Rewrite mixed_vector.h

.gitignore

@@ -1,3 +1,9 @@
+paddle/operators/check_t.save
+paddle/operators/check_tensor.ls
+paddle/operators/tensor.save
+python/paddle/v2/fluid/tests/book/image_classification_resnet.inference.model/
+python/paddle/v2/fluid/tests/book/image_classification_vgg.inference.model/
+python/paddle/v2/fluid/tests/book/label_semantic_roles.inference.model/
 *.DS_Store
 build/
 build_doc/

cmake/cuda.cmake

@@ -181,7 +181,8 @@ elseif(CMAKE_BUILD_TYPE  STREQUAL "Release")
 elseif(CMAKE_BUILD_TYPE  STREQUAL "RelWithDebInfo")
     list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_RELWITHDEBINFO})
 elseif(CMAKE_BUILD_TYPE  STREQUAL "MinSizeRel")
-    list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_MINSIZEREL})
+    # nvcc 9 does not support -Os. Use Release flags instead
+    list(APPEND CUDA_NVCC_FLAGS  ${CMAKE_CXX_FLAGS_RELEASE})
 endif()
 
 mark_as_advanced(CUDA_BUILD_CUBIN CUDA_BUILD_EMULATION CUDA_VERBOSE_BUILD)

paddle/framework/lod_tensor.h

@@ -46,29 +46,7 @@ namespace framework {
  *    0 2 4 7
  *    0 2 5 7 10 12 15 20
  */
-struct LoD : public std::vector<Vector<size_t>> {
-  using std::vector<Vector<size_t>>::vector;
-  platform::Place place() const {
-    if (this->size() == 0) {
-      // Not Initialze Yet.
-      return platform::CPUPlace();
-    } else {
-      return this->front().place();
-    }
-  }
-
-  void CopyFromCUDA() {
-    for (auto it = this->begin(); it != this->end(); ++it) {
-      it->CopyFromCUDA();
-    }
-  }
-
-  void CopyToPeer(platform::Place place) {
-    for (auto it = this->begin(); it != this->end(); ++it) {
-      it->CopyToPeer(place);
-    }
-  }
-};
+using LoD = std::vector<Vector<size_t>>;
 
 std::ostream& operator<<(std::ostream& os, const LoD& lod);
 std::ostream& operator<<(std::ostream& os, const LoDTensor& t);

paddle/framework/lod_tensor_test.cu

@@ -20,6 +20,7 @@
 #include "paddle/platform/assert.h"
 
 #include <gtest/gtest.h>
+#include <paddle/platform/place.h>
 
 __global__ void test(size_t* a, int size) {
   for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < size;
@@ -36,10 +37,9 @@ TEST(LoD, data) {
   lod.push_back(std::vector<size_t>({0, 1, 6, 8, 10, 11}));
 
   auto& v = lod[0];
-  test<<<1, 1>>>(v.cuda_data(), v.size());
+  paddle::platform::CUDAPlace gpu(0);
+  test<<<1, 1>>>(v.CUDAMutableData(gpu), v.size());
   cudaDeviceSynchronize();
-
-  v.CopyFromCUDA();
   for (size_t i = 0; i < v.size(); ++i) {
     EXPECT_EQ(v[i], i * 2);
   }
@@ -63,9 +63,8 @@ TEST(LoDTensor, LoDInGPU) {
 
   auto lod = lod_tensor.lod();
 
-  test<<<1, 8>>>(lod[0].cuda_data(), lod[0].size());
+  test<<<1, 8>>>(lod[0].CUDAMutableData(place), lod[0].size());
   cudaDeviceSynchronize();
-  lod.CopyFromCUDA();
 
   for (size_t i = 0; i < src_lod[0].size(); ++i) {
     EXPECT_EQ(lod[0].data()[i], src_lod[0].data()[i] * 2);

