Refine split tensorrt plugin

hjchen2 · hjchen2 · commit 3e3599f3d937 · 2018-11-21T11:31:04.000Z
diff --git a/paddle/fluid/inference/tensorrt/convert/split_op.cc b/paddle/fluid/inference/tensorrt/convert/split_op.cc
@@ -40,15 +40,14 @@ class SplitOpConverter : public OpConverter {
     int axis = boost::get<int>(op_desc.GetAttr("axis"));
     std::vector<int> output_lengths =
         boost::get<std::vector<int>>(op_desc.GetAttr("sections"));
-    PADDLE_ENFORCE(axis != 0);
+    // PADDLE_ENFORCE(axis != 0);
     if (axis < 0) {
       axis += input_dims.nbDims;
     } else {
       axis -= 1;
     }
 
     PADDLE_ENFORCE(output_lengths.size() == output_num);
-
     //
     plugin::SplitPlugin* plugin = new plugin::SplitPlugin(axis, output_lengths);
     nvinfer1::IPluginLayer* layer =
diff --git a/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.cu b/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.cu
@@ -12,13 +12,61 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <cuda_fp16.h>
+#include <algorithm>
 #include "paddle/fluid/inference/tensorrt/plugin/split_op_plugin.h"
 
 namespace paddle {
 namespace inference {
 namespace tensorrt {
 namespace plugin {
 
+// copied from operators::math::SplitFunctor
+template <typename T>
+__global__ void SplitKernel(const T* input_data, const int in_row,
+                            const int in_col, const int* out_cols,
+                            int out_cols_size, T** outputs_data) {
+  int tid_x = blockIdx.x * blockDim.x + threadIdx.x;
+  int curr_segment = 0;
+  int curr_offset = out_cols[0];
+  for (; tid_x < in_col; tid_x += blockDim.x * gridDim.x) {
+    int curr_col_offset = out_cols[curr_segment + 1];
+    while (curr_col_offset <= tid_x) {
+      curr_offset = curr_col_offset;
+      ++curr_segment;
+      curr_col_offset = out_cols[curr_segment + 1];
+    }
+
+    int local_col = tid_x - curr_offset;
+    int segment_width = curr_col_offset - curr_offset;
+    T* output_ptr = outputs_data[curr_segment];
+    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];
+    }
+  }
+}
+
+template <typename T>
+__global__ void SplitKernel(const T* input_data, const int in_row,
+                            const int in_col, const int fixed_out_col,
+                            T** outputs_data) {
+  int tid_x = blockIdx.x * blockDim.x + threadIdx.x;
+  for (; tid_x < in_col; tid_x += blockDim.x * gridDim.x) {
+    int split = tid_x / fixed_out_col;
+    int in_offset = tid_x - split * fixed_out_col;
+    T* output_ptr = outputs_data[split];
+    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];
+    }
+  }
+}
+
 nvinfer1::Dims SplitPlugin::getOutputDimensions(
     int index, const nvinfer1::Dims* input_dims, int num_inputs) {
   PADDLE_ENFORCE_EQ(num_inputs, 1);
@@ -31,48 +79,95 @@ nvinfer1::Dims SplitPlugin::getOutputDimensions(
 
 int SplitPlugin::initialize() {
   PADDLE_ENFORCE_LE(axis_, nvinfer1::Dims::MAX_DIMS);
-
+  // notice input dims is [C, H, W]
+  nvinfer1::Dims dims = this->getInputDims(0);
+  outer_rows_ = 1;
+  inner_cols_ = 1;
+  for (int i = 0; i < axis_; ++i) {
+    outer_rows_ *= dims.d[i];
+  }
+  for (int i = axis_ + 1; i < dims.nbDims; ++i) {
+    inner_cols_ *= dims.d[i];
+  }
+  same_shape_ = true;
   std::vector<int> segment_offsets(1, 0);
   for (int i = 0; i < this->getNbOutputs(); ++i) {
-    segment_offsets.push_back(segment_offsets.back() + output_length_[i]);
+    if (output_length_[i] != output_length_[0]) {
+      same_shape_ = false;
+    }
+    segment_offsets.push_back(segment_offsets.back() +
+                              output_length_[i] * inner_cols_);
   }
-  segment_offsets_ = segment_offsets;
-  nvinfer1::Dims dims = this->getInputDims(0);
-  nx_ = 1;
-  for (int i = dims.nbDims - 1; i > axis_; --i) {
-    nx_ *= dims.d[i];
+  inner_cols_ *= dims.d[axis_];
+  d_segment_offsets_ = segment_offsets;
+  segment_offsets_ = std::move(segment_offsets);
+  d_output_ptrs_.resize(this->getNbOutputs(), nullptr);
+  return 0;
+}
+
+template <typename T>
+inline void Split(cudaStream_t stream, const bool same_shape,
+                  const int outer_rows, const int inner_cols,
+                  const std::vector<int>& segment_offsets,
+                  const int* d_segment_offsets, const T* input, T** outputs) {
+  const int kThreadsPerBlock = 1024;
+  const int kMaxBlocks = 65535;
+  int block_cols = kThreadsPerBlock;
+  if (inner_cols < kThreadsPerBlock) {  // block_cols is aligned by 32.
+    block_cols = ((inner_cols + 31) >> 5) << 5;
   }
-  ny_ = dims.d[axis_];
-  nz_ = 1;
-  for (int i = axis_ - 1; i >= 0; --i) {
-    nz_ *= dims.d[i];
+  int block_rows = kThreadsPerBlock / block_cols;
+  dim3 block_size = dim3(block_cols, block_rows, 1);
+
+  int grid_cols =
+      std::min((inner_cols + block_cols - 1) / block_cols, kMaxBlocks);
+  int grid_rows =
+      std::min(kMaxBlocks / grid_cols, std::max(outer_rows / block_rows, 1));
+  dim3 grid_size = dim3(grid_cols, grid_rows, 1);
+
+  if (same_shape) {
+    SplitKernel<<<grid_size, block_size, 0, stream>>>(
+        input, outer_rows, inner_cols, segment_offsets[1], outputs);
+  } else {
+    SplitKernel<<<grid_size, block_size, 0, stream>>>(
+        input, outer_rows, inner_cols, d_segment_offsets,
+        static_cast<int>(segment_offsets.size()), outputs);
   }
-  return 0;
 }
 
