static inline void atomic_add(int me, int iterations, int T)
{
int i;
if (me == 0)
pre_op_check(__func__, target[T], iterations, 0);
target[T] = 0;
shmem_barrier_all();
if (me == 1) {
for (i = 0; i < iterations; i++) {
shmem_int_add(&target[T], 1, 0);
shmem_fence();
}
shmem_int_add(&target[T], 1, 0);
if (debug)
printf("PE 1 done with operation\n");
} else
wait_until(&target[T], (iterations+1), 0);
if (verbose) {
if (me == 1)
printf("SHMEM %s finished\n", __func__);
}
}
int oshmem_shmem_preconnect_all(void)
{
int rc = OSHMEM_SUCCESS;
/* force qp creation and rkey exchange for memheap. Does not force exchange of static vars */
if (oshmem_preconnect_all) {
long val;
int nproc = 0;
int i;
val = 0xdeadbeaf;
if (!preconnect_value) {
rc =
MCA_MEMHEAP_CALL(private_alloc(sizeof(long), (void **)&preconnect_value));
}
if (!preconnect_value || (rc != OSHMEM_SUCCESS)) {
SHMEM_API_ERROR("shmem_preconnect_all failed");
return OSHMEM_ERR_OUT_OF_RESOURCE;
}
nproc = _num_pes();
for (i = 0; i < nproc; i++) {
shmem_long_p(preconnect_value, val, i);
}
shmem_fence();
shmem_barrier_all();
SHMEM_API_VERBOSE(5, "Preconnected all PEs");
}
return OSHMEM_SUCCESS;
}
static inline void fetchatomic_inc(int me, int iterations, int T, int S)
{
int i;
if (me == 0)
pre_op_check(__func__, target[T], iterations, 0);
target[T] = 0;
shmem_barrier_all();
if (me == 1) {
if (debug) {
printf("BEFORE flag PE 1 value of source is %d\n",
source[S]);
}
for (i = 0; i < iterations; i++) {
source[S] = shmem_int_finc(&target[T], 0);
shmem_fence();
}
post_op_check("finc", source[S], (iterations-1), 1);
} else
wait_until(&target[T], iterations, 0);
if (verbose) {
if (me == 1)
printf("SHMEM int_finc finished\n");
}
}
double
message_rate (struct pe_vars v, long * buffer, int size, int iterations)
{
int64_t begin, end;
int i, offset;
/*
* Touch memory
*/
memset(buffer, size, sizeof(long) * MAX_MSG_SZ * ITERS_LARGE);
shmem_barrier_all();
if (v.me < v.pairs) {
begin = TIME();
for (i = 0, offset = 0; i < iterations; i++, offset += size) {
//shmem_putmem(&buffer[offset], &buffer[offset], size, v.nxtpe);
shmem_long_put(&buffer[offset], &buffer[offset], size, v.nxtpe);
}
shmem_fence(v.nxtpe);
//shmem_quiet();
end = TIME();
return ((double)iterations * 1e6) / ((double)end - (double)begin);
}
return 0;
}
static inline void gettest(int me, int iterations, int T, int S, int P)
{
int i;
if (me == 1) {
pre_op_check(__func__, target[T], iterations, 1);
shmem_int_p(&source[S], iterations, 0);
shmem_fence();
for (i = 0; i < iterations; i++)
target[T] = shmem_int_g(&source[S], 0);
shmem_int_p(&sync_pes[P], iterations, 0);
post_op_check("get", target[T], iterations, 1);
} else
wait_until(&sync_pes[P], iterations, 0);
if (verbose) {
if (me == 0)
printf("SHMEM %s finished\n", __func__);
}
}
static inline void fetchatomic_add(int me, int iterations, int T, int S)
{
int i;
if (me == 1)
pre_op_check(__func__, target[T], iterations, 1);
target[T] = 0;
shmem_barrier_all();
if (me == 0) {
if (debug) {
printf("BEFORE flag PE 0 value of source is"
" %d = 0?\n", source[S]);
}
for (i = 0; i < iterations; i++) {
source[S] = shmem_int_fadd(&target[T], 1, 1);
shmem_fence();
}
source[S] = shmem_int_fadd(&target[T], 1, 1);
post_op_check("fadd", source[S], iterations, 0);
} else
wait_until(&target[T], (iterations+1), 1);
if (verbose) {
if (me == 0)
printf("SHMEM %s finished\n", __func__);
}
}
static int test_item1(void)
{
int rc = TC_PASS;
int num_proc = 0;
int my_proc = 0;
int peer;
int size;
char *buf;
int test_byte;
int max_heap_size_per_proc;
num_proc = _num_pes();
my_proc = _my_pe();
peer = (my_proc + 1) % num_proc;
max_heap_size_per_proc = 1L << (sys_log2((memheap_size() * HEAP_USAGE_PERCENT)/ num_proc) - 1);
max_heap_size_per_proc = (max_heap_size_per_proc > MAX_SIZE) ? MAX_SIZE : max_heap_size_per_proc;
buf = (char *)shmalloc(max_heap_size_per_proc * num_proc);
if (!buf)
{
log_error(OSH_TC, "shmalloc(%d)\n", max_heap_size_per_proc * num_proc);
return TC_SETUP_FAIL;
}
size = 1L << sys_log2(num_proc);
size = ((size - 2) > 0) ? size : 4;
log_debug(OSH_TC, "%d: buf = %p size=%d\n", my_proc, buf, size);
for (; size <= max_heap_size_per_proc; size *=2)
{
memset(buf + max_heap_size_per_proc * my_proc, 1 + my_proc % (size - 2), max_heap_size_per_proc);
log_debug(OSH_TC, "\n%d: b4 barrier size = %d\n", my_proc, size);
shmem_barrier_all();
log_debug(OSH_TC, "%d: b4 putmem size = %d %p -> %p\n", my_proc, size,
buf+max_heap_size_per_proc*my_proc, buf + max_heap_size_per_proc * my_proc);
shmem_putmem(buf+max_heap_size_per_proc*my_proc, buf+max_heap_size_per_proc*my_proc, size, peer);
shmem_fence();
test_byte = 0;
log_debug(OSH_TC, "%d: b4 getmem size = %d\n %p <- %p ", my_proc, size,
&test_byte,
buf+max_heap_size_per_proc*peer + size - 1
);
shmem_getmem(&test_byte, buf+max_heap_size_per_proc*my_proc + size - 1, 1, peer);
log_debug(OSH_TC, "%d: after getmem size = %d result=%x\n", my_proc, size, test_byte);
if (test_byte != 1 + my_proc % (size-2))
{
log_error(OSH_TC, "fence failed at size %d got = %x expected = %x\n", size, test_byte, 1 + my_proc % (size-2));
rc = TC_FAIL;
}
}
shfree(buf);
log_debug(OSH_TC, rc == TC_PASS? "passed" : "failed");
return rc;
}
int oshmem_shmem_preconnect_all(void)
{
int mca_value = 0;
int rc = OSHMEM_SUCCESS;
(void) mca_base_var_register("oshmem",
"runtime",
NULL,
"preconnect_all",
"Whether to force SHMEM processes to fully "
"wire-up the connections between SHMEM "
"processes during "
"initialization (vs. making connections lazily -- "
"upon the first SHMEM traffic between each "
"process peer pair)",
MCA_BASE_VAR_TYPE_INT,
NULL,
0,
MCA_BASE_VAR_FLAG_SETTABLE,
OPAL_INFO_LVL_9,
MCA_BASE_VAR_SCOPE_READONLY,
&mca_value);
/* force qp creation and rkey exchange for memheap. Does not force exchange of static vars */
if (mca_value) {
long val;
int nproc = 0;
int i;
val = 0xdeadbeaf;
if (!preconnect_value) {
rc =
MCA_MEMHEAP_CALL(private_alloc(sizeof(long), (void **)&preconnect_value));
}
if (!preconnect_value || (rc != OSHMEM_SUCCESS)) {
SHMEM_API_ERROR("shmem_preconnect_all failed");
return OSHMEM_ERR_OUT_OF_RESOURCE;
}
nproc = _num_pes();
for (i = 0; i < nproc; i++) {
shmem_long_p(preconnect_value, val, i);
}
shmem_fence();
shmem_barrier_all();
SHMEM_API_VERBOSE(5, "Preconnected all PEs");
}
return OSHMEM_SUCCESS;
}
int main(void)
{
long source[10] = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 };
int src = 99;
shmem_init();
if (shmem_my_pe() == 0) {
shmem_long_put(dest, source, 10, 1); /*put1*/
shmem_long_put(dest, source, 10, 2); /*put2*/
shmem_fence();
shmem_int_put(&targ, &src, 1, 1); /*put3*/
shmem_int_put(&targ, &src, 1, 2); /*put4*/
}
shmem_barrier_all(); /* sync sender and receiver */
printf("dest[0] on PE %d is %ld\n", shmem_my_pe(), dest[0]);
return 1;
}
void shmem_barrier_all(void)
{
int rc = OSHMEM_SUCCESS;
#if OSHMEM_SPEC_COMPAT == 1
/* all outstanding puts must be completed */
shmem_fence();
#endif
if (mca_scoll_sync_array) {
rc = oshmem_group_all->g_scoll.scoll_barrier(oshmem_group_all,
mca_scoll_sync_array,
SCOLL_DEFAULT_ALG);
}
RUNTIME_CHECK_RC(rc);
}
/****************************************************************************
* Place for Test Item functions
***************************************************************************/
static int test_item1(void)
{
int rc = TC_PASS;
int myPe = shmem_my_pe();
int myPeer = myPe + ( myPe % 2 ? -1 : 1 ) ;
int nPe = shmem_n_pes();
int remainderPe = nPe - (nPe % 2);
static int statArray[ARRAY_SIZE];
int* dynamicArray = shmalloc( ARRAY_SIZE * sizeof(int) );
int iterate;
for (iterate = 0; iterate < ARRAY_SIZE; iterate++)
{
if (myPe != remainderPe)
{
int tryIterate;
int putNum, getNum;
for (tryIterate = 0; tryIterate < TRY_SIZE; tryIterate++)
{
putNum = iterate + myPe;
shmem_int_put(&statArray[iterate], &putNum, 1, myPeer);
shmem_int_put(&dynamicArray[iterate], &putNum, 1, myPeer);
}
shmem_fence();
shmem_int_get(&getNum, &statArray[iterate], 1, myPeer);
if (getNum != putNum)
{
rc = TC_FAIL;
}
shmem_int_get(&getNum, &dynamicArray[iterate], 1, myPeer);
if (getNum != putNum)
{
rc = TC_FAIL;
}
}
shmem_barrier_all();
}
shfree(dynamicArray);
return rc;
}
static void wait_for_completion(int *wait_variable, int pe, int *rc)
{
pthread_t wait_check_thread;
wait_check_data_t wait_data;
shmem_fence();
*wait_variable = 0;
shmem_barrier_all();
shmem_int_p(wait_variable,1, pe );
wait_data.wait_item = wait_variable;
wait_data.wait_value = 0;
wait_data.rc = rc;
pthread_create(&wait_check_thread,NULL, &check_for_wait, (void *) &wait_data);
shmem_int_wait(wait_variable,0);
if (*rc == TC_PASS)
pthread_cancel(wait_check_thread);
}
int
main(int argc, char* argv[])
{
long source[10] = { 1, 2, 3, 4, 5, 6, 7, 8, 9, 10 };
long *target;
int *flag;
int i, num_pes;
int failed = 0;
shmem_init();
target = (long*) shmem_malloc(sizeof(long) * 10);
flag = (int*) shmem_malloc(sizeof(int));
*flag = 0;
num_pes=shmem_n_pes();
memset(target, 0, sizeof(long)*10);
shmem_barrier_all();
if (shmem_my_pe() == 0) {
for(i = 0; i < num_pes; i++) {
shmem_long_put_nbi(target, source, 10, i);
shmem_fence();
shmem_int_inc(flag, i);
}
}
shmem_int_wait_until(flag, SHMEM_CMP_EQ, 1);
for (i = 0; i < 10; i++) {
if (target[i] != source[i]) {
fprintf(stderr,"[%d] target[%d] = %ld, expected %ld\n",
shmem_my_pe(), i, target[i], source[i]);
failed = 1;
}
}
shmem_free(target);
shmem_free(flag);
shmem_finalize();
return failed;
}
void shmem_barrier(int PE_start, int logPE_stride, int PE_size, long *pSync)
{
int rc = OSHMEM_SUCCESS;
oshmem_group_t* group = NULL;
RUNTIME_CHECK_INIT();
#if OSHMEM_SPEC_COMPAT == 1
/* all outstanding puts must be completed */
shmem_fence();
#endif
if ((0 <= PE_start) && (0 <= logPE_stride)) {
/* Create group basing PE_start, logPE_stride and PE_size */
#if OSHMEM_GROUP_CACHE_ENABLED == 0
group = oshmem_proc_group_create(PE_start, (1 << logPE_stride), PE_size);
if (!group)
rc = OSHMEM_ERROR;
#else
group = find_group_in_cache(PE_start, logPE_stride, PE_size);
if (!group) {
group = oshmem_proc_group_create(PE_start,
(1 << logPE_stride),
PE_size);
if (!group) {
rc = OSHMEM_ERROR;
} else {
cache_group(group, PE_start, logPE_stride, PE_size);
}
}
#endif /* OSHMEM_GROUP_CACHE_ENABLED */
/* Collective operation call */
if (rc == OSHMEM_SUCCESS) {
/* Call barrier operation */
rc = group->g_scoll.scoll_barrier(group, pSync, SCOLL_DEFAULT_ALG);
}
#if OSHMEM_GROUP_CACHE_ENABLED == 0
if ( rc == OSHMEM_SUCCESS )
{
oshmem_proc_group_destroy(group);
}
#endif /* OSHMEM_GROUP_CACHE_ENABLED */
}
RUNTIME_CHECK_RC(rc);
}
static inline void putfence(int me, int iterations, int T)
{
int i;
if (me == 1)
pre_op_check(__func__, target[T], iterations, 1);
if (me == 0) {
for (i = 1; i < iterations; i++) {
shmem_int_p(&target[T], i, 1);
shmem_fence();
}
shmem_int_p(&target[T], i, 1);
} else
wait_until(&target[T], iterations, 1);
if (verbose)
if (me == 0)
printf("SHMEM %s finished\n", __func__);
}
/**
* \brief Check to make sure the test is correct, SHMEM
* \param tst Struct that tells the number of cycles and stages to run the test.
