Make Python interface work for non-SM80 targets (#726)

* Make Python interface work for non-SM80 targets

* Remove line in README

examples/40_cutlass_py/customizable/conv2d.py

@@ -34,6 +34,7 @@ import pycutlass
 from pycutlass import *
 from pycutlass.conv2d_operation import *
 from pycutlass.utils import reference_model
+from pycutlass.utils.device import device_cc
 import sys
 import torch.nn.functional as F
 
@@ -146,6 +147,11 @@ try:
 except:
     sys.exit(0)
 
+cc = device_cc()
+if args.compute_capability != cc:
+    raise Exception(("Parameter --compute-capability of {} "
+                    "does not match that of the device of {}.").format(args.compute_capability, cc))
+
 pycutlass.get_memory_pool(init_pool_size=2**30, max_pool_size=2**32)
 
 np.random.seed(0)

examples/40_cutlass_py/customizable/gemm.py

@@ -34,6 +34,7 @@ import pycutlass
 from pycutlass import *
 import cutlass
 from bfloat16 import bfloat16
+from pycutlass.utils.device import device_cc
 import sys
 
 import argparse
@@ -131,12 +132,16 @@ parser.add_argument("-activ_arg", "--activation_args", default=[], nargs="+", ty
 parser.add_argument('--print_cuda', action="store_true",
                     help="print the underlying CUDA kernel")
 
-
 try:
     args = parser.parse_args()
 except:
     sys.exit(0)
 
+cc = device_cc()
+if args.compute_capability != cc:
+    raise Exception(("Parameter --compute-capability of {} "
+                    "does not match that of the device of {}.").format(args.compute_capability, cc))
+
 pycutlass.get_memory_pool(init_pool_size=2**30, max_pool_size=2**32)
 pycutlass.compiler.nvcc()
 

examples/40_cutlass_py/customizable/gemm_grouped.py

@@ -32,6 +32,7 @@
 import numpy as np
 import pycutlass
 from pycutlass import *
+from pycutlass.utils.device import device_cc
 import csv
 import sys
 
@@ -129,6 +130,11 @@ try:
 except:
     sys.exit(0)
 
+cc = device_cc()
+if args.compute_capability != cc:
+    raise Exception(("Parameter --compute-capability of {} "
+                    "does not match that of the device of {}.").format(args.compute_capability, cc))
+
 pycutlass.get_memory_pool(init_pool_size=2**30, max_pool_size=2**32)
 
 np.random.seed(0)