paddlepaddle--paddle/paddle/phi/kernels/funcs/softmax_impl.h

/* 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 <vector>

#include "paddle/phi/backends/cpu/cpu_context.h"
#include "paddle/phi/backends/cpu/cpu_info.h"
#include "paddle/phi/backends/gpu/gpu_context.h"
#include "paddle/phi/core/dense_tensor.h"
#include "paddle/phi/kernels/funcs/cpu_vec.h"
#include "paddle/phi/kernels/funcs/eigen/common.h"

namespace phi {
namespace funcs {

template <typename T, int MajorType = Eigen::RowMajor>
using EigenMatrix = EigenMatrix<T, MajorType>;

template <typename T>
struct ValueClip {
  HOSTDEVICE T operator()(const T& x) const {
    const T kThreshold = static_cast<T>(-64.);
    return x < kThreshold ? kThreshold : x;
  }
};

template <typename DeviceContext, typename T>
class SoftmaxEigen {
 public:
  void operator()(const DeviceContext& dev_ctx,
                  const int axis_dim,
                  const DenseTensor* X,
                  DenseTensor* Y) {
    constexpr int64_t kBatchDim = 0;
    constexpr int64_t kClassDim = 1;
    constexpr int64_t kAxisDim = 1;

    auto logits = EigenMatrix<T>::From(*X);
    auto softmax = EigenMatrix<T>::From(*Y);

    const int64_t batch_size = logits.dimension(kBatchDim);
    const int64_t num_classes = logits.dimension(kClassDim);
    const int64_t num_remain = num_classes / axis_dim;

    Eigen::DSizes<int64_t, 1> along_axis(kAxisDim);
    Eigen::DSizes<int64_t, 2> batch_classes(batch_size, num_classes);
    Eigen::DSizes<int64_t, 2> batch_by_one(batch_size, 1);
    Eigen::DSizes<int64_t, 2> one_by_class(1, num_classes);
    Eigen::DSizes<int64_t, 3> batch_one_remain(batch_size, 1, num_remain);
    Eigen::DSizes<int64_t, 3> one_axis_one(1, axis_dim, 1);
    Eigen::DSizes<int64_t, 2> one_axis(1, axis_dim);
    Eigen::DSizes<int64_t, 3> batch_axis_remain(
        batch_size, axis_dim, num_remain);

    // For numerical stability, logits should be shifted by maximum number along
    // axis, calculate shifted_logits into softmax tensor for memory reuse.
    if (num_remain == 1) {
      // axis == -1, axis and class in same dimension, calculate along
      // class dimension directly for higher performance
      softmax.device(*dev_ctx.eigen_device()) =
          (logits - logits.maximum(along_axis)
                        .eval()
                        .reshape(batch_by_one)
                        .broadcast(one_by_class))
              .unaryExpr(ValueClip<T>());
    } else {
      // axis != -1, class dimension split into (axis, remain), max and sum
      // should be calculated along axis dimension
      softmax.device(*dev_ctx.eigen_device()) =
          (logits.reshape(batch_classes) - logits.reshape(batch_axis_remain)
                                               .maximum(along_axis)
                                               .eval()
                                               .reshape(batch_one_remain)
                                               .broadcast(one_axis_one)
                                               .reshape(batch_classes))
              .unaryExpr(ValueClip<T>());
    }

    softmax.device(*dev_ctx.eigen_device()) = softmax.exp();
    softmax.device(*dev_ctx.eigen_device()) =
        (softmax * softmax.reshape(batch_axis_remain)
                       .sum(along_axis)
                       .inverse()
                       .eval()
                       .broadcast(one_axis));
  }
};

template <typename DeviceContext>
class SoftmaxEigen<DeviceContext, phi::float16> {
 public:
  void operator()(const DeviceContext& dev_ctx,
                  const int axis_dim,
                  const DenseTensor* X,
                  DenseTensor* Y) {
    constexpr int64_t kBatchDim = 0;
    constexpr int64_t kClassDim = 1;
    constexpr int64_t kAxisDim = 1;

    auto logits = EigenMatrix<phi::float16>::From(*X);
    auto softmax = EigenMatrix<phi::float16>::From(*Y);

    const int64_t batch_size = logits.dimension(kBatchDim);
    const int64_t num_classes = logits.dimension(kClassDim);
    const int64_t num_remain = num_classes / axis_dim;

    Eigen::DSizes<int64_t, 1> along_axis(kAxisDim);
    Eigen::DSizes<int64_t, 2> batch_classes(batch_size, num_classes);
    Eigen::DSizes<int64_t, 2> batch_by_one(batch_size, 1);
    Eigen::DSizes<int64_t, 2> one_by_class(1, num_classes);
    Eigen::DSizes<int64_t, 3> batch_one_remain(batch_size, 1, num_remain);
    Eigen::DSizes<int64_t, 3> one_axis_one(1, axis_dim, 1);
    Eigen::DSizes<int64_t, 2> one_axis(1, axis_dim);
    Eigen::DSizes<int64_t, 3> batch_axis_remain(
        batch_size, axis_dim, num_remain);

