Upgrades for router statistics

* Add p port for network stats tracking
* Add tag field to CSV file
* Remove "trace_en" dependency
* Update testbench
* Add preliminary heatmap script

software/py/router_heatmap.py

@@ -0,0 +1,67 @@
+#!env python3
+# Preliminary script.
+
+# For now, run this script from inside of the directory that contains router_stat.csv
+
+import pandas as pd
+
+# Read the CSV into pandas
+df = pd.read_csv("router_stat.csv")
+
+# Find the minimum and maximum cycle 
+cmin = df.global_ctr.min()
+cmax = df.global_ctr.max()
+# Keep only minimum and maximum entries
+df = df[(df.global_ctr == df.global_ctr.min()) | (df.global_ctr == df.global_ctr.max())]
+# Stringify the minimum and maximum entries
+df["Type"] = df.global_ctr.map({cmax: "End", cmin:"Start"})
+
+# Map the Request and Response directions
+df.XY_order = df.XY_order.map({0:"Response", 1:"Request"})
+
+# Drop unnecessary rows
+p = df.drop(["timestamp", "global_ctr", "tag"], axis = 1)
+duration = cmax - cmin
+p.idle = 100.0 * p.idle / duration
+p.utilized = 100.0 * p.utilized / duration
+p.stalled = 100.0 * p.stalled / duration
+p.arbitrated = 100.0 * p.arbitrated / duration
+# Map output directions to human readable strings
+p.output_dir = p.output_dir.map({0: "0 - P (Router Output)",
+ 1:"1 - W (Router Output)",
+ 2:"2 - E (Router Output)",
+ 3:"3 - N (Router Output)",
+ 4:"4 - S (Router Output)",
+ 5:"5 - RW (Router Output)",
+ 6 :"6 - RE (Router Output)",
+ 15: "0 - P (Router Input)"})
+
+# Map Y indicies to "Index (Type)"
+ys = p.y[p.y != 0].unique()
+type_map = {y_i:f"{y_i:02d} (T)" for y_i in ys}
+type_map[ys.max()] = f"{ys.max():02d} (V)"
+type_map[ys.min()] = f"{ys.min():02d} (V)"
+type_map[0] = "00 (H)"
+p.y = p.y.map(type_map)
+
+# Create a hierarchical index
+p = p.set_index(["Type", "XY_order", "output_dir", "y", "x"])
+
+# Select the columns we will print.
+f = p[["utilized", "stalled", "arbitrated", "idle"]].copy()
+f = f.loc[("End")] - f.loc[("Start")]
+f.columns = f.columns.map(lambda x : f"{x:>11}")
+
+with open("router_heatmap.rpt", "w") as fd:
+ data = f.unstack()
+ for n in list(data.index.get_level_values('XY_order').unique()):
+ for d in list(data.index.get_level_values('output_dir').unique()):
+ row = data.loc[(n, d)]
+ s = row.to_string(float_format="%5.1f")
+ if(d == "0 - P (Router Input)"):
+ s = s.replace("arbitrated", "Stalled by Arbitration")
+ else:
+ s = s.replace("arbitrated", "Utilized by Arbitration")
+ fd.write(f"{d}, {n} Network")
+ fd.write(s)
+ fd.write("\n\n")

testbenches/common/v/bsg_nonsynth_manycore_testbench.v

@@ -782,6 +782,7 @@ if (enable_router_profiling_p) begin
  ,.global_ctr_i($root.`HOST_MODULE_PATH.global_ctr)
  ,.trace_en_i($root.`HOST_MODULE_PATH.trace_en)
  ,.print_stat_v_i($root.`HOST_MODULE_PATH.print_stat_v)
+ ,.print_stat_tag_i($root.`HOST_MODULE_PATH.print_stat_tag)
  );
 end
 

testbenches/common/v/router_profiler.v

@@ -10,8 +10,10 @@
  // {timestamp},{global_ctr},{x},{y},{XY_order_p},{output_dir},{idle},{utilized},{stalled},{arbitrated}
 
 module router_profiler
+ import bsg_noc_pkg::*;
  #(parameter x_cord_width_p="inv"
  , parameter y_cord_width_p="inv"
+ , parameter data_width_p=32
  , parameter dims_p="inv"
  , parameter XY_order_p="inv"
  , parameter dirs_lp = 1+(2*dims_p)
@@ -26,6 +28,11 @@ module router_profiler
  , input reset_i
 
