Avoid accessing GPUs when using CPU only (#2012)

* avoid accessing GPUs when using CPU only

* remove unused import/parameter

qiskit_aer/backends/aer_simulator.py

@@ -702,12 +702,10 @@ def __init__(
 
         # Update available methods and devices for class
         if AerSimulator._AVAILABLE_DEVICES is None:
-            AerSimulator._AVAILABLE_DEVICES = available_devices(
-                self._controller, AerSimulator._SIMULATION_DEVICES
-            )
+            AerSimulator._AVAILABLE_DEVICES = available_devices(self._controller)
         if AerSimulator._AVAILABLE_METHODS is None:
             AerSimulator._AVAILABLE_METHODS = available_methods(
-                self._controller, AerSimulator._SIMULATION_METHODS, AerSimulator._AVAILABLE_DEVICES
+                AerSimulator._SIMULATION_METHODS, AerSimulator._AVAILABLE_DEVICES
             )
 
         # Default configuration

qiskit_aer/backends/backend_utils.py

@@ -19,7 +19,6 @@
 
 import psutil
 from qiskit.circuit import QuantumCircuit
-from qiskit.compiler import assemble
 from qiskit.qobj import QasmQobjInstruction
 from qiskit.result import ProbDistribution
 from qiskit.quantum_info import Clifford
@@ -441,40 +440,23 @@ def cpp_execute_circuits(controller, aer_circuits, noise_model, config):
     return controller.execute(aer_circuits, noise_model, config)
 
 
-def available_methods(controller, methods, devices):
-    """Check available simulation methods by running a dummy circuit."""
-    # Test methods are available using the controller
-    dummy_circ = QuantumCircuit(1)
-    dummy_circ.id(0)
+def available_methods(methods, devices):
+    """Check available simulation methods"""
 
     valid_methods = []
-    for device in devices:
-        for method in methods:
-            if method not in valid_methods:
-                qobj = assemble(
-                    dummy_circ, optimization_level=0, shots=1, method=method, device=device
-                )
-                result = cpp_execute_qobj(controller, qobj)
-                if result.get("success", False):
-                    valid_methods.append(method)
+    for method in methods:
+        if method == "tensor_network":
+            if "GPU" in devices:
+                valid_methods.append(method)
+        else:
+            valid_methods.append(method)
     return tuple(valid_methods)
 
 
-def available_devices(controller, devices):
-    """Check available simulation devices by running a dummy circuit."""
-    # Test methods are available using the controller
-    dummy_circ = QuantumCircuit(1)
-    dummy_circ.id(0)
-
-    valid_devices = []
-    for device in devices:
-        qobj = assemble(
-            dummy_circ, optimization_level=0, shots=1, method="statevector", device=device
-        )
-        result = cpp_execute_qobj(controller, qobj)
-        if result.get("success", False):
-            valid_devices.append(device)
-    return tuple(valid_devices)
+def available_devices(controller):
+    """return available simulation devices"""
+    dev = controller.available_devices()
+    return tuple(dev)
 
 
 def add_final_save_instruction(qobj, state):

qiskit_aer/backends/qasm_simulator.py

@@ -444,7 +444,6 @@ def __init__(self, configuration=None, properties=None, provider=None, **backend
         # Update available methods for class
         if QasmSimulator._AVAILABLE_METHODS is None:
             QasmSimulator._AVAILABLE_METHODS = available_methods(
-                self._controller,
                 QasmSimulator._SIMULATION_METHODS,
                 QasmSimulator._SIMULATION_DEVICES,
             )

qiskit_aer/backends/statevector_simulator.py

@@ -254,9 +254,7 @@ def __init__(self, configuration=None, properties=None, provider=None, **backend
         self._controller = aer_controller_execute()
 
         if StatevectorSimulator._AVAILABLE_DEVICES is None:
-            StatevectorSimulator._AVAILABLE_DEVICES = available_devices(
-                self._controller, StatevectorSimulator._SIMULATION_DEVICES
-            )
+            StatevectorSimulator._AVAILABLE_DEVICES = available_devices(self._controller)
 
         if configuration is None:
             configuration = QasmBackendConfiguration.from_dict(

qiskit_aer/backends/unitary_simulator.py

@@ -240,9 +240,7 @@ def __init__(self, configuration=None, properties=None, provider=None, **backend
         self._controller = aer_controller_execute()
 
