add beam_search_decode npu op (#34967)

paddle/fluid/operators/beam_search_decode_op.cc

@@ -45,9 +45,15 @@ struct BeamSearchDecodeFunctor {
         id_tensor_(id_tensor),
         score_tensor_(score_tensor) {
     tensor_on_gpu_ = false;
+    tensor_on_npu_ = false;
     // First make a copy of GPU data on CPU
+    if (platform::is_gpu_place(step_ids_origin_[0].place()) ||
+        platform::is_npu_place(step_ids_origin_[0].place())) {
       if (platform::is_gpu_place(step_ids_origin_[0].place())) {
         tensor_on_gpu_ = true;
+      } else {
+        tensor_on_npu_ = true;
+      }
       platform::DeviceContextPool& pool =
           platform::DeviceContextPool::Instance();
       auto* dev_ctx = pool.Get(step_ids_origin_[0].place());
@@ -55,7 +61,9 @@ struct BeamSearchDecodeFunctor {
       for (auto& step_id : step_ids_origin_) {
         framework::LoDTensor out;
         if (step_id.numel() > 0) {
+          if (tensor_on_gpu_) {
             dev_ctx->Wait();
+          }
           framework::TensorCopy(step_id, platform::CPUPlace(), *dev_ctx, &out);
           dev_ctx->Wait();
         }
@@ -64,8 +72,13 @@ struct BeamSearchDecodeFunctor {
         step_ids_.push_back(out);
       }
     }
+    if (platform::is_gpu_place(step_scores_origin_[0].place()) ||
+        platform::is_npu_place(step_scores_origin_[0].place())) {
       if (platform::is_gpu_place(step_scores_origin_[0].place())) {
         tensor_on_gpu_ = true;
+      } else {
+        tensor_on_npu_ = true;
+      }
       platform::DeviceContextPool& pool =
           platform::DeviceContextPool::Instance();
       auto* dev_ctx = pool.Get(step_scores_origin_[0].place());
@@ -73,7 +86,9 @@ struct BeamSearchDecodeFunctor {
       for (auto& step_score : step_scores_origin_) {
         framework::LoDTensor out;
         if (step_score.numel() > 0) {
+          if (tensor_on_gpu_) {
             dev_ctx->Wait();
+          }
           framework::TensorCopy(step_score, platform::CPUPlace(), *dev_ctx,
                                 &out);
           dev_ctx->Wait();
@@ -89,6 +104,7 @@ struct BeamSearchDecodeFunctor {
   void apply() const;
 
   bool tensor_on_gpu_;
+  bool tensor_on_npu_;
   size_t beam_size_;
   int end_id_;
   // TODO(Superjomn) Here might result serious performance issue in the
@@ -105,8 +121,8 @@ struct BeamSearchDecodeFunctor {
 template <typename T>
 void BeamSearchDecodeFunctor::apply() const {
   BeamSearchDecoder<T> beam_search_decoder(beam_size_, end_id_);
-  // Check if the tensor is on GPU. If so, use the CPU copy instead
-  if (tensor_on_gpu_) {
+  // Check if the tensor is on GPU or NPU. If so, use the CPU copy instead
+  if (tensor_on_gpu_ || tensor_on_npu_) {
     beam_search_decoder.Backtrace(step_ids_, step_scores_, id_tensor_,
                                   score_tensor_);
   } else {

python/paddle/fluid/tests/unittests/npu/test_beam_search_decode_op_npu.py

+#   Copyright (c) 2021 PaddlePaddle Authors. All Rights Reserved.
+#
+# 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.
+
+from __future__ import print_function
+
+import unittest
+
+import numpy as np
+import paddle
+import paddle.fluid.core as core
+from paddle.fluid.op import Operator
+import paddle.fluid as fluid
+from paddle.fluid.framework import Program, program_guard
+
+
+class TestBeamSearchDecodeNPUOp(unittest.TestCase):
+    """unittest of beam_search_decode npu op"""
+
+    def setUp(self):
+        self.scope = core.Scope()
+        self.place = paddle.NPUPlace(0)
+
+    def append_lod_tensor(self, tensor_array, lod, data):
+        lod_tensor = core.LoDTensor()
+        lod_tensor.set_lod(lod)
+        lod_tensor.set(data, self.place)
+        tensor_array.append(lod_tensor)
+
+    def test_get_set(self):
+        ids = self.scope.var("ids").get_lod_tensor_array()
+        scores = self.scope.var("scores").get_lod_tensor_array()
+        # Construct sample data with 5 steps and 2 source sentences
+        # beam_size = 2, end_id = 1
+        # start with start_id
+        [
+            self.append_lod_tensor(
+                array, [[0, 1, 2], [0, 1, 2]], np.array(
+                    [0, 0], dtype=dtype))
+            for array, dtype in ((ids, "int64"), (scores, "float32"))
+        ]
+        [
+            self.append_lod_tensor(
+                array, [[0, 1, 2], [0, 2, 4]],
+                np.array(
+                    [2, 3, 4, 5], dtype=dtype))
+            for array, dtype in ((ids, "int64"), (scores, "float32"))
+        ]
+        [
+            self.append_lod_tensor(
+                array, [[0, 2, 4], [0, 2, 2, 4, 4]],
+                np.array(
+                    [3, 1, 5, 4], dtype=dtype))
+            for array, dtype in ((ids, "int64"), (scores, "float32"))
+        ]
+        [
+            self.append_lod_tensor(
+                array, [[0, 2, 4], [0, 1, 2, 3, 4]],
+                np.array(
+                    [1, 1, 3, 5], dtype=dtype))
+            for array, dtype in ((ids, "int64"), (scores, "float32"))
+        ]
+        [
+            self.append_lod_tensor(
+                array, [[0, 2, 4], [0, 0, 0, 2, 2]],
+                np.array(
+                    [5, 1], dtype=dtype))
+            for array, dtype in ((ids, "int64"), (scores, "float32"))
+        ]
+
+        sentence_ids = self.scope.var("sentence_ids").get_tensor()
+        sentence_scores = self.scope.var("sentence_scores").get_tensor()
+
+        beam_search_decode_op = Operator(
+            "beam_search_decode",
+            # inputs
+            Ids="ids",
+            Scores="scores",
+            # outputs
+            SentenceIds="sentence_ids",
+            SentenceScores="sentence_scores",
+            beam_size=2,
+            end_id=1, )
+
+        beam_search_decode_op.run(self.scope, self.place)
+
+        expected_lod = [[0, 2, 4], [0, 4, 7, 12, 17]]
+        self.assertEqual(sentence_ids.lod(), expected_lod)
+        self.assertEqual(sentence_scores.lod(), expected_lod)
+
+        expected_data = np.array(
+            [0, 2, 3, 1, 0, 2, 1, 0, 4, 5, 3, 5, 0, 4, 5, 3, 1], "int64")
+        self.assertTrue(np.array_equal(np.array(sentence_ids), expected_data))
+        self.assertTrue(
+            np.array_equal(np.array(sentence_scores), expected_data))
+
+
+if __name__ == '__main__':
+    unittest.main()