Adding warmup steps to coarsening.cu (#46)

Author: zinccat

lecture_008/coarsening.cu

@@ -53,6 +53,13 @@ int main()
     cudaMemcpy(d_a, a, size, cudaMemcpyHostToDevice);
     cudaMemcpy(d_b, b, size, cudaMemcpyHostToDevice);
 
+    // warmup
+    VecAdd<<<(N + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK, THREADS_PER_BLOCK>>>(d_a, d_b, d_c);
+    cudaDeviceSynchronize();
+
+    VecAddCoarsened<<<(N + 2*THREADS_PER_BLOCK - 1) / (2*THREADS_PER_BLOCK), THREADS_PER_BLOCK>>>(d_a, d_b, d_c);
+    cudaDeviceSynchronize();
+
     // Start timer for VecAdd kernel
     cudaEventRecord(start);
     VecAdd<<<(N + THREADS_PER_BLOCK - 1) / THREADS_PER_BLOCK, THREADS_PER_BLOCK>>>(d_a, d_b, d_c);