Merge pull request #3191 from zchen0211/develop

gather function only

paddle/operators/CMakeLists.txt

@@ -41,6 +41,8 @@ function(op_library TARGET)
     endif()
 endfunction()
 
+cc_test(gather_test SRCS gather_test.cc DEPS tensor)
+
 cc_library(net_op SRCS net_op.cc DEPS op_registry)
 cc_test(net_op_test SRCS net_op_test.cc DEPS net_op)
 

paddle/operators/gather.h

@@ -0,0 +1,73 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+   http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+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. */
+
+#pragma once
+#include <memory.h>
+#include <cstring>
+
+#include "paddle/framework/ddim.h"
+#include "paddle/framework/tensor.h"
+#include "paddle/platform/place.h"
+
+namespace paddle {
+namespace operators {
+
+// Implementation of CPU copy
+template <typename T>
+void CPUGather(const T* params, const int* indices, const int slice_size,
+               const int index_size, T* output) {
+  const size_t slice_bytes = slice_size * sizeof(T);
+
+  for (size_t i = 0; i < index_size; ++i) {
+    int index_ = indices[i];
+    memcpy(output + i * slice_size, params + index_ * slice_size, slice_bytes);
+  }
+}
+
+// Implementation of GPU copy:
+template <typename T>
+void GPUGather(const T* src, const int* index, const int slice_size,
+               const int index_size, T* output);
+
+/**
+ * Return a new tensor from source tensor, gathered according to index
+ * input[src]: type-T source Tensor
+ * input[index]: type-int index Tensor (1-D)
+ * return: output tensor
+ */
+template <typename T>
+void Gather(const platform::Place& place, const paddle::framework::Tensor* src,
+            const paddle::framework::Tensor* index,
+            paddle::framework::Tensor* output) {
+  // check index of shape 1-D
+  PADDLE_ENFORCE(index->dims().size() == 1);
+  int index_size = index->dims()[0];
+
+  auto src_dims = src->dims();
+  paddle::framework::DDim output_dims(src_dims);
+  output_dims[0] = index_size;
+
+  // slice size
+  int slice_size = 1;
+  for (size_t i = 1; i < src_dims.size(); ++i) slice_size *= src_dims[i];
+
+  // Gathering
+  if (platform::is_cpu_place(place)) {
+    CPUGather<T>(src->data<T>(), index->data<int>(), slice_size, index_size,
+                 output->data<T>());
+  }
+}
+
+}  // namespace operators
+}  // namespace paddle

paddle/operators/gather_test.cc

@@ -0,0 +1,48 @@
+/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+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 "paddle/operators/gather.h"
+#include "paddle/framework/ddim.h"
+#include "paddle/framework/tensor.h"
+#include "paddle/platform/place.h"
+
+#include <gtest/gtest.h>
+#include <iostream>
+#include <string>
+
+TEST(Gather, GatherData) {
+  using namespace paddle::framework;
+  using namespace paddle::platform;
+  using namespace paddle::operators;
+
+  Tensor* src = new Tensor();
+  Tensor* index = new Tensor();
+  Tensor* output = new Tensor();
+
+  int* p_src = nullptr;
+  int* p_index = nullptr;
+  p_src = src->mutable_data<int>(make_ddim({3, 4}), CPUPlace());
+  p_index = index->mutable_data<int>(make_ddim({2}), CPUPlace());
+
+  for (size_t i = 0; i < 12; ++i) p_src[i] = i;
+  p_index[0] = 1;
+  p_index[1] = 0;
+
+  int* p_output = output->mutable_data<int>(make_ddim({2, 4}), CPUPlace());
+
+  Gather<int>(CPUPlace(), src, index, output);
+
+  for (size_t i = 0; i < 4; ++i) EXPECT_EQ(p_output[i], i + 4);
+  for (size_t i = 4; i < 8; ++i) EXPECT_EQ(p_output[i], i - 4);
+}