* \param m Struct that holds the results of the test.
*/
void bit_SHMEM_test(test_p tst, measurement_p m) {
#ifdef SHMEM
buffer_t *abuf, *bbuf, *cbuf;
int i, j, k, icycle, istage, partner_rank;
unsigned char pattern;
abuf = comm_newbuffer(m->buflen); /* set up exchange buffers */
bbuf = comm_newbuffer(m->buflen);
cbuf = comm_newbuffer(m->buflen);
for (icycle = 0; icycle < tst->num_cycles; icycle++) { /* multiple cycles repeat the test */
for (istage = 0; istage < tst->num_stages; istage++) { /* step through the stage schedule */
partner_rank = my_rank ^ istage; /* who's my partner for this stage? */
shmem_barrier_all();
if ((partner_rank < num_ranks) && (partner_rank != my_rank)) { /* valid pair? proceed with test */
for (k=0x00; k< 0x100; k++) { /* try each byte patter */
pattern=k;
for (i=0; i<m->buflen; i++) ((unsigned char *)(abuf->data))[i]=pattern;
shmem_putmem(bbuf->data, abuf->data, m->buflen, partner_rank);
shmem_fence();
shmem_getmem(cbuf->data, bbuf->data, m->buflen, partner_rank);
for (i=0; i<m->buflen; i++) {
if (((unsigned char *)(cbuf->data))[i] != pattern) {
printf("DATA ERROR DETECTED: node:%20s rank:%10d"
" pattern:0x%2x buflen:%10d position:%10d\n",
nodename, my_rank, (int)pattern, m->buflen, i);
} /* if mismatch */
} /* for buflen */
} /* for pattern */
} /* if valid pairing */
} /* for istage */
} /* for icycle */
shmem_barrier_all();
comm_freebuffer(cbuf);
comm_freebuffer(bbuf);
comm_freebuffer(abuf);
#endif
return;
}
int main(int argc, char ** argv) {
int Num_procs; /* number of ranks */
int Num_procsx, Num_procsy; /* number of ranks in each coord direction */
int my_ID; /* SHMEM rank */
int my_IDx, my_IDy; /* coordinates of rank in rank grid */
int right_nbr; /* global rank of right neighboring tile */
int left_nbr; /* global rank of left neighboring tile */
int top_nbr; /* global rank of top neighboring tile */
int bottom_nbr; /* global rank of bottom neighboring tile */
DTYPE *top_buf_out; /* communication buffer */
DTYPE *top_buf_in[2]; /* " " */
DTYPE *bottom_buf_out; /* " " */
DTYPE *bottom_buf_in[2];/* " " */
DTYPE *right_buf_out; /* " " */
DTYPE *right_buf_in[2]; /* " " */
DTYPE *left_buf_out; /* " " */
DTYPE *left_buf_in[2]; /* " " */
int root = 0;
int n, width, height;/* linear global and local grid dimension */
int i, j, ii, jj, kk, it, jt, iter, leftover; /* dummies */
int istart, iend; /* bounds of grid tile assigned to calling rank */
int jstart, jend; /* bounds of grid tile assigned to calling rank */
DTYPE reference_norm;
DTYPE f_active_points; /* interior of grid with respect to stencil */
int stencil_size; /* number of points in the stencil */
DTYPE flops; /* floating point ops per iteration */
int iterations; /* number of times to run the algorithm */
double avgtime, /* timing parameters */
*local_stencil_time, *stencil_time;
DTYPE * RESTRICT in; /* input grid values */
DTYPE * RESTRICT out; /* output grid values */
long total_length_in; /* total required length to store input array */
long total_length_out;/* total required length to store output array */
int error=0; /* error flag */
DTYPE weight[2*RADIUS+1][2*RADIUS+1]; /* weights of points in the stencil */
int *arguments; /* command line parameters */
int count_case=4; /* number of neighbors of a rank */
long *pSync_bcast; /* work space for collectives */
long *pSync_reduce; /* work space for collectives */
double *pWrk_time; /* work space for collectives */
DTYPE *pWrk_norm; /* work space for collectives */
int *iterflag; /* synchronization flags */
int sw; /* double buffering switch */
DTYPE *local_norm, *norm; /* local and global error norms */
/*******************************************************************************
** Initialize the SHMEM environment
********************************************************************************/
prk_shmem_init();
my_ID=prk_shmem_my_pe();
Num_procs=prk_shmem_n_pes();
pSync_bcast = (long *) prk_shmem_malloc(PRK_SHMEM_BCAST_SYNC_SIZE*sizeof(long));
pSync_reduce = (long *) prk_shmem_malloc(PRK_SHMEM_REDUCE_SYNC_SIZE*sizeof(long));
pWrk_time = (double *) prk_shmem_malloc(PRK_SHMEM_REDUCE_MIN_WRKDATA_SIZE*sizeof(double));
pWrk_norm = (DTYPE *) prk_shmem_malloc(PRK_SHMEM_REDUCE_MIN_WRKDATA_SIZE*sizeof(DTYPE));
local_stencil_time = (double *) prk_shmem_malloc(sizeof(double));
stencil_time = (double *) prk_shmem_malloc(sizeof(double));
local_norm = (DTYPE *) prk_shmem_malloc(sizeof(DTYPE));
norm = (DTYPE *) prk_shmem_malloc(sizeof(DTYPE));
iterflag = (int *) prk_shmem_malloc(2*sizeof(int));
if (!(pSync_bcast && pSync_reduce && pWrk_time && pWrk_norm && iterflag &&
local_stencil_time && stencil_time && local_norm && norm))
{
printf("Could not allocate scalar variables on rank %d\n", my_ID);
error = 1;
}
bail_out(error);
for(i=0;i<PRK_SHMEM_BCAST_SYNC_SIZE;i++)
pSync_bcast[i]=PRK_SHMEM_SYNC_VALUE;
for(i=0;i<PRK_SHMEM_REDUCE_SYNC_SIZE;i++)
pSync_reduce[i]=PRK_SHMEM_SYNC_VALUE;
arguments=(int*)prk_shmem_malloc(2*sizeof(int));
/*******************************************************************************
** process, test, and broadcast input parameters
********************************************************************************/
if (my_ID == root) {
#ifndef STAR
printf("ERROR: Compact stencil not supported\n");
error = 1;
goto ENDOFTESTS;
#endif
if (argc != 3){
printf("Usage: %s <# iterations> <array dimension> \n",
*argv);
error = 1;
goto ENDOFTESTS;
}
iterations = atoi(*++argv);
arguments[0]=iterations;
if (iterations < 1){
printf("ERROR: iterations must be >= 1 : %d \n",iterations);
error = 1;
goto ENDOFTESTS;
}
n = atoi(*++argv);
arguments[1]=n;
long nsquare = (long)n * (long)n;
if (nsquare < Num_procs){
printf("ERROR: grid size must be at least # ranks: %ld\n", nsquare);
error = 1;
goto ENDOFTESTS;
}
if (RADIUS < 0) {
printf("ERROR: Stencil radius %d should be non-negative\n", RADIUS);
error = 1;
goto ENDOFTESTS;
}
if (2*RADIUS +1 > n) {
printf("ERROR: Stencil radius %d exceeds grid size %d\n", RADIUS, n);
error = 1;
goto ENDOFTESTS;
}
ENDOFTESTS:;
}
bail_out(error);
/* determine best way to create a 2D grid of ranks (closest to square, for
best surface/volume ratio); we do this brute force for now
*/
for (Num_procsx=(int) (sqrt(Num_procs+1)); Num_procsx>0; Num_procsx--) {
if (!(Num_procs%Num_procsx)) {
Num_procsy = Num_procs/Num_procsx;
break;
}
}
my_IDx = my_ID%Num_procsx;
my_IDy = my_ID/Num_procsx;
/* compute neighbors; don't worry about dropping off the edges of the grid */
right_nbr = my_ID+1;
left_nbr = my_ID-1;
top_nbr = my_ID+Num_procsx;
bottom_nbr = my_ID-Num_procsx;
iterflag[0] = iterflag[1] = 0;
if(my_IDx==0) count_case--;
if(my_IDx==Num_procsx-1) count_case--;
if(my_IDy==0) count_case--;
if(my_IDy==Num_procsy-1) count_case--;
if (my_ID == root) {
printf("Parallel Research Kernels version %s\n", PRKVERSION);
printf("SHMEM stencil execution on 2D grid\n");
printf("Number of ranks = %d\n", Num_procs);
printf("Grid size = %d\n", n);
printf("Radius of stencil = %d\n", RADIUS);
printf("Tiles in x/y-direction = %d/%d\n", Num_procsx, Num_procsy);
printf("Type of stencil = star\n");
#ifdef DOUBLE
printf("Data type = double precision\n");
#else
printf("Data type = single precision\n");
#endif
#if LOOPGEN
printf("Script used to expand stencil loop body\n");
#else
printf("Compact representation of stencil loop body\n");
#endif
#if SPLITFENCE
printf("Split fence = ON\n");
#else
printf("Split fence = OFF\n");
#endif
printf("Number of iterations = %d\n", iterations);
}
shmem_barrier_all();
shmem_broadcast32(&arguments[0], &arguments[0], 2, root, 0, 0, Num_procs, pSync_bcast);
iterations=arguments[0];
n=arguments[1];
shmem_barrier_all();
prk_shmem_free(arguments);
/* compute amount of space required for input and solution arrays */
width = n/Num_procsx;
leftover = n%Num_procsx;
if (my_IDx<leftover) {
istart = (width+1) * my_IDx;
iend = istart + width + 1;
}
else {
istart = (width+1) * leftover + width * (my_IDx-leftover);
iend = istart + width;
}
width = iend - istart + 1;
if (width == 0) {
printf("ERROR: rank %d has no work to do\n", my_ID);
error = 1;
}
bail_out(error);
height = n/Num_procsy;
