[NPU] add beam_search npu op (#34860)

* add beam_search npu op

* fix CMakeList and add unittest

* fix bug of beam search npu op

* fix unittest

* let input ids become int64

* set output ids to int64_t

* delete check_dygraph

* fix beam_width=1

paddle/fluid/operators/beam_search_op.h

@@ -51,11 +51,11 @@ class BeamSearchOpKernel : public framework::OpKernel<T> {
     PADDLE_ENFORCE_NOT_NULL(
         selected_ids,
         platform::errors::NotFound(
-            "Output(selected_scores) of BeamSearchOp is not found."));
+            "Output(selected_ids) of BeamSearchOp is not found."));
     PADDLE_ENFORCE_NOT_NULL(
         selected_scores,
         platform::errors::NotFound(
-            "Output(parent_idx) of BeamSearchOp is not found."));
+            "Output(selected_scores) of BeamSearchOp is not found."));
 
     math::BeamSearchFunctor<DeviceContext, T> alg;
     alg(context.template device_context<DeviceContext>(), pre_ids, pre_scores,

paddle/fluid/operators/beam_search_op_npu.cc

+/* 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. */
+
+#include "paddle/fluid/operators/beam_search_op.h"
+#include "paddle/fluid/framework/op_registry.h"
+
+namespace ops = paddle::operators;
+REGISTER_OP_NPU_KERNEL(
+    beam_search,
+    ops::BeamSearchOpKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::BeamSearchOpKernel<paddle::platform::NPUDeviceContext, double>,
+    ops::BeamSearchOpKernel<paddle::platform::NPUDeviceContext, int>,
+    ops::BeamSearchOpKernel<paddle::platform::NPUDeviceContext, int64_t>);

paddle/fluid/operators/math/CMakeLists.txt

@@ -39,6 +39,10 @@ function(math_library TARGET)
     endif()
 endfunction()
 
+if (WITH_ASCEND_CL)
+  cc_library(beam_search_npu SRCS beam_search_npu.cc DEPS npu_op_runner)
+endif()
+
 # please add new math_library in alphabetical order
 math_library(concat_and_split)
 math_library(context_project DEPS im2col math_function)
@@ -68,7 +72,11 @@ math_library(sequence_padding)
 math_library(sequence_pooling DEPS math_function jit_kernel_helper)
 math_library(sequence_scale)
 math_library(softmax DEPS math_function jit_kernel_helper)
-math_library(beam_search DEPS math_function)
+if (WITH_ASCEND_CL)
+    math_library(beam_search DEPS math_function beam_search_npu)
+else()
+    math_library(beam_search DEPS math_function)
+endif()
 math_library(fc DEPS blas)
 
 math_library(matrix_bit_code)

