Add NPU kernel for norm Op: float16 and float32 (#34609)

* Add NPU kernel for norm Op: float16 and float32

* fix code for code review

* fix for code review

* add type for paddle_throw

* remove unnecessary head file.\nAdd more testcase

* remove a broadcast

paddle/fluid/operators/norm_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/norm_op.h"
+#include "paddle/fluid/operators/npu_op_runner.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename DeviceContext, typename T>
+class NormNPUKernel : public framework::OpKernel<T> {
+ private:
+  void CheckAxis(int axis, int rank) const {
+    // check the axis is in [-rank, rank-1]
+    if (axis <= rank - 1 && axis >= -rank) return;
+    PADDLE_THROW(platform::errors::InvalidArgument(
+        "axis in norm operator must between (%d) and (%d)"
+        "but got (%d).",
+        -rank, rank - 1, axis));
+  }
+
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    VLOG(4) << "Launch Norm Op Kernel on NPU." << std::endl;
+    auto* in_x = ctx.Input<framework::Tensor>("X");
+    auto* out_y = ctx.Output<framework::Tensor>("Out");
+    auto* out_norm = ctx.Output<framework::Tensor>("Norm");
+    out_y->mutable_data<T>(ctx.GetPlace());
+    out_norm->mutable_data<T>(ctx.GetPlace());
+    auto xdim = in_x->dims();
+    float eps = ctx.Attr<float>("epsilon");
+    int axis = ctx.Attr<int>("axis");
+    CheckAxis(axis, xdim.size());
+    if (axis < 0) axis = xdim.size() + axis;
+
+    framework::NPUAttributeMap attr_input_norm;
+    attr_input_norm["axes"] = std::vector<int>({axis});
+    attr_input_norm["p"] = 2;
+    attr_input_norm["keepdim"] = true;
+    attr_input_norm["epsilon"] = eps;
+    const auto& runner =
+        NpuOpRunner("LpNorm", {*in_x}, {*out_norm}, attr_input_norm);
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+    runner.Run(stream);
+    NpuOpRunner("Div", {*in_x, *out_norm}, {*out_y}, {}).Run(stream);
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+namespace plat = paddle::platform;
+REGISTER_OP_NPU_KERNEL(
+    norm, ops::NormNPUKernel<paddle::platform::NPUDeviceContext, float>,
+    ops::NormNPUKernel<paddle::platform::NPUDeviceContext,
+                       paddle::platform::float16>)

python/paddle/fluid/tests/unittests/npu/test_norm_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 sys
+sys.path.append("..")
+import unittest
+import numpy as np
+import paddle
+import paddle.fluid as fluid
+from op_test import OpTest, skip_check_grad_ci
+
+SEED = 2021
+
+
+def l2_norm(x, axis, epsilon):
+    x2 = x**2
+    s = np.sum(x2, axis=axis, keepdims=True)
+    r = np.sqrt(s) + epsilon
+    y = x / np.broadcast_to(r, x.shape)
+    return y, r
+
+
+class TestNorm(OpTest):
+    def setUp(self):
+        paddle.enable_static()
+        self.set_npu()
+        self.place = paddle.NPUPlace(0)
+        self.op_type = "norm"
+        self.init_dtype()
+
+        x = np.random.random(self.shape).astype(self.dtype)
+        y, norm = l2_norm(x, self.axis, self.epsilon)
+        self.inputs = {'X': x}
+        self.attrs = {'epsilon': self.epsilon, 'axis': self.axis}
+        self.outputs = {'Out': y, 'Norm': norm}
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def init_dtype(self):
+        self.dtype = np.float32
+        self.axis = 1
+        self.epsilon = 1e-10
+        self.shape = (2, 3, 4, 5)
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+
+class TestNormOp2(TestNorm):
+    def init_test_case(self):
+        self.shape = [5, 3, 9, 7]
+        self.axis = 0
+        self.epsilon = 1e-8
+        self.dtype = np.float32
+
+
+class TestNormOp3(TestNorm):
+    def init_test_case(self):
+        self.shape = [5, 3, 2, 7]
+        self.axis = -1
+        self.epsilon = 1e-8
+        self.dtype = np.float32
+
+
+class TestNormOp4(TestNorm):
+    def init_test_case(self):
+        self.shape = [128, 1024, 14, 14]
+        self.axis = 2
+        self.epsilon = 1e-8
+        self.dtype = np.float32
+
+
+class API_NormTest(unittest.TestCase):
+    def test_errors(self):
+        paddle.enable_static()
+        with fluid.program_guard(fluid.Program()):
+
+            def test_norm_x_type():
+                data = fluid.data(name="x", shape=[3, 3], dtype="float64")
+                out = fluid.layers.l2_normalize(data)
+
+            self.assertRaises(TypeError, test_norm_x_type)
+
+
+class TestNormFP16(TestNorm):
+    def set_npu(self):
+        self.__class__.use_npu = True
+        self.__class__.no_need_check_grad = True
+
+    def init_dtype(self):
+        self.dtype = np.float16
+        self.axis = -1
+        self.epsilon = 1e-10
+        self.shape = (2, 3, 100)
+
+
+if __name__ == '__main__':
+    unittest.main()