wafer-core 0.1.29__py3-none-any.whl → 0.1.30__py3-none-any.whl

wafer_core/lib/trace_compare/analyzer.py

@@ -145,7 +145,17 @@ def analyze_traces_from_loaded(
         trace2_total = trace2_agg["total_us"] / 1000
         trace1_count = int(trace1_agg["count"])
         trace2_count = int(trace2_agg["count"])
-        ratio = trace1_avg / trace2_avg if trace2_avg > 0 else 1
+        # Speedup: ratio of total times (not per-call averages)
+        # Shows how many times faster/slower trace1 is compared to trace2
+        # > 1.0 means trace1 is slower, < 1.0 means trace1 is faster
+        # Using total time instead of avg time per call because operations may have
+        # vastly different call counts (e.g., fused vs unfused operations)
+        if trace2_total > 0:
+            ratio = trace1_total / trace2_total
+        elif trace1_total > 0:
+            ratio = float("inf")  # trace2 has no time, trace1 is infinitely slower
+        else:
+            ratio = 1.0  # Both are zero
         gap_ms = trace1_total - trace2_total
         
         trace1_pattern = list(
@@ -446,6 +456,11 @@ def analyze_traces_aligned(
     )
     same_kernel_result = analyze_same_kernels_from_alignment(alignment.layer_alignments)
     
+    # Note: amd_kernels = trace1's kernels (filtered if phase_filter != "all")
+    #       nvidia_kernels = trace2's kernels (filtered if phase_filter != "all")
+    # The variable names are misleading but trace1_* should use amd_kernels,
+    # and trace2_* should use nvidia_kernels to match the filtered kernel counts/totals.
+    
     return {
         "metadata": {
             "amd_gpu": amd_trace.gpu_name,
@@ -462,10 +477,10 @@ def analyze_traces_aligned(
             "trace2_platform": trace2.platform,
             "trace2_gpu": trace2.gpu_name,
             "trace2_device": trace2.device_props,
-            "trace1_kernels": len(amd_trace.kernel_events),
-            "trace2_kernels": len(nvidia_trace.kernel_events),
-            "trace1_total_ms": sum(k.get("dur", 0) for k in amd_trace.kernel_events) / 1000,
-            "trace2_total_ms": sum(k.get("dur", 0) for k in nvidia_trace.kernel_events) / 1000,
+            "trace1_kernels": len(amd_kernels),
+            "trace2_kernels": len(nvidia_kernels),
+            "trace1_total_ms": sum(k.get("dur", 0) for k in amd_kernels) / 1000,
+            "trace2_total_ms": sum(k.get("dur", 0) for k in nvidia_kernels) / 1000,
             "phase": phase_filter,
             "trace1_layers": alignment.num_layers,
             "trace2_layers": alignment.num_layers,
@@ -579,7 +594,17 @@ def analyze_traces_aligned(
         trace2_total = trace2_agg["total_us"] / 1000
         trace1_count = int(trace1_agg["count"])
         trace2_count = int(trace2_agg["count"])
-        ratio = trace1_avg / trace2_avg if trace2_avg > 0 else 1
+        # Speedup: ratio of total times (not per-call averages)
+        # Shows how many times faster/slower trace1 is compared to trace2
+        # > 1.0 means trace1 is slower, < 1.0 means trace1 is faster
+        # Using total time instead of avg time per call because operations may have
+        # vastly different call counts (e.g., fused vs unfused operations)
+        if trace2_total > 0:
+            ratio = trace1_total / trace2_total
+        elif trace1_total > 0:
+            ratio = float("inf")  # trace2 has no time, trace1 is infinitely slower
+        else:
+            ratio = 1.0  # Both are zero
         gap_ms = trace1_total - trace2_total
         
         trace1_pattern = list(

{wafer_core-0.1.29.dist-info → wafer_core-0.1.30.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: wafer-core
-Version: 0.1.29
+Version: 0.1.30
 Summary: Core utilities and environments for Wafer GPU kernel optimization
 Requires-Python: >=3.10
 Requires-Dist: aiohttp>=3.9.0

{wafer_core-0.1.29.dist-info → wafer_core-0.1.30.dist-info}/RECORD

@@ -321,7 +321,7 @@ wafer_core/lib/rocprofiler/systems/sample/profiler.py,sha256=CYZPTzNXd48LoCfmY6h
 wafer_core/lib/trace_compare/PERFORMANCE.md,sha256=jkJh7ApZi8H7NKTcz8v0LNtwSFtIUqY88e3QbL749ww,3823
 wafer_core/lib/trace_compare/__init__.py,sha256=CyUPbPQDYhVLCFFA7S_jNSilG3OgqYjmHSKfR5X11go,1377
 wafer_core/lib/trace_compare/aligner.py,sha256=1S8Ob3RaEsIjN0HdqEx0yGsW5uf_lMrJVSH_MnZhKok,13788
-wafer_core/lib/trace_compare/analyzer.py,sha256=YkuOPA3HFX_7mNUEhE9CMOtEMGLQd12lvUkvqqeQF14,29698
+wafer_core/lib/trace_compare/analyzer.py,sha256=Ou_gooG027YVuYVF5oddAkMsObXrrPQLBPHUzSMA4Vg,31078
 wafer_core/lib/trace_compare/api.py,sha256=JSRTcd7eZK1Z8l18TFEiA5A8ENJS1TMz7oIiw1KBbAs,8796
 wafer_core/lib/trace_compare/architecture.py,sha256=8bqlAJQeJLBHblyXvFV-w55PIKiVQDPjDQZ8Jx4tuGg,2110
 wafer_core/lib/trace_compare/classifier.py,sha256=cYAmDW8S75N6cE3mJNZM-UKCJSX7rFP-8klVrukBvNQ,17504
@@ -697,6 +697,6 @@ wafer_core/utils/modal_execution/modal_app.py,sha256=VfS2cX8gHtnlPXemmMcEwDPeQdh
 wafer_core/utils/modal_execution/modal_config.py,sha256=7cGX9TGqilQ3qxI3OFGXV5orjtyRU-PEDOJ4vP2oxno,4421
 wafer_core/utils/modal_execution/modal_execution.py,sha256=gChjnV6jqA3A7IRP3DfvV5cSfm_MN0X4f7JZufXgdZE,24594
 wafer_core/utils/modal_execution/test_modal.py,sha256=_jqou_hrLs1Daf1590Pnb0a_lXMMa2rczAPpW9HpoNQ,8153
-wafer_core-0.1.29.dist-info/METADATA,sha256=Qjyx92KhI1joutpM8lF0G1zgPou-d8CdzWI80QQqKYg,1477
-wafer_core-0.1.29.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
-wafer_core-0.1.29.dist-info/RECORD,,
+wafer_core-0.1.30.dist-info/METADATA,sha256=YuF3VyyP3tvmv2S-7E8epi1J2_1e2yXJfapS1uGQ0Zs,1477
+wafer_core-0.1.30.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+wafer_core-0.1.30.dist-info/RECORD,,