leftover = n%Num_procsy;
if (my_IDy<leftover) {
jstart = (height+1) * my_IDy;
jend = jstart + height + 1;
}
else {
jstart = (height+1) * leftover + height * (my_IDy-leftover);
jend = jstart + height;
}
height = jend - jstart + 1;
if (height == 0) {
printf("ERROR: rank %d has no work to do\n", my_ID);
error = 1;
}
bail_out(error);
if (width < RADIUS || height < RADIUS) {
printf("ERROR: rank %d has work tile smaller then stencil radius\n",
my_ID);
error = 1;
}
bail_out(error);
total_length_in = (width+2*RADIUS);
total_length_in *= (height+2*RADIUS);
total_length_in *= sizeof(DTYPE);
total_length_out = width;
total_length_out *= height;
total_length_out *= sizeof(DTYPE);
in = (DTYPE *) malloc(total_length_in);
out = (DTYPE *) malloc(total_length_out);
if (!in || !out) {
printf("ERROR: rank %d could not allocate space for input/output array\n",
my_ID);
error = 1;
}
bail_out(error);
/* fill the stencil weights to reflect a discrete divergence operator */
for (jj=-RADIUS; jj<=RADIUS; jj++) for (ii=-RADIUS; ii<=RADIUS; ii++)
WEIGHT(ii,jj) = (DTYPE) 0.0;
stencil_size = 4*RADIUS+1;
for (ii=1; ii<=RADIUS; ii++) {
WEIGHT(0, ii) = WEIGHT( ii,0) = (DTYPE) (1.0/(2.0*ii*RADIUS));
WEIGHT(0,-ii) = WEIGHT(-ii,0) = -(DTYPE) (1.0/(2.0*ii*RADIUS));
}
norm[0] = (DTYPE) 0.0;
f_active_points = (DTYPE) (n-2*RADIUS)*(DTYPE) (n-2*RADIUS);
/* intialize the input and output arrays */
for (j=jstart; j<jend; j++) for (i=istart; i<iend; i++) {
IN(i,j) = COEFX*i+COEFY*j;
OUT(i,j) = (DTYPE)0.0;
}
/* allocate communication buffers for halo values */
top_buf_out=(DTYPE*)malloc(2*sizeof(DTYPE)*RADIUS*width);
if (!top_buf_out) {
printf("ERROR: Rank %d could not allocate output comm buffers for y-direction\n", my_ID);
error = 1;
}
bail_out(error);
bottom_buf_out = top_buf_out+RADIUS*width;
top_buf_in[0]=(DTYPE*)prk_shmem_malloc(4*sizeof(DTYPE)*RADIUS*width);
if(!top_buf_in)
{
printf("ERROR: Rank %d could not allocate input comm buffers for y-direction\n", my_ID);
error=1;
}
bail_out(error);
top_buf_in[1] = top_buf_in[0] + RADIUS*width;
bottom_buf_in[0] = top_buf_in[1] + RADIUS*width;
bottom_buf_in[1] = bottom_buf_in[0] + RADIUS*width;
right_buf_out=(DTYPE*)malloc(2*sizeof(DTYPE)*RADIUS*height);
if (!right_buf_out) {
printf("ERROR: Rank %d could not allocate output comm buffers for x-direction\n", my_ID);
error = 1;
}
bail_out(error);
left_buf_out=right_buf_out+RADIUS*height;
right_buf_in[0]=(DTYPE*)prk_shmem_malloc(4*sizeof(DTYPE)*RADIUS*height);
if(!right_buf_in)
{
printf("ERROR: Rank %d could not allocate input comm buffers for x-dimension\n", my_ID);
error=1;
}
bail_out(error);
right_buf_in[1] = right_buf_in[0] + RADIUS*height;
left_buf_in[0] = right_buf_in[1] + RADIUS*height;
left_buf_in[1] = left_buf_in[0] + RADIUS*height;
/* make sure all symmetric heaps are allocated before being used */
shmem_barrier_all();
for (iter = 0; iter<=iterations; iter++){
/* start timer after a warmup iteration */
if (iter == 1) {
shmem_barrier_all();
local_stencil_time[0] = wtime();
}
/* sw determines which incoming buffer to select */
sw = iter%2;
/* need to fetch ghost point data from neighbors */
if (my_IDy < Num_procsy-1) {
for (kk=0,j=jend-RADIUS; j<=jend-1; j++) for (i=istart; i<=iend; i++) {
top_buf_out[kk++]= IN(i,j);
}
shmem_putmem(bottom_buf_in[sw], top_buf_out, RADIUS*width*sizeof(DTYPE), top_nbr);
#if SPLITFENCE
shmem_fence();
shmem_int_inc(&iterflag[sw], top_nbr);
#endif
}
if (my_IDy > 0) {
for (kk=0,j=jstart; j<=jstart+RADIUS-1; j++) for (i=istart; i<=iend; i++) {
bottom_buf_out[kk++]= IN(i,j);
}
shmem_putmem(top_buf_in[sw], bottom_buf_out, RADIUS*width*sizeof(DTYPE), bottom_nbr);
#if SPLITFENCE
shmem_fence();
shmem_int_inc(&iterflag[sw], bottom_nbr);
#endif
}
if(my_IDx < Num_procsx-1) {
for(kk=0,j=jstart;j<=jend;j++) for(i=iend-RADIUS;i<=iend-1;i++) {
right_buf_out[kk++]=IN(i,j);
}
shmem_putmem(left_buf_in[sw], right_buf_out, RADIUS*height*sizeof(DTYPE), right_nbr);
#if SPLITFENCE
shmem_fence();
shmem_int_inc(&iterflag[sw], right_nbr);
#endif
}
if(my_IDx>0) {
for(kk=0,j=jstart;j<=jend;j++) for(i=istart;i<=istart+RADIUS-1;i++) {
left_buf_out[kk++]=IN(i,j);
}
shmem_putmem(right_buf_in[sw], left_buf_out, RADIUS*height*sizeof(DTYPE), left_nbr);
#if SPLITFENCE
shmem_fence();
shmem_int_inc(&iterflag[sw], left_nbr);
#endif
}
#if SPLITFENCE == 0
shmem_fence();
if(my_IDy<Num_procsy-1) shmem_int_inc(&iterflag[sw], top_nbr);
if(my_IDy>0) shmem_int_inc(&iterflag[sw], bottom_nbr);
if(my_IDx<Num_procsx-1) shmem_int_inc(&iterflag[sw], right_nbr);
if(my_IDx>0) shmem_int_inc(&iterflag[sw], left_nbr);
#endif
shmem_int_wait_until(&iterflag[sw], SHMEM_CMP_EQ, count_case*(iter/2+1));
if (my_IDy < Num_procsy-1) {
for (kk=0,j=jend; j<=jend+RADIUS-1; j++) for (i=istart; i<=iend; i++) {
IN(i,j) = top_buf_in[sw][kk++];
}
}
if (my_IDy > 0) {
for (kk=0,j=jstart-RADIUS; j<=jstart-1; j++) for (i=istart; i<=iend; i++) {
IN(i,j) = bottom_buf_in[sw][kk++];
}
}
if (my_IDx < Num_procsx-1) {
for (kk=0,j=jstart; j<=jend; j++) for (i=iend; i<=iend+RADIUS-1; i++) {
IN(i,j) = right_buf_in[sw][kk++];
}
}
if (my_IDx > 0) {
for (kk=0,j=jstart; j<=jend; j++) for (i=istart-RADIUS; i<=istart-1; i++) {
IN(i,j) = left_buf_in[sw][kk++];
}
}
/* Apply the stencil operator */
for (j=MAX(jstart,RADIUS); j<=MIN(n-RADIUS-1,jend); j++) {
for (i=MAX(istart,RADIUS); i<=MIN(n-RADIUS-1,iend); i++) {
#if LOOPGEN
#include "loop_body_star.incl"
#else
for (jj=-RADIUS; jj<=RADIUS; jj++) OUT(i,j) += WEIGHT(0,jj)*IN(i,j+jj);
for (ii=-RADIUS; ii<0; ii++) OUT(i,j) += WEIGHT(ii,0)*IN(i+ii,j);
for (ii=1; ii<=RADIUS; ii++) OUT(i,j) += WEIGHT(ii,0)*IN(i+ii,j);
#endif
}
}
/* add constant to solution to force refresh of neighbor data, if any */
for (j=jstart; j<jend; j++) for (i=istart; i<iend; i++) IN(i,j)+= 1.0;
}
local_stencil_time[0] = wtime() - local_stencil_time[0];
shmem_barrier_all();
shmem_double_max_to_all(&stencil_time[0], &local_stencil_time[0], 1, 0, 0,
Num_procs, pWrk_time, pSync_reduce);
/* compute L1 norm in parallel */
local_norm[0] = (DTYPE) 0.0;
for (j=MAX(jstart,RADIUS); j<MIN(n-RADIUS,jend); j++) {
for (i=MAX(istart,RADIUS); i<MIN(n-RADIUS,iend); i++) {
local_norm[0] += (DTYPE)ABS(OUT(i,j));
}
}
shmem_barrier_all();
#ifdef DOUBLE
shmem_double_sum_to_all(&norm[0], &local_norm[0], 1, 0, 0, Num_procs, pWrk_norm, pSync_reduce);
#else
shmem_float_sum_to_all(&norm[0], &local_norm[0], 1, 0, 0, Num_procs, pWrk_norm, pSync_reduce);
#endif
/*******************************************************************************
** Analyze and output results.
********************************************************************************/
/* verify correctness */
if (my_ID == root) {
norm[0] /= f_active_points;
if (RADIUS > 0) {
reference_norm = (DTYPE) (iterations+1) * (COEFX + COEFY);
}
else {
reference_norm = (DTYPE) 0.0;
}
if (ABS(norm[0]-reference_norm) > EPSILON) {
printf("ERROR: L1 norm = "FSTR", Reference L1 norm = "FSTR"\n",
norm[0], reference_norm);
error = 1;
}
else {
printf("Solution validates\n");
#ifdef VERBOSE
printf("Reference L1 norm = "FSTR", L1 norm = "FSTR"\n",
reference_norm, norm[0]);
#endif
}
}
bail_out(error);
if (my_ID == root) {
/* flops/stencil: 2 flops (fma) for each point in the stencil,
plus one flop for the update of the input of the array */
flops = (DTYPE) (2*stencil_size+1) * f_active_points;
avgtime = stencil_time[0]/iterations;
printf("Rate (MFlops/s): "FSTR" Avg time (s): %lf\n",
1.0E-06 * flops/avgtime, avgtime);
}
prk_shmem_free(top_buf_in);
prk_shmem_free(right_buf_in);
free(top_buf_out);
free(right_buf_out);
prk_shmem_free(pSync_bcast);
prk_shmem_free(pSync_reduce);
prk_shmem_free(pWrk_time);
prk_shmem_free(pWrk_norm);
prk_shmem_finalize();
exit(EXIT_SUCCESS);
}
int main(int argc, char *argv[])
{
int size, rank, world_rank, my_group;
int num_lsms; // number of parallel LSMS instances
int size_lsms; // number of atoms in a lsms instance
int num_steps; // number of energy calculations
int initial_steps; // number of steps before sampling starts
int stepCount=0; // count the Monte Carlo steps executed
double max_time; // maximum walltime for this run in seconds
bool restrict_time = false; // was the maximum time specified?