paddle/fluid/operators/math/beam_search_npu.cc

+/* 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. */
+
+#include "paddle/fluid/operators/math/beam_search.h"
+#include "paddle/fluid/operators/npu_op_runner.h"
+
+namespace paddle {
+namespace framework {
+class LoDTensor;
+class Tensor;
+}  // namespace framework
+namespace platform {
+class NPUDeviceContext;
+}  // namespace platform
+}  // namespace paddle
+
+namespace paddle {
+namespace operators {
+namespace math {
+
+template <typename T>
+class BeamSearchFunctor<platform::NPUDeviceContext, T> {
+ public:
+  void operator()(const platform::NPUDeviceContext& ctx,
+                  const framework::LoDTensor* pre_ids,
+                  const framework::LoDTensor* pre_scores,
+                  const framework::LoDTensor* ids,
+                  const framework::LoDTensor* scores,
+                  framework::LoDTensor* selected_ids,
+                  framework::LoDTensor* selected_scores,
+                  framework::Tensor* parent_idx, size_t level, size_t beam_size,
+                  int end_id, bool is_accumulated) {
+    auto abs_lod = framework::ToAbsOffset(scores->lod());
+    auto& high_level = abs_lod[level];
+
+    int64_t num_seqs = scores->NumElements(level);
+    // size of the first beam is 1, others are equal to beam_size
+    int64_t real_beam_size = static_cast<int64_t>(scores->dims()[0] / num_seqs);
+    // K
+    int64_t seq_width = 1;
+    for (int i = 1; i < scores->dims().size(); i++) {
+      seq_width *= scores->dims()[i];
+    }
+
+    auto place = ctx.GetPlace();
+    auto stream = ctx.stream();
+
+    int64_t total_length = num_seqs * beam_size;
+    int64_t batch_size = static_cast<int64_t>(scores->dims()[0]);
+    selected_ids->mutable_data<int64_t>(framework::make_ddim({total_length, 1}),
+                                        place);
+    selected_scores->mutable_data<float>(
+        framework::make_ddim({total_length, 1}), place);
+    parent_idx->mutable_data<int64_t>(framework::make_ddim({total_length}),
+                                      place);
+
+    // Step1: Define Tensors and Preprocess the situation that pre_id == end_id
+
+    // cast ids and pre_ids from int to float32
+    Tensor ids_int32(framework::proto::VarType::INT32);
+    if (ids->type() != framework::proto::VarType::INT32) {
+      ids_int32.Resize(ids->dims());
+      ids_int32.mutable_data<int>(ctx.GetPlace());
+      auto dst_dtype_ids_int32 = ConvertToNpuDtype(ids_int32.type());
+      const auto& runner_ids_int32 =
+          NpuOpRunner("Cast", {*ids}, {ids_int32},
+                      {{"dst_type", static_cast<int>(dst_dtype_ids_int32)}});
+      runner_ids_int32.Run(stream);
+    } else {
+      ids_int32.ShareDataWith(*ids);
+    }
+
+    Tensor pre_ids_int32(framework::proto::VarType::INT32);
+    if (pre_ids->type() != framework::proto::VarType::INT32) {
+      pre_ids_int32.Resize(pre_ids->dims());
+      pre_ids_int32.mutable_data<int>(ctx.GetPlace());
+      auto dst_dtype_pre_ids_int32 = ConvertToNpuDtype(pre_ids_int32.type());
+      const auto& runner_pre_ids_int32 = NpuOpRunner(
+          "Cast", {*pre_ids}, {pre_ids_int32},
+          {{"dst_type", static_cast<int>(dst_dtype_pre_ids_int32)}});
+      runner_pre_ids_int32.Run(stream);
+    } else {
+      pre_ids_int32.ShareDataWith(*pre_ids);
+    }
+
+    Tensor expand_pre_ids(pre_ids_int32.type());
+    expand_pre_ids.Resize(framework::make_ddim({batch_size, seq_width}));
+    expand_pre_ids.mutable_data<int>(place);
+    const auto& runner_tile_pre_ids =
+        NpuOpRunner("TileWithAxis", {pre_ids_int32}, {expand_pre_ids},
+                    {{"axis", 1}, {"tiles", seq_width}});
+    runner_tile_pre_ids.Run(stream);
+    expand_pre_ids.Resize(ids_int32.dims());
+
+    Tensor expand_pre_scores(pre_scores->type());
+    expand_pre_scores.Resize(framework::make_ddim({batch_size, seq_width}));
+    expand_pre_scores.mutable_data<float>(place);
+    const auto& runner_tile_pre_scores =
+        NpuOpRunner("TileWithAxis", {*pre_scores}, {expand_pre_scores},
+                    {{"axis", 1}, {"tiles", seq_width}});
+    runner_tile_pre_scores.Run(stream);
+    expand_pre_scores.Resize(scores->dims());
+
+    // End_id Tensors
+    Tensor end_id_tmp_tensor(framework::proto::VarType::INT32);
+    end_id_tmp_tensor.mutable_data<int>({1}, ctx.GetPlace());
+    FillNpuTensorWithConstant<int>(&end_id_tmp_tensor, end_id);
+
+    Tensor end_id_tensors(ids_int32.type());
+    end_id_tensors.mutable_data<int>(ids_int32.dims(), place);
+    const auto& runner_fill_end_id =
+        NpuOpRunner("FillD", {end_id_tmp_tensor}, {end_id_tensors},
+                    {{"dims", framework::vectorize(ids_int32.dims())}});
+    runner_fill_end_id.Run(stream);
+
+    // whether expand_pre_ids == end_ids?
+    Tensor equal_end_ids(framework::proto::VarType::BOOL);
+    equal_end_ids.mutable_data<bool>(ids_int32.dims(), place);
+    const auto& runner_equal_end_ids = NpuOpRunner(
+        "Equal", {expand_pre_ids, end_id_tensors}, {equal_end_ids}, {});
+    runner_equal_end_ids.Run(stream);
+
+    // construct a Tensor with dimension ids->dims():
+    // [[False, True, True, True, ...],
+    //  [False, True, True, True, ...],
+    //  ...]
+    Tensor false_tmp_tensor(framework::proto::VarType::INT32);
+    false_tmp_tensor.mutable_data<int>({1}, ctx.GetPlace());
+    FillNpuTensorWithConstant<int>(&false_tmp_tensor, static_cast<int>(false));
+
+    Tensor first_pos_false_tensors(framework::proto::VarType::INT32);
+    first_pos_false_tensors.Resize(framework::make_ddim({batch_size, 1}));
+    first_pos_false_tensors.mutable_data<int>(place);
+    std::vector<int64_t> fill_dims = {batch_size, 1};
+    framework::NPUAttributeMap fill_attr = {{"dims", fill_dims}};
+    const auto& runner_fill_false_tensors = NpuOpRunner(
