"add sequence kernel"

paddle/fluid/operators/sequence_expand_op.cu

@@ -21,48 +21,89 @@ namespace operators {
 using LoDTensor = framework::LoDTensor;
 
 template <typename T>
-__global__ sequence_expand_kernel(const T* x_data, T* out_data, size_t* lod,
+__global__ void sequence_expand_kernel(const T* x_data, T* out_data,
+                                       const size_t* lod, size_t lod_size,
                                        size_t element_len) {
-  int BLOCK_SIZE = 1024;
-  __shared__ T shm_lod[BLOCK_SIZE];
-  for (int idx = threadIdx.x; idx < BLOCK_SIZE; ++idx) {
-    shm_lod[idx] = lod[idx];
+  int tid_x = blockIdx.x * blockDim.x + threadIdx.x;
+  for (; tid_x < static_cast<int>(lod_size - 1);
+       tid_x += blockDim.x * gridDim.x) {
+    int scale = lod[tid_x + 1] - lod[tid_x];
+    int tid_y = blockIdx.y * blockDim.y + threadIdx.y;
+    for (; tid_y < scale; tid_y += blockDim.y * gridDim.y) {
+      int tid_z = blockIdx.z * blockDim.z + threadIdx.z;
+      int item_start = tid_x / element_len;
+      for (; tid_z < element_len; tid_z += blockDim.z * gridDim.z) {
+        out_data[item_start * scale + tid_z] = x_data[item_start + tid_z];
+      }
     }
-  for (int idx = threadIdx.x + blockIdx.x * blockDim.x; idx < lod.size();
-       idx += blockDim.x * gridDim.x) {
-    int scale = lod[i]
   }
 }
 
 template <typename T>
-void SequenceExpandFunctor<platform::CPUDeviceContext, T>::operator()(
-    const platform::CPUDeviceContext& context, const LoDTensor& x,
-    LoDTensor* out) {
-  x_dims = x.dims();
+__global__ void sequence_expand_grad_kernel(const T* dout_data, T* dx_data,
+                                            const size_t* lod, size_t lod_size,
+                                            size_t element_len,
+                                            size_t dout_size) {
+  extern __shared__ T shm[];
+  int tid_x = blockIdx.x * blockDim.x + threadIdx.x;
+  for (; tid_x < static_cast<int>(lod_size - 1);
+       tid_x += blockDim.x * gridDim.x) {
+    int scale = lod[tid_x + 1] - lod[tid_x];
+    int tid_y = blockIdx.y * blockDim.y + threadIdx.y;
+    for (; tid_y < scale; tid_y += blockDim.y * gridDim.y) {
+      int tid_z = blockIdx.z * blockDim.z + threadIdx.z;
+      int item_start = tid_x / element_len;
+      for (; tid_z < element_len; tid_z += blockDim.z * gridDim.z) {
+        shm[item_start + tid_z] += doutx_data[item_start * scale + tid_z];
+      }
+    }
+  }
+  // synchronize before write to dx
+  __syncthreads();
+  for (int idx = blockDimx * blockIdx.x + threadIdx.x;
+       idx < static_cast<int>(dout_size); idx += blockDim.x * gridDim.x) {
+    dx_data[idx] = shm[idx;]
+  }
+}
+
+template <typename T>
+struct SequenceExpandFunctor<platform::CUDADeviceContext, T> {
+  void operator()(const platform::CUDADeviceContext& context,
+                  const LoDTensor& x, LoDTensor* out) {
+    auto x_dims = x.dims();
     size_t element_len = framework::product(x_dims) / x_dims[0];
     T* out_data = out->mutable_data<T>(context.GetPlace());
     auto out_starts = out->lod().back();
 
-  const int kThreadsPerBlock = 1024;
-  int block_cols = kThreadsPerBlock;
-  if (out_cols < kThreadsPerBlock) {  // block_cols is aligned by 32.
-    block_cols = ((out_cols + 31) >> 5) << 5;
+    dim3 block_size(16, 32, element_len);
+    dim3 grid_size(10, 10);
+    sequence_expand_kernel<<<grid_size, block_size, 0, context.stream()>>>(
+        x.data<T>(), out->mutable_data<T>(context.GetPlace()),
+        out_starts.CUDAData(context.GetPlace()), out_starts.size(),
+        element_len);
   }
-  int block_rows = kThreadsPerBlock / block_cols;
-  dim3 block_size = dim3(block_cols, block_rows, 1);
+};
 
-  int max_threads = context.GetMaxPhysicalThreadCount();
-  int max_blocks = std::max(max_threads / kThreadsPerBlock, 1);
+template <typename T>
+struct SequenceExpandGradFunctor<platform::CUDADeviceContext, T> {
+  void operator()(const platform::CUDADeviceContext& ctx, const LoDTensor& x,
+                  const LoDTensor& out, const LoDTensor& dout, LoDTensor* dx) {
+    auto x_dims = x.dims();
+    size_t element_len = framework::product(x_dims) / x_dims[0];
+    const T* x_data = x->data<T>();
+    T* out_data = out->mutable_data<T>(context.GetPlace());
+    auto out_starts = out->lod().back();
 