paddle/framework/mixed_vector_test.cu

@@ -11,62 +11,3 @@
    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
    See the License for the specific language governing permissions and
    limitations under the License. */
-#include <cuda.h>
-#include <cuda_runtime.h>
-#include "gtest/gtest.h"
-
-#include "paddle/framework/init.h"
-#include "paddle/framework/mixed_vector.h"
-
-using namespace paddle::framework;
-using namespace paddle::platform;
-using namespace paddle::memory;
-
-template <typename T>
-__global__ void test(T* data, int size) {
-  for (int i = blockIdx.x * blockDim.x + threadIdx.x; i < size;
-       i += blockDim.x * gridDim.x) {
-    data[i] *= 2;
-  }
-}
-
-TEST(Vector, Normal) {
-  // fill the device context pool.
-  InitDevices();
-
-  Vector<size_t> vec({1, 2, 3});
-  size_t* ptr = vec.data();
-  for (size_t i = 0; i < vec.size(); ++i) {
-    EXPECT_EQ(vec[i], *(ptr + i));
-  }
-
-  vec.clear();
-  vec.CopyFromCUDA();
-
-  std::vector<size_t> v = {1, 2, 3};
-  for (size_t i = 0; i < v.size(); ++i) {
-    EXPECT_EQ(v[i], vec[i]);
-  }
-}
-
-TEST(Vector, MultipleCopy) {
-  InitDevices();
-  Vector<size_t> vec({1, 2, 3});
-  CUDAPlace place(0);
-  vec.mutable_data(place);
-  auto vec2 = Vector<size_t>(vec);
-  {
-    const size_t* ptr = vec2.data(CPUPlace());
-    for (size_t i = 0; i < vec2.size(); ++i) {
-      EXPECT_EQ(*(ptr + i), vec[i]);
-    }
-  }
-  test<size_t><<<3, 3>>>(vec2.mutable_data(place), vec2.size());
-  vec2.CopyFromCUDA();
-  {
-    const size_t* ptr = vec2.data(CPUPlace());
-    for (size_t i = 0; i < vec2.size(); ++i) {
-      EXPECT_EQ(*(ptr + i), vec[i] * 2);
-    }
-  }
-}

paddle/framework/tensor.h

@@ -128,6 +128,10 @@ class Tensor {
 
   inline void set_layout(const DataLayout layout) { layout_ = layout; }
 
+  size_t capacity() const {
+    return holder_ == nullptr ? 0UL : holder_->size() - offset_;
+  }
+
  private:
   friend class LoDTensor;
 

paddle/framework/tensor_impl.h

@@ -52,7 +52,7 @@ struct SizeOfTypeFunctor<HEAD, TAIL...> {
 };
 
 static inline size_t SizeOfType(std::type_index type) {
-  SizeOfTypeFunctor<int, float, double, int16_t, int64_t, bool> functor;
+  SizeOfTypeFunctor<int, float, double, int16_t, int64_t, bool, size_t> functor;
   size_t size = functor(type);
   PADDLE_ENFORCE(size != 0UL, "Cannot get size of type %s", type.name());
   return size;

paddle/operators/adagrad_op.cu

@@ -101,9 +101,9 @@ struct SparseAdagradFunctor<platform::CUDADeviceContext, T> {
     SparseAdagradFunctorKernel<
         T, 256><<<grid2, threads, 0,
                   reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                      .stream()>>>(grad_merge_data, merge_rows.cuda_data(), lr,
-                                   param_data, moment_data, grad_width,
-                                   epsilon);
+                      .stream()>>>(
+        grad_merge_data, merge_rows.CUDAMutableData(context.GetPlace()), lr,
+        param_data, moment_data, grad_width, epsilon);
   }
 };
 

paddle/operators/adam_op.h

@@ -201,7 +201,7 @@ class AdamOpKernel : public framework::OpKernel<T> {
       const T* grad_data = grad_tensor.template data<T>();
       int64_t* rows = nullptr;
       if (platform::is_gpu_place(ctx.GetPlace())) {
-        rows = grad_merge.mutable_rows()->cuda_data();
+        rows = grad_merge.mutable_rows()->CUDAMutableData(ctx.GetPlace());
       } else {
         rows = grad_merge.mutable_rows()->data();
       }

paddle/operators/ctc_align_op.cu

@@ -69,8 +69,9 @@ class CTCAlignOpCUDAKernel : public framework::OpKernel<T> {
 
     auto stream = ctx.cuda_device_context().stream();
     MergeAndDelCudaKernel<T><<<1, 1, 0, stream>>>(
-        num_tokens, tokens, num_seq, input_lod[level].cuda_data(), blank,
-        merge_repeated, dev_out_lod0_ptr, output_data);
+        num_tokens, tokens, num_seq,
+        input_lod[level].CUDAMutableData(ctx.GetPlace()), blank, merge_repeated,
+        dev_out_lod0_ptr, output_data);
 
