sum op (#39165)

paddle/fluid/operators/sum_op_mlu.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/operators/sum_op.h"
+#include "paddle/fluid/operators/mlu/mlu_baseop.h"
+
+namespace paddle {
+namespace operators {
+
+using Tensor = framework::Tensor;
+
+template <typename DeviceContext, typename T>
+class SumMLUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext &ctx) const override {
+    auto out_var = ctx.OutputVar("Out");
+    if (out_var->IsType<framework::LoDTensor>()) {
+      // init
+      auto *out = out_var->GetMutable<framework::LoDTensor>();
+      auto ins = ctx.MultiInput<Tensor>("X");
+      out->mutable_data<T>(ctx.GetPlace());
+      auto place = ctx.GetPlace();
+      int ins_size = static_cast<int>(ins.size());
+      if (ins_size == 1) {
+        TensorCopy(*ins[0], place, out);
+        return;
+      }
+
+      // MLU shoul do sth
+      std::vector<const void *> inputs;
+      std::vector<MLUCnnlTensorDesc> input_descs;
+      std::vector<cnnlTensorDescriptor_t> desc_vector;
+      for (int i = 0; i < ins_size; i++) {
+        input_descs.emplace_back(MLUCnnlTensorDesc(
+            *ins[i], CNNL_LAYOUT_ARRAY, ToCnnlDataType(ins[i]->type())));
+        desc_vector.push_back(input_descs.back().get());
+        inputs.push_back(GetBasePtr(ins[i]));
+      }
+      // init out tensors
+      MLUCnnlTensorDesc output_desc(*out, CNNL_LAYOUT_ARRAY,
+                                    ToCnnlDataType(out->type()));
+      uint32_t ins_size_t = static_cast<uint32_t>(ins_size);
+      MLUCnnl::AddN(ctx, ins_size_t, desc_vector.data(), inputs.data(),
+                    output_desc.get(), GetBasePtr(out));
+
+    } else {
+      PADDLE_THROW(platform::errors::InvalidArgument(
+          "Expected type of Output(out) must be Tensor or But got "
+          "unsupport type: %s.",
+          framework::ToTypeName(out_var->Type())));
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+
+REGISTER_OP_MLU_KERNEL(
+    sum, ops::SumMLUKernel<paddle::platform::MLUDeviceContext, float>,
+    ops::SumMLUKernel<paddle::platform::MLUDeviceContext,
+                      paddle::platform::float16>);

python/paddle/fluid/tests/unittests/mlu/test_sum_op_mlu.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 numpy as np
+import unittest
+import sys
+sys.path.append("..")
+from op_test import OpTest
+import paddle
+import paddle.fluid as fluid
+import paddle.fluid.core as core
+
+paddle.enable_static()
+SEED = 2021
+
+
+class TestSum1(OpTest):
+    def setUp(self):
+        self.set_mlu()
+        self.init_dtype()
+        self.op_type = "sum"
+        self.place = paddle.MLUPlace(0)
+
+        x0 = np.random.random((3, 40)).astype(self.dtype)
+        x1 = np.random.random((3, 40)).astype(self.dtype)
+        x2 = np.random.random((3, 40)).astype(self.dtype)
+        self.inputs = {'X': [("x0", x0), ("x1", x1), ("x2", x2)]}
+        y = x0 + x1 + x2
+        self.outputs = {'Out': y}
+
+        self.attrs = {'use_mkldnn': False}
+
+    def init_dtype(self):
+        self.dtype = np.float32
+
+    def set_mlu(self):
+        self.__class__.use_mlu = True
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+
+class TestSum2(OpTest):
+    def setUp(self):
+        self.set_mlu()
+        self.init_dtype()
+        self.op_type = "sum"
+        self.place = paddle.MLUPlace(0)
+
+        x0 = np.random.random((3, 3)).astype(self.dtype)
+        x1 = np.random.random((3, 3)).astype(self.dtype)
+        x2 = np.random.random((3, 3)).astype(self.dtype)
+        x3 = np.random.random((3, 3)).astype(self.dtype)
+        self.inputs = {'X': [("x0", x0), ("x1", x1), ("x2", x2), ("x3", x3)]}
+        # There will be a problem if just using `y=x0+x1+x2+x3` to calculate the
+        # summation result as the reference standard result. The reason is that 
+        # numpy's fp16 data has precision loss when doing `add` operation.
+        # For example, the results of `x0+x1+x2+x3` is different from that of
+        # `x3+x2+x1+x0` if the dtype is fp16.
+        # Therefore, converting the input to fp32 for calculation.
+        y = (x0.astype(np.float32) + x1.astype(np.float32) +
+             x2.astype(np.float32) + x3.astype(np.float32)).astype(self.dtype)
+        self.outputs = {'Out': y}
+
+        self.attrs = {'use_mkldnn': False}
+
+    def init_dtype(self):
+        self.dtype = np.float16
+
+    def set_mlu(self):
+        self.__class__.use_mlu = True
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+
+class TestSum3(OpTest):
+    def setUp(self):
+        self.set_mlu()
+        self.init_dtype()
+        self.op_type = "sum"
+        self.place = paddle.MLUPlace(0)
+
+        x0 = np.random.random((3, 3)).astype(self.dtype)
+
+        self.inputs = {'X': [("x0", x0)]}
+        y = x0
+        self.outputs = {'Out': y}
+
+        self.attrs = {'use_mkldnn': False}
+
+    def init_dtype(self):
+        self.dtype = np.float16
+
+    def set_mlu(self):
+        self.__class__.use_mlu = True
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place)
+
+
+if __name__ == '__main__':
+    unittest.main()