audit: conservatively count predictor arithmetic

2 files changed, 20 insertions, 4 deletions

diff --git a/experiments/conv_run.py b/experiments/conv_run.py
index 2e2787b..c25c48c 100644
--- a/experiments/conv_run.py
+++ b/experiments/conv_run.py
@@ -191,7 +191,11 @@ def work_report(net, mode, counters):
             * net.mixed_elementwise_ops_per_example()
             + (counters["predictor_warmup_examples"]
                + counters["predictor_update_examples"])
-            * net.predictor_elementwise_ops_per_example)
+            * net.predictor_elementwise_ops_per_example
+            + counters["traffic_calibration_examples"]
+            * net.traffic_calibration_elementwise_ops_per_example
+            + counters["traffic_audit_examples"]
+            * net.predictor_audit_elementwise_ops_per_example)
     else:
         elementwise_operations = 0
     components = {
@@ -234,7 +238,8 @@ def work_report(net, mode, counters):
             "equals one forward-weight MAC count; BP reverse is estimated as one "
             "weight-gradient plus one activation-gradient convolution per forward "
             "convolution; mixed-traffic/predictor elementwise arithmetic is reported "
-            "separately and is not folded into MACs"),
+            "as a conservative operation estimate separately and is not folded "
+            "into MACs"),
     }
 
 
diff --git a/sdil/conv.py b/sdil/conv.py
index cc4c7c6..ea62e95 100644
--- a/sdil/conv.py
+++ b/sdil/conv.py
@@ -662,8 +662,19 @@ class CIFARKPMixedTrafficResNet(CIFARKPResNet):
 
     @property
     def predictor_elementwise_ops_per_example(self):
-        # Residual, centering, moments, normalization, and affine updates.
-        return 12 * self.mixed_units_per_example
+        # Conservative count for traffic/prediction, centering, reductions,
+        # normalized moments, residual power, and affine parameter updates.
+        return 24 * self.mixed_units_per_example
+
+    @property
+    def traffic_calibration_elementwise_ops_per_example(self):
+        # Two RMS measurements, gain application, and achieved-ratio audit.
+        return 8 * self.mixed_units_per_example
+
+    @property
+    def predictor_audit_elementwise_ops_per_example(self):
+        # Traffic/prediction/residual construction and two power reductions.
+        return 10 * self.mixed_units_per_example
 
     @torch.no_grad()
     def traffic_fields(self, hiddens):