     // set output lod
     std::vector<size_t> host_out_lod0(dev_out_lod0.begin(), dev_out_lod0.end());

paddle/operators/lookup_table_op.cu

@@ -125,7 +125,9 @@ class LookupTableGradCUDAKernel : public framework::OpKernel<T> {
       new_rows.resize(ids_dim[0]);
       auto gpu_place = boost::get<platform::CUDAPlace>(context.GetPlace());
 
-      memory::Copy(platform::CPUPlace(), new_rows.cuda_data(), gpu_place,
+      // TODO(yuyang18): Strange code here.
+      memory::Copy(platform::CPUPlace(),
+                   new_rows.CUDAMutableData(context.GetPlace()), gpu_place,
                    ids_data, ids_dim[0] * sizeof(int64_t), stream);
 
       d_table->set_rows(new_rows);

paddle/operators/math/selected_rows_functor.cc

@@ -128,7 +128,7 @@ struct SelectedRowsAddTo<platform::CPUDeviceContext, T> {
     auto* in2_value = input2->mutable_value();
 
     // concat rows
-    in2_rows.insert(in2_rows.end(), in1_rows.begin(), in1_rows.end());
+    in2_rows.Extend(in1_rows.begin(), in1_rows.end());
 
     auto in1_place = input1.place();
     PADDLE_ENFORCE(platform::is_cpu_place(in1_place));

paddle/operators/math/selected_rows_functor.cu

@@ -126,7 +126,8 @@ struct SelectedRowsAddTensor<platform::CUDADeviceContext, T> {
     dim3 grid(1, in1_rows.size());
     SelectedRowsAddTensorKernel<
         T, block_size><<<grid, threads, 0, context.stream()>>>(
-        in1_data, in1_rows.cuda_data(), out_data, in1_row_numel);
+        in1_data, in1_rows.CUDAData(context.GetPlace()), out_data,
+        in1_row_numel);
 
     auto out_eigen = framework::EigenVector<T>::Flatten(*output);
     auto in2_eigen = framework::EigenVector<T>::Flatten(input2);
@@ -153,7 +154,7 @@ struct SelectedRowsAddTo<platform::CUDADeviceContext, T> {
     auto* in2_value = input2->mutable_value();
 
     // concat rows
-    in2_rows.insert(in2_rows.end(), in1_rows.begin(), in1_rows.end());
+    in2_rows.Extend(in1_rows.begin(), in1_rows.end());
 
     auto in1_place = input1.place();
     PADDLE_ENFORCE(platform::is_gpu_place(in1_place));
@@ -216,7 +217,8 @@ struct SelectedRowsAddToTensor<platform::CUDADeviceContext, T> {
     dim3 grid(1, in1_rows.size());
     SelectedRowsAddToTensorKernel<
         T, block_size><<<grid, threads, 0, context.stream()>>>(
-        in1_data, in1_rows.cuda_data(), in2_data, in1_row_numel);
+        in1_data, in1_rows.CUDAData(context.GetPlace()), in2_data,
+        in1_row_numel);
   }
 };
 
@@ -283,9 +285,10 @@ struct MergeAdd<platform::CUDADeviceContext, T> {
     MergeAddKernel<
         T, 256><<<grid1, threads, 0,
                   reinterpret_cast<const platform::CUDADeviceContext&>(context)
-                      .stream()>>>(input_data, input_rows.cuda_data(), out_data,
-                                   out.mutable_rows()->cuda_data(),
-                                   out.rows().size(), input_width);
+                      .stream()>>>(
+        input_data, input_rows.CUDAData(context.GetPlace()), out_data,
+        out.mutable_rows()->CUDAMutableData(context.GetPlace()),
+        out.rows().size(), input_width);
     return out;
   }
 };

paddle/operators/math/sequence2batch.cu

@@ -45,7 +45,6 @@ class CopyMatrixRowsFunctor<platform::CUDADeviceContext, T> {
                   const framework::Tensor& src,
                   framework::Vector<size_t> index_lod, framework::Tensor& dst,
                   bool is_src_index) {
-    size_t* index = index_lod.cuda_data();
     auto src_dims = src.dims();
     auto dst_dims = dst.dims();
     PADDLE_ENFORCE_EQ(src_dims.size(), 2,
@@ -63,7 +62,8 @@ class CopyMatrixRowsFunctor<platform::CUDADeviceContext, T> {
     dim3 grid(8, 1);
     auto stream = context.stream();
     CopyMatrixRowsKernel<T, 128, 8, 8><<<grid, threads, 0, stream>>>(
-        src_data, dst_data, index, height, width, is_src_index);
+        src_data, dst_data, index_lod.CUDAData(context.GetPlace()), height,
+        width, is_src_index);
   }
 };
 