bool restrict_steps = false; // or the max. numer of steps?
int align; // alignment of lsms_instances
double magnetization;
double energy_accumulator; // accumulates the enegy to calculate the mean
int energies_accumulated;
int new_peid,new_root;
static int op,flag;
double *evec,*r_values;
evec=(double *)shmalloc(sizeof(double)*3*size_lsms);
r_values=(double *)shmalloc(sizeof(double)*(R_VALUE_OFFSET+3*(size_lsms+1)));
energy_accumulator=0.0;
energies_accumulated=0;
double walltime_0,walltime;
double restartWriteFrequency=30.0*60.0;
double nextWriteTime=restartWriteFrequency;
MPI_Comm local_comm;
int *lsms_rank0;
MPI_Status status;
char prefix[40];
char i_lsms_name[64];
char gWL_in_name[64], gWL_out_name[64];
char mode_name[64];
char energy_calculation_name[64];
char stupid[37];
char step_out_name[64];
char wl_step_out_name[128];
char *wl_stepf=NULL;
bool step_out_flag=false;
std::ofstream step_out_file;
typedef enum {Constant, Random, WangLandau_1d, ExhaustiveIsing, WangLandau_2d} EvecGenerationMode;
typedef enum {MagneticMoment, MagneticMomentZ, MagneticMomentX, MagneticMomentY} SecondDimension;
EvecGenerationMode evec_generation_mode = Constant;
SecondDimension second_dimension = MagneticMoment;
double ev0[3];
bool return_moments_flag=true; // true-> return all magnetic moments from lsms run at each step.
bool generator_needs_moment=false;
typedef enum {OneStepEnergy, MultiStepEnergy, ScfEnergy} EnergyCalculationMode;
EnergyCalculationMode energyCalculationMode = OneStepEnergy;
int energyIndex=1; // index for the return value to use for the MC step (0: total energy, 1: band energy)
ev0[0]=ev0[1]=0.0; ev0[2]=1.0;
// size has to be align + size_lsms*num_lsms
align=1;
num_lsms=1;
size_lsms=-1;
my_group=-1;
num_steps=1;
initial_steps=0;
sprintf(i_lsms_name,"i_lsms");
gWL_in_name[0]=gWL_out_name[0]=0;
mode_name[0]=0;
energy_calculation_name[0]=0;
// check command line arguments
for(int i=0; i<argc; i++)
{
if(!strcmp("-num_lsms",argv[i])) num_lsms=atoi(argv[++i]);
if(!strcmp("-size_lsms",argv[i])) size_lsms=atoi(argv[++i]);
if(!strcmp("-align",argv[i])) align=atoi(argv[++i]);
if(!strcmp("-num_steps",argv[i])) {num_steps=atoi(argv[++i]); restrict_steps=true;}
if(!strcmp("-initial_steps",argv[i])) initial_steps=atoi(argv[++i]);
if(!strcmp("-walltime",argv[i])) {max_time=60.0*atof(argv[++i]); restrict_time=true;}
if(!strcmp("-i",argv[i])) strncpy(i_lsms_name,argv[++i],64);
if(!strcmp("-random_dir",argv[i])) {evec_generation_mode = Random;}
if(!strcmp("-step_out",argv[i]))
{strncpy(step_out_name,argv[++i],64); step_out_flag=true;
return_moments_flag=true;}
if(!strcmp("-wl_out", argv[i])) strncpy(gWL_out_name,argv[++i],64);
if(!strcmp("-wl_in", argv[i])) strncpy(gWL_in_name,argv[++i],64);
if(!strcmp("-mode", argv[i])) strncpy(mode_name,argv[++i],64);
if(!strcmp("-energy_calculation",argv[i])) strncpy(energy_calculation_name,argv[++i],64);
}
if(!(restrict_steps || restrict_time)) restrict_steps=true;
if(mode_name[0]!=0)
{
if(!strcmp("constant",mode_name)) evec_generation_mode = Constant;
if(!strcmp("random",mode_name)) evec_generation_mode = Random;
if(!strcmp("1d",mode_name)) evec_generation_mode = WangLandau_1d;
if(!strcmp("ising",mode_name)) evec_generation_mode = ExhaustiveIsing;
if(!strcmp("2d",mode_name)) evec_generation_mode = WangLandau_2d;
if(!strcmp("2d-m",mode_name)) {evec_generation_mode = WangLandau_2d; second_dimension=MagneticMoment;}
if(!strcmp("2d-x",mode_name)) {evec_generation_mode = WangLandau_2d; second_dimension=MagneticMomentX;}
if(!strcmp("2d-y",mode_name)) {evec_generation_mode = WangLandau_2d; second_dimension=MagneticMomentY;}
if(!strcmp("2d-z",mode_name)) {evec_generation_mode = WangLandau_2d; second_dimension=MagneticMomentZ;}
}
if(energy_calculation_name[0]!=0)
{
if(energy_calculation_name[0]=='o') { energyCalculationMode = OneStepEnergy; energyIndex=1; }
if(energy_calculation_name[0]=='m') { energyCalculationMode = MultiStepEnergy; energyIndex=1; }
if(energy_calculation_name[0]=='s') { energyCalculationMode = ScfEnergy; energyIndex=0; }
}
#ifdef USE_PAPI
#define NUM_PAPI_EVENTS 4
int hw_counters = PAPI_num_counters();
if(hw_counters>NUM_PAPI_EVENTS) hw_counters=NUM_PAPI_EVENTS;
int papi_events[NUM_PAPI_EVENTS]; // = {PAPI_TOT_INS,PAPI_TOT_CYC,PAPI_FP_OPS,PAPI_VEC_INS};
char *papi_event_name[] = {"PAPI_TOT_INS","PAPI_FP_OPS",
"RETIRED_SSE_OPERATIONS:DOUBLE_ADD_SUB_OPS:DOUBLE_MUL_OPS:DOUBLE_DIV_OPS:OP_TYPE",
"RETIRED_SSE_OPERATIONS:SINGLE_ADD_SUB_OPS:SINGLE_MUL_OPS:SINGLE_DIV_OPS:OP_TYPE"};
// "RETIRED_INSTRUCTIONS",
// "RETIRED_MMX_AND_FP_INSTRUCTIONS:PACKED_SSE_AND_SSE2",
// "RETIRED_SSE_OPERATIONS:DOUBLE_ADD_SUB_OPS:DOUBLE_MUL_OPS:DOUBLE_DIV_OPS:1",
// "RETIRED_SSE_OPERATIONS:SINGLE_ADD_SUB_OPS:SINGLE_MUL_OPS:SINGLE_DIV_OPS:1"
// get events from names:
for(int i=0; i<NUM_PAPI_EVENTS; i++)
{
if(PAPI_event_name_to_code(papi_event_name[i],&papi_events[i]) != PAPI_OK)
{
// printline("Error in obtaining PAPI event code for: "+ttos(papi_event_name[i]),
// std::cerr,parameters.myrankWorld);
// printline("Skipping all following events",
// std::cerr,parameters.myrankWorld);
if(hw_counters>i) hw_counters=i;
}
}
long long papi_values[NUM_PAPI_EVENTS+4];
// printline("PAPI: "+ttos(hw_counters)+" counters available",std::cout,parameters.myrankWorld);
if(hw_counters>NUM_PAPI_EVENTS) hw_counters=NUM_PAPI_EVENTS;
long long papi_real_cyc_0 = PAPI_get_real_cyc();
long long papi_real_usec_0 = PAPI_get_real_usec();
long long papi_virt_cyc_0 = PAPI_get_virt_cyc();
long long papi_virt_usec_0 = PAPI_get_virt_usec();
PAPI_start_counters(papi_events,hw_counters);
#endif
lsms_rank0=(int *)malloc(sizeof(int)*(num_lsms+1));
// initialize MPI:
MPI_Init(&argc, &argv);
MPI_Comm_rank(MPI_COMM_WORLD, &rank);
world_rank=rank;
MPI_Comm_size(MPI_COMM_WORLD, &size);
walltime_0 = get_rtc();
#ifndef SVN_REV
#define SVN_REV "unknown"
#endif
// make sure 'return_moments_flag' is set correctly
switch(evec_generation_mode)
{
case Constant : break;
case Random : break;
case WangLandau_1d :
return_moments_flag = true;
generator_needs_moment = true;
break;
case ExhaustiveIsing : break;
case WangLandau_2d :
return_moments_flag = true;
generator_needs_moment = true;
break;
default: std::cout<<" ERROR: UNKNOWN EVEC GENERATION MODE\n"; exit(1);
}
if(rank==0)
{
std::cout<<"LSMS_3"<<std::endl;
std::cout<<" SVN revision "<<SVN_REV<<std::endl<<std::endl;
#ifdef USE_PAPI
std::cout<<" Using Papi counters"<<std::endl<<std::endl;
#endif
std::cout<<" Size of LSMS instances = "<<size_lsms<<" atoms\n";
std::cout<<" Number of LSMS instances = "<<num_lsms<<std::endl;
std::cout<<" LSMS Energy calculated using ";
switch(energyCalculationMode)
{
case OneStepEnergy: std::cout<<"oneStepEnergy [frozen potential band energy]"<<std::endl; break;
case MultiStepEnergy: std::cout<<"multiStepEnergy [frozen potential band energy with converged Fermi energy]"<<std::endl; break;
case ScfEnergy: std::cout<<"scfEnergy [self-consistent total energy]"<<std::endl; break;
default: std::cout<<"UNKNOWN ENERGY CALCULATION METHOD"<<std::endl; exit(1);
}
if(restrict_steps) std::cout<<" Number of gWL steps = "<<num_steps<<std::endl;
if(restrict_time) std::cout<<" Maximum walltime = "<<max_time<<"s\n";
std::cout<<" Processor alignment (process allocation quantization) = "<<align<<std::endl;
switch(evec_generation_mode)
{
case Constant : std::cout<<" Constant moments direction along "
<<ev0[0]<<" "<<ev0[1]<<" "<<ev0[2]<<std::endl;
break;
case Random : std::cout<<" Random distribution of moments (no Wang-Landau)"<<std::endl;
break;
case WangLandau_1d : std::cout<<" Wang-Landau for one continuous variable (energy)"<<std::endl;
// return_moments_flag = true;
// generator_needs_moment = true;
break;
case ExhaustiveIsing : std::cout<<" Exhaustive Ising sampling"<<std::endl; break;
case WangLandau_2d : std::cout<<" Wang-Landau for two continuous variable (energy, ";
switch(second_dimension)
{
case MagneticMoment : std::cout<<"magnitude of magnetization)"; break;
case MagneticMomentX : std::cout<<"x component of magnetization)"; break;
case MagneticMomentY : std::cout<<"y component of magnetization)"; break;
case MagneticMomentZ : std::cout<<"z component of magnetization)"; break;
}
std::cout<<std::endl;
// return_moments_flag = true;
// generator_needs_moment = true;
break;
default: std::cout<<" ERROR: UNKNOWN EVEC GENERATION MODE\n"; exit(1);
}
if(step_out_flag) std::cout<<" Step output written to: "<<step_out_name<<std::endl;
std::cout<<std::endl;
if(step_out_flag && (evec_generation_mode==WangLandau_1d))
{
// step_out_flag=false;
snprintf(wl_step_out_name,127,"wl1d_%s",step_out_name);
wl_stepf=wl_step_out_name;
}
if(step_out_flag)
{
step_out_file.open(step_out_name);
step_out_file<<"#";
for(int i=0; i<argc; i++) step_out_file<<" "<<argv[i];
step_out_file<<std::endl<<size_lsms<<std::endl;
}
}
if(generator_needs_moment) return_moments_flag=true;
if(num_lsms==1)
{
SHMEM_activeset local_comm;
local_comm.rank=shmem_my_pe();
local_comm.size=shmem_n_pes();
local_comm.start_pe=0;
local_comm.logPE_stride=0;
LSMS lsms_calc(local_comm,i_lsms_name,"1_");
if(rank==0)
{
std::cout<<"executing LSMS(C++) for "<<lsms_calc.numSpins()<<" atoms\n";
std::cout<<" LSMS version = "<<lsms_calc.version()<<std::endl;
}
if(energyCalculationMode==OneStepEnergy)
std::cout<<"one step Energy = "<<lsms_calc.oneStepEnergy()<<std::endl;
else if(energyCalculationMode==MultiStepEnergy)
std::cout<<"multi-step Energy = "<<lsms_calc.multiStepEnergy()<<std::endl;
else if(energyCalculationMode==ScfEnergy)
std::cout<<"self-consistent Energy = "<<lsms_calc.scfEnergy()<<std::endl;
else
{
printf("ERROR: Unknown energy calculation mode for lsms_calc in wl-lsms main!\n");
// MPI_Abort(MPI_COMM_WORLD,5);
exit(5);
}
}
else
{
// build the communicators
//int color=MPI_UNDEFINED;
//Assuming user passes a power of two while using "-align"
int s = align;
int comm_size=(size-align)/num_lsms;
int world_rank;
for(int i=0; i<num_lsms; i++)
{
if((world_rank>=s) && (world_rank<s+comm_size))
{
my_group=i;
//color=i;
new_peid=world_rank-s;
new_root=s;
}
lsms_rank0[i]=s;
s+=comm_size;
}
if(world_rank==0){
//color=num_lsms;
new_peid=0;
comm_size=1;
new_root=0;
}
//MPI_Comm_split(MPI_COMM_WORLD, color, 0, &local_comm);
SHMEM_activeset local_comm;
local_comm.rank=new_peid;
local_comm.size=comm_size;
local_comm.start_pe=new_root;
local_comm.logPE_stride=0;
std::cout<<"world_rank="<<world_rank<<" -> group="<<my_group<<std::endl;
snprintf(prefix,38,"Group %4d: ",my_group);
// now we get ready to do some calculations...
