[NPU] add npu kernel for concat op (#31695)

* add npu kernel for concat op

* add npu kernel for concat op

* refine code

* update

* refine concat_grad

paddle/fluid/operators/concat_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/concat_op.h"
+#include "paddle/fluid/operators/npu_op_runner.h"
+
+namespace paddle {
+namespace operators {
+
+template <typename T>
+class ConcatNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto ins = ctx.MultiInput<framework::LoDTensor>("X");
+    framework::LoDTensor* out = ctx.Output<framework::LoDTensor>("Out");
+    PADDLE_ENFORCE_NOT_NULL(ins[0],
+                            platform::errors::NotFound(
+                                "The first input tensor is not initalized."));
+    auto axis = ctx.Attr<int>("axis");
+
+    if (ctx.HasInput("AxisTensor")) {
+      PADDLE_THROW(platform::errors::NotFound(
+          "The AxisTensor is not supported on NPU now."));
+    }
+    axis = ComputeAxis(static_cast<int64_t>(axis),
+                       static_cast<int64_t>(ins[0]->dims().size()));
+
+    auto place = ctx.GetPlace();
+    out->mutable_data<T>(place);
+
+    std::vector<framework::Tensor> inputs;
+    std::vector<std::string> names;
+    for (size_t i = 0; i < ins.size(); ++i) {
+      if (ins[i] && ins[i]->numel() > 0) {
+        inputs.push_back(*ins[i]);
+        names.push_back("x" + std::to_string(i));
+      } else {
+        continue;
+      }
+    }
+    auto stream =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>()
+            .stream();
+    auto runner = NpuOpRunner(
+        "ConcatD", {inputs}, {*out},
+        {{"concat_dim", axis}, {"N", static_cast<int>(inputs.size())}});
+    runner.AddInputNames(names);
+    runner.Run(stream);
+  }
+};
+
+template <typename T>
+class ConcatGradNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* out_grad =
+        ctx.Input<framework::Tensor>(framework::GradVarName("Out"));
+    auto ins = ctx.MultiInput<framework::LoDTensor>("X");
+    auto out_var_names = ctx.OutputNames(framework::GradVarName("X"));
+    auto outs =
+        ctx.MultiOutput<framework::LoDTensor>(framework::GradVarName("X"));
+
+    {
+      auto dx = outs;
+      auto x = ins;
+      for (size_t i = 0; i < dx.size(); ++i) {
+        if (dx[i] != nullptr) {
+          dx[i]->set_lod(x[i]->lod());
+        }
+      }
+    }
+    PADDLE_ENFORCE_NOT_NULL(ins[0],
+                            platform::errors::NotFound(
+                                "The first input tensor is not initalized."));
+
+    auto axis = ctx.Attr<int>("axis");
+
+    axis = ComputeAxis(static_cast<int64_t>(axis),
+                       static_cast<int64_t>(ins[0]->dims().size()));
+    // get output tensor that the name is not kEmptyVarName
+    std::vector<framework::Tensor> outputs;
+    std::vector<int> sizes;
+    for (size_t j = 0; j < outs.size(); ++j) {
+      if (out_var_names[j] != framework::kEmptyVarName &&
+          outs[j]->numel() != 0UL) {
+        outs[j]->mutable_data<T>(ctx.GetPlace());
+        outputs.push_back(*outs[j]);
+        sizes.push_back(outs[j]->dims()[axis]);
+      }
+    }
+    auto runner =
+        NpuOpRunner("SplitVD", {*out_grad}, outputs,
+                    {{"split_dim", axis},
+                     {"size_splits", sizes},
+                     {"num_split", static_cast<int>(outputs.size())}});
+    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(concat, ops::ConcatNPUKernel<float>,
+                       ops::ConcatNPUKernel<paddle::platform::float16>,
+                       ops::ConcatNPUKernel<int>);
+
+REGISTER_OP_NPU_KERNEL(concat_grad, ops::ConcatGradNPUKernel<float>,
+                       ops::ConcatGradNPUKernel<paddle::platform::float16>,
+                       ops::ConcatGradNPUKernel<int>);

