!9168 Add ApplyAdagrad for cpu

From: @yang_chun
Reviewed-by: @wuxuejian,@c_34
Signed-off-by: @c_34

mindspore/ccsrc/backend/kernel_compiler/cpu/apply_adagrad_cpu_kernel.cc

@@ -0,0 +1,111 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * 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 "backend/kernel_compiler/cpu/apply_adagrad_cpu_kernel.h"
+
+#include <thread>
+#include <vector>
+
+namespace mindspore {
+namespace kernel {
+namespace {
+constexpr size_t kSizeFloat16 = 2;
+constexpr size_t kSizeFloat32 = 4;
+constexpr size_t kInputSize = 4;
+constexpr size_t kOutputSize = 2;
+}  // namespace
+void ApplyAdagradCPUKernel::InitKernel(const CNodePtr &kernel_node) {
+  MS_EXCEPTION_IF_NULL(kernel_node);
+  update_slots_ = AnfAlgo::GetNodeAttr<bool>(kernel_node, "update_slots");
+  dtype_ = AnfAlgo::GetPrevNodeOutputInferDataType(kernel_node, 0);
+}
+
+bool ApplyAdagradCPUKernel::Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> & /*workspace*/,
+                                   const std::vector<AddressPtr> &outputs) {
+  CheckParam(inputs, outputs);
+
+  if (dtype_ == kNumberTypeFloat16) {
+    LaunchKernel<float16>(inputs);
+  } else if (dtype_ == kNumberTypeFloat32) {
+    LaunchKernel<float>(inputs);
+  }
+
+  return true;
+}
+
+void ApplyAdagradCPUKernel::CheckParam(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &outputs) {
+  // inputs: var, accum, lr, gradient
+  if (inputs.size() != kInputSize) {
+    MS_LOG(EXCEPTION) << "Input number is " << inputs.size() << ", but ApplyAdagrad needs 4 inputs.";
+  }
+
+  // outputs: var, accum
+  if (outputs.size() != kOutputSize) {
+    MS_LOG(EXCEPTION) << "Output number is " << outputs.size() << ", but ApplyAdagrad needs 2 outputs.";
+  }
+
+  if (inputs[0]->size != inputs[1]->size || inputs[0]->size != inputs[3]->size) {
+    MS_LOG(EXCEPTION) << "Error input data size!";
+  }
+
+  if (inputs[2]->size != kSizeFloat16 && inputs[2]->size != kSizeFloat32) {
+    MS_LOG(EXCEPTION) << "The attribute lr and grad must be float16 or float32!";
+  }
+}
+
+template <typename T>
+void ApplyAdagradCPUKernel::LaunchKernel(const std::vector<AddressPtr> &inputs) {
+  auto var = reinterpret_cast<T *>(inputs[0]->addr);
+  auto accum = reinterpret_cast<T *>(inputs[1]->addr);
+  auto lr = reinterpret_cast<T *>(inputs[2]->addr);
+  auto gradient = reinterpret_cast<T *>(inputs[3]->addr);
+
+  // multithreading
+  size_t length = inputs[0]->size / sizeof(T);
+  size_t max_thread_num = std::thread::hardware_concurrency();
+  size_t use_thread_num = length < 128 * max_thread_num ? std::ceil(length / 128.0) : max_thread_num;
+  std::vector<std::thread> threads;
+  threads.reserve(use_thread_num);
+  size_t start = 0;
+  size_t batch_size = (length + use_thread_num - 1) / use_thread_num;
+
+  while (start < length) {
+    size_t end = (start + batch_size) > length ? length : (start + batch_size);
+    threads.emplace_back(
+      std::thread(&ApplyAdagradCPUKernel::LaunchApplyAdagrad<T>, this, var, accum, *lr, gradient, start, end));
+    start += batch_size;
+  }
+
+  for (auto &it : threads) {
+    it.join();
+  }
+}
+
+template <typename T>
+void ApplyAdagradCPUKernel::LaunchApplyAdagrad(T *var, T *accum, T lr, T *gradient, size_t start, size_t end) {
+  const T one = T(1);
+  const T eps = T(1e-6);
+  for (size_t i = start; i < end; ++i) {
+    // update accum: accum += grad * grad
+    if (update_slots_) {
+      accum[i] += gradient[i] * gradient[i];
+    }
+    // update var: var -= lr * grad * \frac{1}{\sqrt{accum}}
+    var[i] -= lr * gradient[i] * (one / sqrt(accum[i] + eps));
+  }
+}
+}  // namespace kernel
+}  // namespace mindspore