+        "FillD", {false_tmp_tensor}, {first_pos_false_tensors}, fill_attr);
+    runner_fill_false_tensors.Run(stream);
+
+    Tensor pos_tensors(framework::proto::VarType::INT32);
+    if (seq_width > 1) {
+      pos_tensors.Resize(framework::make_ddim({batch_size, seq_width}));
+      pos_tensors.mutable_data<int>(place);
+
+      Tensor true_tmp_tensor(framework::proto::VarType::INT32);
+      true_tmp_tensor.mutable_data<int>({1}, ctx.GetPlace());
+      FillNpuTensorWithConstant<int>(&true_tmp_tensor, static_cast<int>(true));
+
+      Tensor second_pos_true_tensors(framework::proto::VarType::INT32);
+      second_pos_true_tensors.Resize(
+          framework::make_ddim({batch_size, seq_width - 1}));
+      second_pos_true_tensors.mutable_data<int>(place);
+      std::vector<int64_t> fill_dims2 = {batch_size, seq_width - 1};
+      framework::NPUAttributeMap fill_attr2 = {{"dims", fill_dims2}};
+      const auto& runner_fill_true_tensors = NpuOpRunner(
+          "FillD", {true_tmp_tensor}, {second_pos_true_tensors}, fill_attr2);
+      runner_fill_true_tensors.Run(stream);
+
+      std::vector<framework::Tensor> concat_inputs = {first_pos_false_tensors,
+                                                      second_pos_true_tensors};
+      std::vector<std::string> concat_names = {"x0", "x1"};
+      NpuOpRunner runner_concat_false_true{"ConcatD",
+                                           {concat_inputs},
+                                           {pos_tensors},
+                                           {{"concat_dim", 1}, {"N", 2}}};
+      runner_concat_false_true.AddInputNames(concat_names);
+      runner_concat_false_true.Run(stream);
+      pos_tensors.Resize(ids_int32.dims());
+    } else {
+      pos_tensors.ShareDataWith(first_pos_false_tensors);
+    }
+
+    Tensor cast_pos_tensors_bool(framework::proto::VarType::BOOL);
+    cast_pos_tensors_bool.Resize(pos_tensors.dims());
+    cast_pos_tensors_bool.mutable_data<bool>(ctx.GetPlace());
+    auto dst_dtype = ConvertToNpuDtype(cast_pos_tensors_bool.type());
+    const auto& runner_cast_pos_tensors =
+        NpuOpRunner("Cast", {pos_tensors}, {cast_pos_tensors_bool},
+                    {{"dst_type", static_cast<int>(dst_dtype)}});
+    runner_cast_pos_tensors.Run(stream);
+
+    // if pre_ids == end_ids, save only one score, and others become -inf
+    // construct pre_ids == end_ids and save only one score
+    Tensor save_one_end_score(framework::proto::VarType::BOOL);
+    save_one_end_score.mutable_data<bool>(ids_int32.dims(), place);
+    const auto& runner_logical_and =
+        NpuOpRunner("LogicalAnd", {equal_end_ids, cast_pos_tensors_bool},
+                    {save_one_end_score}, {});
+    runner_logical_and.Run(stream);
+
+    // if save_one_end_score is True, set score to -inf
+    // define -Inf Tensors
+    Tensor ninf_tmp_tensor(scores->type());
+    ninf_tmp_tensor.mutable_data<float>({1}, ctx.GetPlace());
+    float ninf_value =
+        static_cast<float>(-std::numeric_limits<float>::infinity());
+    FillNpuTensorWithConstant<float>(&ninf_tmp_tensor, ninf_value);
+
+    Tensor ninf_tensors(scores->type());
+    ninf_tensors.mutable_data<float>(scores->dims(), place);
+    const auto& runner_fill_ninf =
+        NpuOpRunner("FillD", {ninf_tmp_tensor}, {ninf_tensors},
+                    {{"dims", framework::vectorize(scores->dims())}});
+    runner_fill_ninf.Run(stream);
+
+    // Step2: calculate topk scores
+
+    // get scores used in topk op
+    Tensor tmp_scores(scores->type());
+    tmp_scores.mutable_data<float>(scores->dims(), place);
+    if (!is_accumulated) {
+      // if pre_id == end_id, cal_scores = pre_score, and id = end_id
+      // else, cal_score = pre_score + log(score)
+
+      // calculate log(scores)
+      Tensor log_scores(scores->type());
+      log_scores.mutable_data<float>(scores->dims(), place);
+
+      Tensor one(scores->type());
+      one.mutable_data<float>(scores->dims(), place);
+      const auto& runner_one = NpuOpRunner("OnesLike", {*scores}, {one}, {});
+      runner_one.Run(stream);
+
+      Tensor sub(scores->type());
+      sub.mutable_data<float>(scores->dims(), place);
+      const auto& runner_sub = NpuOpRunner("Sub", {*scores, one}, {sub}, {});
+      runner_sub.Run(stream);
+
+      const auto& runner_log_scores =
+          NpuOpRunner("Log1p", {sub}, {log_scores}, {});
+      runner_log_scores.Run(stream);
+
+      // tmp_scores = pre_score + log(scores)
+      const auto& runner_add_scores =
+          NpuOpRunner("Add", {log_scores, *pre_scores}, {tmp_scores}, {});
+      runner_add_scores.Run(stream);
+
+      // if pre_ids == end_ids, use pre_score rather than score
+      const auto& runner_select_equal_end_score =
+          NpuOpRunner("Select", {equal_end_ids, expand_pre_scores, tmp_scores},
+                      {tmp_scores}, {});
+      runner_select_equal_end_score.Run(stream);
+    } else {
+      // if pre_ids == end_ids, use pre_score rather than score
+      const auto& runner_select_equal_end_score2 =
+          NpuOpRunner("Select", {equal_end_ids, expand_pre_scores, *scores},
+                      {tmp_scores}, {});
+      runner_select_equal_end_score2.Run(stream);
+    }
+
+    // if pre_ids == end_ids, save only one score, and others become -inf
+    Tensor cal_scores(scores->type());
+    cal_scores.mutable_data<float>(scores->dims(), place);
+    const auto& runner_select_inf_score =
+        NpuOpRunner("Select", {save_one_end_score, ninf_tensors, tmp_scores},
+                    {cal_scores}, {});
+    runner_select_inf_score.Run(stream);
+
+    // resize scores from [num_seqs * beam_size, K] to [num_seqs, beam_size * K]
+    // real_beam_size = 1 or beam_size
+    cal_scores.Resize(
+        framework::make_ddim({num_seqs, real_beam_size * seq_width}));
+
+    Tensor topk_scores(scores->type());
+    topk_scores.Resize(
