Crossbar latency

Author: tanner-andrulis

hwcomponents_neurosim/main.py

@@ -750,7 +750,9 @@ class FlipFlop(_NeurosimPlugInComponent):
     _get_stats_func = staticmethod(neurointerface.flip_flop_stats)
     _params = ["tech_node", "cycle_period", "n_bits"]
 
-    def __init__(self, tech_node: float, cycle_period: float, n_bits: int, n_instances: int = 1):
+    def __init__(
+        self, tech_node: float, cycle_period: float, n_bits: int, n_instances: int = 1
+    ):
         super().__init__(
             tech_node=tech_node,
             cycle_period=cycle_period,
@@ -860,7 +862,9 @@ class Adder(_NeurosimPlugInComponent):
     _get_stats_func = staticmethod(neurointerface.adder_stats)
     _params = ["tech_node", "cycle_period", "n_bits"]
 
-    def __init__(self, tech_node: float, cycle_period: float, n_bits: int, n_instances: int = 1):
+    def __init__(
+        self, tech_node: float, cycle_period: float, n_bits: int, n_instances: int = 1
+    ):
         super().__init__(
             tech_node=tech_node,
             cycle_period=cycle_period,
@@ -999,7 +1003,12 @@ class MaxPool(_NeurosimPlugInComponent):
     _params = ["tech_node", "cycle_period", "n_bits", "pool_window"]
 
     def __init__(
-        self, tech_node: float, cycle_period: float, n_bits: int, pool_window: int, n_instances: int = 1
+        self,
+        tech_node: float,
+        cycle_period: float,
+        n_bits: int,
+        pool_window: int,
+        n_instances: int = 1,
     ):
         super().__init__(
             tech_node=tech_node,

hwcomponents_neurosim/neurointerface.py

@@ -354,6 +354,8 @@ def __init__(
         self.n_adder_tree_inputs = n_adder_tree_inputs
         self.n_mux_inputs = n_mux_inputs
         self.max_activation_time = 2**self.temporal_dac_bits - 1
+        self.min_latency = None
+        self.cols_active_at_once = None
 
     def run_neurosim(
         self,
@@ -438,6 +440,25 @@ def read_pipe_thread(pipe, write_to: list):
         self.comps = [
             Component(line) for line in results.split("\n") if "<COMPONENT>" in line
         ]
+
+        columns_at_once = nvsimget("Columns read at once:", results, False)
+        min_latency = nvsimget("Minimum latency per read:", results, False)
+        logger.info(
+            "Crossbar minimum latency is %s ns to read %s columns at once.",
+            min_latency,
+            columns_at_once,
+        )
+        if min_latency < self.cycle_period * 1e9:
+            logger.warning(
+                "Minimum crossbar latency of %s ns is less than the cycle "
+                "time of %s ns.",
+                min_latency,
+                self.cycle_period * 1e9,
+            )
+
+        self.min_latency = min_latency
+        self.cols_active_at_once = columns_at_once
+
         for c in self.comps:
             if "adc" in c.name:
                 c.area *= self.adc_area_scale
@@ -449,28 +470,12 @@ def read_pipe_thread(pipe, write_to: list):
             logger.error("NeuroSIM returned no components. NeuroSIM output below.")
             logger.error("\n\nNeuroSIM returned no components. NeuroSIM output below.")
             logger.error("| " + results.replace("\n", "\n| ") + "  ")
-            if err:
-                logger.error("| " + err.decode("utf-8").replace("\n", "\n| ") + "  ")
             logger.error("NeuroSIM returned no components. NeuroSIM output above.")
             raise ValueError(
                 "NeuroSIM returned no components. Check the generated Neurosim input"
                 " config and make sure all values are populated."
             )
 
-        columns_at_once = nvsimget("Columns read at once:", results, False)
-        min_latency = nvsimget("Minimum latency per read:", results, False)
-        logger.info(
-            "Crossbar minimum latency is %s ns to read %s columns at once.",
-            min_latency,
-            columns_at_once,
-        )
-        if min_latency < self.cycle_period * 1e9:
-            logger.warning(
-                "Minimum crossbar latency of %s ns is less than the cycle "
-                "time of %s ns.",
-                min_latency,
-                self.cycle_period * 1e9,
-            )
         # Remove the config file
         os.remove(inputpath)
 
@@ -596,7 +601,13 @@ def rowcol_stats(
     read = read_off + (read_on - read_off) * avg_input
     write = write_on  # Can't gate writes becasuse we still have to reset cells
 
-    latency = crossbar.latency(kind)
+    latency = max(
+        crossbar.min_latency, crossbar.read_pulse_width * crossbar.max_activation_time
+    )
+    if kind == "row":  # Latency per row
+        latency /= crossbar.rows
+    if kind == "col":  # Latency per column
+        latency /= crossbar.cols_active_at_once
 
     return stats2dict(read, write, area, leakage * crossbar.cycle_period, latency)
 
@@ -618,8 +629,6 @@ def row_stats(
     dac_scale = 2 ** (crossbar.voltage_dac_bits - 1) - 1
     stats["Area"] = stats["Area"] + stats_dac["Area"] * dac_scale
     stats["Leakage"] = stats["Leakage"] + stats_dac["Leakage"] * dac_scale
-    stats["Latency"] = max(stats["Latency"], crossbar.read_pulse_width)
-    stats["Latency"] *= crossbar.max_activation_time
     return stats
 
 
@@ -649,12 +658,15 @@ def cell_stats(
     rlscale = crossbar.cell_read_leak_energy_scale
     act_time = crossbar.max_activation_time
     wscale = crossbar.cell_write_energy_scale
+    latency = max(
+        crossbar.min_latency, crossbar.read_pulse_width * crossbar.max_activation_time
+    )
     return stats2dict(
         read_memcell_energy * act_time,
         write_memcell_energy * wscale,
         crossbar.area_per_cell(),
         crossbar.leakage_per_cell() * rlscale,
-        crossbar.max_activation_time * crossbar.read_pulse_width,
+        latency,
     )