#include <string.h>
#include <dcmf.h>
#include <dcmf_globalcollectives.h>
#include <tbb/task_scheduler_init.h>
#include <tbb/blocked_range.h>
#include <tbb/parallel_reduce.h>
#include <tbb/cache_aligned_allocator.h>
using namespace tbb;
#include <xmmintrin.h>
DCMF_Protocol_t barrier_prot, control0_prot, control1_prot,
put_prot, reduce_prot;
void cb_decr(void *data) {
unsigned *val = (unsigned*)data;
(*val)--;
}
void cb_recv(void *data, const DCMF_Control_t *info, unsigned) {
memcpy((DCMF_Memregion_t*)data, info, sizeof(DCMF_Memregion_t));
}
void barrier() {
DCMF_CriticalSection_enter(0);
volatile unsigned active = 1;
DCMF_Callback_t cb = { cb_decr, (void *) &active };
DCMF_Request_t req;
DCMF_GlobalBarrier(&barrier_prot, &req, cb);
while (active)
DCMF_Messager_advance();
DCMF_CriticalSection_exit(0);
}
struct thread1 {
float ee;
float *x, *y;
thread1(float *xx, float *yy) : ee(0), x(xx), y(yy) {}
thread1(thread1& s, split) { ee = 0; x = s.x; y = s.y; }
void operator() (const blocked_range<int> & r) {
// do computation
float ve[4] = {0, 0, 0, 0};
float e = ee;
for (int i=r.begin(); i!=r.end(); i+=4) {
float half =.5;
_mm_store_ps(&x[i],
_mm_mul_ps(_mm_load1_ps(&half),
_mm_add_ps(_mm_loadu_ps(&y[i+1]),
_mm_loadu_ps(&y[i-1]))));
_mm_store_ps(&ve[0],
_mm_add_ps(_mm_load_ps(&ve[0]),
_mm_mul_ps(_mm_load_ps(&y[i]),
_mm_load_ps(&y[i]))));
}
e += ve[0] + ve[1] + ve[2] + ve[3];
ee = e;
}
void join(thread1& s) { ee += s.ee; }
};
int main(int argc, char *argv[]) {
int n = ...;
DCMF_Messager_initialize();
{ // init barrier, put, reduce
DCMF_GlobalBarrier_Configuration_t barrier_conf =
{DCMF_DEFAULT_GLOBALBARRIER_PROTOCOL};
DCMF_Put_Configuration_t put_conf =
{DCMF_DEFAULT_PUT_PROTOCOL};
DCMF_GlobalAllreduce_Configuration_t reduce_conf =
{DCMF_TREE_GLOBALALLREDUCE_PROTOCOL};
DCMF_CriticalSection_enter(0);
DCMF_GlobalBarrier_register(&barrier_prot, &barrier_conf);
DCMF_Put_register(&put_prot, &put_conf);
DCMF_GlobalAllreduce_register(&reduce_prot, &reduce_conf);
DCMF_CriticalSection_exit(0);
}
unsigned me = DCMF_Messager_rank();
unsigned numproc = DCMF_Messager_size();
int p_left = -1, p_right = -1;
if (me > 0)
p_left = me-1;
if (me < numproc-1)
p_right = me+1;
int n_local0 = 1 + (me * (n-1)) / numproc;
int n_local1 = 1 + ((me+1) * (n-1)) / numproc;
// allocate only local part + ghost zone of the arrays x,y
float *x, *y;
x = cache_aligned_allocator<float>().allocate(n_local1 - n_local0 + 2);
y = cache_aligned_allocator<float>().allocate(n_local1 - n_local0 + 2);
x -= (n_local0 - 1);
y -= (n_local0 - 1);
// ghost zones
DCMF_Memregion_t memregion0, memregion1,
memregion_left, memregion_right;
size_t bytes;
DCMF_CriticalSection_enter(0);
DCMF_Memregion_create(&memregion0, &bytes,
2 * sizeof(float), &y[n_local0-1], 0);
DCMF_Memregion_create(&memregion1, &bytes,
2 * sizeof(float), &y[n_local1-1], 0);
// set memregion_left, memregion_right
DCMF_Control_Configuration_t c0_conf =
{ DCMF_DEFAULT_CONTROL_PROTOCOL, cb_recv, &memregion_right};
DCMF_Control_Configuration_t c1_conf =
{ DCMF_DEFAULT_CONTROL_PROTOCOL, cb_recv, &memregion_left};
DCMF_Control_register(&control0_prot, &c0_conf);
DCMF_Control_register(&control1_prot, &c1_conf);
barrier();
if (p_left != -1)
DCMF_Control(&control0_prot, DCMF_MATCH_CONSISTENCY,
p_left, (DCMF_Control_t*) &memregion0);
if (p_right != -1)
DCMF_Control(&control1_prot, DCMF_MATCH_CONSISTENCY,
p_right, (DCMF_Control_t*) &memregion1);
barrier();
DCMF_CriticalSection_exit(0);
... // fill x, y
{ // fill ghost zone
volatile unsigned active0 = 1, active1 = 1;
DCMF_Callback_t cb0 = { cb_decr, (void*)&active0 },
cb1 = { cb_decr, (void*)&active1 };
DCMF_Request_t req0, req1;
DCMF_CriticalSection_enter(0);
if (p_left != -1)
DCMF_Put(&put_prot, &req0, cb0, DCMF_SEQUENTIAL_CONSISTENCY,
p_left, sizeof(float), &memregion0, &memregion_left,
sizeof(float), sizeof(float));
if (p_right != -1)
DCMF_Put(&put_prot, &req1, cb1, DCMF_SEQUENTIAL_CONSISTENCY,
p_right, sizeof(float), &memregion1, &memregion_right,
sizeof(float), 0);
if (p_left != -1)
while (active0)
DCMF_Messager_advance();
if (p_right != -1)
while (active1)
DCMF_Messager_advance();
DCMF_CriticalSection_exit(0);
barrier();
}
thread1 t(x, y);
parallel_reduce(blocked_range<int>(n_local0, n_local1, 1024), t);
float e = t.ee;
{ // reduction
DCMF_CriticalSection_enter(0);
float e_local = e;
volatile unsigned active = 1;
DCMF_Callback_t cb = { cb_decr, (void*)&active };
DCMF_Request_t req;
DCMF_GlobalAllreduce(&reduce_prot, &req, cb,
DCMF_MATCH_CONSISTENCY, -1,
(char*)&e_local, (char*)&e, 1, DCMF_FLOAT, DCMF_SUM);
while (active)
DCMF_Messager_advance();
DCMF_CriticalSection_exit(0);
}
... // output x, e
barrier();
DCMF_Memregion_destroy(&memregion0);
DCMF_Memregion_destroy(&memregion1);
x += (n_local0 - 1);
y += (n_local0 - 1);
cache_aligned_allocator<float>().deallocate(x, n_local1 - n_local0 + 2);
cache_aligned_allocator<float>().deallocate(y, n_local1 - n_local0 + 2);
DCMF_Messager_finalize();
return 0;
}
|