Slight optimize the default injective schedule (apache#7158)

python/tvm/topi/cuda/injective.py

@@ -57,6 +57,7 @@ def schedule_injective_from_existing(sch, out):
         need_block_split = const_size > max_block * num_thread * vector_width
     except ValueError:
         need_block_split = False
+        const_size = 0
 
     if vector_width > 1:
         fused, v = sch[out].split(fused, vector_width)
@@ -72,7 +73,10 @@ def schedule_injective_from_existing(sch, out):
         # Use less threads for dynamic shape ops to avoid runtime error.
         if is_dynamic_output:
             num_thread //= 2
-        bx, tx = sch[out].split(fused, factor=num_thread)
+        if const_size != 0 and const_size < num_thread:
+            bx, tx = sch[out].split(fused, factor=const_size)
+        else:
+            bx, tx = sch[out].split(fused, factor=num_thread)
         sch[out].bind(tx, te.thread_axis("threadIdx.x"))
         sch[out].bind(bx, te.thread_axis("blockIdx.x"))