[NPU] add NPU add topk (#31596)

* add topk op

* add cmake

* update topk npu op

* refactor func

* fix test not go npu TopKD bug

* NPUPlace(4) to NPUPlace(0)

* update comment
Co-authored-by: Noyjxer <1728722986@qq.com>

paddle/fluid/operators/top_k_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 <memory>
+#include <string>
+
+#include "paddle/fluid/operators/top_k_op.h"
+#include "paddle/fluid/operators/npu_op_runner.h"
+
+namespace paddle {
+namespace operators {
+
+void gen_assist_seq(framework::Tensor* assit_tensor,
+                     int64_t dim, const framework::ExecutionContext& ctx) {
+  const int64_t dimx2 = dim;
+  std::vector<paddle::platform::float16> assit;
+  assit.resize(2 * dimx2);
+  for (int64_t i = 0; i < dimx2; i++) {
+    // for i in range [0, dim]
+    assit[i] = static_cast<paddle::platform::float16>(i);
+
+    // for i in range [dim, dimx2]
+    int64_t idx = static_cast<int64_t>(
+                        static_cast<paddle::platform::float16>(i));
+    int64_t gap = i - idx;
+    assit[i + dim] = static_cast<paddle::platform::float16>(gap);
+  }
+  framework::TensorFromVector(assit, ctx.device_context(), assit_tensor);
+}
+
+
+template <typename DeviceContext, typename T>
+class TopkNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    // read input
+    auto* input = ctx.Input<framework::LoDTensor>("X");
+    auto* output = ctx.Output<framework::LoDTensor>("Out");
+    auto* indices = ctx.Output<framework::LoDTensor>("Indices");
+
+    size_t k = static_cast<int>(ctx.Attr<int>("k"));
+
+    output->mutable_data<T>(ctx.GetPlace());
+    indices->mutable_data<int>(ctx.GetPlace());
+
+    // prepare assit
+    auto dim = input->dims().size();
+    framework::Tensor assist_seq_tensor;
+    assist_seq_tensor.Resize({2 * dim});
+    assist_seq_tensor.mutable_data<T>(ctx.GetPlace());
+    gen_assist_seq(&assist_seq_tensor, dim, ctx);
+
+    framework::NPUAttributeMap attr_input = {{"sorted", "true"},
+                                             {"k", static_cast<int>(k)},
+                                             {"dim", -1},
+                                             {"largest", true}};
+
+    // run ascend
+    auto runner = NpuOpRunner("TopKD",
+                              {*input, assist_seq_tensor},
+                              {*output, *indices},
+                              attr_input);
+
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+
+    runner.Run(stream);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+// Ascend Op TopKD only support input float 16 dtype
+REGISTER_OP_NPU_KERNEL(
+    top_k,
+    ops::TopkNPUKernel<paddle::platform::NPUDeviceContext,
+                                  paddle::platform::float16>);

python/paddle/fluid/tests/unittests/npu/test_top_k_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 numpy as np
+import unittest
+import sys
+sys.path.append("..")
+from op_test import OpTest
+import paddle
+import paddle.fluid as fluid
+from paddle.fluid import core
+
+paddle.enable_static()
+SEED = 2021
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestTopk(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.place = paddle.NPUPlace(0)
+        self.op_type = "top_k"
+        self.init_dtype()
+
+        x = np.array([[0.78104149, 0.88745828, 0.32362268],
+                      [0.82196718, 0.48763277, 0.42826136],
+                      [0.96527182, 0.34851612, 0.12959783]]).astype(self.dtype)
+
+        self.inputs = {'X': x}
+        np_out = np.array([[0.88745828], [0.82196718], [0.96527182]]).astype(self.dtype)
+        np_indices = np.array([[1], [0], [0]])
+
+        self.attrs = {'k': 1, "axis": -1}
+        self.outputs = {'Out': np_out, 'Indices':np_indices}
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+        self.__class__.no_need_check_grad = True
+
+    def init_dtype(self):
+        self.dtype = np.float16
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place, check_dygraph=False)
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestTopkV2(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.place = paddle.NPUPlace(0)
+        self.op_type = "top_k"
+        self.init_dtype()
+
+        x = np.array([[0.78104149, 0.88745828, 0.32362268],
+                      [0.82196718, 0.48763277, 0.42826136],
+                      [0.96527182, 0.34851612, 0.12959783]]).astype(self.dtype)
+
+        self.inputs = {'X': x}
+        np_out = np.array([[0.88745828, 0.78104149], [0.82196718, 0.48763277], [0.96527182, 0.34851612]]).astype(self.dtype)
+        np_indices = np.array([[1, 0], [0, 1], [0, 1]])
+
+        self.attrs = {'k': 2, "axis": -1}
+        self.outputs = {'Out': np_out, 'Indices':np_indices}
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+        self.__class__.no_need_check_grad = True
+
+    def init_dtype(self):
+        self.dtype = np.float16
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place, check_dygraph=False)
+
+
+if __name__ == '__main__':
+    unittest.main()
+
+