if(my_group>=0)
{
double energy;
double band_energy;
int static i_values[10];
double static r_values[10];
static int op;
//MPI_Comm_rank(local_comm, &rank);
rank = local_comm.rank;
snprintf(prefix,38,"%d_",my_group);
// to use the ramdisk on jaguarpf:
// snprintf(prefix,38,"/tmp/ompi/%d_",my_group);
LSMS lsms_calc(local_comm,i_lsms_name,prefix);
snprintf(prefix,38,"Group %4d: ",my_group);
if(rank==0 && my_group==0)
{
std::cout<<prefix<<"executing LSMS(C++) for "<<lsms_calc.numSpins()<<" atoms\n";
std::cout<<prefix<<" LSMS version = "<<lsms_calc.version()<<std::endl;
}
// wait for commands from master
bool finished=false;
while(!finished)
{
if(rank==0)
{
//MPI_Recv(evec,3*size_lsms,MPI_DOUBLE,0,MPI_ANY_TAG,MPI_COMM_WORLD,&status);
//op =status.MPI_TAG;
if (lsms_rank0[0]==world_rank)
shmem_barrier(0, lsms_rank0[0], 2, pSync1);
}
//MPI_Bcast(&op,1,MPI_INT,0,local_comm);
shmem_broadcast32(&op, &op, 1, local_comm.start_pe, local_comm.start_pe, local_comm.logPE_stride, local_comm.size, pSync2);
/* recognized opcodes:
5: calculate energy
recognized energy calculation modes:
OneStepEnergy : calclulate frozen potential band energy in one step (don't converge Ef)
use only if the Fermi energy will not change due to MC steps!
The only method available in LSMS_1.9
MultiStepEnergy : calculate frozen potential band energy after converging Fermi energy
This should be the new default method. If the Fermi energy doesn't change
multiStepEnergy only performs one step and should be equivalent to oneStepEnergy
The tolerance for Ef convergence can be set with LSMS::setEfTol(Real).
The default tolerance is set in the LSMS::LSMS constructor (currently 1.0e-6).
The maximum number of steps is read from the LSMS input file 'nscf' parameter.
ScfEnergy : this will calculate the selfconsistent total energy.
The maximum number of steps is read from the LSMS input file 'nscf' parameter.
NOT IMPLEMENTED YET!!!
10: get number of sites
*/
if(op==5)
{
lsms_calc.setEvec(evec);
if(energyCalculationMode==OneStepEnergy)
energy=lsms_calc.oneStepEnergy(&band_energy);
else if(energyCalculationMode==MultiStepEnergy)
band_energy=energy=lsms_calc.multiStepEnergy();
else if(energyCalculationMode==ScfEnergy)
energy=lsms_calc.scfEnergy(&band_energy);
else
{
printf("ERROR: Unknown energy calculation mode for lsms_calc in wl-lsms main!\n");
//MPI_Abort(MPI_COMM_WORLD,5);
exit(5);
}
r_values[0]=energy;
r_values[1]=band_energy;
if(return_moments_flag)
{
lsms_calc.getMag(&r_values[R_VALUE_OFFSET]);
}
if(rank==0)
{
if(return_moments_flag)
{
//MPI_Send(r_values,R_VALUE_OFFSET+3*size_lsms,MPI_DOUBLE,0,1005,MPI_COMM_WORLD);
shmem_double_put(r_values, r_values, R_VALUE_OFFSET+3*size_lsms, 0);
} else {
//MPI_Send(r_values,R_VALUE_OFFSET,MPI_DOUBLE,0,1005,MPI_COMM_WORLD);
shmem_double_put(r_values, r_values, R_VALUE_OFFSET, 0);
}
shmem_fence();
shmem_int_swap(&flag, world_rank, 0);
}
} else if(op==10) {
i_values[0]=lsms_calc.numSpins();
//MPI_Send(i_values,10,MPI_INT,0,1010,MPI_COMM_WORLD);
shmem_int_put(i_values, i_values, 10, 0);
} else {
// printf("world rank %d: recieved exit\n",world_rank);
finished=true;
}
}
shfree(evec);
//shfree(r_values);
}
else if(world_rank==0)
{
int running;
double **evecs;
//double *r_values;
//int i_values[10];
int *init_steps;
int total_init_steps;
bool accepted;
char *wl_inf=NULL;
char *wl_outf=NULL;
if(gWL_in_name) wl_inf=gWL_in_name;
if(gWL_out_name) wl_outf=gWL_out_name;
EvecGenerator *generator;
/*
// get number of spins from first LSMS instance
// temp r_values:
r_values=(double *)malloc(sizeof(double)*10);
MPI_Send(r_values,1,MPI_DOUBLE, lsms_rank0[0], 10, MPI_COMM_WORLD);
free(r_values);
MPI_Recv(i_values,10,MPI_INT,lsms_rank0[0],1010,MPI_COMM_WORLD,&status);
if(i_values[0]!=size_lsms)
{
printf("Size specified for Wang-Landau and in LSMS input file don't match!\n");
size_lsms=i_values[0];
}
*/
evecs=(double **)shmalloc(sizeof(double *)*num_lsms);
init_steps=(int *)shmalloc(sizeof(int)*num_lsms);
for(int i=0; i<num_lsms; i++)
{
evecs[i]=(double *)shmalloc(sizeof(double)*3*size_lsms);
init_steps[i]=initial_steps;
}
total_init_steps=num_lsms*initial_steps;
// Initialize the correct evec generator
switch(evec_generation_mode)
{
case Random : generator = new RandomEvecGenerator(size_lsms);
break;
case Constant: generator = new ConstantEvecGenerator(size_lsms, ev0, num_lsms);
break;
//case WangLandau_1d : generator = new WL1dEvecGenerator<std::mt19937>(size_lsms, num_lsms,
// evecs, wl_inf, wl_outf, wl_stepf);
case WangLandau_1d : generator = new WL1dEvecGenerator<boost::mt19937>(size_lsms, num_lsms,
evecs, wl_inf, wl_outf, wl_stepf);
break;
case ExhaustiveIsing : generator = new ExhaustiveIsing1dEvecGenerator(size_lsms, num_lsms,
evecs, wl_inf, wl_outf);
break;
//case WangLandau_2d : generator = new WL2dEvecGenerator<std::mt19937>(size_lsms, num_lsms,
// evecs, wl_inf, wl_outf, wl_stepf);
case WangLandau_2d : generator = new WL2dEvecGenerator<boost::mt19937>(size_lsms, num_lsms,
evecs, wl_inf, wl_outf, wl_stepf);
break;
default: std::cerr<<"The code should never arrive here: UNKNOWN EVEC GENERATION MODE\n";
exit(1);
}
for(int i=0; i<num_lsms; i++)
{
generator->initializeEvec(i,evecs[i]);
}
std::cout<<"This is the master node\n";
// issue initial commands to all LSMS instances
running=0;
bool more_work=true;
if(total_init_steps>0)
{
for(int i=0; i<num_lsms; i++)
{
std::cout<<"starting initial calculation in group "<<i<<std::endl;
//MPI_Send(evecs[i], 3*size_lsms, MPI_DOUBLE, lsms_rank0[i], 5, MPI_COMM_WORLD);
shmem_double_put(evec, evecs[i], 3*size_lsms, lsms_rank0[i]);
shmem_int_p(&op, 5, lsms_rank0[i]);
shmem_fence();
num_steps--; running++; stepCount++;
if(restrict_steps) std::cout<<" "<<num_steps<<" steps remaining\n";
}
shmem_barrier(0, lsms_rank0[0], 2, pSync1);
// first deal with the initial steps:
while(running>0)
{
//if(return_moments_flag)
// MPI_Recv(r_values,R_VALUE_OFFSET+3*size_lsms,MPI_DOUBLE,MPI_ANY_SOURCE,MPI_ANY_TAG,MPI_COMM_WORLD,&status);
//else
// MPI_Recv(r_values,R_VALUE_OFFSET,MPI_DOUBLE,MPI_ANY_SOURCE,MPI_ANY_TAG,MPI_COMM_WORLD,&status);
shmem_int_wait(&flag,-1);
running--;
// std::cout<<"received energy E_tot ="<<r_values[0]<<std::endl;
// std::cout<<" band energy E_band="<<r_values[1]<<std::endl;
if(total_init_steps>0)
{
//int r_group=(status.MPI_SOURCE-align)/comm_size;
int r_group=(flag-align)/comm_size;
std::cout<<"starting additional calculation in group "<<r_group<<std::endl;
if(init_steps[r_group]>0)
{
more_work = !(generator->generateUnsampledEvec(r_group,evecs[r_group],r_values[energyIndex]));
init_steps[r_group]--; total_init_steps--;
}
//MPI_Send(evecs[r_group], 3*size_lsms, MPI_DOUBLE, lsms_rank0[r_group], 5, MPI_COMM_WORLD);
shmem_double_put(r_values, evecs[r_group], 3*size_lsms, lsms_rank0[r_group]); //TODO check this
shmem_fence();
num_steps--; running++; stepCount++;
if(restrict_steps && num_steps<=0) more_work=false;
if(restrict_steps) std::cout<<" "<<num_steps<<" steps remaining\n";
walltime = get_rtc() - walltime_0;
if(restrict_time && walltime>=max_time) more_work=false;
if(restrict_time) std::cout<<" "<<max_time-walltime<<" seconds remaining\n";
}
}
}
more_work=true;
running=0;
for(int i=0; i<num_lsms; i++)
{
std::cout<<"starting main calculation in group "<<i<<std::endl;
//MPI_Send(evecs[i], 3*size_lsms, MPI_DOUBLE, lsms_rank0[i], 5, MPI_COMM_WORLD);
shmem_double_put(evec, evecs[i], 3*size_lsms, lsms_rank0[i]);
shmem_int_p(&op, 5, lsms_rank0[i]);
shmem_fence();
num_steps--; running++; stepCount++;
if(restrict_steps) std::cout<<" "<<num_steps<<" steps remaining\n";
}
shmem_barrier(0, lsms_rank0[0], 2, pSync1);
generator->startSampling();
// wait for results and issue new commands or wind down
while(running>0)
{
//MPI_Recv(r_values,R_VALUE_OFFSET+3*size_lsms,MPI_DOUBLE,MPI_ANY_SOURCE,MPI_ANY_TAG,MPI_COMM_WORLD,&status);
shmem_int_wait(&flag,-1);
running--;
