Merge branch 'master' of https://github.com/microideax/FeedforwardNet

codeGenerator/local_search.py

@@ -182,6 +182,7 @@ def per_die_config_dse(sub_conv_N, sub_conv_M, sub_conv_r, sub_conv_R, sub_conv_
 
     return pair_list, lat_list, util_list
 
+# by Yao: fix the number of accelerators to the number of layers
 def per_die_config_dse_multiAcc(sub_conv_N, sub_conv_M, sub_conv_r, sub_conv_R, sub_conv_K, sub_conv_S, sub_flag):
     DSP = 6840 / 3
     dsp_list = []
@@ -237,6 +238,88 @@ def per_die_config_dse_multiAcc(sub_conv_N, sub_conv_M, sub_conv_r, sub_conv_R,
 
     return pair_list, lat_list, util_list
 
+
+# by John: find the optimal number of accelerators in each sub-net
+def per_die_config_dse_multiAcc_flex(sub_conv_N, sub_conv_M, sub_conv_r, sub_conv_R, sub_conv_K, sub_conv_S, sub_flag):
+    start_index = 0
+    min_cycle = sys.maxint
+    for i in range(0, len(sub_conv_N)):
+        for j in range(1, 3 + 1):
+            for k in split_sub_net(start_index, start_index + len(sub_conv_N[i]), j):
+                DSP = 6840 / 3
+                dsp_list = []
+                pair_list = []
+                lat_list = []
+                util_list = []
+                factor = 1
+
+                # re-caculate sub_conv_N, sub_conv_M, sub_conv_R, sub_conv_K
+                sub_conv_N_new = []
+                sub_conv_M_new = []
+                sub_conv_R_new = []
+                sub_conv_K_new = []
+
+                sub_conv_N_new.append(sub_conv_N[i:j] for i, j in zip([0] + k, k + [None]))
+                sub_conv_M_new.append(sub_conv_N[i:j] for i, j in zip([0] + k, k + [None]))
+                sub_conv_R_new.append(sub_conv_N[i:j] for i, j in zip([0] + k, k + [None]))
+                sub_conv_K_new.append(sub_conv_N[i:j] for i, j in zip([0] + k, k + [None]))
+
+                # start_index_new = 0
+                # for t in range(0, j):
+                #     sub_conv_N_new.append(sub_conv_N[i][start_index_new:k[t]])
+                #     sub_conv_M_new.append(sub_conv_M[i][start_index_new:k[t]])
+                #     sub_conv_R_new.append(sub_conv_R[i][start_index_new:k[t]])
+                #     sub_conv_K_new.append(sub_conv_K[i][start_index_new:k[t]])
+                #     start_index_new += k[t]
+
+                dsp_list.append([])
+                sub_net_gop = gop_calculate(sub_conv_N_new[i], sub_conv_M_new[i], sub_conv_R_new[i], sub_conv_K_new[i])
+
+                # allocate_dsp by layer gops
+                dsp_list[i].append(DSP * (
+                            sub_conv_N_new[i][j] * sub_conv_M_new[i][j] * sub_conv_R_new[i][j] * sub_conv_R_new[i][j] *
+                            sub_conv_K_new[i][j] * sub_conv_K_new[i][j]) / sub_net_gop)
+                # do constrained dse for layer
+                pair, cycle, cycle_per_layer = constrained_dse_layer(sub_conv_N[i][j], sub_conv_M[i][j],
+                                                                     sub_conv_r[i][j], sub_conv_R[i][j],
+                                                                     sub_conv_K[i][j],
+                                                                     sub_conv_S[i][j], sub_flag[i][j],
+                                                                     int(dsp_list[i][j]), int(37),
+                                                                     factor)
+                # find the minimum cycles
+                if cycle < min_cycle:
+                    return None
+                    # record and return
+                    # pair_list.append(pair)
+                    # lat_list.append(cycle)
+                    # util_list.append(pair[0] * pair[1] / float(int(dsp_list[i][j])))
+
+        start_index += len(sub_conv_N[i])
+
+    for i in range(0, len(sub_conv_N)):
+        pair, cycle, cycle_per_layer = constrained_dse(sub_conv_N[i], sub_conv_M[i], sub_conv_r[i], sub_conv_R[i],
+                                                       sub_conv_K[i],
+                                                       sub_conv_S[i], sub_flag[i], int(DSP), int(37),
+                                                       factor)
+
+        if len(pair_list) > len(sub_conv_N):
+            for remove_cnt in range(0, len(sub_conv_N)):
+                pair_list.remove(pair_list[0])
+                lat_list.remove(lat_list[0])
+                util_list.remove(util_list[0])
+
+    return pair_list, lat_list, util_list
+
+
+def split_sub_net(start_index, end_index, k):
+    if k == 2:
+       for i in range(start_index, end_index):
+           yield i
+    if k== 3:
+        for i in range(start_index, end_index - 1):
+            for j in range(i + 1, end_index):
+                yield [i, j]
+
 def local_search(sub_conv_N, sub_conv_M, sub_conv_r, sub_conv_R, sub_conv_K, sub_conv_S, sub_flag):
     """
     :param sub_conv_N: the input sub_conv_N is already splitted into several sub-nets

codeGenerator/tm_tn_multiAcc.py

@@ -87,6 +87,8 @@ def multiAcc_dse():
     # kmeans=clusters_layers_kmeans(conv_N, conv_M, conv_r, conv_R, conv_K, conv_S, 2)
     # print kmeans
     partition_location, diff_ratio = model_partition_by_gop(conv_N, conv_M, conv_r, conv_R, conv_K, conv_S, flag)
+    print "test point 1", partition_location, diff_ratio
+
     sub_conv_N, sub_conv_M, sub_conv_r, sub_conv_R, sub_conv_K, sub_conv_S, sub_flag \
         =model_partition_ordered(conv_N, conv_M, conv_r, conv_R, conv_K, conv_S, flag, partition_location[0]+1, partition_location[1]+1)
     print "Best partition output: ", partition_location, diff_ratio