Merge pull request #48 from oneapi-src/quicksilver/update_block_size

Anudeep Kambapu · web-flow · commit 396e7e9b9551 · 2024-02-06T10:20:00.000-08:00
Quicksilver - Update Block sizes
diff --git a/QuickSilver/CUDA/src/main.cc b/QuickSilver/CUDA/src/main.cc
@@ -311,7 +311,7 @@ void cycleTracking(MonteCarlo *monteCarlo, uint64_cu *tallies, uint64_cu *tallie
 #if defined(HAVE_CUDA)
 
                         const size_t N = numParticles;
-                        unsigned int wg_size = 16;
+                        unsigned int wg_size = 256;
                         unsigned int num_wgs = (N + wg_size - 1) / wg_size;
                         CycleTrackingKernel<<<num_wgs, wg_size, NUM_TALLIES * replications * sizeof(int), 0>>>(monteCarlo, numParticles, processingVault, processedVault, tallies_d);
                         checkMsg("CycleTrackingKernel, execution failed\n");
diff --git a/QuickSilver/SYCL/src/main.cc.dp.cpp b/QuickSilver/SYCL/src/main.cc.dp.cpp
@@ -360,7 +360,7 @@ void cycleTracking(MonteCarlo *monteCarlo, uint64_cu *tallies, uint64_cu *tallie
 #if defined(HAVE_SYCL)
 
                         const size_t N = numParticles;
-                        unsigned int wg_size = 16;
+                        unsigned int wg_size = 256;
                         unsigned int num_wgs = (N + wg_size - 1) / wg_size;
 
                         sycl_device_queue.submit([&](sycl::handler &cgh)