std::cout<<"received energy E_tot ="<<r_values[0]<<std::endl;
std::cout<<" band energy E_band="<<r_values[1]<<std::endl;
// printf("from status.MPI_SOURCE=%d\n",status.MPI_SOURCE);
energy_accumulator+=r_values[0]; energies_accumulated++;
if(more_work)
{
int r_group=(status.MPI_SOURCE-align)/comm_size;
std::cout<<"starting additional calculation in group "<<r_group<<std::endl;
if(generator_needs_moment)
{
double m0,m1,m2;
m0=0.0; m1=0.0; m2=0.0;
for(int i=0; i<3*size_lsms; i+=3)
{
m0+=r_values[R_VALUE_OFFSET+i];
m1+=r_values[R_VALUE_OFFSET+i+1];
m2+=r_values[R_VALUE_OFFSET+i+2];
}
switch(second_dimension)
{
case MagneticMoment : magnetization=std::sqrt(m0*m0+m1*m1+m2*m2); break;
case MagneticMomentX : magnetization=m0; break;
case MagneticMomentY : magnetization=m1; break;
case MagneticMomentZ : magnetization=m2; break;
}
if(generator->generateEvec(r_group,evecs[r_group],r_values[energyIndex],magnetization, &accepted))
more_work=false;
} else {
if(generator->generateEvec(r_group,evecs[r_group],r_values[energyIndex], &accepted)) more_work=false;
}
//MPI_Send(evecs[r_group], 3*size_lsms, MPI_DOUBLE, lsms_rank0[r_group], 5, MPI_COMM_WORLD);
shmem_double_put(r_values, evecs[r_group], 3*size_lsms, lsms_rank0[r_group]); //TODO check this
shmem_fence();
num_steps--; running++; stepCount++;
if(restrict_steps && num_steps<=0) more_work=false;
if(restrict_steps) std::cout<<" "<<num_steps<<" steps remaining\n";
walltime = get_rtc() - walltime_0;
if(restrict_time && walltime>=max_time) more_work=false;
if(restrict_time) std::cout<<" "<<max_time-walltime<<" seconds remaining\n";
}
else
{
// send an exit message to this instance of LSMS
int r_group=(status.MPI_SOURCE-align)/comm_size;
MPI_Send(evecs[r_group], 3*size_lsms, MPI_DOUBLE, lsms_rank0[r_group], 2, MPI_COMM_WORLD);
}
if(step_out_flag && accepted)
{
step_out_file<<"# iteration "<<energies_accumulated<<std::endl;
step_out_file.precision(15);
step_out_file<<energies_accumulated<<std::endl;
step_out_file<<r_values[0]<<" "<<r_values[1]<<std::endl;
for(int j=0; j<3*size_lsms; j+=3)
{
step_out_file<<r_values[j+R_VALUE_OFFSET]<<" "<<r_values[j+R_VALUE_OFFSET+1]
<<" "<<r_values[j+R_VALUE_OFFSET+2]<<std::endl;
}
}
// write restart file every restartWriteFrequency seconds
if(walltime>nextWriteTime)
{
generator->writeState("WLrestart.jsn");
nextWriteTime+=restartWriteFrequency;
}
}
generator->writeState("WLrestart.jsn");
/*
if(evec_generation_mode==WangLandau_1d)
(static_cast<WL1dEvecGenerator<std::mt19937> *>(generator))->writeState("WLrestart.state");
if(evec_generation_mode==ExhaustiveIsing)
(static_cast<ExhaustiveIsing1dEvecGenerator *>(generator))->writeState("WLrestart.state");
*/
for(int i=0; i<num_lsms; i++) free(evecs[i]);
shfree(evecs);
//shfree(r_values);
}
}
if(world_rank==0)
{
if(step_out_flag)
{
step_out_file<<"# end\n-1\n"
<<energy_accumulator/double(energies_accumulated)<<std::endl;
step_out_file.close();
}
std::cout<<"Finished all scheduled calculations. Freeing resources.\n";
std::cout<<"Energy mean = "<<energy_accumulator/double(energies_accumulated)<<"Ry\n";
}
if(num_lsms>1)
{
// make sure averyone arrives here:
MPI_Bcast(stupid,37,MPI_CHAR,0,MPI_COMM_WORLD);
if(world_rank==0)
{
MPI_Comm_free(&local_comm);
}
else if(my_group>=0)
{
MPI_Comm_free(&local_comm);
}
}
if(world_rank==0)
{
double walltime = get_rtc() - walltime_0;
std::cout<<" WL-LSMS finished in "<<walltime<<" seconds.\n";
std::cout<<" Monte-Carlo steps / walltime = "
<<double(stepCount)/walltime<<"/sec\n";
}
#ifdef USE_PAPI
PAPI_stop_counters(papi_values,hw_counters);
papi_values[hw_counters ] = PAPI_get_real_cyc()-papi_real_cyc_0;
papi_values[hw_counters+1] = PAPI_get_real_usec()-papi_real_usec_0;
papi_values[hw_counters+2] = PAPI_get_virt_cyc()-papi_virt_cyc_0;
papi_values[hw_counters+3] = PAPI_get_virt_usec()-papi_virt_usec_0;
long long accumulated_counters[NUM_PAPI_EVENTS+4];
/*
for(int i=0; i<hw_counters; i++)
{
printline(ttos(papi_event_name[i])+" = "+ttos(papi_values[i]),
std::cout,parameters.myrankWorld);
}
printline("PAPI real cycles : "+ttos(papi_values[hw_counters]),
std::cout,parameters.myrankWorld);
printline("PAPI real usecs : "+ttos(papi_values[hw_counters+1]),
std::cout,parameters.myrankWorld);
printline("PAPI user cycles : "+ttos(papi_values[hw_counters+2]),
std::cout,parameters.myrankWorld);
printline("PAPI user usecs : "+ttos(papi_values[hw_counters+3]),
std::cout,parameters.myrankWorld);
*/
//MPI_Reduce(papi_values,accumulated_counters,hw_counters+4,
// MPI_LONG,MPI_SUM,0,MPI_COMM_WORLD);
shmem_long_sum_to_all(accumulated_counters, papi_values, hw_counters+4,
comm.pestart, comm.logPE_stride, comm.size, pWrk_i, pSync2);
if(world_rank==0)
{
for(int i=0; i<hw_counters; i++)
{
std::cout<<"Accumulated: "<<(papi_event_name[i])<<" = "<<(accumulated_counters[i])<<"\n";
}
std::cout<<"PAPI accumulated real cycles : "<<(accumulated_counters[hw_counters])<<"\n";
std::cout<<"PAPI accumulated user cycles : "<<(accumulated_counters[hw_counters+2])<<"\n";
double gflops_papi = ((double)accumulated_counters[1])/
(1000.0*(double)papi_values[hw_counters+1]);
double gflops_hw_double = ((double)accumulated_counters[2])/
(1000.0*(double)papi_values[hw_counters+1]);
double gflops_hw_single = ((double)accumulated_counters[3])/
(1000.0*(double)papi_values[hw_counters+1]);
double gips = ((double)accumulated_counters[0])/(1000.0*(double)papi_values[hw_counters+1]);
std::cout<<"PAPI_FP_OPS real GFLOP/s : "<<(gflops_papi)<<"\n";
std::cout<<"PAPI hw double real GFLOP/s : "<<(gflops_hw_double)<<"\n";
std::cout<<"PAPI hw single real GFLOP/s : "<<(gflops_hw_single)<<"\n";
std::cout<<"PAPI real GINST/s : "<<(gips)<<"\n";
}
#endif
//MPI_Finalize();
return 0;
}
void _PERM_IR(_permmap* const pm) {
const int one = 1;
int * rindbase;
int * lindbase;
int * const restrict lsize = (int *)shmalloc(_PROCESSORS * sizeof(int));
int * const restrict rsize = (int *)shmalloc(_PROCESSORS * sizeof(int));
int * restrict * const restrict lind = pm->lind;
int * restrict * const restrict rind = pm->rind;
char * restrict * const restrict rptr = pm->rptr;
int * const restrict rflag = pm->rflag;
int* addr;
int i, j;
for (i = 0; i < _PROCESSORS; i++) {
lsize[i] = lind[i] ? lind[i][0] : 0;
rsize[i] = 0;
}
shmem_barrier_all();
for (i = (_INDEX == _PROCESSORS - 1) ? 0 : _INDEX+1;
i != _INDEX;
i = (i == _PROCESSORS - 1) ? 0 : i++) {
if (lsize[i] > 0) {
#ifdef _SHMEM_PERMUTE_DEBUG
printf("%d sending count to %d\n", _INDEX, i);
fflush(stdout);
#endif
shmem_int_put(&(rsize[_INDEX]), &(lsize[i]), 1, i);
}
}
rsize[_INDEX] = lsize[_INDEX];
shmem_barrier_all();
#ifdef _SHMEM_PERMUTE_DEBUG
sleep(_PROCESSORS);
#endif
_PERM_CleanIndices(lsize, rsize, lind, rind, &lindbase, &rindbase);
#ifdef _SHMEM_PERMUTE_DEBUG
sleep(_INDEX);
printf("FROM PROCESSOR %d\n", _INDEX);
printf(" LSIZE = ");
for (i = 0; i < _PROCESSORS; i++) {
printf("%d ", lsize[i]);
}
printf("\n");
printf(" RSIZE = ");
for (i = 0; i < _PROCESSORS; i++) {
printf("%d ", rsize[i]);
}
printf("\n");
printf(" PROCMAP: size = %d, # elts = %d, encoded = %d :: ", pm->procmap[0], pm->procmap[1], pm->procmap[2]);
for (j = 3; j < pm->procmap[0]; j++) {
printf("%d ", pm->procmap[j]);
}
printf("\n");
for (i = 0; i < _PROCESSORS; i++) {
if (lind[i] != 0) {
printf(" TO PROCESSOR %d: ", i);
printf("size = %d, # elts = %d, encoded = %d :: ", lind[i][0], lind[i][1], lind[i][2]);
for (j = 3; j < lind[i][0]; j++) {
printf("%d ", lind[i][j]);
}
printf("\n");
}
}
printf("\n");
fflush(stdout);
sleep(_PROCESSORS-_INDEX);
#endif
for (i = (_INDEX == _PROCESSORS - 1) ? 0 : _INDEX+1;
i != _INDEX;
i = (i == _PROCESSORS - 1) ? 0 : i++) {
if (rsize[i] > 0) {
#ifdef _SHMEM_PERMUTE_DEBUG
printf("%d sending rind address to %d\n", _INDEX, i);
fflush(stdout);
#endif
rflag[_INDEX] = 0;
shmem_put((void*)&(rptr[_INDEX]), (void*)&(rind[i]), 1, i);
}
}
#ifdef _SHMEM_PERMUTE_DEBUG
sleep(_PROCESSORS);
#endif