 int SplitPlugin::enqueue(int batchSize, const void* const* inputs,
                          void** outputs, void* workspace, cudaStream_t stream) {
-  auto const& input_dims = this->getInputDims(0);
-  int input_size = 0;
-  float const* idata = reinterpret_cast<float const*>(inputs[0]);
-  float** odatas = reinterpret_cast<float**>(outputs);
-
-  // kernel impl here.
-  int inputBatchOffset = nx_ * ny_ * nz_;
-  for (size_t i = 0; i < this->getNbOutputs(); i++) {
-    for (size_t j = 0; j < batchSize; j++) {
-      cudaMemcpyAsync(
-          odatas[i] +
-              j * (segment_offsets_[i + 1] - segment_offsets_[i]) * nx_ *
-                  sizeof(float),
-          inputs[0] +
-              (inputBatchOffset * j + segment_offsets_[i] * nx_) *
-                  sizeof(float),
-          (segment_offsets_[i + 1] - segment_offsets_[i]) * nx_ * sizeof(float),
-          cudaMemcpyDeviceToDevice, stream);
+  float const* input_ptr = reinterpret_cast<float const*>(inputs[0]);
+  if (axis_ == -1 && this->getNbOutputs() < 10) {
+    float** output_ptrs = reinterpret_cast<float**>(outputs);
+    int data_type_size = (this->getDataType() == nvinfer1::DataType::kFLOAT)
+                             ? sizeof(__half)
+                             : sizeof(float);
+    for (int i = 0; i < this->getNbOutputs(); ++i) {
+      PADDLE_ENFORCE(
+          cudaMemcpyAsync(
+              output_ptrs[i], input_ptr + segment_offsets_[i],
+              (segment_offsets_[i + 1] - segment_offsets_[i]) * data_type_size,
+              cudaMemcpyDeviceToDevice, stream) == cudaSuccess);
+    }
+  } else {
+    outer_rows_ *= batchSize;
+    const int* d_segment_offsets_ptr =
+        thrust::raw_pointer_cast(&d_segment_offsets_[0]);
+    float** output_ptrs = thrust::raw_pointer_cast(&d_output_ptrs_[0]);
+    PADDLE_ENFORCE(cudaMemcpyAsync(output_ptrs, outputs,
+                                   this->getNbOutputs() * sizeof(float*),
+                                   cudaMemcpyHostToDevice,
+                                   stream) == cudaSuccess);
+    if (this->getDataType() == nvinfer1::DataType::kFLOAT) {
+      Split(stream, same_shape_, outer_rows_, inner_cols_, segment_offsets_,
+            d_segment_offsets_ptr, input_ptr, output_ptrs);
+    } else {
+      Split(stream, same_shape_, outer_rows_, inner_cols_, segment_offsets_,
+            d_segment_offsets_ptr, (__half*)input_ptr,  // NOLINT
+            (__half**)output_ptrs);                     // NOLINT
     }
   }
-
   return cudaGetLastError() != cudaSuccess;
 }
 
diff --git a/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.h b/paddle/fluid/inference/tensorrt/plugin/split_op_plugin.h
@@ -14,6 +14,7 @@
 
 #pragma once
 
+#include <thrust/device_vector.h>
 #include <vector>
 #include "paddle/fluid/inference/tensorrt/plugin/trt_plugin.h"
 
@@ -25,7 +26,7 @@ namespace plugin {
 class SplitPlugin : public PluginTensorRT {
  public:
   SplitPlugin(int axis, std::vector<int> const &output_lengths)
-      : axis_(axis), output_length_(output_lengths) {}
+      : axis_(axis), same_shape_(true), output_length_(output_lengths) {}
 
   SplitPlugin(void const *serial_data, size_t serial_length) {
     deserializeBase(serial_data, serial_length);
@@ -60,9 +61,13 @@ class SplitPlugin : public PluginTensorRT {
   }
 
   int axis_;
+  int outer_rows_;
+  int inner_cols_;
+  bool same_shape_;
   std::vector<int> output_length_;
-  int nx_, ny_, nz_;
   std::vector<int> segment_offsets_;
+  thrust::device_vector<int> d_segment_offsets_;
+  thrust::device_vector<float *> d_output_ptrs_;
 };
 
 }  // namespace plugin