         if UnitarySimulator._AVAILABLE_DEVICES is None:
-            UnitarySimulator._AVAILABLE_DEVICES = available_devices(
-                self._controller, UnitarySimulator._SIMULATION_DEVICES
-            )
+            UnitarySimulator._AVAILABLE_DEVICES = available_devices(self._controller)
 
         if configuration is None:
             configuration = QasmBackendConfiguration.from_dict(

qiskit_aer/backends/wrappers/aer_controller_binding.hpp

@@ -56,6 +56,11 @@ class ControllerExecutor {
     return AerToPy::to_python(
         controller_execute<T>(circuits, noise_model, config));
   }
+
+  py::object available_devices() {
+    T controller;
+    return AerToPy::to_python(controller.available_devices());
+  }
 };
 
 template <typename T>
@@ -100,6 +105,11 @@ void bind_aer_controller(MODULE m) {
                  return self.execute(circuits, noise_model_native, config);
                });
 
+  aer_ctrl.def("available_devices",
+               [aer_ctrl](ControllerExecutor<Controller> &self) {
+                 return self.available_devices();
+               });
+
   py::class_<Config> aer_config(m, "AerConfig");
   aer_config.def(py::init());
   aer_config.def_readwrite("shots", &Config::shots);

releasenotes/notes/fix_simulator_init_gpu-597afb3d5241ed66.yaml

@@ -0,0 +1,8 @@
+---
+fixes:
+  - |
+    This fix changes `device` query method from running simple circuit to
+    search devices from C++ binary to prevent initializing GPUs at
+    initialization phase and simulation methods are listed in Python code.
+    Aer built with GPU support will not initialize when `device=CPU` is used.
+    And only initialize and access GPUs defined in `target_gpus` option.

src/controllers/aer_controller.hpp

@@ -74,7 +74,7 @@ namespace AER {
 
 class Controller {
 public:
-  Controller() { clear_parallelization(); }
+  Controller() {}
 
   //-----------------------------------------------------------------------
   // Execute qobj
@@ -96,8 +96,8 @@ class Controller {
   // config settings will be passed to the State and Data classes
   void set_config(const Config &config);
 
-  // Clear the current config
-  void clear_config();
+  // return available devicess
+  std::vector<std::string> available_devices();
 
 protected:
   //-----------------------------------------------------------------------
@@ -162,9 +162,6 @@ class Controller {
   // Parallelization Config
   //-----------------------------------------------------------------------
 
-  // Set OpenMP thread settings to default values
-  void clear_parallelization();
-
   // Set parallelization for experiments
   void set_parallelization_experiments(const reg_t &required_memory_list);
 
@@ -175,18 +172,18 @@ class Controller {
   size_t get_gpu_memory_mb();
 
   // The maximum number of threads to use for various levels of parallelization
-  int max_parallel_threads_;
+  int max_parallel_threads_ = 0;
 
   // Parameters for parallelization management in configuration
-  int max_parallel_experiments_;
-  size_t max_memory_mb_;
-  size_t max_gpu_memory_mb_;
+  int max_parallel_experiments_ = 1;
+  size_t max_memory_mb_ = 0;
+  size_t max_gpu_memory_mb_ = 0;
 
   // use explicit parallelization
-  bool explicit_parallelization_;
+  bool explicit_parallelization_ = false;
 
   // Parameters for parallelization management for experiments
-  int parallel_experiments_;
+  int parallel_experiments_ = 1;
 
   bool parallel_nested_ = false;
 
@@ -197,6 +194,8 @@ class Controller {
 
   // runtime parameter binding
   bool runtime_parameter_bind_ = false;
+
+  reg_t target_gpus_; // GPUs to be used
 };
 
 //=========================================================================
@@ -231,6 +230,8 @@ void Controller::set_config(const Config &config) {
 
   if (config.max_memory_mb.has_value())
     max_memory_mb_ = config.max_memory_mb.value();
+  else
+    max_memory_mb_ = get_system_memory_mb();
 