for (i = (_INDEX == 0) ? _PROCESSORS-1 : _INDEX-1;
i != _INDEX;
i = (i == 0) ? _PROCESSORS-1 : i--) {
if (lsize[i] > 0) {
#ifdef _SHMEM_PERMUTE_DEBUG
printf("%d waiting for rind address from %d, sending lind\n", _INDEX, i);
fflush(stdout);
#endif
shmem_wait((long*)&(rptr[i]), 0);
addr = (int*)rptr[i];
rptr[i] = 0;
shmem_int_put(addr, lind[i], lsize[i], i);
}
}
#ifdef _SHMEM_PERMUTE_DEBUG
sleep(_PROCESSORS);
#endif
shmem_fence();
for (i = (_INDEX == 0) ? _PROCESSORS-1 : _INDEX-1;
i != _INDEX;
i = (i == 0) ? _PROCESSORS-1 : i--) {
if (lsize[i] > 0) {
#ifdef _SHMEM_PERMUTE_DEBUG
printf("IR %d sending one to %d\n", _INDEX, i);
fflush(stdout);
#endif
shmem_int_put(&(rflag[_INDEX]), &one, 1, i);
}
}
if (lsize[_INDEX] > 0) {
memcpy(rind[_INDEX], lind[_INDEX], lsize[_INDEX]*sizeof(int));
}
pm->lindbase = lindbase;
pm->rindbase = rindbase;
shfree(lsize);
shfree(rsize);
}
void _PERM_DR(const _permmap* const pm, _permdata* const pd, const int scatter, _array_fnc dst, _array_fnc src) {
const int one = 1;
const int eltsize = pd->eltsize;
int i;
int * const restrict ldecnt = (int*)_zmalloc(_PROCESSORS*sizeof(int), "perm dr lcnt");
int * const restrict rdecnt = (int*)_zmalloc(_PROCESSORS*sizeof(int), "perm dr rcnt");
char * const restrict * const restrict ldata = pd->ldata;
char * const restrict * const restrict rdata = pd->rdata;
char * restrict * const restrict rptr = pm->rptr;
int * restrict rflag = pm->rflag;
char* addr;
#ifdef _SHMEM_PERMUTE_DEBUG
printf("DR start %d\n", _INDEX);
fflush(stdout);
sleep(5);
#endif
for (i=0; i<_PROCESSORS; i++) {
ldecnt[i] = (scatter ? _PERM_LCNT(pm, i) : _PERM_RCNT(pm, i)) * eltsize;
rdecnt[i] = (scatter ? _PERM_RCNT(pm, i) : _PERM_LCNT(pm, i)) * eltsize;
}
for (i = (_INDEX == _PROCESSORS - 1) ? 0 : _INDEX+1;
i != _INDEX;
i = (i == _PROCESSORS - 1) ? 0 : i++) {
if (rdecnt[i] > 0) {
#ifdef _SHMEM_PERMUTE_DEBUG
printf("DR %d sending addr, %d, to %d\n", _INDEX, (int)&(rdata[i]), i);
fflush(stdout);
sleep(5);
#endif
rflag[i] = 0;
shmem_put((void*)&(rptr[_INDEX]), (void*)&(rdata[i]), 1, i);
}
}
for (i = (_INDEX == 0) ? _PROCESSORS-1 : _INDEX-1;
i != _INDEX;
i = (i == 0) ? _PROCESSORS-1 : i--) {
if (ldecnt[i] > 0) {
shmem_wait((long*)&(rptr[i]), 0);
addr = rptr[i];
rptr[i] = 0;
#ifdef _SHMEM_PERMUTE_DEBUG
printf("DR %d waiting addr, %d, from %d\n", _INDEX, (int)addr, i);
fflush(stdout);
sleep(5);
#endif
shmem_putmem(addr, ldata[i], ldecnt[i], i);
}
}
shmem_fence();
for (i = (_INDEX == 0) ? _PROCESSORS-1 : _INDEX-1;
i != _INDEX;
i = (i == 0) ? _PROCESSORS-1 : i--) {
if (ldecnt[i] > 0) {
#ifdef _SHMEM_PERMUTE_DEBUG
printf("DR %d sending flag to %d\n", _INDEX, i);
fflush(stdout);
sleep(5);
#endif
shmem_int_put(&(rflag[_INDEX]), &one, 1, i);
}
}
if (ldecnt[_INDEX] > 0) {
memcpy(rdata[_INDEX], ldata[_INDEX], ldecnt[_INDEX]);
}
_zfree(ldecnt, "perm dr lcnt");
_zfree(rdecnt, "perm dr rcnt");
pd->count = -1;
}
int main(int argc, char ** argv)
{
long Block_order; /* number of columns owned by rank */
int Block_size; /* size of a single block */
int Colblock_size; /* size of column block */
int Tile_order=32; /* default Tile order */
int tiling; /* boolean: true if tiling is used */
int Num_procs; /* number of ranks */
int order; /* order of overall matrix */
int bufferCount; /* number of input buffers */
int targetBuffer; /* buffer with which to communicate */
int send_to, recv_from; /* ranks with which to communicate */
long bytes; /* combined size of matrices */
int my_ID; /* rank */
int root=0; /* rank of root */
int iterations; /* number of times to do the transpose */
long i, j, it, jt, istart;/* dummies */
int iter; /* index of iteration */
int phase; /* phase inside staged communication */
int colstart; /* starting column for owning rank */
int error; /* error flag */
double *A_p; /* original matrix column block */
double *B_p; /* transposed matrix column block */
double **Work_in_p; /* workspace for the transpose function */
double *Work_out_p; /* workspace for the transpose function */
double epsilon = 1.e-8; /* error tolerance */
double avgtime; /* timing parameters */
long *pSync_bcast; /* work space for collectives */
long *pSync_reduce; /* work space for collectives */
double *pWrk; /* work space for SHMEM collectives */
double *local_trans_time,
*trans_time; /* timing parameters */
double *abserr,
*abserr_tot; /* local and aggregate error */
int *send_flag,
*recv_flag; /* synchronization flags */
int *arguments; /* command line arguments */
/*********************************************************************
** Initialize the SHMEM environment
*********************************************************************/
prk_shmem_init();
my_ID=prk_shmem_my_pe();
Num_procs=prk_shmem_n_pes();
if (my_ID == root) {
printf("Parallel Research Kernels version %s\n", PRKVERSION);
printf("SHMEM matrix transpose: B = A^T\n");
}
// initialize sync variables for error checks
pSync_bcast = (long *) prk_shmem_align(prk_get_alignment(),PRK_SHMEM_BCAST_SYNC_SIZE*sizeof(long));
pSync_reduce = (long *) prk_shmem_align(prk_get_alignment(),PRK_SHMEM_REDUCE_SYNC_SIZE*sizeof(long));
pWrk = (double *) prk_shmem_align(prk_get_alignment(),sizeof(double) * PRK_SHMEM_REDUCE_MIN_WRKDATA_SIZE);
local_trans_time = (double *) prk_shmem_align(prk_get_alignment(),sizeof(double));
trans_time = (double *) prk_shmem_align(prk_get_alignment(),sizeof(double));
arguments = (int *) prk_shmem_align(prk_get_alignment(),4*sizeof(int));
abserr = (double *) prk_shmem_align(prk_get_alignment(),2*sizeof(double));
abserr_tot = abserr + 1;
if (!pSync_bcast || !pSync_reduce || !pWrk || !local_trans_time ||
!trans_time || !arguments || !abserr) {
printf("Rank %d could not allocate scalar work space on symm heap\n", my_ID);
error = 1;
goto ENDOFTESTS;
}
for(i=0;i<PRK_SHMEM_BCAST_SYNC_SIZE;i++)
pSync_bcast[i]=PRK_SHMEM_SYNC_VALUE;
for(i=0;i<PRK_SHMEM_REDUCE_SYNC_SIZE;i++)
pSync_reduce[i]=PRK_SHMEM_SYNC_VALUE;
/*********************************************************************
** process, test and broadcast input parameters
*********************************************************************/
error = 0;
if (my_ID == root) {
if (argc != 4 && argc != 5){
printf("Usage: %s <# iterations> <matrix order> <# buffers> [Tile size]\n",
*argv);
error = 1; goto ENDOFTESTS;
}
iterations = atoi(*++argv);
arguments[0]=iterations;
if(iterations < 1){
printf("ERROR: iterations must be >= 1 : %d \n",iterations);
error = 1; goto ENDOFTESTS;
}
order = atoi(*++argv);
arguments[1]=order;
if (order < Num_procs) {
printf("ERROR: matrix order %d should at least # procs %d\n",
order, Num_procs);
error = 1; goto ENDOFTESTS;
}
if (order%Num_procs) {
printf("ERROR: matrix order %d should be divisible by # procs %d\n",
order, Num_procs);
error = 1; goto ENDOFTESTS;
}
bufferCount = atoi(*++argv);
arguments[2]=bufferCount;
if (Num_procs > 1) {
if ((bufferCount < 1) || (bufferCount >= Num_procs)) {
printf("ERROR: bufferCount must be >= 1 and < # procs : %d\n", bufferCount);
error = 1; goto ENDOFTESTS;
}
}
if (argc == 5) Tile_order = atoi(*++argv);
arguments[3]=Tile_order;
ENDOFTESTS:;
}
bail_out(error);
if (my_ID == root) {
printf("Number of ranks = %d\n", Num_procs);
printf("Matrix order = %d\n", order);
printf("Number of iterations = %d\n", iterations);
printf("Number of buffers = %d\n", bufferCount);
if ((Tile_order > 0) && (Tile_order < order))
printf("Tile size = %d\n", Tile_order);
else printf("Untiled\n");
}
shmem_barrier_all();
/* Broadcast input data to all ranks */
shmem_broadcast32(&arguments[0], &arguments[0], 4, root, 0, 0, Num_procs, pSync_bcast);
iterations=arguments[0];
order=arguments[1];
bufferCount=arguments[2];
Tile_order=arguments[3];
shmem_barrier_all();
prk_shmem_free(arguments);
/* a non-positive tile size means no tiling of the local transpose */
tiling = (Tile_order > 0) && (Tile_order < order);
bytes = 2 * sizeof(double) * order * order;
