Support npu kernel for expand_as_v2 op (#34620)

* Support npu kernel for expand_as_v2 op

* mofify the registry data type name

* fix test unit

* fix npu compile error, test=develop

* fix compute function
Co-authored-by: Nqili93 <qili93@qq.com>

paddle/fluid/operators/expand_as_v2_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/expand_as_v2_op.h"
+#include "paddle/fluid/operators/npu_op_runner.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename DeviceContext, typename T>
+class ExpandAsV2NPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto rank = context.Input<Tensor>("X")->dims().size();
+    auto target_shape = context.Attr<std::vector<int>>("target_shape");
+    auto target_rank = target_shape.size();
+    PADDLE_ENFORCE_GE(target_rank, rank,
+                      platform::errors::InvalidArgument(
+                          "The rank (%d) of the input 'target_tensor' for "
+                          "expand_as_v2 op must be greater than or equal to "
+                          "the rank (%d) of the input 'x'.",
+                          target_rank, rank));
+    PADDLE_ENFORCE_GE(rank, 1, platform::errors::InvalidArgument(
+                                   "The rank (%d) of the input 'x' for "
+                                   "expand_as_v2 op must be positive.",
+                                   rank));
+    PADDLE_ENFORCE_LE(target_rank, MAX_RANK_SUPPORTED,
+                      platform::errors::InvalidArgument(
+                          "The rank (%d) of the input 'target_tensor' for "
+                          "expand_as_v2 op must be less than or equal to %d.",
+                          target_rank, MAX_RANK_SUPPORTED));
+    ExpandAs(context);
+  }
+
+ protected:
+  void ExpandAs(const framework::ExecutionContext& context) const {
+    auto* in0 = context.Input<framework::Tensor>("X");
+    auto in_dims = in0->dims();
+    auto target_shape = context.Attr<std::vector<int>>("target_shape");
+    auto vec_in_dims = framework::vectorize<int>(in_dims);
+    auto diff = target_shape.size() - vec_in_dims.size();
+    vec_in_dims.insert(vec_in_dims.begin(), diff, 1);
+
+    for (size_t i = 0; i < vec_in_dims.size(); ++i) {
+      PADDLE_ENFORCE_NE(target_shape[i], 0,
+                        platform::errors::InvalidArgument(
+                            "The value of target shape cannot be zero."));
+      if (vec_in_dims[i] != 1) {
+        PADDLE_ENFORCE_EQ(
+            vec_in_dims[i], target_shape[i],
+            platform::errors::InvalidArgument(
+                "The value (%d) of the non-singleton dimension does not match"
+                " the corresponding value (%d) in "
+                "target tensor for expand_as_v2 op.",
+                vec_in_dims[i], target_shape[i]));
+      }
+    }
+    auto* out0 = context.Output<framework::Tensor>("Out");
+
+    framework::DDim out_dims = framework::make_ddim(target_shape);
+
+    out0->Resize(out_dims);
+    out0->mutable_data<T>(context.GetPlace());
+
+    const auto& runner =
+        NpuOpRunner("ExpandD", {*in0}, {*out0}, {{"shape", target_shape}});
+
+    auto stream =
+        context.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+
+    runner.Run(stream);
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_NPU_KERNEL(
+    expand_as_v2,
+    ops::ExpandAsV2NPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::ExpandAsV2NPUKernel<paddle::platform::NPUDeviceContext, int>,
+    ops::ExpandAsV2NPUKernel<paddle::platform::NPUDeviceContext, int8_t>,
+    ops::ExpandAsV2NPUKernel<paddle::platform::NPUDeviceContext, uint8_t>,
+    ops::ExpandAsV2NPUKernel<paddle::platform::NPUDeviceContext,
+                             paddle::platform::float16>);

python/paddle/fluid/tests/unittests/npu/test_expand_as_v2_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
+
+paddle.enable_static()
+np.random.seed(10)
+
+
+class TestExpandAsOpRank1(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.place = paddle.NPUPlace(0)
+        self.op_type = "expand_as_v2"
+        x = np.random.rand(100).astype("float32")
+        target_tensor = np.random.rand(2, 100).astype("float32")
+        self.inputs = {'X': x}
+        self.attrs = {'target_shape': target_tensor.shape}
+        bcast_dims = [2, 1]
+        output = np.tile(self.inputs['X'], bcast_dims)
+        self.outputs = {'Out': output}
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+    def test_check_grad(self):
+        pass
+
+
+class TestExpandAsOpRank2(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.place = paddle.NPUPlace(0)
+        self.op_type = "expand_as_v2"
+        x = np.random.rand(10, 12).astype("float32")
+        target_tensor = np.random.rand(10, 12).astype("float32")
+        self.inputs = {'X': x}
+        self.attrs = {'target_shape': target_tensor.shape}
+        bcast_dims = [1, 1]
+        output = np.tile(self.inputs['X'], bcast_dims)
+        self.outputs = {'Out': output}
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+    def test_check_grad(self):
+        pass
+
+
+class TestExpandAsOpRank3(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.place = paddle.NPUPlace(0)
+        self.op_type = "expand_as_v2"
+        x = np.random.rand(2, 3, 20).astype("float32")
+        target_tensor = np.random.rand(2, 3, 20).astype("float32")
+        self.inputs = {'X': x}
+        self.attrs = {'target_shape': target_tensor.shape}
+        bcast_dims = [1, 1, 1]
+        output = np.tile(self.inputs['X'], bcast_dims)
+        self.outputs = {'Out': output}
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+    def test_check_grad(self):
+        pass
+
+
+class TestExpandAsOpRank4(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.place = paddle.NPUPlace(0)
+        self.op_type = "expand_as_v2"
+        x = np.random.rand(1, 1, 7, 16).astype("float32")
+        target_tensor = np.random.rand(4, 6, 7, 16).astype("float32")
+        self.inputs = {'X': x}
+        self.attrs = {'target_shape': target_tensor.shape}
+        bcast_dims = [4, 6, 1, 1]
+        output = np.tile(self.inputs['X'], bcast_dims)
+        self.outputs = {'Out': output}
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+    def test_check_grad(self):
+        pass
+
+
+# Test python API
+class TestExpandAsV2API(unittest.TestCase):
+    def test_api(self):
+        input1 = np.random.random([12, 14]).astype("float32")
+        input2 = np.random.random([2, 12, 14]).astype("float32")
+        x = fluid.layers.data(
+            name='x', shape=[12, 14], append_batch_size=False, dtype="float32")
+
+        y = fluid.layers.data(
+            name='target_tensor',
+            shape=[2, 12, 14],
+            append_batch_size=False,
+            dtype="float32")
+
+        out_1 = paddle.expand_as(x, y=y)
+
+        exe = fluid.Executor(place=fluid.NPUPlace(0))
+        res_1 = exe.run(fluid.default_main_program(),
+                        feed={"x": input1,
+                              "target_tensor": input2},
+                        fetch_list=[out_1])
+        assert np.array_equal(res_1[0], np.tile(input1, (2, 1, 1)))
+
+
+if __name__ == '__main__':
+    unittest.main()