Fix DALI paddle plugin stream synchronization error (NVIDIA#3487)

Signed-off-by: sneaxiy <[email protected]>

dali/python/nvidia/dali/plugin/paddle.py

@@ -17,6 +17,7 @@
 import math
 
 import numpy as np
+import paddle
 
 from nvidia.dali import types
 from nvidia.dali.backend import TensorListCPU, TensorGPU, TensorListGPU
@@ -72,6 +73,8 @@ def feed_ndarray(dali_tensor, ptr, cuda_stream=None):
                     (if not provided, an internal user stream will be selected)
     """
 
+    if cuda_stream is None:
+        cuda_stream = paddle.device.cuda.current_stream().cuda_stream 
     cuda_stream = types._raw_cuda_stream(cuda_stream)
 
     c_type_pointer = ctypes.c_void_p(ptr)
@@ -315,12 +318,6 @@ def __next__(self):
                                          category_pd_type[cat])
             data_batches[i] = pd_tensors
 
-            # due to https://github.com/PaddlePaddle/Paddle/issues/35555 the tensors are allocated
-            # first by setting shape and calling _mutable_data, then the device is synchronized
-            # and after that, another call to _mutable_data is made to obtain the pointer and
-            # copy data from the pipeline to the tensor
-            # when the issue is solved the copy can be moved to the preceding loop
-            fluid.core._cuda_synchronize(pd_gpu_place)
             for cat, tensor in category_tensors.items():
                 ptr = pd_tensors[cat]._mutable_data(category_place[cat],
                                                category_pd_type[cat])

qa/setup_packages.py

@@ -448,8 +448,8 @@ def get_pyvers_name(self, url, cuda_version):
                         { "101" : ["0.9.0"],
                           "111" : ["0.9.0"] }, extra_index="https://download.pytorch.org/whl/cu{cuda_v}/"),
                 CudaPackage("paddlepaddle-gpu",
-                        { "100" : ["2.0.2"],
-                          "110" : ["2.0.2"]})
+                        { "100" : ["2.2.0"],
+                          "110" : ["2.2.0"]})
                ]
 
 all_packages_keys = [pckg.key for pckg in all_packages]