Merge pull request #11527 from jacquesqiao/concat-grad-support-data-input

jacquesqiao · web-flow · commit 69827f305b45 · 2018-06-19T12:54:17.000+08:00
concat support data as input
diff --git a/paddle/fluid/operators/concat_op.h b/paddle/fluid/operators/concat_op.h
@@ -60,34 +60,45 @@ template <typename DeviceContext, typename T>
 class ConcatGradKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const {
-    auto* in = ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
+    auto* out_grad =
+        ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
+    auto ins = ctx.MultiInput<framework::Tensor>("X");
+    auto out_var_names = ctx.Outputs(framework::GradVarName("X"));
     auto outs = ctx.MultiOutput<framework::Tensor>(framework::GradVarName("X"));
     int64_t axis = static_cast<int64_t>(ctx.Attr<int>("axis"));
 
+    // get output tensor that the name is not kEmptyVarName
+    std::vector<framework::Tensor*> outputs;
+    for (size_t j = 0; j < outs.size(); ++j) {
+      if (out_var_names[j] != framework::kEmptyVarName) {
+        outs[j]->mutable_data<T>(ctx.GetPlace());
+        outputs.push_back(outs[j]);
+      } else {
+        outputs.push_back(nullptr);
+      }
+    }
+
     // Sometimes direct copies will be faster, this maybe need deeply analysis.
     if (axis == 0 && outs.size() < 10) {
       size_t input_offset = 0;
-      auto in_stride = framework::stride_numel(in->dims());
+      const auto in_stride = framework::stride_numel(out_grad->dims());
 
-      for (auto& out : outs) {
-        out->mutable_data<T>(ctx.GetPlace());
-        auto out_stride = framework::stride_numel(out->dims());
-        StridedNumelCopyWithAxis<T>(ctx.device_context(), axis, out->data<T>(),
-                                    out_stride, in->data<T>() + input_offset,
-                                    in_stride, out_stride[axis]);
+      for (size_t i = 0; i < outs.size(); ++i) {
+        auto out_stride = framework::stride_numel(ins[i]->dims());
+        auto* out = outputs[i];
+        if (out != nullptr) {
+          StridedNumelCopyWithAxis<T>(
+              ctx.device_context(), axis, out->data<T>(), out_stride,
+              out_grad->data<T>() + input_offset, in_stride, out_stride[axis]);
+        }
         input_offset += out_stride[axis];
       }
     } else {
-      std::vector<framework::Tensor> outputs(outs.size());
-      for (size_t j = 0; j < outs.size(); ++j) {
-        outs[j]->mutable_data<T>(ctx.GetPlace());
-        outputs[j] = *outs[j];
-      }
-
       auto& dev_ctx = ctx.template device_context<DeviceContext>();
       paddle::operators::math::ConcatGradFunctor<DeviceContext, T>
           concat_grad_functor;
-      concat_grad_functor(dev_ctx, *in, static_cast<int>(axis), &outputs);
+      concat_grad_functor(dev_ctx, *out_grad, ins, static_cast<int>(axis),
+                          &outputs);
     }
   }
 };
diff --git a/paddle/fluid/operators/math/concat.cc b/paddle/fluid/operators/math/concat.cc
@@ -70,35 +70,40 @@ template <typename T>
 class ConcatGradFunctor<platform::CPUDeviceContext, T> {
  public:
   void operator()(const platform::CPUDeviceContext& context,
-                  const framework::Tensor& input, const int axis,
-                  std::vector<framework::Tensor>* outputs) {
+                  const framework::Tensor& input,
+                  const std::vector<const framework::Tensor*>& ref_inputs,
+                  const int axis, std::vector<framework::Tensor*>* outputs) {
     // TODO(zcd): Add input data validity checking
-    int num = outputs->size();
+    size_t num = outputs->size();
 
     int input_rows = 1;
-    auto dim_0 = outputs->at(0).dims();
+    auto dim_0 = ref_inputs[0]->dims();
     for (int i = 0; i < axis; ++i) {
       input_rows *= dim_0[i];
     }
+
     int input_cols = 0;
 
     std::vector<int64_t> output_cols(outputs->size());
-    for (int i = 0; i < num; ++i) {
-      int t_cols = outputs->at(i).numel() / input_rows;
+    for (size_t i = 0; i < num; ++i) {
+      int t_cols = ref_inputs[i]->numel() / input_rows;
       input_cols += t_cols;
       output_cols[i] = t_cols;
     }
     auto cpu_place = boost::get<platform::CPUPlace>(context.GetPlace());
 
