frank921 / sum.cpp

#include <algorithm> #include <iostream> #include <chrono> #include <math.h> //------------------------------------------------ double Sum(double* arr, size_t N) { double result = 0.; for (size_t i = 0; i < N; ++i) { result += arr[i]; } return result; } //------------------------------------------------- __global__ void pairwiseSum(double *x, double *y, double *res, size_t N) { size_t index = blockIdx.x * blockDim.x + threadIdx.x; size_t sumIndex = index * 2; if (sumIndex < N) { res[index] = x[sumIndex] + ((sumIndex + 1) < N ? y[sumIndex + 1] : 0); } } /** * @brief Computes the sum of the array * @param arr - the pointer to the beginning of an array * @param N - the length of the array */ double SumGPU(double *arr, size_t N) { double *currentArr_d, *nextArr_d; size_t currentSize = N, nextSize; cudaMalloc(&currentArr_d, currentSize * sizeof(double)); cudaMemcpy(currentArr_d, arr, currentSize * sizeof(double), cudaMemcpyHostToDevice); while (currentSize > 20) { nextSize = (currentSize + 1) / 2; cudaMalloc(&nextArr_d, nextSize * sizeof(double)); size_t threadsPerBlock = 1024; size_t blocksPerGrid = (nextSize + threadsPerBlock - 1) / threadsPerBlock; pairwiseSum<<<blocksPerGrid, threadsPerBlock>>>(currentArr_d, currentArr_d, nextArr_d, currentSize); cudaFree(currentArr_d); currentArr_d = nextArr_d; currentSize = nextSize; } double *currentArr_h = (double *) malloc(currentSize * sizeof(double)); cudaMemcpy(currentArr_h, currentArr_d, currentSize * sizeof(double), cudaMemcpyDeviceToHost); double finalResult = 0; for (size_t i = 0; i < currentSize; i++) { finalResult += currentArr_h[i]; } free(currentArr_h); cudaFree(currentArr_d); return finalResult; } //--------------------------------------------------- int main(int argc, char* argv[]) { // setting the random seed to get the same result each time srand(42); // taking as input, which algo to run int alg_ind = std::stoi(argv[1]); // Generating data size_t N = 1 << 26; double* arr = (double*) malloc(N * sizeof(double)); for (size_t i = 0; i < N; ++i) { arr[i] = static_cast <double> (rand()) / static_cast <double> (RAND_MAX); } double result = 0.; auto start = std::chrono::steady_clock::now(); switch (alg_ind) { case 0: result = Sum(arr, N); break; case 1: result = SumGPU(arr, N); break; } auto finish = std::chrono::steady_clock::now(); auto elapsed = std::chrono::duration_cast<std::chrono::microseconds>(finish - start).count(); std::cout << "Elapsed time: " << elapsed << std::endl; std::cout << "Total result: " << result << std::endl; delete[] arr; return 0; } // total computation time on CPU: 198,664 ms // total computation time on GPU: 149,512 ms // actual computation time on GPU: 2,451 ms