paddle/fluid/operators/npu_op_runner.cc

@@ -256,6 +256,9 @@ aclTensorDesc *NpuOpRunner::CreateTensorDesc(Tensor tensor) {
   auto *desc = aclCreateTensorDesc(dtype, dims.size(), dims.data(), format);
   PADDLE_ENFORCE_NOT_NULL(
       desc, platform::errors::External("Call aclCreateTensorDesc failed."));
+  PADDLE_ENFORCE_NPU_SUCCESS(aclSetTensorStorageFormat(desc, format));
+  PADDLE_ENFORCE_NPU_SUCCESS(
+      aclSetTensorStorageShape(desc, dims.size(), dims.data()));
   return desc;
 }
 

python/paddle/fluid/tests/unittests/npu/test_concat_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
+
+paddle.enable_static()
+SEED = 2021
+
+
+@unittest.skipIf(not paddle.is_compiled_with_npu(),
+                 "core is not compiled with NPU")
+class TestConcat(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.op_type = "concat"
+        self.place = paddle.NPUPlace(4)
+        self.init_dtype()
+        self.init_test_data()
+
+        self.inputs = {'X': [('x0', self.x0), ('x1', self.x1), ('x2', self.x2)]}
+        self.attrs = {'axis': self.axis}
+        if self.axis < 0:
+            self.actual_axis = self.axis + len(self.x0.shape)
+            self.actual_axis = self.actual_axis if self.actual_axis > 0 else 0
+        else:
+            self.actual_axis = self.axis
+
+        self.outputs = {
+            'Out': np.concatenate(
+                (self.x0, self.x1, self.x2), axis=self.actual_axis)
+        }
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def init_dtype(self):
+        self.dtype = np.float32
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place, check_dygraph=False)
+
+    def init_test_data(self):
+        self.x0 = np.random.random((1, 4, 50)).astype(self.dtype)
+        self.x1 = np.random.random((2, 4, 50)).astype(self.dtype)
+        self.x2 = np.random.random((3, 4, 50)).astype(self.dtype)
+        self.axis = 0
+
+    def test_check_grad(self):
+        self.check_grad_with_place(
+            self.place, ['x0'], 'Out', check_dygraph=False)
+        self.check_grad_with_place(
+            self.place, ['x1'], 'Out', check_dygraph=False)
+        self.check_grad_with_place(
+            self.place, ['x2'], 'Out', check_dygraph=False)
+
+
+class TestConcatFP16(OpTest):
+    def setUp(self):
+        self.set_npu()
+        self.op_type = "concat"
+        self.place = paddle.NPUPlace(4)
+        self.init_dtype()
+        self.init_test_data()
+
+        self.inputs = {'X': [('x0', self.x0), ('x1', self.x1), ('x2', self.x2)]}
+        self.attrs = {'axis': self.axis}
+        if self.axis < 0:
+            self.actual_axis = self.axis + len(self.x0.shape)
+            self.actual_axis = self.actual_axis if self.actual_axis > 0 else 0
+        else:
+            self.actual_axis = self.axis
+
+        self.outputs = {
+            'Out': np.concatenate(
+                (self.x0, self.x1, self.x2), axis=self.actual_axis)
+        }
+
+    def set_npu(self):
+        self.__class__.use_npu = True
+        self.__class__.no_need_check_grad = True
+
+    def init_dtype(self):
+        self.dtype = np.float16
+
+    def test_check_output(self):
+        self.check_output_with_place(self.place, check_dygraph=False)
+
+    def init_test_data(self):
+        self.x0 = np.random.random((1, 4, 50)).astype(self.dtype)
+        self.x1 = np.random.random((2, 4, 50)).astype(self.dtype)
+        self.x2 = np.random.random((3, 4, 50)).astype(self.dtype)
+        self.axis = 0
+
+
+if __name__ == '__main__':
+    unittest.main()