[NPU] add pad and pad_grad (#38658)

[NPU] add pad and pad_grad

paddle/fluid/operators/pad_op_npu.cc

+/* Copyright (c) 2022 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/framework/op_registry.h"
+#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/platform/device/npu/npu_op_runner.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename T>
+class PadNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* x = context.Input<Tensor>("X");
+    auto* out = context.Output<Tensor>("Out");
+    auto paddings = context.Attr<std::vector<int>>("paddings");
+    float pad_value = context.Attr<float>("pad_value");
+
+    PADDLE_ENFORCE_LT(abs(pad_value), 1e-5,
+                      platform::errors::Unimplemented(
+                          "Ascend npu only support pad_value=0 right now,"
+                          "but received pad_value is %f .",
+                          pad_value));
+
+    out->mutable_data<T>(context.GetPlace());
+
+    NpuOpRunner runner;
+    runner.SetType("Pad")
+        .AddInput(*x)
+        .AddInput(std::move(paddings))
+        .AddOutput(*out);
+
+    auto stream =
+        context.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+    runner.Run(stream);
+  }
+};
+
+template <typename T>
+class PadGradNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* d_out = context.Input<Tensor>(framework::GradVarName("Out"));
+    auto* d_x = context.Output<Tensor>(framework::GradVarName("X"));
+    auto paddings = context.Attr<std::vector<int>>("paddings");
+
+    d_x->mutable_data<T>(context.GetPlace());
+
+    auto d_x_dims = d_x->dims();
+    auto size = paddle::framework::vectorize(d_x_dims);
+    std::vector<int> offsets(0);
+    int i = 0;
+    for (auto iter = paddings.begin(); iter < paddings.end(); ++iter, ++i) {
+      if (i % 2 == 0) {
+        offsets.push_back(*iter);
+      }
+    }
+
+    auto stream =
+        context.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+
+    const auto& runner = NpuOpRunner("SliceD", {*d_out}, {*d_x},
+                                     {{"offsets", offsets}, {"size", size}});
+    runner.Run(stream);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+namespace plat = paddle::platform;
+
+REGISTER_OP_NPU_KERNEL(pad, ops::PadNPUKernel<plat::float16>,
+                       ops::PadNPUKernel<float>, ops::PadNPUKernel<int>);
+
+REGISTER_OP_NPU_KERNEL(pad_grad, ops::PadNPUKernel<plat::float16>,
+                       ops::PadGradNPUKernel<float>);

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

+#   Copyright (c) 2022 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 sys
+sys.path.append("..")
+from op_test import OpTest
+import paddle
+import paddle.fluid.core as core
+import paddle.fluid as fluid
+from paddle.fluid import Program, program_guard
+
+paddle.enable_static()
+
+
+class TestPadOp(OpTest):
+    def setUp(self):
+        self.op_type = "pad"
+        self.set_npu()
+        self.init_dtype()
+        self.initTestCase()
+
+        self.inputs = {'X': np.random.random(self.shape).astype(self.dtype), }
+        self.attrs = {}
+        self.attrs['paddings'] = np.array(self.paddings).flatten()
+        self.attrs['pad_value'] = self.pad_value
+        self.outputs = {
+            'Out': np.pad(self.inputs['X'],
+                          self.paddings,
+                          mode='constant',
+                          constant_values=self.pad_value)
+        }
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+    def test_check_grad_normal(self):
+        if self.dtype == np.float16:
+            return
+
+        self.check_grad_with_place(self.place, ['X'], 'Out')
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+        self.place = paddle.NPUPlace(0)
+
+    def init_dtype(self):
+        self.dtype = np.float32
+
+    def initTestCase(self):
+        self.shape = (16, 16)
+        self.paddings = [(1, 1), (2, 3)]
+        self.pad_value = 0.0
+
+
+class TestCase1(TestPadOp):
+    def initTestCase(self):
+        self.shape = (2, 3, 4, 5)
+        self.paddings = [(0, 1), (2, 3), (2, 1), (1, 1)]
+        self.pad_value = 0.0
+
+
+class TestCase2(TestPadOp):
+    def initTestCase(self):
+        self.shape = (5, 5, 5)
+        self.paddings = [(0, 0), (0, 0), (1, 2)]
+        self.pad_value = 0.0
+
+
+class TestCase3(TestPadOp):
+    def initTestCase(self):
+        self.shape = (100)
+        self.paddings = [(0, 1)]
+        self.pad_value = 0.0
+
+
+#----------------Pad Fp16----------------
+
+
+def create_test_fp16(parent):
+    class TestPadFp16(parent):
+        def init_dtype(self):
+            self.dtype = np.float16
+
+    cls_name = "{0}_{1}".format(parent.__name__, "Fp16")
+    TestPadFp16.__name__ = cls_name
+    globals()[cls_name] = TestPadFp16
+
+
+create_test_fp16(TestPadOp)
+create_test_fp16(TestCase1)
+create_test_fp16(TestCase2)
+create_test_fp16(TestCase3)
+
+
+class TestPadOpError(unittest.TestCase):
+    def test_errors(self):
+        with program_guard(Program(), Program()):
+            input_data = np.random.random((2, 2)).astype("float32")
+
+            def test_Variable():
+                fluid.layers.pad(x=input_data, paddings=[1, 1, 1, 1])
+
+            self.assertRaises(TypeError, test_Variable)
+
+            data = fluid.data(name='data', shape=[4], dtype='float16')
+            fluid.layers.pad(x=data, paddings=[0, 1])
+
+
+if __name__ == '__main__':
+    unittest.main()