Improve how CUDAEnsemble reports failure to find a CUDA device.

This aligns closer to CUDASimulation's behaviour. Closes #858
FLAMEGPU · Jun 1, 2022 · 376f37d · 376f37d
1 parent 620e904
commit 376f37d
Show file tree

Hide file tree

Showing 2 changed files with 26 additions and 4 deletions.
diff --git a/src/flamegpu/gpu/CUDAEnsemble.cu b/src/flamegpu/gpu/CUDAEnsemble.cu
@@ -71,14 +71,26 @@ unsigned int CUDAEnsemble::simulate(const RunPlanVector &plans) {
     run_logs.clear();
     run_logs.resize(plans.size());
     // Workout how many devices and runner we will be executing
-    int ct = -1;
-    gpuErrchk(cudaGetDeviceCount(&ct));
+    int device_count = -1;
+    cudaError_t cudaStatus = cudaGetDeviceCount(&device_count);
+    if (cudaStatus != cudaSuccess) {
+        THROW exception::InvalidCUDAdevice("Error finding CUDA devices!  Do you have a CUDA-capable GPU installed?, in CUDAEnsemble::simulate()");
+    }
+    if (device_count == 0) {
+        THROW exception::InvalidCUDAdevice("Error no CUDA devices found!, in CUDAEnsemble::simulate()");
+    }
+    for (const int &id : config.devices) {
+        if (id >= device_count) {
+            THROW exception::InvalidCUDAdevice("Requested CUDA device %d is not valid, only %d CUDA devices available!, in CUDAEnsemble::simulate()", id, device_count);
+        }
+    }
+
     std::set<int> devices;
     if (config.devices.size()) {
         devices = config.devices;
     } else {
-        for (int i = 0; i < ct; ++i) {
-        devices.emplace(i);
+        for (int i = 0; i < device_count; ++i) {
+            devices.emplace(i);
         }
     }
     // Check that each device is capable, and init cuda context

diff --git a/tests/test_cases/gpu/test_cuda_ensemble.cu b/tests/test_cases/gpu/test_cuda_ensemble.cu
@@ -110,6 +110,16 @@ TEST(TestCUDAEnsemble, initialise_devices) {
     ensemble.initialise(sizeof(argv) / sizeof(char*), argv);
     EXPECT_EQ(ensemble.getConfig().devices, std::set<int>({0}));
 }
+TEST(TestCUDAEnsemble, initialise_devices_wrong) {
+    flamegpu::ModelDescription model("test");
+    flamegpu::CUDAEnsemble ensemble(model);
+    int device_ct = -1;
+    EXPECT_EQ(cudaGetDeviceCount(&device_ct), cudaSuccess);
+    ensemble.Config().devices = { device_ct };
+    RunPlanVector plan(model, 1);
+    // Sim with out of bounds device ID, get exception
+    EXPECT_THROW(ensemble.simulate(plan), exception::InvalidCUDAdevice);
+}
 TEST(TestCUDAEnsemble, initialise_quiet) {
     // Create a model
     flamegpu::ModelDescription model("test");