+        framework::make_ddim({num_seqs, static_cast<int64_t>(beam_size)}));
+    topk_scores.mutable_data<float>(ctx.GetPlace());
+
+    Tensor tmp_indices(framework::proto::VarType::INT32);
+    tmp_indices.Resize(
+        framework::make_ddim({num_seqs, static_cast<int64_t>(beam_size)}));
+    tmp_indices.mutable_data<int>(ctx.GetPlace());
+
+    // run topk op
+    NpuOpRunner runner_topk;
+    runner_topk.SetType("TopKV2")
+        .AddInput(cal_scores)
+        .AddInput(std::vector<int>{static_cast<int>(beam_size)})
+        .AddOutput(topk_scores)
+        .AddOutput(tmp_indices)
+        .AddAttr("sorted", true)
+        .AddAttr("dim", -1)
+        .AddAttr("largest", true);
+    runner_topk.Run(stream);
+
+    // cast tmp_indices from int to float32 for Sort op
+    Tensor cast_tmp_indices(framework::proto::VarType::FP32);
+    cast_tmp_indices.Resize(tmp_indices.dims());
+    cast_tmp_indices.mutable_data<float>(ctx.GetPlace());
+    auto dst_dtype_tmp_indices_fp32 =
+        ConvertToNpuDtype(cast_tmp_indices.type());
+    const auto& runner_cast_tmp_indices = NpuOpRunner(
+        "Cast", {tmp_indices}, {cast_tmp_indices},
+        {{"dst_type", static_cast<int>(dst_dtype_tmp_indices_fp32)}});
+    runner_cast_tmp_indices.Run(stream);
+
+    // sort tmp_indices
+    Tensor sorted_tmp_indices(framework::proto::VarType::FP32);
+    sorted_tmp_indices.Resize(tmp_indices.dims());
+    sorted_tmp_indices.mutable_data<float>(ctx.GetPlace());
+    Tensor sorted_score_indices(framework::proto::VarType::INT32);
+    sorted_score_indices.Resize(tmp_indices.dims());
+    sorted_score_indices.mutable_data<int>(ctx.GetPlace());
+    const auto& runner_sort_tmp_indices = NpuOpRunner(
+        "Sort", {cast_tmp_indices}, {sorted_tmp_indices, sorted_score_indices},
+        {{"axis", 1}, {"descending", false}});
+    runner_sort_tmp_indices.Run(stream);
+
+    // cast sorted_tmp_indices from float32 to int
+    Tensor cast_sort_tmp_indices(framework::proto::VarType::INT32);
+    cast_sort_tmp_indices.Resize(sorted_tmp_indices.dims());
+    cast_sort_tmp_indices.mutable_data<int>(ctx.GetPlace());
+    auto dst_dtype_tmp_indices_int32 =
+        ConvertToNpuDtype(cast_sort_tmp_indices.type());
+    const auto& runner_cast_sort_tmp_indices = NpuOpRunner(
+        "Cast", {sorted_tmp_indices}, {cast_sort_tmp_indices},
+        {{"dst_type", static_cast<int>(dst_dtype_tmp_indices_int32)}});
+    runner_cast_sort_tmp_indices.Run(stream);
+
+    // Step 3: infer selected ids from tmp_indices and ids
+
+    // if pre_ids == end_ids, use pre_ids rather than ids
+    Tensor cal_ids(ids_int32.type());
+    cal_ids.mutable_data<int>(ids_int32.dims(), place);
+    const auto& runner_select_equal_end_id = NpuOpRunner(
+        "Select", {equal_end_ids, expand_pre_ids, ids_int32}, {cal_ids}, {});
+    runner_select_equal_end_id.Run(stream);
+
+    // resize ids from [num_seqs * real_beam_size, K] to [num_seqs,
+    // real_beam_size * K]
+    // real_beam_size = 1 or beam_size
+    cal_ids.Resize(
+        framework::make_ddim({num_seqs, real_beam_size * seq_width}));
+
+    // construct batch_ids like [[0, 0, 0], [1, 1, 1], ..., [bs-1, bs-1, bs-1]]
+    // construct arange(num_seqs*beam_size).reshape((num_seqs, beam_size)) //
+    // beam_size
+    Tensor batch_ids(framework::proto::VarType::INT32);
+    batch_ids.Resize(
+        framework::make_ddim({num_seqs, static_cast<int64_t>(beam_size), 1}));
+    batch_ids.mutable_data<int>(place);
+
+    std::vector<int> vector_batch_ids;
+    for (int i = 0; i < num_seqs * static_cast<int>(beam_size); ++i) {
+      vector_batch_ids.push_back(static_cast<int>(i / beam_size));
+    }
+    framework::TensorFromVector(vector_batch_ids, ctx, &batch_ids);
+    batch_ids.Resize(
+        framework::make_ddim({num_seqs, static_cast<int64_t>(beam_size), 1}));
+
+    // sort topk_scores to get selected_scores
+    // get indices of gather_nd op for calculating selected_scores
+    Tensor gather_nd_score_indices(framework::proto::VarType::INT32);
+    gather_nd_score_indices.Resize(
+        framework::make_ddim({num_seqs, static_cast<int64_t>(beam_size), 2}));
+    gather_nd_score_indices.mutable_data<int>(place);
+
+    sorted_score_indices.Resize(
+        framework::make_ddim({num_seqs, static_cast<int64_t>(beam_size), 1}));
+    std::vector<framework::Tensor> concat_inputs2 = {batch_ids,
+                                                     sorted_score_indices};
+    std::vector<std::string> concat_names = {"x0", "x1"};
+    NpuOpRunner runner_concat_score_indices{"ConcatD",
+                                            {concat_inputs2},
+                                            {gather_nd_score_indices},
+                                            {{"concat_dim", 2}, {"N", 2}}};
+    runner_concat_score_indices.AddInputNames(concat_names);
+    runner_concat_score_indices.Run(stream);
+
+    // use gather_nd to get selected_scores
+    const auto& runner_gather_nd_scores =
+        NpuOpRunner("GatherNd", {topk_scores, gather_nd_score_indices},
+                    {*selected_scores}, {});
+    runner_gather_nd_scores.Run(stream);
+
+    // get indices of gather_nd op
+    cast_sort_tmp_indices.Resize(
+        framework::make_ddim({num_seqs, static_cast<int64_t>(beam_size), 1}));
+    Tensor gather_nd_id_indices(framework::proto::VarType::INT32);
+    gather_nd_id_indices.Resize(
+        framework::make_ddim({num_seqs, static_cast<int64_t>(beam_size), 2}));
+    gather_nd_id_indices.mutable_data<int>(place);
+
+    std::vector<framework::Tensor> concat_inputs3 = {batch_ids,
+                                                     cast_sort_tmp_indices};
+    NpuOpRunner runner_concat_id_indices{"ConcatD",
+                                         {concat_inputs3},
+                                         {gather_nd_id_indices},
+                                         {{"concat_dim", 2}, {"N", 2}}};
