[NPU] add one_hot_op_npu and tests (#34258)

* add one_hot_op and tests

* update

* make code clear

paddle/fluid/operators/npu_op_runner.cc

@@ -240,6 +240,38 @@ NpuOpRunner &NpuOpRunner::AddInput(std::vector<int64_t> &&dims) {
   return *this;
 }
 
+NpuOpRunner &NpuOpRunner::AddInput(std::vector<float> &&values) {
+  platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
+  auto *dev_ctx =
+      static_cast<platform::CPUDeviceContext *>(pool.Get(platform::CPUPlace()));
+  Tensor host_tensor;
+  TensorFromVector(values, *dev_ctx, &host_tensor);
+  host_tensors_.emplace_back(host_tensor);
+
+  // create aclTensorDesc
+  input_descs_.emplace_back(CreateTensorDesc(host_tensor, ACL_MEMTYPE_HOST));
+  // create aclDataBuffer
+  input_buffers_.emplace_back(CreateDataBuffer(host_tensor));
+
+  return *this;
+}
+
+NpuOpRunner &NpuOpRunner::AddInput(std::vector<double> &&values) {
+  platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
+  auto *dev_ctx =
+      static_cast<platform::CPUDeviceContext *>(pool.Get(platform::CPUPlace()));
+  Tensor host_tensor;
+  TensorFromVector(values, *dev_ctx, &host_tensor);
+  host_tensors_.emplace_back(host_tensor);
+
+  // create aclTensorDesc
+  input_descs_.emplace_back(CreateTensorDesc(host_tensor, ACL_MEMTYPE_HOST));
+  // create aclDataBuffer
+  input_buffers_.emplace_back(CreateDataBuffer(host_tensor));
+
+  return *this;
+}
+
 NpuOpRunner &NpuOpRunner::AddOutput(const Tensor &tensor) {
   // create aclTensorDesc
   output_descs_.emplace_back(CreateTensorDesc(tensor));

paddle/fluid/operators/npu_op_runner.h

@@ -71,6 +71,10 @@ class NpuOpRunner {
 
   NpuOpRunner &AddInput(std::vector<int64_t> &&dims);
 
+  NpuOpRunner &AddInput(std::vector<float> &&values);
+
+  NpuOpRunner &AddInput(std::vector<double> &&values);
+
   NpuOpRunner &AddOutput(const Tensor &tensor);
 
   NpuOpRunner &AddInputs(const std::vector<Tensor> &tensors);

paddle/fluid/operators/one_hot_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/one_hot_op.h"
+
+#include "paddle/fluid/operators/npu_op_runner.h"
+
+namespace paddle {
+namespace operators {
+using Tensor = framework::Tensor;
+
+template <typename T>
+class OneHotNPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto& dev_ctx =
+        ctx.template device_context<paddle::platform::NPUDeviceContext>();
+    auto* in = ctx.Input<LoDTensor>("X");
+    auto* out = ctx.Output<LoDTensor>("Out");
+    int depth = ctx.Attr<int>("depth");
+
+    if (ctx.HasInput("depth_tensor")) {
+      auto* depth_tensor = ctx.Input<Tensor>("depth_tensor");
+      std::vector<int32_t> depth_data;
+      framework::TensorToVector(*depth_tensor, dev_ctx, &depth_data);
+      depth = depth_data[0];
+      auto in_dims = in->dims();
+      framework::DDim out_dims(in_dims);
+      out_dims[out_dims.size() - 1] = depth;
+      out->Resize(out_dims);
+    }
+    out->mutable_data<float>(ctx.GetPlace());
+
+    float on_value = 1.0f, off_value = 0.0f;
+    if (in->type() == framework::proto::VarType::INT32) {
+      NpuOpRunner runner;
+      runner.SetType("OneHot")
+          .AddInput(*in)
+          .AddInput(std::vector<int32_t>({static_cast<int32_t>(depth)}))
+          .AddInput(std::vector<float>({on_value}))
+          .AddInput(std::vector<float>({off_value}))
+          .AddAttr("axis", -1)
+          .AddOutput(*out);
+      runner.Run(dev_ctx.stream());
+    } else {
+      Tensor transformed_in;
+      transformed_in.mutable_data<int32_t>(in->dims(), dev_ctx.GetPlace());
+      const auto& cast_runner = NpuOpRunner("Cast", {*in}, {transformed_in},
+                                            {{"dst_type", ACL_INT32}});
+      cast_runner.Run(dev_ctx.stream());
+      NpuOpRunner runner;
+      runner.SetType("OneHot")
+          .AddInput(transformed_in)
+          .AddInput(std::vector<int32_t>({static_cast<int32_t>(depth)}))
+          .AddInput(std::vector<float>({on_value}))
+          .AddInput(std::vector<float>({off_value}))
+          .AddAttr("axis", -1)
+          .AddOutput(*out);
+      runner.Run(dev_ctx.stream());
+    }
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+
+namespace ops = paddle::operators;
+namespace plat = paddle::platform;
+
+REGISTER_OP_NPU_KERNEL(one_hot, ops::OneHotNPUKernel<int32_t>,
+                       ops::OneHotNPUKernel<int64_t>);

