square performance done

optimization_algo/rScripts/square-latency-tp_gain.r

@@ -10,84 +10,87 @@ ts=25
 # plot_2 > latency_penalty v partition
 pdf("Square Performance.pdf", height=9, width=15)
 
-sq_data <- read.csv("../resulting_csv/cma_logmore_sq.csv", header=TRUE, sep=",")
-sq_data_subset <- subset(sq_data, feasable == 1 & target=="DRAM_cycle" & seeding_type=="allzeros")
+sq_data <- read.csv("../resulting_csv/cma_logmore_sq_full_evo.csv", header=TRUE, sep=",")
+sq_data_subset <- subset(sq_data, feasable == 1 & target=="DRAM_cycle" & seeding_type=="allzeros" & popsize == 500)
 
 topo_unique<-unique(sq_data_subset$topology)
-partition_unique<-unique(sq_data_subset$partitions)
-res_unique<-unique(sq_data_subset$res_unit)
-
-column_data <- data.frame(matrix(ncol=7,nrow=0))
-col_headings <- c("topology", "partitions", "sq_tp_gain", "sq_latency_penalty", "tp_gain_latency_metrics","res", "tp_partition")
-names(column_data) <- col_headings
-
+print(topo_unique)
 for (topo in topo_unique) {
+    topo_subset = subset(sq_data_subset, topology==topo)
+    partition_unique<-unique(topo_subset$partitions)
+    print(partition_unique)
+    res_unique<-unique(topo_subset$res_unit)
+    print(res_unique)
+    
+    column_data <- data.frame(matrix(ncol=7,nrow=0))
+    col_headings <- c("topology", "partitions", "sq_tp_gain", "sq_latency_penalty", "tp_gain_latency_metrics","res", "tp_partition")
+    names(column_data) <- col_headings
+    
     for (part in partition_unique) {
         for (res in res_unique) {
-            sq_data_uni = subset(sq_data_subset, topology==topo & partitions==part & res_unit==res)
-
-            if (dim(sq_data_uni)[1]!=0) {
-                sq_max_evo = max(sq_data_uni$evo_counter)
-                sq_row = which(sq_data_uni$evo_counter==sq_max_evo)
-
+            topo_subset_uni = subset(topo_subset, partitions==part & res_unit==res)
+            
+            if (dim(topo_subset_uni)[1]!=0) {
+                sq_max_evo = max(topo_subset_uni$evo_counter)
+                sq_row = which(topo_subset_uni$evo_counter==sq_max_evo)[1]
                 column_data <- rbind(column_data, data.frame(
                     "topology"=topo,
                     "partitions"=part,
-                    "sq_tp_gain"=sq_data_uni$tp_gain[sq_row], 
-                    "sq_latency_penalty"=sq_data_uni$latency_penalty[sq_row],
-                    "tp_gain_latency_metrics"=sq_data_uni$tp_gain[sq_row]/sq_data_uni$latency_penalty[sq_row],
-                    "res"=toString(sq_data_uni$res_unit[sq_row]),
-                    "tp_partition"=sq_data_uni$tp_partition[sq_row]))
+                    "sq_tp_gain"=topo_subset_uni$tp_gain[sq_row], 
+                    "sq_latency_penalty"=topo_subset_uni$latency_penalty[sq_row],
+                    "tp_gain_latency_metrics"=topo_subset_uni$tp_gain[sq_row]/topo_subset_uni$latency_penalty[sq_row],
+                    "res"=toString(topo_subset_uni$res_unit[sq_row]),
+                    "tp_partition"=topo_subset_uni$tp_partition[sq_row]))
             }
         }
     }
+    
+    d <- ggplot(column_data,aes(y=as.numeric(tp_partition), x=as.numeric(partitions), group=res))
+    d <- d + geom_point(size=ps, aes(color=res, shape=res))
+    d <- d + geom_line(size=ls, aes(color=res))
+    d <- d + labs(x="No of Partitions", y="Throughput")
+    d <- d + ggtitle(paste(topo, "Square size array Throughtput by size tier"))
+    d <- d + theme_bw()
+    d <- d + theme(axis.text=element_text(size=ats))
+    d <- d + theme(text=element_text(size=ts))
+    d <- d + theme(legend.position="bottom")
+    
+    p <- ggplot(column_data,aes(y=as.numeric(sq_tp_gain), x=as.numeric(partitions), group=res))
+    p <- p + geom_point(size=ps, aes(color=res, shape=res))
