int ID, ID_nb, ID_donor, nrounds, error, strlength;
RCCE_FLAG flag_sent, flag_ack;
double *cbuffer, *buffer, sum;
char msg[RCCE_MAX_ERROR_STRING];
RCCE_init(&argc, &argv);
ID = RCCE_ue();
ID_nb = (ID+1)%RCCE_num_ues();
ID_donor = (ID-1+RCCE_num_ues())%RCCE_num_ues();
if (argc != 2) {
if (ID==0) printf("Executable requires one parameter (number of rounds): %d\n",argc-1);
return(1);
}
nrounds = atoi(*++argv);
if (nrounds < 0) {
if (ID==0) printf("Number of rounds should be non-negative: %d\n", nrounds);
return(1);
}
/* allocate private memory and comm buffer space */
buffer = (double *) malloc(BUFSIZE*sizeof(double));
if (!buffer) printf("Mark 01: Failed to allocate private buffer on proc %d\n", ID);
cbuffer = (double *) RCCE_malloc(BUFSIZE*sizeof(double));
if (!buffer) printf("Mark 02:RCCE failed to allocate %d doubles on proc %d\n",
BUFSIZE, ID);
/* initialize buffer with UE-specific data */
for (int i=0; i<BUFSIZE; i++) buffer[i] = (double)(ID+1+i);
sum = 0.0;
for (int i=0; i<BUFSIZE; i++) sum += buffer[i];
printf("Initial sum on UE %03d equals %f\n", ID, sum);
/* create and initialize flag variables */
if (error=RCCE_flag_alloc(&flag_sent))
printf("Mark 03a: Could not allocate flag_sent on %d, error=%d\n", ID, error);
if (error=RCCE_flag_alloc(&flag_ack))
printf("Mark 03b: Could not allocate flag_ack on %d, error=%d\n", ID, error);
if(error=RCCE_flag_write(&flag_sent, RCCE_FLAG_UNSET, ID))
printf("Mark 04: Could not initialize flag_sent on %d, error=%d\n", ID, error);
if(error=RCCE_flag_write(&flag_ack, RCCE_FLAG_SET, ID_donor))
printf("Mark 05: Could not initialize flag_ack on %d, error=%d\n", ID_donor, error);
for (int round=0; round<nrounds; round++) {
int size = BUFSIZE*sizeof(double);
RCCE_wait_until(flag_ack, RCCE_FLAG_SET);
RCCE_flag_write(&flag_ack, RCCE_FLAG_UNSET, ID);
RCCE_put((t_vcharp)cbuffer, (t_vcharp)buffer, size, ID_nb);
RCCE_flag_write(&flag_sent, RCCE_FLAG_SET, ID_nb);
RCCE_wait_until(flag_sent, RCCE_FLAG_SET);
RCCE_flag_write(&flag_sent, RCCE_FLAG_UNSET, ID);
RCCE_get((t_vcharp)buffer, (t_vcharp)cbuffer, size, ID);
RCCE_flag_write(&flag_ack, RCCE_FLAG_SET, ID_donor);
}
/* compute local sum */
sum = 0.0;
for (int i=0; i<BUFSIZE; i++) sum += buffer[i];
printf("Final sum on UE %03d equals %f\n", ID, sum);
RCCE_finalize();
return(0);
}
int RCCE_APP(int argc, char **argv) {
/* statically allocated space sits in off-chip private memory */
float a[NXNY], *buff;
int i, offset, iter=10, tile;
int MY_ID;
int NTILES1;
double time;
RCCE_FLAG flag0, flag1;
RCCE_init(&argc, &argv);
NTILES1 = RCCE_num_ues()-1;
MY_ID = RCCE_ue();
if (NX%8) {
printf("Grid width should be multiple of 8: %d\n", NX);
exit(1);
}
if (argc>1) iter=atoi(*++argv);
if (MY_ID==0) printf("Executing %d iterations\n", iter);
/* allocate space on the comm buffer */
buff = (float *) RCCE_malloc(sizeof(float)*2*NX);
/* Allocate flags to coordinate comm. */
if (RCCE_flag_alloc(&flag0)) return(1);
if (RCCE_flag_alloc(&flag1)) return(1);
/* initialize array a on all tiles; this stuffs a into private caches */
for (offset=0, i=0; i<NXNY; i++) a[i+offset] = 0.0;
if (MY_ID == 0)
for (offset=0, i=0; i<NX; i++) a[i+offset] = 1.0;
if (MY_ID == NTILES1)
for (offset=NXNY1,i=0; i<NX; i++) a[i+offset] = 2.0;
/* put in a barrier so everybody can be sure to have initialized */
RCCE_barrier(&RCCE_COMM_WORLD);
/* main loop */
if (MY_ID==0) time = RCCE_wtime();
while ((iter--)>0){
/* start with copying fringe data to neighboring tiles */
if (MY_ID!=NTILES1) {
/* Initialize neighbor flag to zero */
RCCE_flag_write(&flag0, RCCE_FLAG_UNSET, MY_ID+1);
/* copy private data to shared comm buffer of neighbor */
RCCE_put((t_vcharp)(&buff[0]), (t_vcharp)(&a[NXNY2]), NX*sizeof(float), MY_ID+1);
RCCE_flag_write(&flag0, RCCE_FLAG_SET, MY_ID+1);
}
if (MY_ID != 0) {
/* Initialize neighbor flag to zero */
RCCE_flag_write(&flag1, 0, MY_ID-1);
/* copy private data to shared comm buffer of neighbor */
RCCE_put((t_vcharp)(&buff[NX]), (t_vcharp)(&a[NX]), NX*sizeof(float), MY_ID-1);
RCCE_flag_write(&flag1, RCCE_FLAG_SET, MY_ID-1);
}
/* Make sure the data has been recvd and copy data out of buffer(s) */
if (MY_ID!=NTILES1) {
RCCE_wait_until(flag1, RCCE_FLAG_SET);
RCCE_get((t_vcharp)(&a[NXNY1]), (t_vcharp)(&buff[NX]), NX*sizeof(float),MY_ID);
}
if (MY_ID!=0) {
RCCE_wait_until(flag0, RCCE_FLAG_SET);
RCCE_get((t_vcharp)(&a[0]), (t_vcharp)(&buff[0]), NX*sizeof(float),MY_ID);
}
/* apply the stencil operation */
for (i=0; i<NXNY2; i++) {
a[i+O3] +=
W1*a[i+O1] + W2*a[i+O2] + W3*a[i+O3] + W4*a[i+O4] + W5*a[i+O5];
}
}
RCCE_barrier(&RCCE_COMM_WORLD);
if (MY_ID==0) {
time = RCCE_wtime()-time;
}
/* print result strip by strip; this would not be done on RC */
for (int id=0; id<=NTILES1; id++) {
RCCE_barrier(&RCCE_COMM_WORLD);
if (MY_ID==id) {
int start = NX; int end = NXNY1;
if (MY_ID==0) start = 0;
if (MY_ID == NTILES1) end = NXNY;
for (offset=0, i=start; i<end; i++) {
if (!(i%NX)) printf("\n");
// comment out next line and uncomment subsequent three to print error
printf("%f ",a[i+offset]);
// int jj=i/NX+(MY_ID*(NY-1));
// double aexact=1.0+(double)jj/((NTILES1+1)*(NY-1));
// printf("%f ",a[i+offset]-aexact);
}
}
}
RCCE_barrier(&RCCE_COMM_WORLD);
if (MY_ID==0) {
printf("\nTotal time: %lf\n", time);
}
RCCE_finalize();
return(0);
}