   // for debugging
   if (config._parallel_experiments.has_value()) {
@@ -307,7 +308,21 @@ void Controller::set_config(const Config &config) {
       cudaGetLastError();
       throw std::runtime_error("No CUDA device available!");
     }
+    if (config.target_gpus.has_value()) {
+      target_gpus_ = config.target_gpus.value();
+
+      if (nDev < target_gpus_.size()) {
+        throw std::invalid_argument(
+            "target_gpus has more GPUs than available.");
+      }
+    } else {
+      target_gpus_.resize(nDev);
+      for (int_t i = 0; i < nDev; i++)
+        target_gpus_[i] = i;
+    }
     sim_device_ = Device::GPU;
+
+    max_gpu_memory_mb_ = get_gpu_memory_mb();
 #endif
   } else {
     throw std::runtime_error(std::string("Invalid simulation device (\"") +
@@ -338,27 +353,6 @@ void Controller::set_config(const Config &config) {
     runtime_parameter_bind_ = config.runtime_parameter_bind_enable.value();
 }
 
-void Controller::clear_config() {
-  clear_parallelization();
-  method_ = Method::automatic;
-  sim_device_ = Device::CPU;
-  sim_precision_ = Precision::Double;
-}
-
-void Controller::clear_parallelization() {
-  max_parallel_threads_ = 0;
-  max_parallel_experiments_ = 1;
-
-  parallel_experiments_ = 1;
-  parallel_nested_ = false;
-
-  num_process_per_experiment_ = 1;
-
-  explicit_parallelization_ = false;
-  max_memory_mb_ = get_system_memory_mb();
-  max_gpu_memory_mb_ = get_gpu_memory_mb();
-}
-
 void Controller::set_parallelization_experiments(
     const reg_t &required_memory_mb_list) {
 
@@ -420,14 +414,9 @@ size_t Controller::get_system_memory_mb() {
 size_t Controller::get_gpu_memory_mb() {
   size_t total_physical_memory = 0;
 #ifdef AER_THRUST_GPU
-  int iDev, nDev, j;
-  if (cudaGetDeviceCount(&nDev) != cudaSuccess) {
-    cudaGetLastError();
-    nDev = 0;
-  }
-  for (iDev = 0; iDev < nDev; iDev++) {
+  for (int_t iDev = 0; iDev < target_gpus_.size(); iDev++) {
     size_t freeMem, totalMem;
-    cudaSetDevice(iDev);
+    cudaSetDevice(target_gpus_[iDev]);
     cudaMemGetInfo(&freeMem, &totalMem);
     total_physical_memory += totalMem;
   }
@@ -444,6 +433,20 @@ size_t Controller::get_gpu_memory_mb() {
   return total_physical_memory >> 20;
 }
 
+std::vector<std::string> Controller::available_devices() {
+  std::vector<std::string> ret;
+
+  ret.push_back(std::string("CPU"));
+#ifdef AER_THRUST_GPU
+  ret.push_back(std::string("GPU"));
+#else
+#ifdef AER_THRUST_CPU
+  ret.push_back(std::string("Thrust"));
+#endif
+#endif
+  return ret;
+}
+
 //-------------------------------------------------------------------------
 // Qobj execution
 //-------------------------------------------------------------------------

src/simulators/circuit_executor.hpp

@@ -315,9 +315,11 @@ void Executor<state_t>::set_config(const Config &config) {
   // set target GPUs
 #ifdef AER_THRUST_GPU
   int nDev = 0;
-  if (cudaGetDeviceCount(&nDev) != cudaSuccess) {
-    cudaGetLastError();
-    nDev = 0;
+  if (sim_device_ == Device::GPU) {
+    if (cudaGetDeviceCount(&nDev) != cudaSuccess) {
+      cudaGetLastError();
+      nDev = 0;
+    }
   }
   if (config.target_gpus.has_value()) {
     target_gpus_ = config.target_gpus.value();
@@ -457,7 +459,8 @@ void Executor<state_t>::set_parallelization(const Config &config,
 
   if (max_memory_mb_ == 0)
     max_memory_mb_ = get_system_memory_mb();
-  max_gpu_memory_mb_ = get_gpu_memory_mb();
+  if (sim_device_ == Device::GPU && num_gpus_ > 0)
+    max_gpu_memory_mb_ = get_gpu_memory_mb();
 
   // number of threads for parallel loop of experiments
   parallel_experiments_ = omp_get_num_threads();