+    runner_concat_id_indices.AddInputNames(concat_names);
+    runner_concat_id_indices.Run(stream);
+
+    // use gather_nd to get selected_ids
+    Tensor topk_ids(framework::proto::VarType::INT32);
+    topk_ids.Resize(
+        framework::make_ddim({num_seqs, static_cast<int64_t>(beam_size)}));
+    topk_ids.mutable_data<int>(ctx.GetPlace());
+
+    const auto& runner_gather_nd_ids = NpuOpRunner(
+        "GatherNd", {cal_ids, gather_nd_id_indices}, {topk_ids}, {});
+    runner_gather_nd_ids.Run(stream);
+
+    // cast topk_ids from int to int64 to get selected_ids
+    auto dst_dtype_selected_ids = ConvertToNpuDtype(selected_ids->type());
+    const auto& runner_cast_selected_ids =
+        NpuOpRunner("Cast", {topk_ids}, {*selected_ids},
+                    {{"dst_type", static_cast<int>(dst_dtype_selected_ids)}});
+    runner_cast_selected_ids.Run(stream);
+
+    // TODO(pangyoki): PruneEndBeams
+
+    // Step 4: set lod of output Tensor
+    // define Tensor with value `seq_width`
+    Tensor seq_width_tensor(framework::proto::VarType::INT32);
+    seq_width_tensor.mutable_data<int>({1}, ctx.GetPlace());
+    FillNpuTensorWithConstant<int>(&seq_width_tensor,
+                                   static_cast<int>(seq_width));
+
+    // beam_ids = tmp_indices // seq_width
+    Tensor beam_ids(framework::proto::VarType::INT32);
+    beam_ids.Resize(
+        framework::make_ddim({num_seqs, static_cast<int64_t>(beam_size)}));
+    beam_ids.mutable_data<int>(ctx.GetPlace());
+    cast_sort_tmp_indices.Resize(
+        framework::make_ddim({num_seqs, static_cast<int64_t>(beam_size)}));
+
+    const auto& runner_div = NpuOpRunner(
+        "Div", {cast_sort_tmp_indices, seq_width_tensor}, {beam_ids}, {});
+    runner_div.Run(stream);
+
+    // get parent_idx by adding batch_ids to beam_ids
+    // construct scale_batch_ids like [[0, 0, 0], [bw, bw, bw], ..., [bs-1*bw,
+    // bs-1*bw, bs-1*bw]]
+    batch_ids.Resize(
+        framework::make_ddim({num_seqs, static_cast<int64_t>(beam_size)}));
+
+    // cast batch_ids from int to float32
+    Tensor cast_batch_ids(framework::proto::VarType::FP32);
+    cast_batch_ids.Resize(batch_ids.dims());
+    cast_batch_ids.mutable_data<float>(ctx.GetPlace());
+    auto dst_dtype1 = ConvertToNpuDtype(cast_batch_ids.type());
+    const auto& runner_cast_batch_ids =
+        NpuOpRunner("Cast", {batch_ids}, {cast_batch_ids},
+                    {{"dst_type", static_cast<int>(dst_dtype1)}});
+    runner_cast_batch_ids.Run(stream);
+
+    // scale batch_ids with beam_size
+    Tensor scale_batch_ids(framework::proto::VarType::FP32);
+    scale_batch_ids.Resize(batch_ids.dims());
+    scale_batch_ids.mutable_data<float>(place);
+    const auto& runner_power =
+        NpuOpRunner("Power", {cast_batch_ids}, {scale_batch_ids},
+                    {{"power", static_cast<float>(1.0)},
+                     {"scale", static_cast<float>(beam_size)},
+                     {"shift", static_cast<float>(0.0)}});
+    runner_power.Run(stream);
+
+    // cast cast_scale_batch_ids from float32 to int
+    Tensor cast_scale_batch_ids(framework::proto::VarType::INT32);
+    cast_scale_batch_ids.Resize(scale_batch_ids.dims());
+    cast_scale_batch_ids.mutable_data<int>(ctx.GetPlace());
+    auto dst_dtype2 = ConvertToNpuDtype(cast_scale_batch_ids.type());
+    const auto& runner_cast_scale_batch_ids =
+        NpuOpRunner("Cast", {scale_batch_ids}, {cast_scale_batch_ids},
+                    {{"dst_type", static_cast<int>(dst_dtype2)}});
+    runner_cast_scale_batch_ids.Run(stream);
+
+    // calculate parent_idx
+    Tensor tmp_parent_idx(framework::proto::VarType::INT32);
+    tmp_parent_idx.Resize(parent_idx->dims());
+    tmp_parent_idx.mutable_data<int>(place);
+    const auto& runner_add_beam_id = NpuOpRunner(
+        "Add", {beam_ids, cast_scale_batch_ids}, {tmp_parent_idx}, {});
+    runner_add_beam_id.Run(stream);
+
+    // cast tmp_parent_idx from int to int64 to get parent_idx
+    auto dst_dtype_parent_idx = ConvertToNpuDtype(parent_idx->type());
+    const auto& runner_cast_parent_idx =
+        NpuOpRunner("Cast", {tmp_parent_idx}, {*parent_idx},
+                    {{"dst_type", static_cast<int>(dst_dtype_parent_idx)}});
+    runner_cast_parent_idx.Run(stream);
+
+    std::vector<int> vector_parent_idx;
+    framework::TensorToVector(tmp_parent_idx, ctx, &vector_parent_idx);
+
+    // set low level, len(low_level) = high_level[-1]
+    std::vector<int> low_level;
+    std::vector<int> num_parent_ids(num_seqs * beam_size,
+                                    static_cast<int64_t>(0));
+    size_t low_level_size = high_level[num_seqs];
+    size_t sum_parent_id = 0;
+
+    // calculate number of every parent_id
+    for (size_t i = 0; i < num_seqs * beam_size; ++i) {
+      num_parent_ids[vector_parent_idx[i]]++;
+    }
+
+    // update low_level
+    low_level.push_back(0);
+    for (size_t i = 0; i < low_level_size; ++i) {
+      sum_parent_id += num_parent_ids[i];
+      low_level.push_back(sum_parent_id);
+    }
+
+    // fill lod
+    framework::LoD lod(2);
+    lod[0].assign(high_level.begin(), high_level.end());
+    lod[1].assign(low_level.begin(), low_level.end());
+    if (!framework::CheckLoD(lod)) {
+      PADDLE_THROW(platform::errors::InvalidArgument(
+          "lod %s is not right in"
+          " beam_search, please check your code.",
+          framework::LoDToString(lod)));
+    }
+    selected_ids->set_lod(lod);
+    selected_scores->set_lod(lod);
+  }
+};
+
+template class BeamSearchFunctor<platform::NPUDeviceContext, int>;
+template class BeamSearchFunctor<platform::NPUDeviceContext, int64_t>;
+template class BeamSearchFunctor<platform::NPUDeviceContext, float>;
+template class BeamSearchFunctor<platform::NPUDeviceContext, double>;
+
+}  // namespace math
+}  // namespace operators
+}  // namespace paddle

