support running DP in circuit training

dreamplace/Params.py

@@ -169,3 +169,70 @@ def solution_file_suffix(self):
             return "def"
         else: # Bookshelf
             return "pl"
+
+def get_bin_size(width, height, num_bins=128*128):
+    """
+    @brief find the power two bin size closest to the canvas ratio.
+    """
+    num_bins_x = math.sqrt(num_bins * width / height)
+    num_bins_x = int(math.pow(2, round(math.log(num_bins_x) / math.log(2))))
+    num_bins_x = max(min(num_bins_x, num_bins), 1) # constrain num_bins_x between 1 and num_bins
+    num_bins_y = int(num_bins / num_bins_x)
+
+    return num_bins_x, num_bins_y
+
+def get_dreamplace_params(
+    iteration, 
+    target_density,
+    learning_rate,
+    canvas_width=None,
+    canvas_height=None,
+    num_bins_x=None,
+    num_bins_y=None,
+    gpu=False,
+    result_dir='results',
+    legalize_flag=False,
+    stop_overflow=0.1,
+    routability_opt_flag=False):
+    """"
+    @brief return the parameters to config Dreamplace in circuit training.
+    """
+    params = Params()
+    params.circuit_training_mode = True
+
+    # set number of bins
+    if num_bins_x and num_bins_y:
+        params.num_bins_x = num_bins_x
+        params.num_bins_y = num_bins_y
+    elif canvas_width and canvas_height:
+        # extract #bins from canvas info
+        params.num_bins_x, params.num_bins_y = get_bin_size(canvas_width,
+                                                            canvas_height)
+    else:
+        num_bins_x = 128
+        num_bins_y = 128
+
+    params.global_place_stages = [{
+        'num_bins_x': params.num_bins_x,
+        'num_bins_y': params.num_bins_y,
+        'iteration': iteration,
+        'learning_rate': learning_rate,
+        'wirelength': 'weighted_average',
+        'optimizer': 'nesterov',
+    }]
+    params.legalize_flag = legalize_flag
+    params.detailed_place_flag = False
+    params.target_density = target_density
+    params.density_weight = 8e-5
+    params.gpu = gpu
+    params.result_dir = result_dir
+    params.stop_overflow = stop_overflow
+
+    # disable regioning
+    params.regioning = False
+
+    # routability related flag
+    params.routability_opt_flag = routability_opt_flag
+    params.adjust_nctugr_area_flag = False
+
+    return params

dreamplace/PlaceDB.py

@@ -87,6 +87,8 @@ def __init__(self):
         self.bin_size_y = None
         self.num_bins_x = None
         self.num_bins_y = None
+
+        self.num_non_movable_macros = None
 
         self.num_movable_pins = None
 
@@ -439,25 +441,26 @@ def print_row(self, row_id):
         """
         logging.debug("row %d %s" % (row_id, self.rows[row_id]))
 
-    #def flatten_nested_map(self, net2pin_map):
-    #    """
-    #    @brief flatten an array of array to two arrays like CSV format
-    #    @param net2pin_map array of array
-    #    @return a pair of (elements, cumulative column indices of the beginning element of each row)
-    #    """
-    #    # flat netpin map, length of #pins
-    #    flat_net2pin_map = np.zeros(len(pin2net_map), dtype=np.int32)
-    #    # starting index in netpin map for each net, length of #nets+1, the last entry is #pins
-    #    flat_net2pin_start_map = np.zeros(len(net2pin_map)+1, dtype=np.int32)
-    #    count = 0
-    #    for i in range(len(net2pin_map)):
-    #        flat_net2pin_map[count:count+len(net2pin_map[i])] = net2pin_map[i]
-    #        flat_net2pin_start_map[i] = count
-    #        count += len(net2pin_map[i])
-    #    assert flat_net2pin_map[-1] != 0
-    #    flat_net2pin_start_map[len(net2pin_map)] = len(pin2net_map)
-
-    #    return flat_net2pin_map, flat_net2pin_start_map
+    ## add this func for circuit training.
+    def flatten_nested_map(self, pin2net_map, net2pin_map):
+       """
+       @brief flatten an array of array to two arrays like CSV format
+       @param net2pin_map array of array
+       @return a pair of (elements, cumulative column indices of the beginning element of each row)
+       """
+       # flat netpin map, length of #pins
+       flat_net2pin_map = np.zeros(len(pin2net_map), dtype=np.int32)
+       # starting index in netpin map for each net, length of #nets+1, the last entry is #pins
+       flat_net2pin_start_map = np.zeros(len(net2pin_map)+1, dtype=np.int32)
+       count = 0
+       for i in range(len(net2pin_map)):
+           flat_net2pin_map[count:count+len(net2pin_map[i])] = net2pin_map[i]
+           flat_net2pin_start_map[i] = count
+           count += len(net2pin_map[i])
+       assert flat_net2pin_map[-1] != 0
+       flat_net2pin_start_map[len(net2pin_map)] = len(pin2net_map)
+
+       return flat_net2pin_map, flat_net2pin_start_map
 
     def read(self, params):
         """
@@ -582,7 +585,9 @@ def __call__(self, params):
         """
         tt = time.time()
 
-        self.read(params)
+        # circuit training will not read local design.
+        if not params.circuit_training_mode:
+            self.read(params)
         self.initialize(params)
 
         logging.info("reading benchmark takes %g seconds" % (time.time()-tt))
@@ -679,14 +684,17 @@ def initialize(self, params):
 
         # shift and scale
         # adjust shift_factor and scale_factor if not set
-        params.shift_factor[0] = self.xl
-        params.shift_factor[1] = self.yl
-        logging.info("set shift_factor = (%g, %g), as original row bbox = (%g, %g, %g, %g)" 
-                % (params.shift_factor[0], params.shift_factor[1], self.xl, self.yl, self.xh, self.yh))
-        if params.scale_factor == 0.0 or self.site_width != 1.0:
-            params.scale_factor = 1.0 / self.site_width
-        logging.info("set scale_factor = %g, as site_width = %g" % (params.scale_factor, self.site_width))
-        self.scale(params.shift_factor, params.scale_factor)
+        if not params.circuit_training_mode:
+            params.shift_factor[0] = self.xl
+            params.shift_factor[1] = self.yl
+            logging.info("set shift_factor = (%g, %g), as original row bbox = (%g, %g, %g, %g)" 
+                    % (params.shift_factor[0], params.shift_factor[1], self.xl, self.yl, self.xh, self.yh))
+            if params.scale_factor == 0.0 or self.site_width != 1.0:
+                params.scale_factor = 1.0 / self.site_width
+            logging.info("set scale_factor = %g, as site_width = %g" % (params.scale_factor, self.site_width))
+            self.scale(params.shift_factor, params.scale_factor)
+        else:
+            params.scale_factor = 1.0
 
         content = """
 ================================= Benchmark Statistics =================================

dreamplace/params.json

@@ -219,5 +219,9 @@
 "deterministic_flag" : {
     "descripton" : "whether require run-to-run determinism, may have efficiency overhead", 
     "default" : 0
+    },
+"circuit_training_mode" : {
+    "descripton" : "whether use Dreamplace for circuit training", 
+    "default" : 0
     }
 }