Merge pull request #6967 from reyoung/feature/optimize_adam_speed

Use for_range to rewrite adam

paddle/operators/adam_op.h

@@ -13,59 +13,113 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
-#include "paddle/framework/eigen.h"
+#include <math.h>  // for sqrt in CPU and CUDA
 #include "paddle/framework/op_registry.h"
+#include "paddle/operators/detail/safe_ref.h"
+#include "paddle/platform/for_range.h"
 
 namespace paddle {
 namespace operators {
 
+template <typename T>
+struct AdamFunctor {
+  T beta1_;
+  T beta2_;
+  T epsilon_;
+
+  const T* beta1_pow_;
+  const T* beta2_pow_;
+  const T* moment1_;
+  T* moment1_out_;
+  const T* moment2_;
+  T* moment2_out_;
+  const T* lr_;
+  const T* grad_;
+  const T* param_;
+  T* param_out_;
+
+  AdamFunctor(T beta1, T beta2, T epsilon, const T* beta1_pow,
+              const T* beta2_pow, const T* mom1, T* mom1_out, const T* mom2,
+              T* mom2_out, const T* lr, const T* grad, const T* param,
+              T* param_out)
+      : beta1_(beta1),
+        beta2_(beta2),
+        epsilon_(epsilon),
+        beta1_pow_(beta1_pow),
+        beta2_pow_(beta2_pow),
+        moment1_(mom1),
+        moment1_out_(mom1_out),
+        moment2_(mom2),
+        moment2_out_(mom2_out),
+        lr_(lr),
+        grad_(grad),
+        param_(param),
+        param_out_(param_out) {}
+
+  inline HOSTDEVICE void operator()(size_t i) const {
+    // Merge all memory access together.
+    T g = grad_[i];
+    T mom1 = moment1_[i];
+    T mom2 = moment2_[i];
+    T lr = *lr_;
+    T beta1_pow = *beta1_pow_;
+    T beta2_pow = *beta2_pow_;
+    T p = param_[i];
+
+    // Calculation
+    lr *= sqrt(1 - beta2_pow) / (1 - beta1_pow);
+    mom1 = beta1_ * mom1 + (1 - beta1_) * g;
+    mom2 = beta2_ * mom2 + (1 - beta2_) * g * g;
+    p -= lr * (mom1 / (sqrt(mom2) + epsilon_));
+
+    // Write back to global memory
+    moment1_out_[i] = mom1;
+    moment2_out_[i] = mom2;
+    param_out_[i] = p;
+  }
+};
+
 template <typename DeviceContext, typename T>
 class AdamOpKernel : public framework::OpKernel<T> {
  public:
   void Compute(const framework::ExecutionContext& ctx) const override {
-    auto param_out_tensor = ctx.Output<framework::Tensor>("ParamOut");
-    auto moment1_out_tensor = ctx.Output<framework::Tensor>("Moment1Out");
-    auto moment2_out_tensor = ctx.Output<framework::Tensor>("Moment2Out");
-
-    param_out_tensor->mutable_data<T>(ctx.GetPlace());
-    moment1_out_tensor->mutable_data<T>(ctx.GetPlace());
-    moment2_out_tensor->mutable_data<T>(ctx.GetPlace());
+    using paddle::framework::LoDTensor;
+    using paddle::operators::detail::Ref;
 
     T beta1 = static_cast<T>(ctx.Attr<float>("beta1"));
     T beta2 = static_cast<T>(ctx.Attr<float>("beta2"));
     T epsilon = static_cast<T>(ctx.Attr<float>("epsilon"));
+    auto& param = Ref(ctx.Input<LoDTensor>("Param"), "Must set Param");
+    auto& grad = Ref(ctx.Input<LoDTensor>("Grad"), "Must set Grad");
+    auto& mom1 = Ref(ctx.Input<LoDTensor>("Moment1"), "Must set Moment1");
+    auto& mom2 = Ref(ctx.Input<LoDTensor>("Moment2"), "Must set Moment2");
+    auto& lr =
+        Ref(ctx.Input<LoDTensor>("LearningRate"), "Must set LearningRate");
+
+    auto& beta1_pow =
+        Ref(ctx.Input<LoDTensor>("Beta1Pow"), "Must set Beta1Pow");
+    auto& beta2_pow =
+        Ref(ctx.Input<LoDTensor>("Beta2Pow"), "Must set Beta2Pow");
+
+    auto& param_out =
+        Ref(ctx.Output<LoDTensor>("ParamOut"), "Must set ParamOut");
+    auto& mom1_out =
+        Ref(ctx.Output<LoDTensor>("Moment1Out"), "Must set Moment1Out");
+    auto& mom2_out =
+        Ref(ctx.Output<LoDTensor>("Moment2Out"), "Must set Moment1Out");
 
