[NPU ]add npu kernel for gaussian random (#33983)

* add npu operator for gaussian random.

* bugfix: add wait after memory copy.

* update gaussian random op: use TensorCopy.

paddle/fluid/operators/gaussian_random_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 <random>
+
+#include "paddle/fluid/framework/generator.h"
+#include "paddle/fluid/framework/op_registry.h"
+#include "paddle/fluid/framework/op_version_registry.h"
+#include "paddle/fluid/operators/fill_constant_op.h"
+#ifdef PADDLE_WITH_MKLDNN
+#include "paddle/fluid/platform/mkldnn_helper.h"
+#endif
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+template <typename T>
+class NPUGaussianRandomKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    float mean = context.Attr<float>("mean");
+    float std = context.Attr<float>("std");
+    auto* tensor = context.Output<framework::Tensor>("Out");
+    tensor->mutable_data<T>(context.GetPlace());
+
+    Tensor cpu_tensor(tensor->type());
+    cpu_tensor.Resize(tensor->dims());
+    T* cpu_data = cpu_tensor.mutable_data<T>(platform::CPUPlace());
+    std::normal_distribution<T> dist(mean, std);
+
+    int64_t size = tensor->numel();
+
+    unsigned int seed = static_cast<unsigned int>(context.Attr<int>("seed"));
+    auto engine = framework::GetCPURandomEngine(seed);
+    for (int64_t i = 0; i < size; ++i) {
+      cpu_data[i] = dist(*engine);
+    }
+    framework::TensorCopy(
+        cpu_tensor, context.GetPlace(),
+        context.template device_context<platform::DeviceContext>(), tensor);
+    context.template device_context<paddle::platform::NPUDeviceContext>()
+        .Wait();
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+REGISTER_OP_NPU_KERNEL(gaussian_random, ops::NPUGaussianRandomKernel<float>);

paddle/fluid/operators/uniform_random_op_npu.cc

@@ -12,11 +12,12 @@ 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/uniform_random_op.h"
 #include <string>
+
 #include "paddle/fluid/framework/generator.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/operators/uniform_random_op.h"
 
 namespace paddle {
 namespace operators {

python/paddle/fluid/tests/unittests/npu/test_gaussian_random_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
+import unittest
+import numpy as np
+sys.path.append("..")
+import paddle
+import paddle.fluid as fluid
+from op_test import OpTest
+from test_gaussian_random_op import TestGaussianRandomOp
+
+paddle.enable_static()
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestNPUGaussianRandomOp(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.op_type = "gaussian_random"
+        self.init_dtype()
+        self.set_attrs()
+        self.inputs = {}
+        self.use_mkldnn = False
+        self.attrs = {
+            "shape": [123, 92],
+            "mean": self.mean,
+            "std": self.std,
+            "seed": 10,
+            "use_mkldnn": self.use_mkldnn
+        }
+        paddle.seed(10)
+
+        self.outputs = {'Out': np.zeros((123, 92), dtype='float32')}
+
+    def set_attrs(self):
+        self.mean = 1.0
+        self.std = 2.
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+        self.place = paddle.NPUPlace(0)
+
+    def init_dtype(self):
+        self.dtype = np.float32
+
+    def test_check_output(self):
+        self.check_output_customized(self.verify_output, self.place)
+
+    def verify_output(self, outs):
+        self.assertEqual(outs[0].shape, (123, 92))
+        hist, _ = np.histogram(outs[0], range=(-3, 5))
+        hist = hist.astype("float32")
+        hist /= float(outs[0].size)
+        data = np.random.normal(size=(123, 92), loc=1, scale=2)
+        hist2, _ = np.histogram(data, range=(-3, 5))
+        hist2 = hist2.astype("float32")
+        hist2 /= float(outs[0].size)
+        self.assertTrue(
+            np.allclose(
+                hist, hist2, rtol=0, atol=0.01),
+            "hist: " + str(hist) + " hist2: " + str(hist2))
+
+
+if __name__ == "__main__":
+    unittest.main()