python/paddle/fluid/tests/unittests/npu/test_one_hot_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("..")
+
+from op_test import OpTest
+import paddle
+import paddle.fluid as fluid
+import paddle.fluid.core as core
+from paddle.fluid.framework import Program, program_guard
+
+paddle.enable_static()
+
+
+class TestOneHotOp(OpTest):
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def setUp(self):
+        self.set_npu()
+        self.op_type = 'one_hot'
+        depth = 10
+        depth_np = np.array(10).astype('int32')
+        dimension = 12
+        x_lod = [[4, 1, 3, 3]]
+        x = [np.random.randint(0, depth - 1) for i in range(sum(x_lod[0]))]
+        x = np.array(x).astype('int32').reshape([sum(x_lod[0]), 1])
+
+        out = np.zeros(shape=(np.product(x.shape[:-1]),
+                              depth)).astype('float32')
+
+        for i in range(np.product(x.shape)):
+            out[i, x[i]] = 1.0
+
+        self.inputs = {'X': (x, x_lod), 'depth_tensor': depth_np}
+        self.attrs = {'dtype': int(core.VarDesc.VarType.FP32)}
+        self.outputs = {'Out': (out, x_lod)}
+
+    def test_check_output(self):
+        self.check_output_with_place(paddle.NPUPlace(0), check_dygraph=False)
+
+
+class TestOneHotOp_attr(OpTest):
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def setUp(self):
+        self.set_npu()
+        self.op_type = 'one_hot'
+        depth = 10
+        dimension = 12
+        x_lod = [[4, 1, 3, 3]]
+        x = [np.random.randint(0, depth - 1) for i in range(sum(x_lod[0]))]
+        x = np.array(x).astype('int32').reshape([sum(x_lod[0]), 1])
+
+        out = np.zeros(shape=(np.product(x.shape[:-1]),
+                              depth)).astype('float32')
+
+        for i in range(np.product(x.shape)):
+            out[i, x[i]] = 1.0
+
+        self.inputs = {'X': (x, x_lod)}
+        self.attrs = {'dtype': int(core.VarDesc.VarType.FP32), 'depth': depth}
+        self.outputs = {'Out': (out, x_lod)}
+
+    def test_check_output(self):
+        self.check_output_with_place(paddle.NPUPlace(0), check_dygraph=False)
+
+
+class TestOneHotOp_default_dtype(OpTest):
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def setUp(self):
+        self.set_npu()
+        self.op_type = 'one_hot'
+        depth = 10
+        depth_np = np.array(10).astype('int32')
+        dimension = 12
+        x_lod = [[4, 1, 3, 3]]
+        x = [np.random.randint(0, depth - 1) for i in range(sum(x_lod[0]))]
+        x = np.array(x).astype('int32').reshape([sum(x_lod[0]), 1])
+
+        out = np.zeros(shape=(np.product(x.shape[:-1]),
+                              depth)).astype('float32')
+
+        for i in range(np.product(x.shape)):
+            out[i, x[i]] = 1.0
+
+        self.inputs = {'X': (x, x_lod), 'depth_tensor': depth_np}
+        self.attrs = {}
+        self.outputs = {'Out': (out, x_lod)}
+
+    def test_check_output(self):
+        self.check_output_with_place(paddle.NPUPlace(0), check_dygraph=False)
+
+
+class TestOneHotOp_default_dtype_attr(OpTest):
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def setUp(self):
+        self.set_npu()
+        self.op_type = 'one_hot'
+        depth = 10
+        dimension = 12
+        x_lod = [[4, 1, 3, 3]]
+        x = [np.random.randint(0, depth - 1) for i in range(sum(x_lod[0]))]
+        x = np.array(x).astype('int32').reshape([sum(x_lod[0]), 1])
+
+        out = np.zeros(shape=(np.product(x.shape[:-1]),
+                              depth)).astype('float32')
+
+        for i in range(np.product(x.shape)):
+            out[i, x[i]] = 1.0
+
+        self.inputs = {'X': (x, x_lod)}
+        self.attrs = {'depth': depth}
+        self.outputs = {'Out': (out, x_lod)}
+
+    def test_check_output(self):
+        self.check_output_with_place(paddle.NPUPlace(0), check_dygraph=False)
+
+
+class TestOneHotOp_out_of_range(OpTest):
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def setUp(self):
+        self.set_npu()
+        self.op_type = 'one_hot'
+        depth = 10
+        x_lod = [[4, 1, 3, 3]]
+        x = [np.random.choice([-1, depth]) for i in range(sum(x_lod[0]))]
+        x = np.array(x).astype('int32').reshape([sum(x_lod[0]), 1])
+
+        out = np.zeros(shape=(np.product(x.shape[:-1]),
+                              depth)).astype('float32')
+
+        self.inputs = {'X': (x, x_lod)}
+        self.attrs = {'depth': depth, 'allow_out_of_range': True}
+        self.outputs = {'Out': (out, x_lod)}
+
+    def test_check_output(self):
+        self.check_output_with_place(paddle.NPUPlace(0), check_dygraph=False)
+
+
+class TestOneHotOp_dtype_int64(OpTest):
+    def set_npu(self):
+        self.__class__.use_npu = True
+
+    def setUp(self):
+        self.set_npu()
+        self.op_type = 'one_hot'
+        depth = 10
+        dimension = 12
+        x_lod = [[4, 1, 3, 3]]
+        x = [np.random.randint(0, depth - 1) for i in range(sum(x_lod[0]))]
+        x = np.array(x).astype('int64').reshape([sum(x_lod[0]), 1])
+
+        out = np.zeros(shape=(np.product(x.shape[:-1]),
+                              depth)).astype('float32')
+
+        for i in range(np.product(x.shape)):
+            out[i, x[i]] = 1.0
+
+        self.inputs = {'X': (x, x_lod)}
+        self.attrs = {'depth': depth}
+        self.outputs = {'Out': (out, x_lod)}
+
+    def test_check_output(self):
+        self.check_output_with_place(paddle.NPUPlace(0), check_dygraph=False)
+
+
+if __name__ == '__main__':
+    paddle.enable_static()
+    unittest.main()