Restore invalid changes

codeplaysoftware · Jul 17, 2024 · 69d5c2a · 69d5c2a
1 parent 3496593
commit 69d5c2a
Showing 1 changed file with 2 additions and 1 deletion.
diff --git a/tools/util/include/cutlass/util/reference/device/tensor_compare.h b/tools/util/include/cutlass/util/reference/device/tensor_compare.h
@@ -96,6 +96,7 @@ __global__ void
   size_t idx = ThreadIdxX() + BlockDimX() * BlockIdxX();
 
   for (; idx < capacity; idx += GridDimX() * BlockDimX()) {
+
     Element a = cutlass::ReferenceFactory<Element>::get(ptr_A, idx);
     Element b = cutlass::ReferenceFactory<Element>::get(ptr_B, idx);
 
@@ -238,7 +239,7 @@ bool BlockCompareRelativelyEqual(
 #if defined (CUTLASS_ENABLE_SYCL)
     block_size = 128;
     grid_size = (capacity + block_size - 1) / block_size;
-    //grid_size = (grid_size < 64 ? grid_size : 64); // limit grid size to avoid out_of_resources runtime error.
+    grid_size = (grid_size < 64 ? grid_size : 64); // limit grid size to avoid out_of_resources runtime error.
 #else
     // if grid_size or block_size are zero, query occupancy using the CUDA Occupancy API
     cudaError_t result = cudaOccupancyMaxPotentialBlockSize(