+    p <- p + geom_line(size=ls, aes(color=res))
+    p <- p + geom_hline(yintercept=1)
+    p <- p + labs(x="No of Partitions", y="Throughput Gain")
+    p <- p + ggtitle(paste(topo, "Square size array Throughtput Gain by size tier"))
+    p <- p + theme_bw()
+    p <- p + theme(axis.text=element_text(size=ats))
+    p <- p + theme(text=element_text(size=ts))
+    p <- p + theme(legend.position="bottom")
+    
+    q <- ggplot(column_data,aes(y=as.numeric(sq_latency_penalty), x=as.numeric(partitions), group=res))
+    q <- q + geom_point(size=ps, aes(color=res, shape=res))
+    q <- q + geom_line(size=ls, aes(color=res))
+    q <- q + geom_hline(yintercept=1)
+    q <- q + labs(x="No of Partitions", y="Latency Penalty")
+    q <- q + ggtitle(paste(topo, "Square size array Latency Penalty by size tier"))
+    q <- q + theme_bw()
+    q <- q + theme(axis.text=element_text(size=ats))
+    q <- q + theme(text=element_text(size=ts))
+    q <- q + theme(legend.position="bottom")
+    
+    z <- ggplot(column_data,aes(y=as.numeric(tp_gain_latency_metrics), x=as.numeric(partitions), group=res))
+    z <- z + geom_point(size=ps, aes(color=res, shape=res))
+    z <- z + geom_line(size=ls, aes(color=res))
+    z <- z + geom_hline(yintercept=1)
+    z <- z + labs(x="No of Partitions", y="Metrics(tp/latency)")
+    z <- z + ggtitle(paste(topo, "Square size array Metrics by size tier"))
+    z <- z + theme_bw()
+    z <- z + theme(axis.text=element_text(size=ats))
+    z <- z + theme(text=element_text(size=ts))
+    z <- z + theme(legend.position="bottom")
+    
+    print(d)
+    print(p)
+    print(q)
+    print(z)
 }
-
-d <- ggplot(column_data,aes(y=as.numeric(tp_partition), x=as.numeric(partitions), group=res))
-d <- d + geom_point(size=ps, aes(color=res, shape=res))
-d <- d + geom_line(size=ls, aes(color=res))
-d <- d + labs(x="No of Partitions", y="Throughput")
-d <- d + ggtitle("Square size array Throughtput by size tier")
-d <- d + theme_bw()
-d <- d + theme(axis.text=element_text(size=ats))
-d <- d + theme(text=element_text(size=ts))
-d <- d + theme(legend.position="bottom")
-
-p <- ggplot(column_data,aes(y=as.numeric(sq_tp_gain), x=as.numeric(partitions), group=res))
-p <- p + geom_point(size=ps, aes(color=res, shape=res))
-p <- p + geom_line(size=ls, aes(color=res))
-p <- p + geom_hline(yintercept=1)
-p <- p + labs(x="No of Partitions", y="Throughput Gain")
-p <- p + ggtitle("Square size array Throughtput Gain by size tier")
-p <- p + theme_bw()
-p <- p + theme(axis.text=element_text(size=ats))
-p <- p + theme(text=element_text(size=ts))
-p <- p + theme(legend.position="bottom")
-
-q <- ggplot(column_data,aes(y=as.numeric(sq_latency_penalty), x=as.numeric(partitions), group=res))
-q <- q + geom_point(size=ps, aes(color=res, shape=res))
-q <- q + geom_line(size=ls, aes(color=res))
-q <- q + geom_hline(yintercept=1)
-q <- q + labs(x="No of Partitions", y="Latency Penalty")
-q <- q + ggtitle("Square size array Latency Penalty by size tier")
-q <- q + theme_bw()
-q <- q + theme(axis.text=element_text(size=ats))
-q <- q + theme(text=element_text(size=ts))
-q <- q + theme(legend.position="bottom")
-
-z <- ggplot(column_data,aes(y=as.numeric(tp_gain_latency_metrics), x=as.numeric(partitions), group=res))
-z <- z + geom_point(size=ps, aes(color=res, shape=res))
-z <- z + geom_line(size=ls, aes(color=res))
-z <- z + geom_hline(yintercept=1)
-z <- z + labs(x="No of Partitions", y="Metrics(tp/latency)")
-z <- z + ggtitle("Square size array Metrics by size tier")
-z <- z + theme_bw()
-z <- z + theme(axis.text=element_text(size=ats))
-z <- z + theme(text=element_text(size=ts))
-z <- z + theme(legend.position="bottom")
-
-print(d)
-print(p)
-print(q)
-print(z)
 dev.off()

