Merge pull request #15639 from PaddlePaddle/revert-15296-async_double_buffered_py_reader

Revert "Async double buffered py reader"

paddle/fluid/operators/reader/buffered_reader.cc

@@ -14,7 +14,6 @@
 
 #include "paddle/fluid/operators/reader/buffered_reader.h"
 #include <vector>
-#include "paddle/fluid/framework/data_type.h"
 
 namespace paddle {
 namespace operators {
@@ -25,12 +24,6 @@ BufferedReader::~BufferedReader() {
     position_.front().wait();
     position_.pop();
   }
-#ifdef PADDLE_WITH_CUDA
-  if (platform::is_gpu_place(place_)) {
-    platform::SetDeviceId(boost::get<platform::CUDAPlace>(place_).device);
-    PADDLE_ENFORCE(cudaStreamDestroy(stream));
-  }
-#endif
 }
 
 BufferedReader::BufferedReader(
@@ -40,12 +33,6 @@ BufferedReader::BufferedReader(
       thread_pool_(1),
       place_(place),
       buffer_size_(buffer_size) {
-#ifdef PADDLE_WITH_CUDA
-  if (platform::is_gpu_place(place_)) {
-    platform::SetDeviceId(boost::get<platform::CUDAPlace>(place_).device);
-    PADDLE_ENFORCE(cudaStreamCreate(&stream));
-  }
-#endif
   cpu_buffer_.resize(buffer_size);
   gpu_buffer_.resize(buffer_size);
   ReadTillBufferFullAsync();
@@ -67,39 +54,14 @@ void BufferedReader::ReadAsync(size_t i) {
       return -1UL;
     }
 
-#ifdef PADDLE_WITH_CUDA
-    // NOTE(liangdun): using async copy instead of TensorCopySync
-    // TensorCopySync would block other stream
     if (platform::is_gpu_place(place_)) {
       TensorVec &gpu = gpu_buffer_[i];
       gpu.resize(cpu.size());
       for (size_t i = 0; i < cpu.size(); ++i) {
-        gpu[i].Resize(cpu[i].dims());
-        gpu[i].set_layout(cpu[i].layout());
-        auto cpu_place = cpu[i].place();
-        auto cpu_ptr = cpu[i].data<void>();
-        auto gpu_ptr = gpu[i].mutable_data(place_, cpu[i].type());
-        auto size =
-            cpu[i].numel() * paddle::framework::SizeOfType(cpu[i].type());
-        if (platform::is_cuda_pinned_place(cpu_place))
-          memory::Copy(boost::get<platform::CUDAPlace>(place_), gpu_ptr,
-                       boost::get<platform::CUDAPinnedPlace>(cpu_place),
-                       cpu_ptr, size, stream);
-        else if ((platform::is_gpu_place(cpu_place)))
-          memory::Copy(boost::get<platform::CUDAPlace>(place_), gpu_ptr,
-                       boost::get<platform::CUDAPlace>(cpu_place), cpu_ptr,
-                       size, stream);
-        else
-          // if cpu place is not pinned, async copy is slower than sync copy,
-          // so we use sync copy instead.
-          memory::Copy(boost::get<platform::CUDAPlace>(place_), gpu_ptr,
-                       boost::get<platform::CPUPlace>(cpu_place), cpu_ptr, size,
-                       0);
+        framework::TensorCopySync(cpu[i], place_, &gpu[i]);
         gpu[i].set_lod(cpu[i].lod());
       }
-      PADDLE_ENFORCE(cudaStreamSynchronize(stream));
     }
-#endif
     return i;
   }));
 }

paddle/fluid/operators/reader/buffered_reader.h

@@ -19,9 +19,6 @@
 #include <vector>
 #include "ThreadPool.h"
 #include "paddle/fluid/framework/reader.h"
-#ifdef PADDLE_WITH_CUDA
-#include "paddle/fluid/platform/gpu_info.h"
-#endif
 
 namespace paddle {
 namespace operators {
@@ -62,9 +59,6 @@ class BufferedReader : public framework::DecoratedReader {
   std::vector<TensorVec> cpu_buffer_;
   std::vector<TensorVec> gpu_buffer_;
   size_t prev_pos_{-1UL};
-#ifdef PADDLE_WITH_CUDA
-  cudaStream_t stream;
-#endif
 };
 
 }  // namespace reader

python/paddle/fluid/layers/io.py

@@ -484,7 +484,7 @@ def _py_reader(capacity,
                name=None,
                use_double_buffer=True,
                feed_list=None):
-    use_cuda_pinned_place = use_double_buffer and core.is_compiled_with_cuda()
+
     if feed_list is not None:
         if not isinstance(feed_list, list):
             raise TypeError("feed_list should be a list of Variable"
@@ -565,10 +565,7 @@ def _py_reader(capacity,
                 for item in tensors:
                     if not isinstance(item, core.LoDTensor):
                         tmp = core.LoDTensor()
-                        if use_cuda_pinned_place:
-                            tmp.set(item, core.CUDAPinnedPlace())
-                        else:
-                            tmp.set(item, core.CPUPlace())
+                        tmp.set(item, core.CPUPlace())
                         item = tmp
 
                     array.append(item)