/*********************************************************************
** The matrix is broken up into column blocks that are mapped one to a
** rank. Each column block is made up of Num_procs smaller square
** blocks of order block_order.
*********************************************************************/
Block_order = order/Num_procs;
colstart = Block_order * my_ID;
Colblock_size = order * Block_order;
Block_size = Block_order * Block_order;
/*********************************************************************
** Create the column block of the test matrix, the row block of the
** transposed matrix, and workspace (workspace only if #procs>1)
*********************************************************************/
A_p = (double *)prk_malloc(Colblock_size*sizeof(double));
if (A_p == NULL){
printf(" Error allocating space for original matrix on node %d\n",my_ID);
error = 1;
}
bail_out(error);
B_p = (double *)prk_malloc(Colblock_size*sizeof(double));
if (B_p == NULL){
printf(" Error allocating space for transpose matrix on node %d\n",my_ID);
error = 1;
}
bail_out(error);
if (Num_procs>1) {
Work_in_p = (double**)prk_malloc(bufferCount*sizeof(double));
Work_out_p = (double *) prk_malloc(Block_size*sizeof(double));
recv_flag = (int*) prk_shmem_align(prk_get_alignment(),bufferCount*sizeof(int));
if ((Work_in_p == NULL)||(Work_out_p==NULL) || (recv_flag == NULL)){
printf(" Error allocating space for work or flags on node %d\n",my_ID);
error = 1;
}
if (bufferCount < (Num_procs - 1)) {
send_flag = (int*) prk_shmem_align(prk_get_alignment(), (Num_procs-1) * sizeof(int));
if (send_flag == NULL) {
printf("Error allocating space for flags on node %d\n", my_ID);
error = 1;
}
}
bail_out(error);
for(i=0;i<bufferCount;i++) {
Work_in_p[i]=(double *) prk_shmem_align(prk_get_alignment(),Block_size*sizeof(double));
if (Work_in_p[i] == NULL) {
printf(" Error allocating space for work on node %d\n",my_ID);
error = 1;
}
bail_out(error);
}
if (bufferCount < (Num_procs - 1)) {
for(i=0;i<(Num_procs-1);i++)
send_flag[i]=0;
}
for(i=0;i<bufferCount;i++)
recv_flag[i]=0;
}
/* Fill the original column matrices */
istart = 0;
for (j=0;j<Block_order;j++)
for (i=0;i<order; i++) {
A(i,j) = (double) (order*(j+colstart) + i);
B(i,j) = 0.0;
}
shmem_barrier_all();
if (bufferCount < (Num_procs - 1)) {
if (Num_procs > 1) {
for ( i = 0; i < bufferCount; i++) {
recv_from = (my_ID + i + 1)%Num_procs;
shmem_int_inc(&send_flag[i], recv_from);
}
}
}
shmem_barrier_all();
for (iter = 0; iter<=iterations; iter++){
/* start timer after a warmup iteration */
if (iter == 1) {
shmem_barrier_all();
local_trans_time[0] = wtime();
}
/* do the local transpose */
istart = colstart;
if (!tiling) {
for (i=0; i<Block_order; i++)
for (j=0; j<Block_order; j++) {
B(j,i) += A(i,j);
A(i,j) += 1.0;
}
}
else {
for (i=0; i<Block_order; i+=Tile_order)
for (j=0; j<Block_order; j+=Tile_order)
for (it=i; it<MIN(Block_order,i+Tile_order); it++)
for (jt=j; jt<MIN(Block_order,j+Tile_order);jt++) {
B(jt,it) += A(it,jt);
A(it,jt) += 1.0;
}
}
for (phase=1; phase<Num_procs; phase++){
recv_from = (my_ID + phase )%Num_procs;
send_to = (my_ID - phase + Num_procs)%Num_procs;
targetBuffer = (iter * (Num_procs - 1) + (phase - 1)) % bufferCount;
istart = send_to*Block_order;
if (!tiling) {
for (i=0; i<Block_order; i++)
for (j=0; j<Block_order; j++){
Work_out(j,i) = A(i,j);
A(i,j) += 1.0;
}
}
else {
for (i=0; i<Block_order; i+=Tile_order)
for (j=0; j<Block_order; j+=Tile_order)
for (it=i; it<MIN(Block_order,i+Tile_order); it++)
for (jt=j; jt<MIN(Block_order,j+Tile_order);jt++) {
Work_out(jt,it) = A(it,jt);
A(it,jt) += 1.0;
}
}
if (bufferCount < (Num_procs - 1))
shmem_int_wait_until(&send_flag[phase-1], SHMEM_CMP_EQ, iter+1);
shmem_double_put(&Work_in_p[targetBuffer][0], &Work_out_p[0], Block_size, send_to);
shmem_fence();
shmem_int_inc(&recv_flag[targetBuffer], send_to);
i = (iter * (Num_procs - 1) + phase) / bufferCount;
if ((iter * (Num_procs - 1) + phase) % bufferCount)
i++;
shmem_int_wait_until(&recv_flag[targetBuffer], SHMEM_CMP_EQ, i);
istart = recv_from*Block_order;
/* scatter received block to transposed matrix; no need to tile */
for (j=0; j<Block_order; j++)
for (i=0; i<Block_order; i++)
B(i,j) += Work_in(targetBuffer, i,j);
if (bufferCount < (Num_procs - 1)) {
if ((phase + bufferCount) < Num_procs)
recv_from = (my_ID + phase + bufferCount) % Num_procs;
else
recv_from = (my_ID + phase + bufferCount + 1 - Num_procs) % Num_procs;
shmem_int_inc(&send_flag[(phase+bufferCount-1)%(Num_procs-1)], recv_from);
}
} /* end of phase loop */
} /* end of iterations */
local_trans_time[0] = wtime() - local_trans_time[0];
shmem_barrier_all();
shmem_double_max_to_all(trans_time, local_trans_time, 1, 0, 0, Num_procs, pWrk, pSync_reduce);
abserr[0] = 0.0;
istart = 0;
double addit = ((double)(iterations+1) * (double) (iterations))/2.0;
for (j=0;j<Block_order;j++) for (i=0;i<order; i++) {
abserr[0] += ABS(B(i,j) - (double)((order*i + j+colstart)*(iterations+1)+addit));
}
shmem_barrier_all();
shmem_double_sum_to_all(abserr_tot, abserr, 1, 0, 0, Num_procs, pWrk, pSync_reduce);
if (my_ID == root) {
if (abserr_tot[0] <= epsilon) {
printf("Solution validates\n");
avgtime = trans_time[0]/(double)iterations;
printf("Rate (MB/s): %lf Avg time (s): %lf\n",1.0E-06*bytes/avgtime, avgtime);
#ifdef VERBOSE
printf("Summed errors: %f \n", abserr[0]);
#endif
}
else {
printf("ERROR: Aggregate squared error %e exceeds threshold %e\n", abserr[0], epsilon);
error = 1;
}
}
bail_out(error);
if (Num_procs>1)
{
if (bufferCount < (Num_procs - 1))
prk_shmem_free(send_flag);
prk_shmem_free(recv_flag);
prk_free(Work_out_p);
for(i=0;i<bufferCount;i++)
prk_shmem_free(Work_in_p[i]);
prk_free(Work_in_p);
}
prk_shmem_free(pSync_bcast);
prk_shmem_free(pSync_reduce);
prk_shmem_free(pWrk);
prk_shmem_finalize();
exit(EXIT_SUCCESS);
} /* end of main */
void
shmemi_broadcast32_tree (void *target, const void *source,
size_t nlong,
int PE_root, int PE_start,
int logPE_stride, int PE_size, long *pSync)
{
int child_l, child_r, parent;
const int step = 1 << logPE_stride;
int my_pe = GET_STATE (mype);
int *target_ptr, *source_ptr;
int no_children;
long is_ready, lchild_ready, rchild_ready;
is_ready = 1;
lchild_ready = -1;
rchild_ready = -1;
shmem_long_wait_until (&pSync[0], SHMEM_CMP_EQ, SHMEM_SYNC_VALUE);
shmem_long_wait_until (&pSync[1], SHMEM_CMP_EQ, SHMEM_SYNC_VALUE);
pSync[0] = 0;
pSync[1] = 0;
target_ptr = (int *) target;
source_ptr = (int *) source;
set_2tree (PE_start, step, PE_size, &parent, &child_l, &child_r, my_pe);
no_children = 0;
build_tree (PE_start, step, PE_root, PE_size,
&parent, &child_l, &child_r, my_pe);
shmemi_trace (SHMEM_LOG_BROADCAST,
"before broadcast, R_child = %d L_child = %d",
child_r, child_l);
/* The actual broadcast */
if (PE_size > 1) {
if (my_pe == (PE_start + step * PE_root)) {
pSync[0] = SHMEM_SYNC_VALUE;
if (child_l != -1) {
shmem_long_get (&lchild_ready, (const long *) &pSync[0],
1, child_l);
while (lchild_ready != 0)
shmem_long_get (&lchild_ready, &pSync[0], 1, child_l);
shmem_int_put (target_ptr, source_ptr, nlong, child_l);
shmem_fence ();
shmem_long_put (&pSync[0], &is_ready, 1, child_l);
no_children = 1;
}
if (child_r != -1) {
shmem_long_get (&rchild_ready, &pSync[0], 1, child_r);
while (rchild_ready != 0)
shmem_long_get (&rchild_ready, &pSync[0], 1, child_r);
shmem_int_put (target_ptr, source_ptr, nlong, child_r);
shmem_fence ();
shmem_long_put (&pSync[0], &is_ready, 1, child_r);
no_children = 2;
}
shmem_long_wait_until (&pSync[1], SHMEM_CMP_EQ,
(long) no_children);
pSync[1] = SHMEM_SYNC_VALUE;
}
else {
shmem_long_wait_until (&pSync[0], SHMEM_CMP_EQ, is_ready);
pSync[0] = SHMEM_SYNC_VALUE;
shmemi_trace (SHMEM_LOG_BROADCAST, "inside else");
memcpy (source_ptr, target_ptr, nlong * sizeof (int));
if (child_l != -1) {
shmem_long_get (&lchild_ready, &pSync[0], 1, child_l);
while (lchild_ready != 0)
shmem_long_get (&lchild_ready, &pSync[0], 1, child_l);
shmem_int_put (target_ptr, source_ptr, nlong, child_l);
shmem_fence ();
shmem_long_put (&pSync[0], &is_ready, 1, child_l);
no_children = 1;
}
if (child_r != -1) {
shmem_long_get (&rchild_ready, &pSync[0], 1, child_r);
while (rchild_ready != 0)
shmem_long_get (&rchild_ready, &pSync[0], 1, child_r);
shmem_int_put (target_ptr, source_ptr, nlong, child_r);
shmem_fence ();
shmem_long_put (&pSync[0], &is_ready, 1, child_r);
no_children = 2;
}
pSync[0] = SHMEM_SYNC_VALUE;
if (no_children == 0) {
pSync[1] = SHMEM_SYNC_VALUE;
/* TO DO: Is check for parents pSync required? */
shmem_long_inc (&pSync[1], parent);
}
else {
shmem_long_wait_until (&pSync[1], SHMEM_CMP_EQ,
(long) no_children);
pSync[1] = SHMEM_SYNC_VALUE;
/* printf("PE %d incrementing child count on PE
%d\n",my_pe,parent); */
shmem_long_inc (&pSync[1], parent);
}
}
shmemi_trace (SHMEM_LOG_BROADCAST, "at the end of bcast32");
/* shmem_barrier(PE_start, logPE_stride, PE_size, pSync); */
}
}
JNIEXPORT void JNICALL Java_shmem_ShMem_fence(JNIEnv *env, jclass clazz)
{
shmem_fence();
}