-    auto param = framework::EigenVector<T>::Flatten(
-        *ctx.Input<framework::Tensor>("Param"));
-    auto grad = framework::EigenVector<T>::Flatten(
-        *ctx.Input<framework::Tensor>("Grad"));
-    auto moment1 = framework::EigenVector<T>::Flatten(
-        *ctx.Input<framework::Tensor>("Moment1"));
-    auto moment2 = framework::EigenVector<T>::Flatten(
-        *ctx.Input<framework::Tensor>("Moment2"));
-    auto lr = framework::EigenVector<T>::Flatten(
-        *ctx.Input<framework::Tensor>("LearningRate"));
-    auto beta1_pow = framework::EigenVector<T>::Flatten(
-        *ctx.Input<framework::Tensor>("Beta1Pow"));
-    auto beta2_pow = framework::EigenVector<T>::Flatten(
-        *ctx.Input<framework::Tensor>("Beta2Pow"));
-    auto param_out = framework::EigenVector<T>::Flatten(*param_out_tensor);
-    auto moment1_out = framework::EigenVector<T>::Flatten(*moment1_out_tensor);
-    auto moment2_out = framework::EigenVector<T>::Flatten(*moment2_out_tensor);
-    auto* place = ctx.template device_context<DeviceContext>().eigen_device();
-
-    moment1_out.device(*place) = beta1 * moment1 + (1 - beta1) * grad;
-    moment2_out.device(*place) = beta2 * moment2 + (1 - beta2) * grad.square();
-
-    // All of these are tensors of 1 element
-    auto lr_t = lr * (1 - beta2_pow).sqrt() / (1 - beta1_pow);
-    // Eigen does not support automatic broadcast
-    // Get dimensions of moment vector to broadcast lr_t
-    Eigen::DSizes<int, 1> m_dsize(moment1_out_tensor->numel());
-    param_out.device(*place) =
-        param -
-        lr_t.broadcast(m_dsize) *
-            (moment1_out / (moment2_out.sqrt() + epsilon));
+    AdamFunctor<T> functor(beta1, beta2, epsilon, beta1_pow.template data<T>(),
+                           beta2_pow.template data<T>(),
+                           mom1.template data<T>(),
+                           mom1_out.template mutable_data<T>(ctx.GetPlace()),
+                           mom2.template data<T>(),
+                           mom2_out.template mutable_data<T>(ctx.GetPlace()),
+                           lr.template data<T>(), grad.template data<T>(),
+                           param.template data<T>(),
+                           param_out.template mutable_data<T>(ctx.GetPlace()));
+    platform::ForRange<DeviceContext> for_range(
+        static_cast<const DeviceContext&>(ctx.device_context()), param.numel());
+    for_range(functor);
   }
 };
 

paddle/platform/for_range.h

@@ -0,0 +1,85 @@
+/* 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. */
+
+#pragma once
+#include "paddle/platform/device_context.h"
+
+namespace paddle {
+namespace platform {
+
+template <typename DeviceContext>
+struct ForRange {
+  ForRange(const DeviceContext& dev_ctx, size_t limit);
+
+  template <typename Function>
+  void operator()(Function func) const;
+};
+
+template <>
+struct ForRange<CPUDeviceContext> {
+  ForRange(const CPUDeviceContext& dev_ctx, size_t limit) : limit_(limit) {}
+
+  template <typename Function>
+  void operator()(Function func) const {
+    for (size_t i = 0; i < limit_; ++i) {
+      func(i);
+    }
+  }
+
+  size_t limit_;
+};
+
+#ifdef __NVCC__
+template <typename Function>
+__global__ static void ForRangeElemwiseOpGridIsOne(Function func) {
+  size_t idx = static_cast<size_t>(threadIdx.x);
+  func(idx);
+}
+
+template <typename Function>
+__global__ static void ForRangeElemwiseOp(Function func, int limit) {
+  size_t idx = static_cast<size_t>(blockIdx.x * blockDim.x + threadIdx.x);
+  if (idx < limit) {
+    func(idx);
+  }
+}
+
+template <>
+struct ForRange<CUDADeviceContext> {
+  ForRange(const CUDADeviceContext& dev_ctx, size_t limit)
+      : dev_ctx_(dev_ctx), limit_(static_cast<int>(limit)) {}
+
+  template <typename Function>
+  inline void operator()(Function func) const {
+    constexpr size_t num_threads = 1024;
+    int block_size = limit_ <= num_threads ? limit_ : num_threads;
+    int grid_size = (limit_ + num_threads - 1) / num_threads;
+
+    if (grid_size == 1) {
+      ForRangeElemwiseOpGridIsOne<<<1, block_size, 0, dev_ctx_.stream()>>>(
+          func);
+    } else {
+      ForRangeElemwiseOp<<<grid_size, block_size, 0, dev_ctx_.stream()>>>(
+          func, limit_);
+    }
+  }
+
+  const CUDADeviceContext& dev_ctx_;
+  int limit_;
+};
+
+#endif
+
+}  // namespace platform
+}  // namespace paddle