+{::nomarkdown}
+
++ The example below shows a C++ benchmark comparing the performance of sorting a large array on a CPU versus a GPU. It provides a clear visual of how parallel processing can drastically outperform traditional sequential execution for data-heavy tasks. +
+ ++ In the first cell we create the unsorted data that is going to be sorted by the CPU and GPU. We have loaded a compiled CUDA .so file beforehand. +
+
+unsigned int N_bench = 1048576;
+std::vector<unsigned int> data_cpu(N_bench);
+std::vector<unsigned int> data_gpu(N_bench);
+
+for (unsigned int i = 0; i < N_bench; i++) {
+ unsigned int val = N_bench - i;
+ data_cpu[i] = val;
+ data_gpu[i] = val;
+}
+
+ + Next we use std::sort and merge_sort_gpu_full function form the loaded CUDA code and measure the time the CPU and GPU sorts the data. +
+
+auto start_cpu = std::chrono::high_resolution_clock::now();
+std::sort(data_cpu.begin(), data_cpu.end());
+auto end_cpu = std::chrono::high_resolution_clock::now();
+
+std::chrono::duration<double, std::milli> cpu_ms = end_cpu - start_cpu;
+
+auto start_gpu = std::chrono::high_resolution_clock::now();
+merge_sort_gpu_full(data_gpu.data(), N_bench);
+auto end_gpu = std::chrono::high_resolution_clock::now();
+
+std::chrono::duration<double, std::milli> gpu_ms = end_gpu - start_gpu;
+
+ + Finally we print both times and compare them to see how much faster parallel processing is. +
+
+double speedup = cpu_ms.count() / gpu_ms.count();
+
+std::cout << "CPU (std::sort) took: " << std::fixed << std::setprecision(4) << cpu_ms.count() << " ms" << std::endl;
+std::cout << "GPU (Merge Sort) took: " << gpu_ms.count() << " ms" << std::endl;
+
+std::cout << std::endl;
+
+std::cout << "GPU Speedup: " << speedup << " times faster than CPU" << std::endl;
+
+