paddle/operators/math/sequence_padding.cu

@@ -121,12 +121,12 @@ class PaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
     if (norm_by_times) {
       SequencePaddingKernel<T, 1, 1><<<grid, threads, 0, context.stream()>>>(
           padding_data, const_cast<T*>(seq_data),
-          abs_offset_lod[level].cuda_data(), sequence_width,
+          abs_offset_lod[level].CUDAData(context.GetPlace()), sequence_width,
           max_sequence_length, num_sequences);
     } else {
       SequencePaddingKernel<T, 0, 1><<<grid, threads, 0, context.stream()>>>(
           padding_data, const_cast<T*>(seq_data),
-          abs_offset_lod[level].cuda_data(), sequence_width,
+          abs_offset_lod[level].CUDAData(context.GetPlace()), sequence_width,
           max_sequence_length, num_sequences);
     }
   }
@@ -196,12 +196,12 @@ class UnpaddingLoDTensorFunctor<platform::CUDADeviceContext, T> {
     if (norm_by_times) {
       SequencePaddingKernel<T, 1, 0><<<grid, threads, 0, context.stream()>>>(
           const_cast<T*>(padding_data), seq_data,
-          abs_offset_lod[level].cuda_data(), sequence_width,
+          abs_offset_lod[level].CUDAData(context.GetPlace()), sequence_width,
           max_sequence_length, num_sequences);
     } else {
       SequencePaddingKernel<T, 0, 0><<<grid, threads, 0, context.stream()>>>(
           const_cast<T*>(padding_data), seq_data,
-          abs_offset_lod[level].cuda_data(), sequence_width,
+          abs_offset_lod[level].CUDAData(context.GetPlace()), sequence_width,
           max_sequence_length, num_sequences);
     }
   }

paddle/operators/math/sequence_pooling.cu

@@ -73,7 +73,8 @@ class MaxSeqPoolFunctor<platform::CUDADeviceContext, T> {
     dim3 grid(num_seq, 1);
     auto stream = context.stream();
     KeMaxSequencePool<T><<<grid, threads, 0, stream>>>(
-        in_data, starts.cuda_data(), out_data, max_index, num_seq, dim);
+        in_data, starts.CUDAData(context.GetPlace()), out_data, max_index,
+        num_seq, dim);
   }
 };
 

paddle/operators/math/sequence_scale.cu

@@ -46,7 +46,8 @@ class ScaleLoDTensorFunctor<platform::CUDADeviceContext, T> {
 
     SequenceScaleKernel<T, PADDLE_CUDA_NUM_THREADS><<<
         num_seq, PADDLE_CUDA_NUM_THREADS, 0, context.stream()>>>(
-        seq_data, abs_offset_lod[level].cuda_data(), scales, seq_width);
+        seq_data, abs_offset_lod[level].CUDAMutableData(context.GetPlace()),
+        scales, seq_width);
   }
 };
 

paddle/operators/parallel_do_op.cc

@@ -79,9 +79,6 @@ inline void CopyOrShare(const framework::Variable &src,
       dst->GetMutable<LoDTensor>()->set_lod(src.Get<LoDTensor>().lod());
     } else {
       Copy(src.Get<LoDTensor>(), dst_place, dst->GetMutable<LoDTensor>());
-      framework::LoD lod(src.Get<LoDTensor>().lod());
-      lod.CopyToPeer(dst_place);
-      dst->GetMutable<LoDTensor>()->set_lod(lod);
     }
   } else if (src.IsType<SelectedRows>()) {
     auto &src_sr = src.Get<SelectedRows>();
@@ -92,9 +89,6 @@ inline void CopyOrShare(const framework::Variable &src,
       dst_sr->set_rows(src_sr.rows());
     } else {
       Copy(src_sr.value(), dst_place, dst_sr->mutable_value());
-      framework::Vector<int64_t> lod(src_sr.rows());
-      lod.CopyToPeer(dst_place);
-      dst_sr->set_rows(lod);
     }
   } else {
     PADDLE_THROW("Expect LoDTensor/SelectedRows, get %s", src.Type().name());
@@ -152,9 +146,6 @@ class ParallelDoOp : public framework::OperatorBase {
         auto *sub_scope = sub_scopes[i];
         auto *dst = sub_scope->Var(param)->GetMutable<LoDTensor>();
         framework::Copy(src, place, dst);
-        framework::LoD lod(src.lod());
-        lod.CopyToPeer(place);
-        dst->set_lod(lod);
       }
     }
     WaitOnPlaces(places);