     // computation
-    for (int k = 0; k < input_rows; ++k) {
+    for (size_t k = 0; k < input_rows; ++k) {
       const T* src_ptr = input.data<T>() + k * input_cols;
       int col_idx = 0;
       for (int j = 0; j < num; ++j) {
         int col_len = output_cols[j];
-        T* dst_ptr = outputs->at(j).data<T>() + k * col_len;
-        memory::Copy(cpu_place, dst_ptr, cpu_place, src_ptr + col_idx,
-                     sizeof(T) * col_len);
+        auto* out_tensor = outputs->at(j);
+        if (out_tensor != nullptr) {
+          T* dst_ptr = out_tensor->data<T>() + k * col_len;
+          memory::Copy(cpu_place, dst_ptr, cpu_place, src_ptr + col_idx,
+                       sizeof(T) * col_len);
+        }
         col_idx += col_len;
       }
     }
diff --git a/paddle/fluid/operators/math/concat.cu b/paddle/fluid/operators/math/concat.cu
@@ -102,10 +102,12 @@ __global__ void KernelConcatGrad(const T* input_data, const int in_row,
     int local_col = tid_x - curr_offset;
     int segment_width = curr_col_offset - curr_offset;
     T* output_ptr = outputs_data[curr_segment];
-    int tid_y = blockIdx.y * blockDim.y + threadIdx.y;
-    for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y)
-      output_ptr[tid_y * segment_width + local_col] =
-          input_data[tid_y * in_col + tid_x];
+    if (output_ptr != nullptr) {
+      int tid_y = blockIdx.y * blockDim.y + threadIdx.y;
+      for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y)
+        output_ptr[tid_y * segment_width + local_col] =
+            input_data[tid_y * in_col + tid_x];
+    }
   }
 }
 
@@ -118,10 +120,12 @@ __global__ void KernelConcatGrad(const T* input_data, const int in_row,
     int split = tid_x / fixed_out_col;
     int in_offset = tid_x - split * fixed_out_col;
     T* output_ptr = outputs_data[split];
-    int tid_y = blockIdx.y * blockDim.y + threadIdx.y;
-    for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y)
-      output_ptr[tid_y * fixed_out_col + in_offset] =
-          input_data[tid_y * in_col + tid_x];
+    if (output_ptr != nullptr) {
+      int tid_y = blockIdx.y * blockDim.y + threadIdx.y;
+      for (; tid_y < in_row; tid_y += blockDim.y * gridDim.y)
+        output_ptr[tid_y * fixed_out_col + in_offset] =
+            input_data[tid_y * in_col + tid_x];
+    }
   }
 }
 
@@ -203,17 +207,18 @@ template <typename T>
 class ConcatGradFunctor<platform::CUDADeviceContext, T> {
  public:
   void operator()(const platform::CUDADeviceContext& context,
-                  const framework::Tensor& input, const int axis,
-                  std::vector<framework::Tensor>* outputs) {
+                  const framework::Tensor& input,
+                  const std::vector<const framework::Tensor*>& ref_inputs,
+                  const int axis, std::vector<framework::Tensor*>* outputs) {
     // TODO(zcd): Add input data validity checking
     int o_num = outputs->size();
     int out_row = 1;
-    auto dim_0 = outputs->at(0).dims();
+    auto dim_0 = ref_inputs[0]->dims();
     for (int i = 0; i < axis; ++i) {
       out_row *= dim_0[i];
     }
 
-    int out_col = outputs->at(0).numel() / out_row;
+    int out0_col = ref_inputs[0]->numel() / out_row;
     int in_col = 0, in_row = out_row;
     bool sameShape = true;
 
@@ -223,13 +228,17 @@ class ConcatGradFunctor<platform::CUDADeviceContext, T> {
 
     outputs_cols[0] = 0;
     for (int i = 0; i < o_num; ++i) {
-      int t_col = outputs->at(i).numel() / out_row;
+      int t_col = outputs->at(i)->numel() / out_row;
       if (sameShape) {
-        if (t_col != out_col) sameShape = false;
+        if (t_col != out0_col) sameShape = false;
       }
       in_col += t_col;
       outputs_cols[i + 1] = in_col;
-      outputs_ptr[i] = outputs->at(i).data<T>();
+      if (outputs->at(i) != nullptr) {
+        outputs_ptr[i] = outputs->at(i)->data<T>();
+      } else {
+        outputs_ptr[i] = nullptr;
+      }
     }
 
     T** dev_out_gpu_data =
@@ -255,7 +264,7 @@ class ConcatGradFunctor<platform::CUDADeviceContext, T> {
 
     if (sameShape) {
       KernelConcatGrad<<<grid_size, block_size, 0, context.stream()>>>(
-          input.data<T>(), in_row, in_col, out_col, dev_out_gpu_data);
+          input.data<T>(), in_row, in_col, out0_col, dev_out_gpu_data);
     } else {
       const int* dev_outs_col_data = outputs_cols.CUDAData(context.GetPlace());
       KernelConcatGrad<<<grid_size, block_size, 0, context.stream()>>>(
diff --git a/paddle/fluid/operators/math/concat.h b/paddle/fluid/operators/math/concat.h
@@ -57,7 +57,8 @@ template <typename DeviceContext, typename T>
 class ConcatGradFunctor {
  public:
   void operator()(const DeviceContext& context, const framework::Tensor& input,
-                  const int axis, std::vector<framework::Tensor>* outputs);
+                  const std::vector<const framework::Tensor*>& ref_inputs,
+                  const int axis, std::vector<framework::Tensor*>* outputs);
 };
 
 }  // namespace math
