adding synchronize for the executor along with the overhead example

Author: youwuyou

examples/ginkgo-overhead/ginkgo-overhead.jl

@@ -0,0 +1,34 @@
+using Ginkgo
+using Printf
+using SparseArrays
+
+# Type alias
+const T = Float64
+num_iters = 1000000
+
+# print ginkgo library version
+version()
+
+# Creates executor for a specific backend
+exec = create(:reference)
+
+# testing solver overhead with a 1x1 system
+A = spzeros(T, 1, 1); A[1, 1] = 1.0; A = GkoCsr(A, exec)
+b = number(NaN, exec)
+x = number(T(0.0), exec)
+
+# time solver overhead
+solver = GkoIterativeSolver(:cg, A, exec, maxiter = num_iters)
+tic = time_ns()
+
+apply!(solver, b, x)
+synchronize(exec)
+
+tac = time_ns()
+
+total_ns = tac - tic
+total_s  = total_ns / 1.0e9
+avg_ns   = total_ns / num_iters
+
+@printf("Running %d iterations of the CG solver took a total of %.5f seconds.\n", num_iters, total_s)
+@printf("\tAverage library overhead:     %.2f [nanoseconds / iteration]\n", avg_ns)

src/Ginkgo.jl

@@ -77,7 +77,8 @@ export
 
     # Executor
     create,
-    
+    synchronize,  
+
     # CPU executors only
     get_num_cores,
     get_num_threads_per_core,

src/base/Executor.jl

@@ -187,6 +187,15 @@ function create(executor_type::Symbol; device_id::Integer = 0, master::GkoCPUExe
     end
 end
 
+"""
+    synchronize(executor::GkoExecutor)
+
+Synchronize the given executor. This function blocks until all operations associated with the executor are complete.
+"""
+function synchronize(executor::GkoExecutor = EXECUTOR[])
+    API.gko_executor_synchronize(executor.ptr)
+    return nothing
+end
 
 # Getters for all CPU executors
 get_num_cores(exec::GkoCPUExecutor) = Int32(API.gko_executor_cpu_get_num_cores(exec.ptr))