added working implementation of h1 optim£ (04c3417b) · Commits · Luca Tornatore / dADP

src/adp/adp.c

+40 −15

Original line number	Original line	Diff line number	Diff line
	@@ -510,8 +510,6 @@ datapoint_info_t find_possibly_halo_datapoint_rma(global_context_t* ctx, idx_t i
	else		else
	{		{
	datapoint_info_t tmp_dp;		datapoint_info_t tmp_dp;
	#pragma omp critical
	{
	idx_t i = idx - ctx -> rank_idx_start[owner];		idx_t i = idx - ctx -> rank_idx_start[owner];
	MPI_Request request;		MPI_Request request;
	MPI_Status status;		MPI_Status status;
	@@ -520,8 +518,6 @@ datapoint_info_t find_possibly_halo_datapoint_rma(global_context_t* ctx, idx_t i
	i * sizeof(datapoint_info_t), sizeof(datapoint_info_t), MPI_BYTE, win_datapoints, &request);		i * sizeof(datapoint_info_t), sizeof(datapoint_info_t), MPI_BYTE, win_datapoints, &request);
	MPI_Wait(&request, MPI_STATUS_IGNORE);		MPI_Wait(&request, MPI_STATUS_IGNORE);

	}

	return tmp_dp;		return tmp_dp;
	}		}
	}		}
	@@ -680,9 +676,11 @@ clusters_t Heuristic1(global_context_t *ctx)

	struct timespec start_tot, finish_tot;		struct timespec start_tot, finish_tot;
	double elapsed_tot;		double elapsed_tot;
			double elapsed_time;

	TIME_DEF;		TIME_DEF;

			TIME_START;
	lu_dynamic_array_t all_centers, removed_centers, actual_centers, max_rho;		lu_dynamic_array_t all_centers, removed_centers, actual_centers, max_rho;

	lu_dynamic_array_allocate(&all_centers);		lu_dynamic_array_allocate(&all_centers);
	@@ -698,7 +696,7 @@ clusters_t Heuristic1(global_context_t *ctx)
	MPI_Win_create(ctx -> local_datapoints, ctx -> local_n_points * sizeof(datapoint_info_t),		MPI_Win_create(ctx -> local_datapoints, ctx -> local_n_points * sizeof(datapoint_info_t),
	1, MPI_INFO_NULL, ctx -> mpi_communicator, &win_datapoints);		1, MPI_INFO_NULL, ctx -> mpi_communicator, &win_datapoints);
	MPI_Win_fence(0, win_datapoints);		MPI_Win_fence(0, win_datapoints);
	MPI_Win_lock_all(0, win_datapoints);		//MPI_Win_lock_all(0, win_datapoints);

	#if !defined(THREAD_FUNNELED)		#if !defined(THREAD_FUNNELED)
	#pragma omp parallel for		#pragma omp parallel for
	@@ -744,7 +742,6 @@ clusters_t Heuristic1(global_context_t *ctx)
	*		*
	* optimized v2 use a queue of center removal and then exchange them		* optimized v2 use a queue of center removal and then exchange them
	*/		*/

	heap_node* to_remove_mask = (heap_node)MY_MALLOC(nsizeof(heap_node));		heap_node* to_remove_mask = (heap_node)MY_MALLOC(nsizeof(heap_node));

	for(idx_t p = 0; p < n; ++p)		for(idx_t p = 0; p < n; ++p)
	@@ -752,6 +749,9 @@ clusters_t Heuristic1(global_context_t *ctx)
	to_remove_mask[p].array_idx = MY_SIZE_MAX;		to_remove_mask[p].array_idx = MY_SIZE_MAX;
	to_remove_mask[p].value = -9999999;		to_remove_mask[p].value = -9999999;
	}		}

			// sort by density

	qsort(dp_info_ptrs, n, sizeof(datapoint_info_t*), cmpPP);		qsort(dp_info_ptrs, n, sizeof(datapoint_info_t*), cmpPP);