mindspore/ccsrc/backend/kernel_compiler/cpu/apply_adagrad_cpu_kernel.h

@@ -0,0 +1,67 @@
+/**
+ * Copyright 2020 Huawei Technologies Co., Ltd
+ *
+ * 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.
+ */
+#ifndef MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_CPU_APPLY_ADAGRAD_CPU_KERNEL_H_
+#define MINDSPORE_CCSRC_BACKEND_KERNEL_COMPILER_CPU_APPLY_ADAGRAD_CPU_KERNEL_H_
+
+#include <thread>
+#include <vector>
+#include "backend/kernel_compiler/cpu/cpu_kernel.h"
+#include "backend/kernel_compiler/cpu/cpu_kernel_factory.h"
+
+namespace mindspore {
+namespace kernel {
+class ApplyAdagradCPUKernel : public CPUKernel {
+ public:
+  ApplyAdagradCPUKernel() = default;
+  ~ApplyAdagradCPUKernel() override = default;
+
+  void InitKernel(const CNodePtr &kernel_node) override;
+
+  bool Launch(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> & /*workspace*/,
+              const std::vector<AddressPtr> &outputs) override;
+
+ private:
+  static void CheckParam(const std::vector<AddressPtr> &inputs, const std::vector<AddressPtr> &outputs);
+  template <typename T>
+  void LaunchKernel(const std::vector<AddressPtr> &inputs);
+  template <typename T>
+  void LaunchApplyAdagrad(T *var, T *accum, T lr, T *gradient, size_t start, size_t end);
+  bool update_slots_{true};
+  TypeId dtype_{kTypeUnknown};
+};
+
+MS_REG_CPU_KERNEL(ApplyAdagrad,
+                  KernelAttr()
+                    .AddInputAttr(kNumberTypeFloat32)
+                    .AddInputAttr(kNumberTypeFloat32)
+                    .AddInputAttr(kNumberTypeFloat32)
+                    .AddInputAttr(kNumberTypeFloat32)
+                    .AddOutputAttr(kNumberTypeFloat32)
+                    .AddOutputAttr(kNumberTypeFloat32),
+                  ApplyAdagradCPUKernel);
+
+MS_REG_CPU_KERNEL(ApplyAdagrad,
+                  KernelAttr()
+                    .AddInputAttr(kNumberTypeFloat16)
+                    .AddInputAttr(kNumberTypeFloat16)
+                    .AddInputAttr(kNumberTypeFloat16)
+                    .AddInputAttr(kNumberTypeFloat16)
+                    .AddOutputAttr(kNumberTypeFloat16)
+                    .AddOutputAttr(kNumberTypeFloat16),
+                  ApplyAdagradCPUKernel);
+}  // namespace kernel
+}  // namespace mindspore
+#endif

tests/st/ops/cpu/test_apply_adagrad_op.py

@@ -0,0 +1,61 @@
+# Copyright 2020 Huawei Technologies Co., Ltd
+#
+# 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.
+# ============================================================================
+
+import numpy as np
+import pytest
+
+import mindspore.context as context
+import mindspore.nn as nn
+from mindspore import Tensor, Parameter
+from mindspore.ops import operations as P
+import mindspore.common.dtype as mstype
+
+context.set_context(mode=context.GRAPH_MODE, device_target="CPU")
+
+var_np = np.random.rand(3, 3).astype(np.float32)
+accum_np = np.random.rand(3, 3).astype(np.float32)
+
+
+class Net(nn.Cell):
+    def __init__(self):
+        super(Net, self).__init__()
+        self.apply_adagrad = P.ApplyAdagrad()
+        self.var = Parameter(Tensor(var_np), name="var")
+        self.accum = Parameter(Tensor(accum_np), name="accum")
+
+    def construct(self, lr, grad):
+        self.apply_adagrad(self.var, self.accum, lr, grad)
+        return self.var, self.accum
+
+
+@pytest.mark.level0
+@pytest.mark.platform_x86_cpu
+@pytest.mark.env_onecard
+def test_apply_adagrad():
+    # numpy op
+    grident_np = np.random.rand(3, 3).astype(np.float32)
+    expect_accum_np = accum_np + grident_np * grident_np
+    expect_var_np = var_np - (0.001 * grident_np * (1 / np.sqrt(expect_accum_np + 1e-6)))
+
+    net = Net()
+    lr = Tensor(0.001, mstype.float32)
+    grad = Tensor(grident_np)
+    out = net(lr, grad)
+    res_var_mindspore = out[0].asnumpy()
+    res_accum_mindspore = out[1].asnumpy()
+    eps = np.array([1e-6 for i in range(9)]).reshape(3, 3)
+
+    assert np.all(expect_var_np - res_var_mindspore < eps)
+    assert np.all(expect_accum_np - res_accum_mindspore < eps)