-  int grid_cols =
-      std::min((out_cols + block_cols - 1) / block_cols, max_blocks);
-  int grid_rows =
-      std::min(max_blocks / grid_cols, std::max(out_rows / block_rows, 1));
-  dim3 grid_size = dim3(grid_cols, grid_rows, 1);
-  sequence_expand_kernel<<<grid_size, block_size, 0, context.stream()>>>(
-      x.data<T>(), out->mutable_data<T>(context.GetPlace()),
-      out_starts.CUDAData(context.GetPlace()), element_len);
-}
+    dim3 block_size(16, 32, element_len);
+    dim3 grid_size(10, 10);
+    size_t out_size = framework::product(dx->dims());
+    sequence_expand_kernel<<<grid_size, block_size, out_size * sizeof(T),
+                             context.stream()>>>(
+        dout.data<T>(), dx->mutable_data<T>(context.GetPlace()),
+        out_starts.CUDAData(context.GetPlace()), out_starts.size(), element_len,
+        out_size);
+  }
+};
 
 }  // namespace operators
 }  // namespace paddle

paddle/fluid/operators/sequence_expand_op.h

@@ -28,15 +28,19 @@ struct SequenceExpandFunctor {
   void operator()(const DeviceContext& ctx, const LoDTensor& x, LoDTensor* out);
 };
 
-// template <typename DeviceContext, typename T>
-// struct SequenceExpandGradFunctor {};
+template <typename DeviceContext, typename T>
+struct SequenceExpandGradFunctor {
+  void operator()(const DeviceContext& ctx, const LoDTensor& x,
+                  const LoDTensor& out, const LoDTensor& dout, LoDTensor* dx);
+};
 
 template <typename T>
-void SequenceExpandFunctor<platform::CPUDeviceContext, T>::operator()(
-    const platform::CPUDeviceContext& context, const LoDTensor& x,
+struct SequenceExpandFunctor<platform::CPUDeviceContext, T> {
+  void operator()(const platform::CPUDeviceContext& context, const LoDTensor& x,
                   LoDTensor* out) {
-  x_dims = x.dims();
+    auto x_dims = x.dims();
     size_t element_len = framework::product(x_dims) / x_dims[0];
+    const T* x_data = x->data<T>();
     T* out_data = out->mutable_data<T>(context.GetPlace());
     auto out_starts = out->lod().back();
 
@@ -52,7 +56,8 @@ void SequenceExpandFunctor<platform::CPUDeviceContext, T>::operator()(
       x_data += element_len;
       out_data += element_len * scale;
     }
-}
+  }
+};
 
 template <typename DeviceContext, typename T>
 class SequenceExpandKernel : public framework::OpKernel<T> {
@@ -60,7 +65,6 @@ class SequenceExpandKernel : public framework::OpKernel<T> {
   void Compute(const framework::ExecutionContext& context) const override {
     auto* x = context.Input<LoDTensor>("X");
     auto* out = context.Output<LoDTensor>("Out");
-    const T* x_data = x->data<T>();
     auto x_dims = x->dims();
     auto* y = context.Input<LoDTensor>("Y");
     PADDLE_ENFORCE(!y->lod().empty(), "y should have lod");
@@ -86,19 +90,14 @@ class SequenceExpandKernel : public framework::OpKernel<T> {
  *    Grad(X).lod = Input(X).lod
  *
  * */
-template <typename DeviceContext, typename T>
-class SequenceExpandGradKernel : public framework::OpKernel<T> {
- public:
-  void Compute(const framework::ExecutionContext& context) const override {
-    auto* d_out = context.Input<LoDTensor>(framework::GradVarName("Out"));
-    auto* x = context.Input<LoDTensor>("X");
-    auto* out = context.Input<LoDTensor>("Out");
-    auto* d_x = context.Output<LoDTensor>(framework::GradVarName("X"));
-    auto out_last_level = out->lod().back();
-    d_x->set_lod(x->lod());
-    const T* d_out_data = d_out->data<T>();
+template <typename T>
+struct SequenceExpandGradFunctor<platform::CPUDeviceContext, T> {
+  void operator()(const platform::CPUDeviceContext& ctx, const LoDTensor& x,
+                  const LoDTensor& out, const LoDTensor& dout, LoDTensor* dx) {
+    auto out_last_level = out.lod().back();
+    const T* d_out_data = d_out.data<T>();
     T* d_x_data = d_x->mutable_data<T>(context.GetPlace());
-    size_t element_len = d_out->numel() / d_out->dims()[0];
+    size_t element_len = d_out.numel() / d_out.dims()[0];
     for (size_t i = 0; i < out_last_level.size() - 1; ++i) {
       size_t repeat = out_last_level[i + 1] - out_last_level[i];
       Eigen::TensorMap<
@@ -106,14 +105,27 @@ class SequenceExpandGradKernel : public framework::OpKernel<T> {
       d_out_t(d_out_data, static_cast<int>(repeat), element_len);
       Eigen::TensorMap<Eigen::Tensor<T, 1, Eigen::RowMajor, Eigen::DenseIndex>>
       d_x_t(d_x_data, static_cast<int>(element_len));
-      auto place =
-          context.template device_context<DeviceContext>().eigen_device();
-      d_x_t.device(*place) = d_out_t.sum(Eigen::array<int, 1>({{0}}));
+      d_x_t.device(*context.eigen_device()) =
+          d_out_t.sum(Eigen::array<int, 1>({{0}}));
       d_out_data += (repeat * element_len);
       d_x_data += element_len;
     }
   }
 };
 
+template <typename DeviceContext, typename T>
+class SequenceExpandGradKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* d_out = context.Input<LoDTensor>(framework::GradVarName("Out"));
+    auto* x = context.Input<LoDTensor>("X");
+    auto* out = context.Input<LoDTensor>("Out");
+    auto* d_x = context.Output<LoDTensor>(framework::GradVarName("X"));
+    d_x->set_lod(x->lod());
+    SequenceExpandGradFunctor(context.template device_context(), *x, *out,
+                              d_out, d_x);
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle