Add dropout and log_loss for kunlun

paddle/fluid/operators/dropout_op_xpu.cc

@@ -0,0 +1,126 @@
+/* Copyright (c) 2016 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/dropout_op.h"
+#include <memory>
+#include <string>
+#include "paddle/fluid/platform/xpu_header.h"
+namespace paddle {
+namespace operators {
+
+#ifdef PADDLE_WITH_XPU
+static std::map<int, float*> mask_data_tables;
+static const int max_data_size = 32 * 1024 * 1024;
+static std::mutex s_mask_data_table_lock;
+template <typename DeviceContext, typename T>
+class DropoutXPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    auto* x = context.Input<Tensor>("X");
+    auto* y = context.Output<Tensor>("Out");
+    const auto* x_data = x->data<T>();
+    auto* y_data = y->mutable_data<T>(context.GetPlace());
+    float dropout_prob = context.Attr<float>("dropout_prob");
+    auto dropout_implementation =
+        context.Attr<std::string>("dropout_implementation");
+    float* mask_data_table = nullptr;
+    PADDLE_ENFORCE_EQ(!context.HasInput("Seed"), true,
+                      platform::errors::InvalidArgument(
+                          ("Input(Seed) not supported on XPU")));
+    if (!context.Attr<bool>("is_test")) {
+      int dev_id = boost::get<platform::XPUPlace>(context.GetPlace()).device;
+      int prop = static_cast<int>(dropout_prob * 100);
+      int is_upscale = (dropout_implementation == "upscale_in_train");
+      /* mask_data_tables key contains 3 part:
+       *  | 31-16  | 15-8 | 7-0        |
+       *  | dev_id | prob | is_upscale |
+       */
+      int index = (dev_id << 16) + (prop << 8) + is_upscale;
+      std::lock_guard<std::mutex> lock(s_mask_data_table_lock);
+      if (mask_data_tables.find(index) == mask_data_tables.end()) {
+        float* mask_data_host = new float[max_data_size];
+        std::random_device rnd;
+        std::minstd_rand engine;
+        int seed =
+            context.Attr<bool>("fix_seed") ? context.Attr<int>("seed") : rnd();
+        engine.seed(seed);
+        std::uniform_real_distribution<float> dist(0, 1);
+        for (size_t i = 0; i < max_data_size; ++i) {
+          if (dist(engine) < dropout_prob) {
+            mask_data_host[i] = 0.0f;
+          } else {
+            if (is_upscale) {
+              mask_data_host[i] = 1.0f / static_cast<T>(1.0f - dropout_prob);
+            } else {
+              mask_data_host[i] = 1.0;
+            }
+          }
+        }
+        PADDLE_ENFORCE(
+            xpu_malloc(reinterpret_cast<void**>(&mask_data_table),
+                       max_data_size * sizeof(float)) == xpu::Error_t::SUCCESS,
+            "XPU no enough memory");
+        memory::Copy(boost::get<platform::XPUPlace>(context.GetPlace()),
+                     mask_data_table, platform::CPUPlace(), mask_data_host,
+                     max_data_size * sizeof(float));
+        mask_data_tables[index] = mask_data_table;
+        free(mask_data_host);
+      } else {
+        mask_data_table = mask_data_tables[index];
+      }
+    }
+    if (!context.Attr<bool>("is_test")) {  // Train
+      auto* mask = context.Output<Tensor>("Mask");
+      auto* mask_data = mask->mutable_data<T>(context.GetPlace());
+      size_t size = framework::product(mask->dims());
+      auto& dev_ctx = context.template device_context<DeviceContext>();
+      int r = xpu::dropout(dev_ctx.x_context(), mask_data_table, x_data,
+                           mask_data, y_data, max_data_size, size);
+      PADDLE_ENFORCE(r == xpu::Error_t::SUCCESS, "XPU kernel error!");
+    } else {  // Infer
+      float scale = 0.0f;
+      if (dropout_implementation == "upscale_in_train") {
+        scale = 1.0f;
+      } else {
+        scale = static_cast<T>(1.0f - dropout_prob);
+      }
+      auto& dev_ctx = context.template device_context<DeviceContext>();
+      int r = xpu::scale(dev_ctx.x_context(), x->numel(), scale, 0.0f, 0,
+                         x_data, y_data);
+      PADDLE_ENFORCE(r == xpu::Error_t::SUCCESS, "XPU kernel error!");
+    }
+  }
+};
+template <typename DeviceContext, typename T>
+class DropoutGradXPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& context) const override {
+    PADDLE_ENFORCE(!context.Attr<bool>("is_test"),
+                   "GradOp is only callable when is_test is false");
+    auto* grad_x = context.Output<Tensor>(framework::GradVarName("X"));
+    auto* grad_y = context.Input<Tensor>(framework::GradVarName("Out"));
+    auto* mask = context.Input<Tensor>("Mask");
+    grad_x->mutable_data<T>(context.GetPlace());
+    auto& dev_ctx = context.template device_context<DeviceContext>();
+    int r = xpu::elementwise_mul(dev_ctx.x_context(), grad_y->data<T>(),
+                                 mask->data<T>(), grad_x->data<T>(),
+                                 grad_y->numel());
+    PADDLE_ENFORCE(r == xpu::Error_t::SUCCESS, "XPU kernel error!");
+  }
+};
+}  // namespace operators
+}  // namespace paddle
+namespace ops = paddle::operators;
+REGISTER_OP_XPU_KERNEL(
+    dropout, ops::DropoutXPUKernel<paddle::platform::XPUDeviceContext, float>);
+REGISTER_OP_XPU_KERNEL(
+    dropout_grad,
+    ops::DropoutGradXPUKernel<paddle::platform::XPUDeviceContext, float>);
+#endif

paddle/fluid/operators/log_loss_op_xpu.cc

@@ -0,0 +1,66 @@
+/* Copyright (c) 2016 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. */
+#ifdef PADDLE_WITH_XPU
+
+#include "paddle/fluid/operators/log_loss_op.h"
+#include <memory>
+namespace paddle {
+namespace operators {
+
+template <typename DeviceContext, typename T, typename AttrType = T>
+class LogLossXPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* predict = ctx.Input<Tensor>("Predicted");
+    auto* labels = ctx.Input<Tensor>("Labels");
+    auto* loss = ctx.Output<Tensor>("Loss");
+    auto epsilon = static_cast<T>(ctx.Attr<AttrType>("epsilon"));
+    loss->mutable_data<T>(ctx.GetPlace());
+    int n = predict->numel();
+    auto& dev_ctx = ctx.template device_context<DeviceContext>();
+    int r =
+        xpu::log_loss_fwd(dev_ctx.x_context(), n, epsilon, predict->data<T>(),
+                          labels->data<T>(), loss->data<T>());
+    PADDLE_ENFORCE(r == xpu::Error_t::SUCCESS, "XPU kernel error!");
+  }
+};
+template <typename DeviceContext, typename T, typename AttrType = T>
+class LogLossGradXPUKernel : public framework::OpKernel<T> {
+ public:
+  void Compute(const framework::ExecutionContext& ctx) const override {
+    auto* predict = ctx.Input<Tensor>("Predicted");
+    auto* labels = ctx.Input<Tensor>("Labels");
+    auto* dloss = ctx.Input<Tensor>(framework::GradVarName("Loss"));
+    auto* dpred = ctx.Output<Tensor>(framework::GradVarName("Predicted"));
+    if (!dpred) {
+      return;
+    }
+    auto epsilon = static_cast<T>(ctx.Attr<AttrType>("epsilon"));
+    dpred->mutable_data<T>(ctx.GetPlace());
+    int n = predict->numel();
+    auto& dev_ctx = ctx.template device_context<DeviceContext>();
+    int r = xpu::log_loss_bwd(dev_ctx.x_context(), n, epsilon,
+                              predict->data<T>(), labels->data<T>(),
+                              dloss->data<T>(), dpred->data<T>());
+    PADDLE_ENFORCE(r == xpu::Error_t::SUCCESS, "XPU kernel error!");
+  }
+};
+
+}  // namespace operators
+}  // namespace paddle
+namespace ops = paddle::operators;
+REGISTER_OP_XPU_KERNEL(
+    log_loss, ops::LogLossXPUKernel<paddle::platform::XPUDeviceContext, float>);
+REGISTER_OP_XPU_KERNEL(
+    log_loss_grad,
+    ops::LogLossGradXPUKernel<paddle::platform::XPUDeviceContext, float>);
+
+#endif

python/paddle/fluid/tests/unittests/xpu/test_dropout_op_xpu.py

@@ -0,0 +1,187 @@
+#   Copyright (c) 2018 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.fluid.core as core
+from op_test import OpTest, skip_check_grad_ci
+import paddle
+import paddle.fluid as fluid
+from paddle.fluid import Program, program_guard
+
+
+@unittest.skipIf(not paddle.is_compiled_with_xpu(),
+                 "core is not compiled with XPU")
+class TestDropoutOp(OpTest):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.inputs = {'X': np.random.random((32, 64)).astype("float32")}
+        self.attrs = {'dropout_prob': 0.0, 'fix_seed': True, 'is_test': False}
+        self.outputs = {
+            'Out': self.inputs['X'],
+            'Mask': np.ones((32, 64)).astype('uint8')
+        }
+
+    def test_check_output(self):
+        if paddle.is_compiled_with_xpu():
+            paddle.enable_static()
+            place = paddle.XPUPlace(0)
+            self.check_output_with_place(place)
+
+    def test_check_grad_normal(self):
+        if paddle.is_compiled_with_xpu():
+            paddle.enable_static()
+            place = paddle.XPUPlace(0)
+            self.check_grad_with_place(place, ['X'], 'Out')
+
+
+class TestDropoutOpInput1d(OpTest):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.inputs = {'X': np.random.random((2000, )).astype("float32")}
+        self.attrs = {'dropout_prob': 0.0, 'fix_seed': True, 'is_test': False}
+        self.outputs = {
+            'Out': self.inputs['X'],
+            'Mask': np.ones((2000)).astype('uint8')
+        }
+
+    def test_check_output(self):
+        if paddle.is_compiled_with_xpu():
+            paddle.enable_static()
+            place = paddle.XPUPlace(0)
+            self.check_output_with_place(place)
+
+    def test_check_grad_normal(self):
+        if paddle.is_compiled_with_xpu():
+            paddle.enable_static()
+            place = paddle.XPUPlace(0)
+            self.check_grad_with_place(place, ['X'], 'Out')
+
+
+class TestDropoutOp2(TestDropoutOp):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.inputs = {'X': np.random.random((32, 64)).astype("float32")}
+        self.attrs = {'dropout_prob': 1.0, 'fix_seed': True, 'is_test': False}
+        self.outputs = {
+            'Out': np.zeros((32, 64)).astype('float32'),
+            'Mask': np.zeros((32, 64)).astype('uint8')
+        }
+
+
+class TestDropoutOp3(TestDropoutOp):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.inputs = {'X': np.random.random((32, 64, 2)).astype("float32")}
+        self.attrs = {'dropout_prob': 0.0, 'fix_seed': True, 'is_test': False}
+        self.outputs = {
+            'Out': self.inputs['X'],
+            'Mask': np.ones((32, 64, 2)).astype('uint8')
+        }
+
+
+@skip_check_grad_ci(reason="For inference, check_grad is not required.")
+class TestDropoutOp4(OpTest):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.inputs = {'X': np.random.random((32, 64)).astype("float32")}
+        self.attrs = {'dropout_prob': 0.35, 'fix_seed': True, 'is_test': True}
+        self.outputs = {
+            'Out': self.inputs['X'] * (1.0 - self.attrs['dropout_prob'])
+        }
+
+    def test_check_output(self):
+        if paddle.is_compiled_with_xpu():
+            paddle.enable_static()
+            place = paddle.XPUPlace(0)
+            self.check_output_with_place(place)
+
+
+@skip_check_grad_ci(reason="For inference, check_grad is not required.")
+class TestDropoutOp5(OpTest):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.inputs = {'X': np.random.random((32, 64, 3)).astype("float32")}
+        self.attrs = {'dropout_prob': 0.75, 'is_test': True}
+        self.outputs = {
+            'Out': self.inputs['X'] * (1.0 - self.attrs['dropout_prob'])
+        }
+
+    def test_check_output(self):
+        if paddle.is_compiled_with_xpu():
+            paddle.enable_static()
+            place = paddle.XPUPlace(0)
+            self.check_output_with_place(place)
+
+
+class TestDropoutOp6(TestDropoutOp):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.inputs = {'X': np.random.random((32, 64, 2)).astype("float32")}
+        self.attrs = {
+            'dropout_prob': 0.0,
+            'fix_seed': True,
+            'is_test': False,
+            'dropout_implementation': 'upscale_in_train'
+        }
+        self.outputs = {
+            'Out': self.inputs['X'],
+            'Mask': np.ones((32, 64, 2)).astype('uint8')
+        }
+
+
+@skip_check_grad_ci(reason="For inference, check_grad is not required.")
+class TestDropoutOp7(OpTest):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.inputs = {'X': np.random.random((32, 64)).astype("float32")}
+        self.attrs = {
+            'dropout_prob': 0.35,
+            'fix_seed': True,
+            'is_test': True,
+            'dropout_implementation': 'upscale_in_train'
+        }
+        self.outputs = {'Out': self.inputs['X']}
+
+    def test_check_output(self):
+        if paddle.is_compiled_with_xpu():
+            paddle.enable_static()
+            place = paddle.XPUPlace(0)
+            self.check_output_with_place(place)
+
+
+@skip_check_grad_ci(reason="For inference, check_grad is not required.")
+class TestDropoutOp8(OpTest):
+    def setUp(self):
+        self.op_type = "dropout"
+        self.inputs = {'X': np.random.random((32, 64, 3)).astype("float32")}
+        self.attrs = {
+            'dropout_prob': 0.75,
+            'is_test': True,
+            'dropout_implementation': 'upscale_in_train'
+        }
+        self.outputs = {'Out': self.inputs['X']}
+
+    def test_check_output(self):
+        if paddle.is_compiled_with_xpu():
+            paddle.enable_static()
+            place = paddle.XPUPlace(0)
+            self.check_output_with_place(place)
+
+
+if __name__ == '__main__':
+    unittest.main()