	/**		/**
	@@ -774,6 +774,13 @@ clusters_t Heuristic1(global_context_t *ctx)
	omp_init_lock(lock_array + i);		omp_init_lock(lock_array + i);
	}		}

			elapsed_time = TIME_STOP;
			LOG_WRITE("Putative centers", elapsed_time);

			TIME_START;

			MPI_Win_fence(MPI_MODE_NOPUT, win_datapoints);

	#if !defined(THREAD_FUNNELED)		#if !defined(THREAD_FUNNELED)
	#pragma omp parallel for schedule(dynamic)		#pragma omp parallel for schedule(dynamic)
	#endif		#endif
	@@ -793,9 +800,10 @@ clusters_t Heuristic1(global_context_t *ctx)
	// actually is the p-th point		// actually is the p-th point
	int owner = foreign_owner(ctx, jidx);		int owner = foreign_owner(ctx, jidx);
	//if local process it		//if local process it
			idx_t jpos = jidx - ctx -> idx_start;
	if(owner == ctx -> mpi_rank)		if(owner == ctx -> mpi_rank)
	{		{
	idx_t jpos = jidx - ctx -> idx_start;		//acquire the lock
	omp_set_lock(lock_array + jpos);		omp_set_lock(lock_array + jpos);
	if(i_point.g > to_remove_mask[jpos].value)		if(i_point.g > to_remove_mask[jpos].value)
	{		{
	@@ -815,8 +823,14 @@ clusters_t Heuristic1(global_context_t *ctx)
	}		}
	}		}

			MPI_Win_fence(MPI_MODE_NOPUT, win_datapoints);

	//assemble arrays into a single buffer		//assemble arrays into a single buffer

			elapsed_time = TIME_STOP;
			LOG_WRITE("Finding centers to prune", elapsed_time);
			TIME_START;

	idx_t tot_removal = 0;		idx_t tot_removal = 0;
	for(idx_t p = 0; p < n; ++p)		for(idx_t p = 0; p < n; ++p)
	{		{
	@@ -964,7 +978,11 @@ clusters_t Heuristic1(global_context_t *ctx)

	// merge into the mask		// merge into the mask

	#pragma omp parallel for		elapsed_time = TIME_STOP;
			LOG_WRITE("Communicating eliminations", elapsed_time);
			TIME_START;

			#pragma omp parallel for schedule(dynamic)
	for(idx_t i = 0; i < tot_recv_counts; ++i)		for(idx_t i = 0; i < tot_recv_counts; ++i)
	{		{
	idx_t el_pos = recv_removals[i].target_id - ctx -> idx_start;		idx_t el_pos = recv_removals[i].target_id - ctx -> idx_start;
	@@ -1046,6 +1064,11 @@ clusters_t Heuristic1(global_context_t *ctx)
	free(lock_array);		free(lock_array);
	free(recv_removals);		free(recv_removals);

			elapsed_time = TIME_STOP;
			LOG_WRITE("Merging", elapsed_time);

			TIME_START;

	int n_centers = (int)actual_centers.count;		int n_centers = (int)actual_centers.count;
	int tot_centers;		int tot_centers;
	MPI_Allreduce(&n_centers, &tot_centers, 1, MPI_INT, MPI_SUM, ctx -> mpi_communicator);		MPI_Allreduce(&n_centers, &tot_centers, 1, MPI_INT, MPI_SUM, ctx -> mpi_communicator);
	@@ -1188,7 +1211,7 @@ clusters_t Heuristic1(global_context_t *ctx)

	}		}

	MPI_Win_unlock_all(win_datapoints);		//MPI_Win_unlock_all(win_datapoints);
	MPI_Win_fence(0, win_datapoints);		MPI_Win_fence(0, win_datapoints);
	MPI_Win_free(&win_datapoints);		MPI_Win_free(&win_datapoints);

	@@ -1209,6 +1232,8 @@ clusters_t Heuristic1(global_context_t *ctx)
	free(ks);		free(ks);
	#endif		#endif

			elapsed_time = TIME_STOP;
			LOG_WRITE("Cluster assign", elapsed_time);

	free(actual_centers.data);		free(actual_centers.data);
	actual_centers.size = tot_centers;		actual_centers.size = tot_centers;