Support npu kernel for Crop op (#34542)

* add crop npu op

* code style check

* update code to remove usless headers

* add testcase and upsate code

paddle/fluid/operators/crop_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/crop_op.h"
+#include "paddle/fluid/operators/npu_op_runner.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename DeviceContext, typename T>
+class CropNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* x = ctx.Input<framework::Tensor>("X");
+
+    std::vector<int> offset_list;
+    if (ctx.HasInput("Offsets")) {
+      auto* offsets_tensor = ctx.Input<framework::Tensor>("Offsets");
+      TensorToVector(*offsets_tensor, ctx.device_context(), &offset_list);
+      if (offset_list.empty()) {
+        offset_list.resize(x->dims().size(), 0);
+      }
+    } else {
+      auto res = ctx.Attr<std::vector<int>>("offsets");
+      if (res.empty()) {
+        offset_list.resize(x->dims().size(), 0);
+      } else {
+        offset_list.insert(offset_list.end(), res.begin(), res.end());
+      }
+    }
+
+    PADDLE_ENFORCE_EQ(
+        static_cast<int64_t>(offset_list.size()), x->dims().size(),
+        platform::errors::InvalidArgument(
+            "The shape (%d) of CropOp's "
+            "'offset' attribute should be equal to the shape of dims "
+            "(%d) of the Input(X).",
+            offset_list.size(), x->dims().size()));
+
+    int axis_int = 0;
+    framework::NPUAttributeMap attr_input = {{"offsets", offset_list},
+                                             {"axis", axis_int}};
+    auto* out = ctx.Output<framework::Tensor>("Out");
+    out->mutable_data<T>(ctx.GetPlace());
+
+    if (ctx.HasInput("Y")) {
+      auto* shape = ctx.Input<framework::Tensor>("Y");
+      PADDLE_ENFORCE_EQ(shape->dims().size(), x->dims().size(),
+                        platform::errors::InvalidArgument(
+                            "The shape of dims of (%d) of CropOp's "
+                            "Input(shape) should be equal to the shape of dims "
+                            "(%d) of the Input(X).",
+                            shape->dims().size(), x->dims().size()));
+
+      const auto& runner =
+          NpuOpRunner("Crop", {*x, *shape}, {*out}, attr_input);
+      auto stream =
+          ctx.template device_context<paddle::platform::NPUDeviceContext>()
+              .stream();
+      runner.Run(stream);
+    } else {
+      auto shape_size = ctx.Attr<std::vector<int>>("shape");
+      PADDLE_ENFORCE_EQ(shape_size.size(), x->dims().size(),
+                        platform::errors::InvalidArgument(
+                            "The shape of dims of (%d) of CropOp's "
+                            "Input(shape) should be equal to the shape of dims "
+                            "(%d) of the Input(X).",
+                            shape_size.size(), x->dims().size()));
+      Tensor tmp_shape(x->type());
+      tmp_shape.Resize(framework::make_ddim(shape_size));
+      tmp_shape.mutable_data<T>(ctx.GetPlace());
+      const auto& runner =
+          NpuOpRunner("Crop", {*x, tmp_shape}, {*out}, attr_input);
+      auto stream =
+          ctx.template device_context<paddle::platform::NPUDeviceContext>()
+              .stream();
+      runner.Run(stream);
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_NPU_KERNEL(
+    crop, ops::CropNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::CropNPUKernel<paddle::platform::NPUDeviceContext, int>,
+    ops::CropNPUKernel<paddle::platform::NPUDeviceContext,
+                       paddle::platform::float16>);

python/paddle/fluid/tests/unittests/npu/test_crop_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
+from test_crop_op import crop
+
+paddle.enable_static()
+np.random.seed(10)
+
+
+class TestCropOp(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.place = paddle.NPUPlace(0)
+        self.op_type = "crop"
+        self.attrs = {}
+        self.offset_by_input = False
+        self.crop_by_input = False
+        self.dtype = np.float32
+        self.initTestCase()
+        if self.crop_by_input:
+            self.inputs = {
+                'X': np.random.random(self.x_shape).astype(self.dtype),
+                'Y': np.random.random(self.crop_shape).astype(self.dtype)
+            }
+        else:
+            self.attrs['shape'] = self.crop_shape
+            self.inputs = {
+                'X': np.random.random(self.x_shape).astype(self.dtype),
+            }
+
+        if self.offset_by_input:
+            self.inputs['Offsets'] = np.array(self.offsets).astype('int32')
+        else:
+            self.attrs['offsets'] = self.offsets
+
+        if len(self.offsets) == 0:
+            self.offsets = np.zeros_like(self.crop_shape)
+
+        self.outputs = {
+            'Out': crop(self.inputs['X'], self.offsets, self.crop_shape)
+        }
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def initTestCase(self):
+        self.x_shape = (10, 10)
+        self.crop_shape = [2, 2]
+        self.offsets = [1, 2]
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+
+class TestCase1(TestCropOp):
+    def initTestCase(self):
+        self.x_shape = (16, 8, 32)
+        self.crop_shape = [2, 2, 3]
+        self.offsets = [1, 5, 3]
+
+
+class TestCase2(TestCropOp):
+    def initTestCase(self):
+        self.x_shape = (15, 8)
+        self.crop_shape = [15, 8]
+        self.offsets = [0, 0]
+
+
+class TestCase3(TestCropOp):
+    def initTestCase(self):
+        self.x_shape = (4, 10)
+        self.crop_shape = [2, 3]
+        self.offsets = [0, 2]
+        self.offset_by_input = True
+
+
+class TestCase4(TestCropOp):
+    def initTestCase(self):
+        self.x_shape = (10, 9, 14)
+        self.crop_shape = [3, 3, 5]
+        self.offsets = []
+
+
+class TestCase5(TestCropOp):
+    def initTestCase(self):
+        self.x_shape = (10, 9, 14)
+        self.crop_shape = [3, 3, 5]
+        self.offsets = [3, 5, 4]
+        self.offset_by_input = True
+
+
+class TestCase6(TestCropOp):
+    def initTestCase(self):
+        self.x_shape = (10, 9, 14)
+        self.crop_shape = [3, 3, 5]
+        self.offsets = [3, 5, 4]
+        self.offset_by_input = True
+        self.__class__.no_need_check_grad = True
+        self.dtype = np.float16
+
+
+class TestCase7(TestCropOp):
+    def initTestCase(self):
+        self.x_shape = (10, 9, 14)
+        self.crop_shape = [3, 3, 5]
+        self.offsets = [3, 5, 4]
+        self.offset_by_input = True
+        self.dtype = np.int32
+
+
+class TestCase8(TestCropOp):
+    def initTestCase(self):
+        self.x_shape = (10, 9, 14)
+        self.crop_shape = [3, 3, 5]
+        self.offsets = []
+        self.offset_by_input = True
+
+
+class TestCase9(TestCropOp):
+    def initTestCase(self):
+        self.x_shape = (10, 9, 14)
+        self.crop_shape = [3, 3, 5]
+        self.offsets = [3, 5, 4]
+        self.crop_by_input = True
+
+
+class TestCase10(TestCropOp):
+    def initTestCase(self):
+        self.x_shape = (10, 9, 14)
+        self.crop_shape = [3, 3, 5]
+        self.offsets = [3, 5, 4]
+        self.crop_by_input = True
+        self.offset_by_input = True
+
+
+if __name__ == '__main__':
+    unittest.main()