Fix wrong bundle parallelisation

brian2cuda/templates/synapses.cu

@@ -130,24 +130,18 @@ num_loops = 1;
 {% elif synaptic_effects == "target" %}
 // Synaptic effects modify target group variables but NO source group variables.
 num_blocks = num_parallel_blocks;
-// if no bundles: num_threads = 1
-num_threads_per_bundle = 1;
 num_loops = 1;
-if ({{pathway.name}}_scalar_delay)
-{
-	if ({{owner.name}}_multiple_pre_post)
-		num_threads = 1;
-	else
-		num_threads = max_threads_per_block;
-}
-else  // heterogeneous delays
-{
-	num_threads = {{pathway.name}}_max_num_bundles;
-	if (num_threads > max_threads_per_block)
-	{
+num_threads = 1;
+if (!{{owner.name}}_multiple_pre_post){
+	if ({{pathway.name}}_scalar_delay)
 		num_threads = max_threads_per_block;
-	}
+	else  // heterogeneous delays
+		num_threads = {{pathway.name}}_max_bundle_size;
 }
+if (num_threads > max_threads_per_block)
+	num_threads = max_threads_per_block;
+// num_threads_per_bundle only used for heterogeneous delays
+num_threads_per_bundle = num_threads;
 {% elif synaptic_effects == "source" %}
 // Synaptic effects modify source group variables.
 num_blocks = 1;

brian2cuda/templates/synapses_initialise_queue.cu

@@ -359,6 +359,8 @@ void _run_{{pathobj}}_initialise_queue()
 					num_synapses = h_unique_delay_start_idx_by_pre_id[i][bundle_idx + 1] - synapses_start_idx;
 				//h_size_by_bundle_id_by_pre[i].push_back(num_synapses);
 				h_size_by_bundle_id.push_back(num_synapses);
+				if (num_synapses > {{pathobj}}_max_bundle_size)
+					{{pathobj}}_max_bundle_size = num_synapses;
 				int32_t* synapse_bundle = new int32_t[num_synapses];
 				// TODO: don't copy synapses to synapse_bundle, just cudaMemcpy it directly to the device with
 				// CudaSafeCall( cudaMemcpy(d_..., h_synapses_by_pre_id[i] + synapses_start_idx, ...)
@@ -443,7 +445,7 @@ void _run_{{pathobj}}_initialise_queue()
 				sum_num_unique_elements += num_unique_elements;
             }  // end if(num_elements < 0)
 		}
-		long unsigned int num_bundle_ids = sum_num_unique_elements;
+		num_bundle_ids = sum_num_unique_elements;
 		// floor(mean(h_size_by_bundle_id))
 		{{pathobj}}_mean_bundle_size = sum_num_synapses / num_bundle_ids;
 		//delete [] h_size_by_bundle_id_by_pre;