carm-roofline 1.0.1.dev0__tar.gz → 1.0.1.dev1__tar.gz

{carm_roofline-1.0.1.dev0 → carm_roofline-1.0.1.dev1}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: carm-roofline
-Version: 1.0.1.dev0
+Version: 1.0.1.dev1
 Summary: CARM: Cache-Aware Roofline Model benchmarking and visualization toolkit
 Author: CHaMP Team
 License:                   GNU LESSER GENERAL PUBLIC LICENSE

{carm_roofline-1.0.1.dev0 → carm_roofline-1.0.1.dev1}/carm_roofline.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: carm-roofline
-Version: 1.0.1.dev0
+Version: 1.0.1.dev1
 Summary: CARM: Cache-Aware Roofline Model benchmarking and visualization toolkit
 Author: CHaMP Team
 License:                   GNU LESSER GENERAL PUBLIC LICENSE

{carm_roofline-1.0.1.dev0 → carm_roofline-1.0.1.dev1}/profiling/backends.py

@@ -126,6 +126,8 @@ class PAPIHLBackend(ProfilerBackend):
         detail(f"Available PAPI events: {len(self._available_events)}")
         for metric_type, impl in self._resolved_metrics.items():
             detail(f"  {metric_type.name} -> {impl.description}")
+            if impl.warning is not None:
+                detail(f"    Note: {impl.warning}")
 
         return any(impl.priority < 100 for impl in self._resolved_metrics.values())
 

{carm_roofline-1.0.1.dev0 → carm_roofline-1.0.1.dev1}/profiling/metrics.py

@@ -106,6 +106,7 @@ class MetricDefinition:
     priority: int = 0
     priority_modifier: Callable[[MetricResolutionConfig], int] = lambda _: 0
     description: str = ""
+    warning: str | None = None
 
 
 # ---------------------------------------------------------------------------
@@ -168,30 +169,36 @@ def _build_metric_definitions() -> dict[MetricType, list[MetricDefinition]]:
         """Return +15 when config.data_type matches *dt*, -15 otherwise."""
         return lambda cfg: 15 if cfg.data_type is dt else -15
 
+    def _make_fmadd_warning(papi_event: str) -> str:
+        return f"{papi_event} may count multiply-add instructions as 1 or 2 ops depending on the platform"
+
     return {
         MetricType.FLOPS: [
+            MetricDefinition(
+                type=MetricType.FLOPS,
+                required_events=frozenset({"PAPI_FP_OPS"}),
+                compute=lambda e, ctx: e["PAPI_FP_OPS"],
+                priority=100,
+                description="Exact value from PAPI_FP_OPS (both precisions)",
+                warning=_make_fmadd_warning("PAPI_FP_OPS"),
+            ),
             MetricDefinition(
                 type=MetricType.FLOPS,
                 required_events=frozenset({"PAPI_DP_OPS"}),
                 compute=lambda e, ctx: e["PAPI_DP_OPS"],
-                priority=100,
+                priority=90,
                 priority_modifier=_data_type_match(DataType.f64),
                 description="Exact value from PAPI_DP_OPS",
+                warning=_make_fmadd_warning("PAPI_DP_OPS"),
             ),
             MetricDefinition(
                 type=MetricType.FLOPS,
                 required_events=frozenset({"PAPI_SP_OPS"}),
                 compute=lambda e, ctx: e["PAPI_SP_OPS"],
-                priority=100,
+                priority=90,
                 priority_modifier=_data_type_match(DataType.f32),
                 description="Exact value from PAPI_SP_OPS",
-            ),
-            MetricDefinition(
-                type=MetricType.FLOPS,
-                required_events=frozenset({"PAPI_FP_OPS"}),
-                compute=lambda e, ctx: e["PAPI_FP_OPS"],
-                priority=50,
-                description="Approximated from PAPI_FP_OPS (has all precisions)",
+                warning=_make_fmadd_warning("PAPI_SP_OPS"),
             ),
             MetricDefinition(
                 type=MetricType.FLOPS,
@@ -199,6 +206,7 @@ def _build_metric_definitions() -> dict[MetricType, list[MetricDefinition]]:
                 compute=lambda e, ctx: e["PAPI_FP_INS"] * ctx.ops_per_instruction,
                 priority=10,
                 description="Approximated from PAPI_FP_INS (FP instr count)",
+                warning="PAPI_FP_INS counts fused multiply-add instructions as 1 operation",
             ),
         ],
         MetricType.BYTES: [
@@ -235,12 +243,28 @@ def _build_metric_definitions() -> dict[MetricType, list[MetricDefinition]]:
                 priority=85,
                 description="Approximated from PAPI_LD_INS + PAPI_SR_INS and bytes per instruction",
             ),
+            # Relevant for Zen3 (no PAPI loads/store events, only DCA, native forwarding event for higher accuracy)
+            MetricDefinition(
+                type=MetricType.BYTES,
+                required_events=frozenset({"PAPI_L1_DCA", "STORE_TO_LOAD_FORWARD"}),
+                compute=lambda e, ctx: (e["PAPI_L1_DCA"] + e["STORE_TO_LOAD_FORWARD"]) * ctx.bytes_per_instruction,
+                priority=40,
+                description="Approximated from PAPI_L1_DCA + STORE_TO_LOAD_FORWARD and bytes per instruction",
+                warning=(
+                    "PAPI_L1_DCA counts cache accesses: if the microarchitecture coalesces requests into fewer cache "
+                    "accesses, this metric may underestimate the number of bytes."
+                ),
+            ),
             MetricDefinition(
                 type=MetricType.BYTES,
                 required_events=frozenset({"PAPI_L1_DCA"}),
                 compute=lambda e, ctx: e["PAPI_L1_DCA"] * ctx.bytes_per_instruction,
                 priority=30,
                 description="Approximated from PAPI_L1_DCA and bytes per instruction",
+                warning=(
+                    "PAPI_L1_DCA counts cache accesses: if the microarchitecture coalesces requests into fewer cache "
+                    "accesses or forwards stores to loads, this metric may underestimate the number of bytes."
+                ),
             ),
         ],
     }

{carm_roofline-1.0.1.dev0 → carm_roofline-1.0.1.dev1}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "carm-roofline"
-version = "1.0.1.dev0"
+version = "1.0.1.dev1"
 description = "CARM: Cache-Aware Roofline Model benchmarking and visualization toolkit"
 readme = "README.md"
 requires-python = ">=3.9"