 
+ , input [dirs_lp-1:0] v_i
+ , input [dirs_lp-1:0] ready_i
+ , input [dirs_lp-1:0] yumi_o
+
+ , input [dirs_lp-1:0][dirs_lp-1:0] req
  , input [dirs_lp-1:0][dirs_lp-1:0] req_t
  , input [dirs_lp-1:0][dirs_lp-1:0] yumi_lo
 
@@ -34,6 +41,7 @@ module router_profiler
 
  , input trace_en_i
  , input print_stat_v_i
+ , input [data_width_p-1:0] print_stat_tag_i
  , input [31:0] global_ctr_i
  );
 
@@ -48,13 +56,19 @@ module router_profiler
 
 
  router_stat_s [dirs_lp-1:0] stat_r;
+ router_stat_s stat_p_r;
 
 
  always_ff @ (posedge clk_i) begin
  if (reset_i) begin
  stat_r <= '0;
+ stat_p_r <= '0;
  end
  else begin
+ stat_p_r.idle <= stat_p_r.idle + (v_i[P] == 0);
+ stat_p_r.utilized <= stat_p_r.utilized + (v_i[P] & yumi_o[P]);
+ stat_p_r.stalled <= stat_p_r.stalled + (v_i[P] & !yumi_o[P]);
+ stat_p_r.arbitrated <= stat_p_r.arbitrated + (v_i[P] & !yumi_o[P] & ((req[P] & ready_i) != '0));
  for (integer i = 0; i < dirs_lp; i++) begin
  stat_r[i].idle <= stat_r[i].idle + ($countones(req_t[i]) == 0);
  stat_r[i].utilized <= stat_r[i].utilized + ((req_t[i] & yumi_lo[i]) != '0);
@@ -85,7 +99,7 @@ module router_profiler
  & (XY_order_p == 1)) begin
 
  fd = $fopen(tracefile_p, "a");
- $fwrite(fd,"timestamp,global_ctr,x,y,XY_order,output_dir,idle,utilized,stalled,arbitrated\n");
+ $fwrite(fd,"timestamp,global_ctr,tag,x,y,XY_order,output_dir,idle,utilized,stalled,arbitrated\n");
  $fclose(fd);
 
  end
@@ -96,17 +110,26 @@ module router_profiler
 
  // when there is print_stat_v_i signal received, it dumps the stats.
  always @ (posedge clk_i) begin
- if (~reset_i & trace_en_i & print_stat_v_i) begin
+ if (~reset_i & print_stat_v_i) begin
  fd = $fopen(tracefile_p, "a");
  for (integer i = 0; i < dirs_lp; i++) begin
- $fwrite(fd, "%0t,%0d,%0d,%0d,%0d,%0d,%0d,%0d,%0d,%0d\n", 
- $time, global_ctr_i, my_x_i, my_y_i, XY_order_p, i,
+ $fwrite(fd, "%0t,%0d,%0x,%0d,%0d,%0d,%0d,%0d,%0d,%0d,%0d\n",
+ $time, global_ctr_i, print_stat_tag_i, my_x_i, my_y_i, XY_order_p, i,
  stat_r[i].idle,
  stat_r[i].utilized,
  stat_r[i].stalled,
  stat_r[i].arbitrated
  );
  end
+ // Use output_dir = 15 so that it doesn't overlap with potential full ruche implementations
+ $fwrite(fd, "%0t,%0d,%0x,%0d,%0d,%0d,%0d,%0d,%0d,%0d,%0d\n",
+ $time, global_ctr_i, print_stat_tag_i, my_x_i, my_y_i, XY_order_p, 15,
+ stat_p_r.idle,
+ stat_p_r.utilized,
+ stat_p_r.stalled,
+ stat_p_r.arbitrated
+ );
+
  $fclose(fd); 
  end
  end