Limit max block size in ReduceInnerSmall kernel

NVIDIA · Oct 25, 2022 · 8bd64cc · 8bd64cc
1 parent c4bbdc5
commit 8bd64cc
Showing 1 changed file with 3 additions and 1 deletion.
diff --git a/dali/kernels/reduce/reduce_gpu_impl.cuh b/dali/kernels/reduce/reduce_gpu_impl.cuh
@@ -1202,12 +1202,14 @@ class ReduceImplGPU {
 
     auto launch_params = [&](auto kernel, int nsamples, int shm_size) {
       int preferred_block_size = 256;
+      int max_block_size = 256;  // yields better results compared to larger blocks
       int preferred_grid_size;  // unused
       CUDA_CALL(cudaOccupancyMaxPotentialBlockSize(
         &preferred_grid_size,
         &preferred_block_size,
         kernel,
-        shm_size));
+        shm_size,
+        max_block_size));
 
       dim3 block(32, preferred_block_size / 32);
       int gridx = std::max(32, 512/nsamples);