New CUDA simulator.

jakurzak · Aug 10, 2022 · 2f475ae · 2f475ae
1 parent a1e53ff
commit 2f475ae
Show file tree

Hide file tree

Showing 16 changed files with 1,292 additions and 11,456 deletions.
diff --git a/apps/qsim_base_cuda.cu b/apps/qsim_base_cuda.cu
@@ -112,46 +112,42 @@ int main(int argc, char* argv[]) {
     return 1;
   }
 
-  Circuit<GateQSim<float>> circuit;
+  using fp_type = float;
+
+  Circuit<GateQSim<fp_type>> circuit;
   if (!CircuitQsimParser<IOFile>::FromFile(opt.maxtime, opt.circuit_file,
                                            circuit)) {
     return 1;
   }
 
   struct Factory {
-    using Simulator = qsim::SimulatorCUDA<float>;
+    using Simulator = qsim::SimulatorCUDA<fp_type>;
     using StateSpace = Simulator::StateSpace;
 
-    Factory(const StateSpace::Parameter& param1,
-            const Simulator::Parameter& param2)
-        : param1(param1), param2(param2) {}
+    Factory(const StateSpace::Parameter& param) : param(param) {}
 
     StateSpace CreateStateSpace() const {
-      return StateSpace(param1);
+      return StateSpace(param);
     }
 
     Simulator CreateSimulator() const {
-      return Simulator(param2);
+      return Simulator();
     }
 
-    const StateSpace::Parameter& param1;
-    const Simulator::Parameter& param2;
+    const StateSpace::Parameter& param;
   };
 
   using Simulator = Factory::Simulator;
   using StateSpace = Simulator::StateSpace;
   using State = StateSpace::State;
-  using Fuser = MultiQubitGateFuser<IO, GateQSim<float>>;
+  using Fuser = MultiQubitGateFuser<IO, GateQSim<fp_type>>;
   using Runner = QSimRunner<IO, Fuser, Factory>;
 
   StateSpace::Parameter param1;
   param1.num_threads = opt.num_threads;
   param1.num_dblocks = opt.num_dblocks;
 
-  Simulator::Parameter param2;
-  param2.num_threads = opt.num_threads;
-
-  Factory factory(param1, param2);
+  Factory factory(param1);
 
   StateSpace state_space = factory.CreateStateSpace();
   State state = state_space.Create(circuit.num_qubits);

diff --git a/apps/qsim_qtrajectory_cuda.cu b/apps/qsim_qtrajectory_cuda.cu
@@ -190,23 +190,20 @@ int main(int argc, char* argv[]) {
   using fp_type = float;
 
   struct Factory {
-    using Simulator = qsim::SimulatorCUDA<fp_type>;
+    using Simulator = qsim::SimulatorCUDA<float>;
     using StateSpace = Simulator::StateSpace;
 
-    Factory(const StateSpace::Parameter& param1,
-            const Simulator::Parameter& param2)
-        : param1(param1), param2(param2) {}
+    Factory(const StateSpace::Parameter& param) : param(param) {}
 
     StateSpace CreateStateSpace() const {
-      return StateSpace(param1);
+      return StateSpace(param);
     }
 
     Simulator CreateSimulator() const {
-      return Simulator(param2);
+      return Simulator();
     }
 
-    const StateSpace::Parameter& param1;
-    const Simulator::Parameter& param2;
+    const StateSpace::Parameter& param;
   };
 
   using Simulator = Factory::Simulator;
@@ -235,8 +232,7 @@ int main(int argc, char* argv[]) {
   }
 
   StateSpace::Parameter param1;
-  Simulator::Parameter param2;
-  Factory factory(param1, param2);
+  Factory factory(param1);
 
   Simulator simulator = factory.CreateSimulator();
   StateSpace state_space = factory.CreateStateSpace();

diff --git a/docs/cirq_interface.md b/docs/cirq_interface.md
@@ -190,11 +190,9 @@ is required to enable GPU execution:
 library if set to any other value.
 
 If `use_gpu` is set and `gpu_mode` is set to 0, the remaining parameters can
-optionally be set to fine-tune perfomance for a specific device or circuit.
+optionally be set to fine-tune StateSpace perfomance for a specific device.
 In most cases, the default values provide good performance.
-* `gpu_sim_threads`: number of threads per CUDA block to use for the GPU
-Simulator. This must be a power of 2 in the range [32, 256].
 * `gpu_state_threads`: number of threads per CUDA block to use for the GPU
 StateSpace. This must be a power of 2 in the range [32, 1024].
-* `gpu_data_blocks`: number of data blocks to use on GPU. Below 16 data blocks,
-performance is noticeably reduced.
+* `gpu_data_blocks`: number of data blocks to use for the GPU StateSpace.
+Below 16 data blocks, performance is noticeably reduced.
diff --git a/lib/run_qsim.h b/lib/run_qsim.h
@@ -133,6 +133,7 @@ struct QSimRunner final {
       }
 
       if (param.verbosity > 3) {
+        state_space.DeviceSync();
         double t2 = GetTime();
         IO::messagef("gate %lu done in %g seconds.\n", i, t2 - t1);
       }
@@ -147,6 +148,7 @@ struct QSimRunner final {
     }
 
     if (param.verbosity > 0) {
+      state_space.DeviceSync();
       double t2 = GetTime();
       IO::messagef("time is %g seconds.\n", t2 - t0);
     }
@@ -221,12 +223,14 @@ struct QSimRunner final {
       }
 
       if (param.verbosity > 3) {
+        state_space.DeviceSync();
         double t2 = GetTime();
         IO::messagef("gate %lu done in %g seconds.\n", i, t2 - t1);
       }
     }
 
     if (param.verbosity > 0) {
+      state_space.DeviceSync();
       double t2 = GetTime();
       IO::messagef("simu time is %g seconds.\n", t2 - t0);
     }