python/paddle/fluid/tests/unittests/npu/test_beam_search_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 paddle
+import sys
+sys.path.append("..")
+from op_test import OpTest
+import unittest
+import numpy as np
+import paddle.fluid as fluid
+
+paddle.enable_static()
+
+
+class TestBeamSearchNPUOp(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.place = paddle.NPUPlace(0)
+        self.op_type = "beam_search"
+        self.init_data()
+        self.inputs = {
+            'pre_ids': (self.pre_ids, self.lod),
+            'pre_scores': (self.pre_score, self.lod),
+            'ids': (self.ids, self.lod),
+            'scores': (self.score, self.lod)
+        }
+        # The `target_lod` attribute is still based on offset
+        self.attrs = {
+            'level': 0,
+            'beam_size': self.beam_size,
+            'end_id': 0,
+            'is_accumulated': self.is_accumulated
+        }
+        self.outputs = {
+            'selected_ids': (self.selected_ids, self.out_lod),
+            'selected_scores': (self.selected_scores, self.out_lod),
+            'parent_idx': self.parent_idx
+        }
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def init_data(self):
+        self.beam_size = 2
+        self.is_accumulated = True
+        self.pre_ids = np.array([[1], [2], [3], [4]], dtype='int64')
+        self.ids = np.array(
+            [[4, 2, 5], [2, 1, 3], [3, 5, 2], [8, 2, 1]], dtype='int64')
+        self.lod = [[2, 2], [1, 1, 1, 1]]
+        self.out_lod = [[2, 2], [1, 1, 1, 1]]
+        self.offset_lod = [[0, 2, 4], [0, 1, 2, 3, 4]]
+        self.score = np.array(
+            [
+                [0.5, 0.3, 0.2],
+                [0.6, 0.3, 0.1],
+                [0.9, 0.5, 0.1],
+                [0.7, 0.5, 0.1],
+            ],
+            dtype='float32')
+        self.pre_score = np.array([[0.1], [0.2], [0.3], [0.4]], dtype='float32')
+        self.selected_ids = np.array([4, 2, 3, 8])[:, np.newaxis]
+        self.selected_scores = np.array([0.5, 0.6, 0.9, 0.7])[:, np.newaxis]
+        self.parent_idx = np.array([0, 1, 2, 3])
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place, atol=1e-3)
+
+
+class TestBeamSearchNPUOp2(TestBeamSearchNPUOp):
+    def init_data(self):
+        self.beam_size = 2
+        self.is_accumulated = True
+        self.pre_ids = np.array([[1], [2], [3], [4]], dtype='int64')
+        self.ids = np.array([[4, 2], [7, 3], [3, 5], [8, 1]], dtype='int64')
+        self.lod = [[2, 2], [1, 1, 1, 1]]
+        self.out_lod = [[2, 2], [2, 0, 1, 1]]
+        self.offset_lod = [[0, 2, 4], [0, 2, 2, 3, 4]]
+        self.score = np.array(
+            [
+                [0.6, 0.9],
+                [0.5, 0.3],
+                [0.9, 0.5],
+                [0.1, 0.7],
+            ], dtype='float32')
+        self.pre_score = np.array([[0.1], [0.2], [0.3], [0.4]], dtype='float32')
+        self.selected_ids = np.array([4, 2, 3, 1])[:, np.newaxis]
+        self.selected_scores = np.array([0.6, 0.9, 0.9, 0.7])[:, np.newaxis]
+        self.parent_idx = np.array([0, 0, 2, 3])
+
+
+class TestBeamSearchNPUOp3(TestBeamSearchNPUOp):
+    def init_data(self):
+        # end_id = 0
+        self.beam_size = 2
+        self.is_accumulated = True
+        self.pre_ids = np.array([[1], [0], [0], [4]], dtype='int64')
+        self.ids = np.array([[4, 2], [7, 3], [3, 5], [8, 1]], dtype='int64')
+        self.lod = [[2, 2], [1, 1, 1, 1]]
+        self.out_lod = [[2, 2], [1, 1, 0, 2]]
+        self.offset_lod = [[0, 2, 4], [0, 1, 2, 2, 4]]
+        self.score = np.array(
+            [
+                [0.6, 0.9],
+                [0.5, 0.3],
+                [0.9, 0.5],
+                [0.6, 0.7],
+            ], dtype='float32')
+        self.pre_score = np.array([[0.1], [1.2], [0.5], [0.4]], dtype='float32')
+        self.selected_ids = np.array([2, 0, 8, 1])[:, np.newaxis]
+        self.selected_scores = np.array([0.9, 1.2, 0.6, 0.7])[:, np.newaxis]
+        self.parent_idx = np.array([0, 1, 3, 3])
+
+
+class TestBeamSearchNPUOp4(TestBeamSearchNPUOp):
+    def init_data(self):
+        # is_accumulated = False
+        self.beam_size = 2
+        self.is_accumulated = False
+        self.pre_ids = np.array([[1], [2], [3], [4]], dtype='int64')
+        self.ids = np.array([[4, 2], [7, 3], [3, 5], [8, 1]], dtype='int64')
+        self.lod = [[2, 2], [1, 1, 1, 1]]
+        self.out_lod = [[2, 2], [0, 2, 1, 1]]
+        self.offset_lod = [[0, 2, 4], [0, 0, 2, 3, 4]]
+        self.score = np.array(
+            [
+                [0.6, 0.9],
+                [0.5, 0.3],
+                [0.9, 0.5],
+                [0.1, 0.7],
+            ], dtype='float32')
+        self.pre_score = np.array([[0.1], [2.2], [0.3], [0.4]], dtype='float32')
+        self.selected_ids = np.array([7, 3, 3, 1])[:, np.newaxis]
+        self.selected_scores = np.array(
+            [1.50685, 0.996027, 0.194639, 0.043325])[:, np.newaxis]
+        self.parent_idx = np.array([1, 1, 2, 3])
+
+
+class TestBeamSearchNPUOp5(TestBeamSearchNPUOp):
+    def init_data(self):
+        # beam_size = 1
+        self.beam_size = 1
+        self.is_accumulated = True
+        self.pre_ids = np.array([[1], [2], [3], [4]], dtype='int64')
+        self.ids = np.array([[4, 2], [7, 3], [3, 5], [8, 1]], dtype='int64')
+        self.lod = [[1, 1, 1, 1], [1, 1, 1, 1]]
+        self.out_lod = [[1, 1, 1, 1], [1, 1, 1, 1]]
+        self.offset_lod = [[0, 1, 2, 3, 4], [0, 1, 2, 3, 4]]
+        self.score = np.array(
+            [
+                [0.6, 0.9],
+                [0.5, 0.3],
+                [0.9, 0.5],
+                [0.1, 0.7],
+            ], dtype='float32')
+        self.pre_score = np.array([[0.1], [0.2], [0.3], [0.4]], dtype='float32')
+        self.selected_ids = np.array([2, 7, 3, 1])[:, np.newaxis]
+        self.selected_scores = np.array([0.9, 0.5, 0.9, 0.7])[:, np.newaxis]
+        self.parent_idx = np.array([0, 1, 2, 3])
+
+
+if __name__ == '__main__':
+    unittest.main()

