int searchSIMDTree(int32_t **tree, int *fanout, int levels, int32_t value) {
int iLevel = 0;
int lOffset = 0;
int pOffset = 0;
int32_t cmpmask = 0;
int32_t eqmask = 0;
__m128i key = _mm_cvtsi32_si128(value);
key = _mm_shuffle_epi32(key, _MM_SHUFFLE(0,0,0,0));
while (iLevel < levels) {
int f = fanout[iLevel];
pOffset = lOffset;
lOffset *= f - 1;
int iter = 0;
int position = 0;
while (iter < f/4) {
__m128i delimiters = _mm_load_si128((__m128i const*)&tree[iLevel][lOffset + iter*4]);
__m128i compare = _mm_cmpgt_epi32(key, delimiters);
cmpmask = _mm_movemask_ps(_mm_castsi128_ps(compare));
cmpmask ^= 0x0F;
if (cmpmask) {
position = _bit_scan_forward(cmpmask);
break;
}
iter++;
}
int offset = lOffset + iter*4 + position;
lOffset = offset + pOffset;
iLevel++;
}
return lOffset;
}
BOOST_FORCEINLINE unsigned find_lsb(unsigned mask, mpl::int_<1> const&)
{
return _bit_scan_forward(mask);
}
void run_graph_program(GraphProgram<T,U,V>* gp, Graph<V>& g, int iterations=1, struct run_graph_program_temp_structure<T,U,V>* rgpts=NULL) { //iterations = -1 ==> until convergence
int it = 0;
int converged = 1;
unsigned long long int init_start = __rdtsc();
auto act = gp->getActivity();
SparseInVector<T>* px;
SparseOutVector<U>* py;
if (rgpts == NULL) {
px = new SparseInVector<T>(g.nvertices);
py = new SparseOutVector<U>(g.nvertices);
}
SparseInVector<T>&x = (rgpts==NULL)?(*px):*(rgpts->px);
SparseOutVector<U>& y = (rgpts==NULL)?(*py):*(rgpts->py);
#ifdef __TIMING
printf("Nvertices = %d numints = %d \n", g.nvertices, y.numInts);
#endif
unsigned long long int start, end;
int* start_vertex = new int[nthreads+1];
//divide numInts to start_vertex
//divide the active vertices in each into start_index
start_vertex[nthreads] = g.nvertices;
#pragma omp parallel num_threads(nthreads)
{
int tid = omp_get_thread_num();
int ints_per_th = (y.numInts/nthreads)*32;
int sv = ints_per_th*tid;
sv = (((sv/32)/4)*4)*32; //sv is multiple of 32 and sv/32 is a multiple of 4
sv = (((sv/32)/SIMD_WIDTH)*SIMD_WIDTH)*32; //sv is multiple of 32 and sv/32 is a multiple of SIMD_WIDTH
if (sv >= g.nvertices) sv = g.nvertices;
if (sv == 0) sv = 0;
start_vertex[tid] = sv;
}
unsigned long long int init_end = __rdtsc();
#ifdef __TIMING
printf("GraphMat init time = %f ms \n", (init_end-init_start)/(CPU_FREQ)*1e3);
#endif
while(1) {
unsigned long long int iteration_start = __rdtsc();
x.clear();
y.clear();
converged = 1;
start = __rdtsc();
//check active vector and set message vector
int count = 0;
#pragma omp parallel num_threads(nthreads) reduction(+:count)
{
int tid = omp_get_thread_num();
for (int i = start_vertex[tid]; i < start_vertex[tid+1]; i++){
if (g.active[i]) {
T message;
bool msg_opt = gp->send_message(g.vertexproperty[i], message);
if (msg_opt) {
x.set(i, message);
count++;
}
}
}
}
x.length = count;
#ifdef __TIMING
printf("x.length = %d \n", x.length);
#endif
end = __rdtsc();
#ifdef __TIMING
printf("Send message time = %.3f ms \n", (end-start)/(CPU_FREQ)*1e3);
#endif
start = __rdtsc();
//do SpMV
if (gp->getOrder() == OUT_EDGES) {
SpMTSpV(g, gp, x, y);
} else if (gp->getOrder() == IN_EDGES) {
SpMSpV(g, gp, x, y);
} else if (gp->getOrder() == ALL_EDGES) {
SpMTSpV(g, gp, x, y);
SpMSpV(g, gp, x, y);
} else {
printf("Unrecognized option \n");
exit(1);
}
end = __rdtsc();
#ifdef __TIMING
printf("SPMV time = %.3f ms \n", (end-start)/(CPU_FREQ)*1e3);
#endif
start = __rdtsc();
g.setAllInactive();
//update state and activity and check for convergence if needed
int nout = 0;
int total_search = 0;
converged = 1;
#pragma omp parallel num_threads(nthreads) reduction(+:nout) reduction(&:converged) reduction(+:total_search) //schedule(static)
{
int zero = 0;
SIMDINTTYPE xmm_zero = _MM_SET1(zero);
int tid = omp_get_thread_num();
int count_ones = 0;
int end_of_numInts = start_vertex[tid+1]/32;
if (tid == nthreads-1) end_of_numInts = y.numInts;
for (int ii = start_vertex[tid]/32; ii < end_of_numInts; ii+=SIMD_WIDTH) {
__m128i xmm_local_bitvec = _mm_loadu_si128((__m128i*)(y.bitvector + ii));
__m128 xmm_cmp_mask = _mm_castsi128_ps(_mm_cmpeq_epi32((xmm_local_bitvec), (xmm_zero)));
int mask_value_0 = _mm_movemask_ps(xmm_cmp_mask);
if(mask_value_0 == 15)
{
continue;
}
for(int i = ii; i < ii+SIMD_WIDTH; i++)
{
unsigned int value = y.bitvector[i];
while (value != 0) {
int last_bit = _bit_scan_forward(value);
int idx = i*32 + last_bit;
V old_prop;
old_prop = g.vertexproperty[idx];
gp->apply(y.value[idx], g.vertexproperty[idx]);
nout++;
if (old_prop != g.vertexproperty[idx]) {
g.setActive(idx);
count_ones++;
converged = 0;
total_search++;
}
value &= (~(1<<last_bit));
}
}
}
}
if (act == ALL_VERTICES) {
g.setAllActive();
}
#ifdef __TIMING
printf("Number of vertices that changed state = %d \n", total_search);
#endif
end = __rdtsc();
#ifdef __TIMING
printf("Apply time = %.3f ms \n", (end-start)/(CPU_FREQ)*1e3);
#endif
gp->do_every_iteration(it);
unsigned long long int iteration_end = __rdtsc();
#ifdef __TIMING
printf("Iteration %d :: %f msec :: updated %d vertices \n", it, (iteration_end-iteration_start)/(CPU_FREQ)*1e3, nout);
#endif
it++;
if (it == iterations) {
break;
}
if (iterations <= 0 && converged == 1) {
break;
}
}
unsigned long long int clear_start = __rdtsc();
delete [] start_vertex;
if (rgpts == NULL) {
delete px;
delete py;
}
unsigned long long int clear_end = __rdtsc();
#ifdef __TIMING
printf("GraphMat clear time = %f msec \n", (clear_end-clear_start)/(CPU_FREQ)*1e3);
#endif
printf("Completed %d iterations \n", it);
}
