add klo bench

python/jittor/__init__.py

@@ -9,7 +9,7 @@
 # file 'LICENSE.txt', which is part of this source code package.
 # ***************************************************************
 
-__version__ = '1.3.5.40'
+__version__ = '1.3.5.41'
 from jittor_utils import lock
 with lock.lock_scope():
     ori_int = int

python/jittor/utils/bench_klo.py

@@ -0,0 +1,14 @@
+import jittor as jt
+
+jt.flags.use_device = 1
+n = 100000
+
+with jt.profile_scope(10, 10) as rep:
+    jt.code([2], "float32", [],
+    cuda_header='''__global__ void kernel(float* a) {}''',
+    cuda_src=f'''
+    for (int i=0; i<{n}; i++) kernel<<<1,32>>>(out0_p);
+    ''').sync()
+
+avg_ns = float(rep[1][4]) / n
+print("kernel launch overhead(ns):", avg_ns)