python/paddle/fluid/tests/unittests/test_beam_search_op.py

@@ -38,6 +38,7 @@ class BeamSearchOpTester(unittest.TestCase):
         self._create_pre_scores()
         self._create_scores()
         self._create_pre_ids()
+        self.set_outputs()
         self.scope.var('selected_ids').get_tensor()
         self.scope.var('selected_scores').get_tensor()
         self.scope.var('parent_idx').get_tensor()
@@ -53,22 +54,19 @@ class BeamSearchOpTester(unittest.TestCase):
             selected_scores='selected_scores',
             parent_idx='parent_idx',
             level=0,
-            beam_size=2,
-            end_id=0, )
+            beam_size=self.beam_size,
+            end_id=0,
+            is_accumulated=self.is_accumulated)
         op.run(self.scope, core.CPUPlace())
         selected_ids = self.scope.find_var("selected_ids").get_tensor()
         selected_scores = self.scope.find_var("selected_scores").get_tensor()
         parent_idx = self.scope.find_var("parent_idx").get_tensor()
+        self.assertTrue(np.allclose(np.array(selected_ids), self.output_ids))
         self.assertTrue(
-            np.allclose(
-                np.array(selected_ids), np.array([4, 2, 3, 8])[:, np.newaxis]))
+            np.allclose(np.array(selected_scores), self.output_scores))
+        self.assertEqual(selected_ids.lod(), self.output_lod)
         self.assertTrue(
-            np.allclose(
-                np.array(selected_scores),
-                np.array([0.5, 0.6, 0.9, 0.7])[:, np.newaxis]))
-        self.assertEqual(selected_ids.lod(), [[0, 2, 4], [0, 1, 2, 3, 4]])
-        self.assertTrue(
-            np.allclose(np.array(parent_idx), np.array([0, 1, 2, 3])))
+            np.allclose(np.array(parent_idx), self.output_parent_idx))
 
     def _create_pre_ids(self):
         np_data = np.array([[1, 2, 3, 4]], dtype='int64')
@@ -97,6 +95,194 @@ class BeamSearchOpTester(unittest.TestCase):
         tensor = create_tensor(self.scope, "scores", np_data)
         tensor.set_lod(self.lod)
 
