dicp support dipu native memory format

dicp/dicp/vendor/AscendGraph/codegen/ascend.py

@@ -77,14 +77,11 @@ def placeholder(self, name, target, args, kwargs):
         self.input_args.append(self.cur_node)
 
         fake_tensor = self.cur_node.meta['val']
-
-        format = "NCHW"
         index = -1
 
         if isinstance(fake_tensor, torch.SymInt):
             dims = [1]
             data_type = "INT32"
-            format = "ND"
             self.sym_to_inputs[fake_tensor.node.str()] = name
         elif symint_in_shape(fake_tensor.shape):
             # mention symint position in args
@@ -110,14 +107,12 @@ def placeholder(self, name, target, args, kwargs):
             dims = list(fake_tensor.shape)
             data_type = get_ascend_dtype(fake_tensor.dtype).upper()
 
-        if 'format' in self.cur_node.meta:
-            format = self.cur_node.meta['format']
         # gen data_nodes
         self.data_nodes.append({
             "op_name": self.args_dict[name],
             "op_type": "Data",
             "dims": dims,
-            "format": format,
+            "format": self.cur_node.meta['native_memory_format'],
             "data_type": data_type,
             "cpp_data_type": data_type,
             "index": index

dicp/dicp/vendor/AscendGraph/opset_convert.py

@@ -1,4 +1,5 @@
 import torch
+import torch_dipu
 from dicp.dynamo_bridge.op_transformer import BackendPatternMatcherTransformer
 from dicp.vendor.AscendGraph.ascend_op import MatMul, CastToCpu, IdentityInp
 from dicp.vendor.AscendGraph.conversion import AtenToAscendTransformer
@@ -9,16 +10,20 @@
 )
 
 
-# 该pass需要在FuseTransposeMatmul之后
 class ArgsTransDataPass:
     def transform(self, gm: torch.fx.graph_module):
         for n in gm.graph.nodes:
-            if n.op != 'call_function':
-                continue
-            if type(n.target) in [MatMul]:
-                for arg in n.args:
-                    if arg.op == 'placeholder':
-                        arg.meta['format'] = 'FRACTAL_NZ'
+            if n.op == 'placeholder':
+                fake_tensor = n.meta['val']
+                if not hasattr(torch_dipu, 'get_native_memory_format'):
+                    n.meta['native_memory_format'] = 'ND'
+                    continue
+                memo = fake_tensor.fake_mode.fake_tensor_converter.tensor_memo     
+                for key in memo:
+                    if id(memo[key].fake_device) == id(fake_tensor.fake_device):
+                        memory_format = torch_dipu.get_native_memory_format(key())
+                        n.meta['native_memory_format'] = str(memory_format.value)
+                        break
         return gm
 
 
@@ -84,5 +89,5 @@ def ascendgraph_opset_convert(
         gm = BackendPatternMatcherTransformer(
             ascend_pattern_matcher, ascend_patterns_cls_list).transform(gm)
     gm = OutputMarkPass().transform(gm)
-    # gm = ArgsTransDataPass().transform(gm)
+    gm = ArgsTransDataPass().transform(gm)
     return gm