paddle/operators/row_conv_op.cu

@@ -307,7 +307,7 @@ class RowConvKernel<platform::CUDADeviceContext, T>
     int input_dim = X->dims()[1];
     int num_sequence = batch_indices.size() - 1;
     int future_context = Filter->dims()[0];
-    size_t *idx = batch_indices.cuda_data();
+    size_t *idx = batch_indices.CUDAMutableData(context.GetPlace());
     auto stream = context.cuda_device_context().stream();
 
     if (future_context <= 32) {
@@ -345,7 +345,7 @@ class RowConvGradKernel<platform::CUDADeviceContext, T>
     int input_dim = X->dims()[1];
     int num_sequence = batch_indices.size() - 1;
     int future_context = Filter->dims()[0];
-    size_t *idx = batch_indices.cuda_data();
+    size_t *idx = batch_indices.CUDAMutableData(context.GetPlace());
 
     auto &device_ctx = context.cuda_device_context();
     math::SetConstant<platform::CUDADeviceContext, T> zero;

paddle/operators/sequence_erase_op.cu

@@ -87,8 +87,7 @@ class SequenceEraseOpCUDAKernel : public framework::OpKernel<T> {
     // Copy LoD to GPU
     auto lod0 = lod[0];
     auto lod_len = lod0.size();
-    thrust::device_vector<size_t> dev_in_lod = lod0;
-    size_t* dev_in_lod_ptr = thrust::raw_pointer_cast(dev_in_lod.data());
+    const size_t* dev_in_lod_ptr = lod0.CUDAData(ctx.GetPlace());
 
     // Calc output LoD
     thrust::device_vector<size_t> dev_out_lod(lod_len);

paddle/operators/sgd_op.cu

@@ -102,8 +102,8 @@ class SGDOpCUDAKernel : public framework::OpKernel<T> {
       dim3 grid(1, in_rows.size());
       SparseSGDFunctorKernel<
           T, 256><<<grid, threads, 0, ctx.cuda_device_context().stream()>>>(
-          in_data, in_rows.cuda_data(), learning_rate->data<T>(), out_data,
-          in_row_numel);
+          in_data, in_rows.CUDAData(ctx.GetPlace()), learning_rate->data<T>(),
+          out_data, in_row_numel);
 
     } else {
       PADDLE_THROW("Unsupported Variable Type of Grad");

paddle/operators/target_assign_op.h

@@ -137,8 +137,8 @@ class TargetAssignKernel : public framework::OpKernel<T> {
       PADDLE_ENFORCE_EQ(gt_lod.data()[i], gt_label_lod.data()[i]);
     }
 
-    size_t* gt_lod_data = gt_lod.data(ctx.GetPlace());
-    size_t* neg_lod_data = neg_lod.data(ctx.GetPlace());
+    size_t* gt_lod_data = gt_lod.MutableData(ctx.GetPlace());
+    size_t* neg_lod_data = neg_lod.MutableData(ctx.GetPlace());
 
     TargetAssignFunctor<T> functor(box_data, label_data, match_idx_data,
                                    gt_lod_data, background_label, num,

paddle/testing/paddle_gtest_main.cc

@@ -20,6 +20,7 @@ limitations under the License. */
 #include "paddle/memory/memory.h"
 
 int main(int argc, char** argv) {
+  testing::InitGoogleTest(&argc, argv);
   std::vector<char*> new_argv;
   std::string gflags_env;
   for (int i = 0; i < argc; ++i) {
@@ -35,7 +36,6 @@ int main(int argc, char** argv) {
   int new_argc = static_cast<int>(new_argv.size());
   char** new_argv_address = new_argv.data();
   google::ParseCommandLineFlags(&new_argc, &new_argv_address, false);
-  testing::InitGoogleTest(&argc, argv);
   paddle::memory::Used(paddle::platform::CPUPlace());
 
 #ifdef PADDLE_WITH_CUDA