+    def set_outputs(self):
+        self.beam_size = 2
+        self.is_accumulated = True
+        self.output_ids = np.array([4, 2, 3, 8])[:, np.newaxis]
+        self.output_scores = np.array([0.5, 0.6, 0.9, 0.7])[:, np.newaxis]
+        self.output_lod = [[0, 2, 4], [0, 1, 2, 3, 4]]
+        self.output_parent_idx = np.array([0, 1, 2, 3])
+
+
+class BeamSearchOpTester2(BeamSearchOpTester):
+    def _create_pre_ids(self):
+        np_data = np.array([[1], [2], [3], [4]], dtype='int64')
+        tensor = create_tensor(self.scope, 'pre_ids', np_data)
+
+    def _create_pre_scores(self):
+        np_data = np.array([[0.1, 0.2, 0.3, 0.4]], dtype='float32')
+        tensor = create_tensor(self.scope, 'pre_scores', np_data)
+
+    def _create_ids(self):
+        self.lod = [[0, 2, 4], [0, 1, 2, 3, 4]]
+        np_data = np.array([[4, 2], [7, 3], [3, 5], [8, 1]], dtype='int64')
+        tensor = create_tensor(self.scope, "ids", np_data)
+        tensor.set_lod(self.lod)
+
+    def _create_scores(self):
+        np_data = np.array(
+            [
+                [0.6, 0.9],
+                [0.5, 0.3],
+                [0.9, 0.5],
+                [0.1, 0.7],
+            ], dtype='float32')
+        tensor = create_tensor(self.scope, "scores", np_data)
+        tensor.set_lod(self.lod)
+
+    def set_outputs(self):
+        self.beam_size = 2
+        self.is_accumulated = True
+        self.output_ids = np.array([2, 4, 3, 1])[:, np.newaxis]
+        self.output_scores = np.array([0.9, 0.6, 0.9, 0.7])[:, np.newaxis]
+        self.output_lod = [[0, 2, 4], [0, 2, 2, 3, 4]]
+        self.output_parent_idx = np.array([0, 0, 2, 3])
+
+
+class BeamSearchOpTester3(BeamSearchOpTester):
+    # pre_id = end_id
+    def _create_pre_ids(self):
+        np_data = np.array([[1], [0], [0], [4]], dtype='int64')
+        tensor = create_tensor(self.scope, 'pre_ids', np_data)
+
+    def _create_pre_scores(self):
+        np_data = np.array([[0.1], [1.2], [0.5], [0.4]], dtype='float32')
+        tensor = create_tensor(self.scope, 'pre_scores', np_data)
+
+    def _create_ids(self):
+        self.lod = [[0, 2, 4], [0, 1, 2, 3, 4]]
+        np_data = np.array([[4, 2], [7, 3], [3, 5], [8, 1]], dtype='int64')
+        tensor = create_tensor(self.scope, "ids", np_data)
+        tensor.set_lod(self.lod)
+
+    def _create_scores(self):
+        np_data = np.array(
+            [
+                [0.6, 0.9],
+                [0.5, 0.3],
+                [0.9, 0.5],
+                [0.6, 0.7],
+            ], dtype='float32')
+        tensor = create_tensor(self.scope, "scores", np_data)
+        tensor.set_lod(self.lod)
+
+    def set_outputs(self):
+        self.beam_size = 2
+        self.is_accumulated = True
+        self.output_ids = np.array([2, 0, 1, 8])[:, np.newaxis]
+        self.output_scores = np.array([0.9, 1.2, 0.7, 0.6])[:, np.newaxis]
+        self.output_lod = [[0, 2, 4], [0, 1, 2, 2, 4]]
+        self.output_parent_idx = np.array([0, 1, 3, 3])
+
+
+class BeamSearchOpTester4(BeamSearchOpTester):
+    # prune beam search while pre_id of in all beams is end_id
+    def _create_pre_ids(self):
+        np_data = np.array([[0], [0], [0], [4]], dtype='int64')
+        tensor = create_tensor(self.scope, 'pre_ids', np_data)
+
+    def _create_pre_scores(self):
+        np_data = np.array([[0.1], [1.2], [0.5], [0.4]], dtype='float32')
+        tensor = create_tensor(self.scope, 'pre_scores', np_data)
+
+    def _create_ids(self):
+        self.lod = [[0, 2, 4], [0, 1, 2, 3, 4]]
+        np_data = np.array([[4, 2], [7, 3], [3, 5], [8, 1]], dtype='int64')
+        tensor = create_tensor(self.scope, "ids", np_data)
+        tensor.set_lod(self.lod)
+
+    def _create_scores(self):
+        np_data = np.array(
+            [
+                [0.6, 0.9],
+                [0.5, 0.3],
+                [0.9, 0.5],
+                [0.6, 0.7],
+            ], dtype='float32')
+        tensor = create_tensor(self.scope, "scores", np_data)
+        tensor.set_lod(self.lod)
+
+    def set_outputs(self):
+        self.beam_size = 2
+        self.is_accumulated = True
+        self.output_ids = np.array([1, 8])[:, np.newaxis]
+        self.output_scores = np.array([0.7, 0.6])[:, np.newaxis]
+        self.output_lod = [[0, 2, 4], [0, 0, 0, 0, 2]]
+        self.output_parent_idx = np.array([3, 3])
+
+
+class BeamSearchOpTester5(BeamSearchOpTester):
+    # is_accumulated = False
+    def _create_pre_ids(self):
+        np_data = np.array([[1], [2], [3], [4]], dtype='int64')
+        tensor = create_tensor(self.scope, 'pre_ids', np_data)
+
+    def _create_pre_scores(self):
+        np_data = np.array([[0.1, 2.2, 0.3, 0.4]], dtype='float32')
+        tensor = create_tensor(self.scope, 'pre_scores', np_data)
+
+    def _create_ids(self):
+        self.lod = [[0, 2, 4], [0, 1, 2, 3, 4]]
+        np_data = np.array([[4, 2], [7, 3], [3, 5], [8, 1]], dtype='int64')
+        tensor = create_tensor(self.scope, "ids", np_data)
+        tensor.set_lod(self.lod)
+
+    def _create_scores(self):
+        np_data = np.array(
+            [
+                [0.6, 0.9],
+                [0.5, 0.3],
+                [0.9, 0.5],
+                [0.1, 0.7],
+            ], dtype='float32')
+        tensor = create_tensor(self.scope, "scores", np_data)
+        tensor.set_lod(self.lod)
+
+    def set_outputs(self):
+        self.beam_size = 2
+        self.is_accumulated = False
+        self.output_ids = np.array([7, 3, 3, 1])[:, np.newaxis]
+        self.output_scores = np.array(
+            [1.50685, 0.996027, 0.194639, 0.043325])[:, np.newaxis]
+        self.output_lod = [[0, 2, 4], [0, 0, 2, 3, 4]]
+        self.output_parent_idx = np.array([1, 1, 2, 3])
+
+
+class BeamSearchOpTester6(BeamSearchOpTester):
+    # beam_size = 1
+    def _create_pre_ids(self):
+        np_data = np.array([[1], [2], [3], [4]], dtype='int64')
+        tensor = create_tensor(self.scope, 'pre_ids', np_data)
+
+    def _create_pre_scores(self):
+        np_data = np.array([[0.1, 0.2, 0.3, 0.4]], dtype='float32')
+        tensor = create_tensor(self.scope, 'pre_scores', np_data)
+
+    def _create_ids(self):
+        self.lod = [[0, 1, 2, 3, 4], [0, 1, 2, 3, 4]]
+        np_data = np.array([[4, 2], [7, 3], [3, 5], [8, 1]], dtype='int64')
+        tensor = create_tensor(self.scope, "ids", np_data)
+        tensor.set_lod(self.lod)
+
+    def _create_scores(self):
+        np_data = np.array(
+            [
+                [0.6, 0.9],
+                [0.5, 0.3],
+                [0.9, 0.5],
+                [0.1, 0.7],
+            ], dtype='float32')
+        tensor = create_tensor(self.scope, "scores", np_data)
+        tensor.set_lod(self.lod)
+
+    def set_outputs(self):
+        self.beam_size = 1
+        self.is_accumulated = True
+        self.output_ids = np.array([2, 7, 3, 1])[:, np.newaxis]
+        self.output_scores = np.array([0.9, 0.5, 0.9, 0.7])[:, np.newaxis]
+        self.output_lod = [[0, 1, 2, 3, 4], [0, 1, 2, 3, 4]]
+        self.output_parent_idx = np.array([0, 1, 2, 3])
+
 
 class TestBeamSearchOpError(unittest.TestCase):
     def test_errors(self):