Improved sample 18

Bernhard Kerbl · Bernhard Kerbl · commit c4325e6700ac · 2021-05-03T02:20:42.000+02:00
diff --git a/18_StandardLibrary/src/main.cu b/18_StandardLibrary/src/main.cu
@@ -5,7 +5,7 @@
 #include <random>
 #include <cuda/std/atomic>
 
-__host__ __device__ void reduceBasic(int tId, int numThreads, int N, const int* input, cuda::std::atomic<int>* result)
+__host__ __device__ void reduceAtomic(int tId, int numThreads, int N, const int* input, cuda::std::atomic<int>* result)
 {
 	if (tId >= N)
 		return;
@@ -15,16 +15,23 @@ __host__ __device__ void reduceBasic(int tId, int numThreads, int N, const int*
 	int myEnd = (tId == numThreads - 1) ? N : myStart + perThread;
 
 	for (int i = myStart; i < myEnd; i++)
-	{
-		int val = input[i];
-		result->fetch_add(val);
-	}
+		result->fetch_add(input[i]);
 }
 
-__global__ void launchReduceBasic(int N, const int* input, cuda::std::atomic<int>* result)
+__global__ void launchReductionGPU(int N, const int* input, cuda::std::atomic<int>* result)
 {
 	int tId = blockIdx.x * blockDim.x + threadIdx.x;
-	reduceBasic(tId, N, N, input, result);
+	reduceAtomic(tId, N, N, input, result);
+}
+
+template<unsigned int NUM_THREADS>
+__host__ void launchReductionCPU(int N, int* mNumbers, cuda::std::atomic<int>* result)
+{
+	std::vector<std::thread> threads(NUM_THREADS);
+	for (int i = 0; i < threads.size(); i++)
+		threads[i] = std::thread(reduceAtomic, i, NUM_THREADS, N, mNumbers, result);
+	for (std::thread& t : threads)
+		t.join();
 }
 
 int main()
@@ -34,35 +41,24 @@ int main()
 	of the standard in C++20. Soon!
 	*/
 	constexpr int N = 1<<16;
-	constexpr int cpuNumThreads = 4;
-	constexpr int gpuBlockSize = 256;
-
-	int* mNumbers;
-	cudaMallocManaged((void**)&mNumbers, sizeof(int) * N);
 
 	std::default_random_engine eng(42);
 	std::uniform_int_distribution<int> dist(10, 42);
 
-	cuda::std::atomic<int>* mResults;
-	cudaMallocManaged((void**)&mResults, 2 * sizeof(cuda::std::atomic<int>));
-	new (mResults) cuda::std::atomic<int>(0);
-	new (mResults+1) cuda::std::atomic<int>(0);
-
+	int* mNumbers;
+	cudaMallocManaged((void**)&mNumbers, sizeof(int) * N);
 	for (int i = 0; i < N; i++)
 		mNumbers[i] = dist(eng);
 
-	std::vector<std::thread> threads(cpuNumThreads);
-	for (int i = 0; i < threads.size(); i++)
-		threads[i] = std::thread(reduceBasic, i, cpuNumThreads, N, mNumbers, &mResults[0]);
-
-	for (std::thread& t : threads)
-		t.join();
-
-	std::cout << "Reduction result CPU: " << mResults[0] << "\n" << std::endl;
+	cuda::std::atomic<int>* mResults;
+	cudaMallocManaged((void**)&mResults, 2 * sizeof(cuda::std::atomic<int>));
+	mResults[0] = mResults[1] = 0;
 
-	launchReduceBasic<<<(N + gpuBlockSize - 1) / gpuBlockSize, gpuBlockSize>>>(N, mNumbers, &mResults[1]);
+	launchReductionCPU<4>(N, mNumbers, &mResults[0]);
+	launchReductionGPU<<<(N + 255) / 256, 256>>>(N, mNumbers, &mResults[1]);
 	cudaDeviceSynchronize();
 
+	std::cout << "Reduction result CPU: " << mResults[0] << "\n" << std::endl;
 	std::cout << "Reduction result GPU: " << mResults[1] << "\n" << std::endl;
 
 	cudaFree(mNumbers);
