Merge pull request #106 from cosmo-wander-ai/codex/rv1126b-npu-load-metrics

fix(rknn): parse RV1126B aggregate NPU load

Author: samuel--hu

src/service/system/impl/AcceleratorMetricsProviderRknn.cc

@@ -51,16 +51,30 @@ namespace {
                 return std::nullopt;
             cores_by_id[*core_id] = static_cast<double>(*percent) / 100.0;
         }
-        if (cores_by_id.empty())
+        if (!cores_by_id.empty()) {
+            NpuLoadSnapshot result;
+            result.cores.reserve(cores_by_id.size());
+            for (const auto& [core_id, load] : cores_by_id) {
+                (void)core_id;
+                result.cores.push_back(load);
+                result.aggregate = std::max(result.aggregate, load);
+            }
+            return result;
+        }
+
+        // Single-core Rockchip parts such as RV1126B expose only an aggregate
+        // counter ("NPU load: 37%") instead of RK3576-style CoreN fields.
+        static const std::regex aggregate_pattern(R"(NPU\s+load\s*:\s*([0-9]+)\s*%)", std::regex::icase);
+        std::smatch aggregate_match;
+        if (!std::regex_search(text, aggregate_match, aggregate_pattern))
+            return std::nullopt;
+        const auto percent = ParseUnsigned(aggregate_match[1]);
+        if (!percent || *percent > 100)
             return std::nullopt;
 
         NpuLoadSnapshot result;
-        result.cores.reserve(cores_by_id.size());
-        for (const auto& [core_id, load] : cores_by_id) {
-            (void)core_id;
-            result.cores.push_back(load);
-            result.aggregate = std::max(result.aggregate, load);
-        }
+        result.aggregate = static_cast<double>(*percent) / 100.0;
+        result.cores.push_back(result.aggregate);
         return result;
     }
 

test/test_accelerator_metrics_provider.cc

@@ -67,6 +67,22 @@ TEST_CASE("RKNN accelerator metrics provider reports per-core busy-time load", "
     CHECK(metrics.gpudevusage.front().gpuusage == Catch::Approx(0.42));
 }
 
+TEST_CASE("RKNN accelerator metrics provider reports aggregate-only busy-time load",
+          "[system][metrics][rknn]") {
+    ScopedNpuLoadFixture fixture("NPU load:  37%\n");
+    auto provider = cosmo::service::detail::CreateAcceleratorMetricsProvider();
+    REQUIRE(provider != nullptr);
+
+    const auto metrics = provider->QueryUtilization();
+    CHECK(metrics.gpuusageAvailable);
+    CHECK(metrics.gpuusage == Catch::Approx(0.37));
+    CHECK(metrics.utilizationMetric == "busy-time-load");
+    REQUIRE(metrics.coreUtilizations.size() == 1);
+    CHECK(metrics.coreUtilizations.front() == Catch::Approx(0.37));
+    REQUIRE(metrics.gpudevusage.size() == 1);
+    CHECK(metrics.gpudevusage.front().gpuusage == Catch::Approx(0.37));
+}
+
 TEST_CASE("RKNN accelerator metrics provider rejects invalid load", "[system][metrics][rknn]") {
     ScopedNpuLoadFixture fixture("NPU load:  Core0: 105%, Core1: 0%,\n");
     auto provider = cosmo::service::detail::CreateAcceleratorMetricsProvider();
@@ -77,4 +93,16 @@ TEST_CASE("RKNN accelerator metrics provider rejects invalid load", "[system][me
     CHECK(metrics.gpuusage == 0.0);
     CHECK(metrics.coreUtilizations.empty());
 }
+
+TEST_CASE("RKNN accelerator metrics provider rejects invalid aggregate-only load",
+          "[system][metrics][rknn]") {
+    ScopedNpuLoadFixture fixture("NPU load: 101%\n");
+    auto provider = cosmo::service::detail::CreateAcceleratorMetricsProvider();
+    REQUIRE(provider != nullptr);
+
+    const auto metrics = provider->QueryUtilization();
+    CHECK_FALSE(metrics.gpuusageAvailable);
+    CHECK(metrics.gpuusage == 0.0);
+    CHECK(metrics.coreUtilizations.empty());
+}
 #endif