optimization_algo/rScripts/square-vs-rect.r

@@ -10,7 +10,7 @@ ts=25
 # plot_2 > latency_penalty v partition
 pdf(file="square_vs_rect.pdf", height=9, width=15)
 
-sq_data <- read.csv("../resulting_csv/cma_logmore_sq_2.csv", header=TRUE, sep=",")
+sq_data <- read.csv("../resulting_csv/cma_logmore_sq_full_evo.csv", header=TRUE, sep=",")
 sq_data_subset <- subset(sq_data, feasable == 1 & target=="DRAM_cycle" & seeding_type=="optimised")
 
 data <- read.csv("../resulting_csv/cma.csv", header=TRUE, sep=",")

optimization_algo/resulting_csv/cma_logmore_sq_full_evo.csv

-D,R,A,M,_,c,y,c,l,e
+target, evo_counter,partitions,topology, feasable, tp_partition, latency_partition, tp_fullmap, latency_fm, tp_gain, latency_penalty, best_layer_partition, best_resource_partition, time_taken, sigma, seed, valid_sampling_percentage, trial, popsize, res_unit, seeding_type
 DRAM_cycle,0,3,FasterRCNN.csv,1,8.259773170109203e-07,3632061,4.2228695412021157e-07,2368058,1.9559621933662459,1.5337719768688098,"[['Conv1', 'CB2a_1', 'CB2a_2', 'CB2a_3', 'CB2s', 'IB2b_1', 'IB2b_2', 'IB2b_3', 'IB2c_1', 'IB2c_2', 'IB2c_3', 'CB3a_1', 'CB3a_2', 'CB3a_3', 'CB3s'], ['IB3b_1', 'IB3b_2', 'IB3b_3', 'IB3c_1', 'IB3c_2', 'IB3c_3', 'IB3d_1', 'IB3d_2', 'IB3d_3', 'CB4a_1', 'CB4a_2', 'CB4a_3', 'CB4s', 'IB4b_1', 'IB4b_2'], ['IB4b_3', 'IB4c_1', 'IB4c_2', 'IB4c_3', 'IB4d_1', 'IB4d_2', 'IB4d_3', 'IB4e_1', 'IB4e_2', 'IB4e_3', 'IB4f_1', 'IB4f_2', 'IB4f_3', 'RPN_Conv1', 'RPN_Conv_bbox', 'RPN_Conv3_cls']]","[105, 48, 66]",0.05727791786193848,0.6682246239536419,"[0, 0]",22.0,2,100,17280,allzeros
 DRAM_cycle,1,3,FasterRCNN.csv,1,9.160285764274702e-07,3275007,4.2228695412021157e-07,2368058,2.1692087986376825,1.3829927307523717,"[['Conv1', 'CB2a_1', 'CB2a_2', 'CB2a_3', 'CB2s', 'IB2b_1', 'IB2b_2', 'IB2b_3', 'IB2c_1', 'IB2c_2', 'IB2c_3', 'CB3a_1'], ['CB3a_2', 'CB3a_3', 'CB3s', 'IB3b_1', 'IB3b_2', 'IB3b_3', 'IB3c_1', 'IB3c_2', 'IB3c_3', 'IB3d_1'], ['IB3d_2', 'IB3d_3', 'CB4a_1', 'CB4a_2', 'CB4a_3', 'CB4s', 'IB4b_1', 'IB4b_2', 'IB4b_3', 'IB4c_1', 'IB4c_2', 'IB4c_3', 'IB4d_1', 'IB4d_2', 'IB4d_3', 'IB4e_1', 'IB4e_2', 'IB4e_3', 'IB4f_1', 'IB4f_2', 'IB4f_3', 'RPN_Conv1', 'RPN_Conv_bbox', 'RPN_Conv3_cls']]","[66, 39, 108]",0.07065033912658691,0.7054671136953392,"[0, 0]",25.0,2,100,17280,allzeros
 DRAM_cycle,2,3,FasterRCNN.csv,1,9.160285764274702e-07,3275007,4.2228695412021157e-07,2368058,2.1692087986376825,1.3829927307523717,"[['Conv1', 'CB2a_1', 'CB2a_2', 'CB2a_3', 'CB2s', 'IB2b_1', 'IB2b_2', 'IB2b_3', 'IB2c_1', 'IB2c_2', 'IB2c_3', 'CB3a_1'], ['CB3a_2', 'CB3a_3', 'CB3s', 'IB3b_1', 'IB3b_2', 'IB3b_3', 'IB3c_1', 'IB3c_2', 'IB3c_3', 'IB3d_1'], ['IB3d_2', 'IB3d_3', 'CB4a_1', 'CB4a_2', 'CB4a_3', 'CB4s', 'IB4b_1', 'IB4b_2', 'IB4b_3', 'IB4c_1', 'IB4c_2', 'IB4c_3', 'IB4d_1', 'IB4d_2', 'IB4d_3', 'IB4e_1', 'IB4e_2', 'IB4e_3', 'IB4f_1', 'IB4f_2', 'IB4f_3', 'RPN_Conv1', 'RPN_Conv_bbox', 'RPN_Conv3_cls']]","[66, 39, 108]",0.10741591453552246,0.6719565027292101,"[0, 0]",63.0,2,100,17280,allzeros

scaleSim/scale.py

@@ -279,7 +279,7 @@ class scale:
         
         print(all_arr_dim_list)
 
-        pool = Pool(processes = 12)
+        pool = Pool(processes = 4)
         for pro in pool.imap_unordered(self.run_mp_once, all_arr_dim_list):
             self.run_name = net_name + "_" + self.dataflow + "_" + str(pro[0]) + "x" + str(pro[1])
             self.cleanup(pro)