fixed cmd arguement

Author: masahi

python/tvm/contrib/cutlass/gen_conv2d.py

@@ -157,10 +157,10 @@ def profile(
         self,
         d_shape,
         w_shape,
+        out_shape,
         padding,
         stride,
         dilation,
-        out_shape,
         out_dtype,
         profile_all=True,
         use_multiprocessing=False,
@@ -188,13 +188,13 @@ def profile(
                 "--k=%d" % OC,
                 "--c=%d" % IC,
                 "--r=%d" % R,
-                "--s=%d" % IC,
+                "--s=%d" % S,
                 "--pad_h=%d" % padding[0],
-                "--pad_w=%d," % padding[1],
+                "--pad_w=%d" % padding[1],
                 "--stride_h=%d" % stride[0],
-                "--stride_w=%d" % stride[0],
+                "--stride_w=%d" % stride[1],
                 "--dilation_h=%d" % dilation[0],
-                "--dilation_w=%d" % dilation[0],
+                "--dilation_w=%d" % dilation[1],
             ]
             for op in ops:
                 out = self.engine.evaluate(op, args)
@@ -224,6 +224,4 @@ def profile(
             alignment = gemm_profile_result["alignment"]
             data_type = gemm_profile_result["data_type"]
 
-            out = create_conv2d_operator([tile_description], data_type, [alignment])[0]
-            # print(out["src"])
-            return out
+            return create_conv2d_operator([tile_description], data_type, [alignment])[0]

python/tvm/contrib/cutlass/gen_tensor_op.py

@@ -190,7 +190,6 @@ def _compile(self, op):
         fi.write(op["src"])
         fi.close()
         cmd = self.cmd.format(cflags=self.cflags, src=fi.name, output=opath)
-        print(cmd)
         os.system(cmd)
         os.unlink(fi.name)
 
@@ -213,7 +212,6 @@ def evaluate(self, op, args):
         for arg in args:
             cmd.append(str(arg))
         try:
-            print("".join(cmd))
             sp = subprocess.run(cmd, capture_output=True, check=True)
             rt = float(sp.stdout)
             logger.info("%s, %f", op_name, rt)