    // For numerical stability, logits should be shifted by maximum number along
    // axis, calculate shifted_logits into softmax tensor for memory reuse.
    if (num_remain == 1) {
      // axis == -1, axis and class in same dimension, calculate along
      // class dimension directly for higher performance
      softmax.device(*dev_ctx.eigen_device()) =
          (logits - logits.maximum(along_axis)
                        .reshape(batch_by_one)
                        .broadcast(one_by_class))
              .unaryExpr(ValueClip<phi::float16>());
    } else {
      // axis != -1, class dimension split into (axis, remain), max and sum
      // should be calculated along axis dimension
      softmax.device(*dev_ctx.eigen_device()) =
          (logits.reshape(batch_axis_remain) - logits.reshape(batch_axis_remain)
                                                   .maximum(along_axis)
                                                   .reshape(batch_one_remain)
                                                   .broadcast(one_axis_one)
                                                   .reshape(batch_classes))
              .unaryExpr(ValueClip<phi::float16>());
    }

    softmax.device(*dev_ctx.eigen_device()) = softmax.exp();
    softmax.device(*dev_ctx.eigen_device()) =
        (softmax * softmax.reshape(batch_axis_remain)
                       .sum(along_axis)
                       .inverse()
                       .broadcast(one_axis));
  }
};

template <typename DeviceContext>
class SoftmaxEigen<DeviceContext, phi::bfloat16> {
 public:
  void operator()(const DeviceContext& dev_ctx,
                  const int axis_dim,
                  const DenseTensor* X,
                  DenseTensor* Y) {
    constexpr int64_t kBatchDim = 0;
    constexpr int64_t kClassDim = 1;
    constexpr int64_t kAxisDim = 1;

    auto logits = EigenMatrix<phi::bfloat16>::From(*X);
    auto softmax = EigenMatrix<phi::bfloat16>::From(*Y);

    const int64_t batch_size = logits.dimension(kBatchDim);
    const int64_t num_classes = logits.dimension(kClassDim);
    const int64_t num_remain = num_classes / axis_dim;

    Eigen::DSizes<int64_t, 1> along_axis(kAxisDim);
    Eigen::DSizes<int64_t, 2> batch_classes(batch_size, num_classes);
    Eigen::DSizes<int64_t, 2> batch_by_one(batch_size, 1);
    Eigen::DSizes<int64_t, 2> one_by_class(1, num_classes);
    Eigen::DSizes<int64_t, 3> batch_one_remain(batch_size, 1, num_remain);
    Eigen::DSizes<int64_t, 3> one_axis_one(1, axis_dim, 1);
    Eigen::DSizes<int64_t, 2> one_axis(1, axis_dim);
    Eigen::DSizes<int64_t, 3> batch_axis_remain(
        batch_size, axis_dim, num_remain);

    // For numerical stability, logits should be shifted by maximum number along
    // axis, calculate shifted_logits into softmax tensor for memory reuse.
    if (num_remain == 1) {
      // axis == -1, axis and class in same dimension, calculate along
      // class dimension directly for higher performance
      softmax.device(*dev_ctx.eigen_device()) =
          (logits - logits.maximum(along_axis)
                        .reshape(batch_by_one)
                        .broadcast(one_by_class))
              .unaryExpr(ValueClip<phi::bfloat16>());
    } else {
      // axis != -1, class dimension split into (axis, remain), max and sum
      // should be calculated along axis dimension
      softmax.device(*dev_ctx.eigen_device()) =
          (logits.reshape(batch_axis_remain) - logits.reshape(batch_axis_remain)
                                                   .maximum(along_axis)
                                                   .reshape(batch_one_remain)
                                                   .broadcast(one_axis_one)
                                                   .reshape(batch_classes))
              .unaryExpr(ValueClip<phi::bfloat16>());
    }

    softmax.device(*dev_ctx.eigen_device()) = softmax.exp();
    softmax.device(*dev_ctx.eigen_device()) =
        (softmax * softmax.reshape(batch_axis_remain)
                       .sum(along_axis)
                       .inverse()
                       .broadcast(one_axis));
  }
};

template <typename DeviceContext, typename T, typename Enable>
void SoftmaxFunctor<DeviceContext, T, Enable>::operator()(
    const DeviceContext& dev_ctx,
    const int axis_dim,
    const DenseTensor* X,
    DenseTensor* Y) {
  SoftmaxEigen<DeviceContext, T>()(dev_ctx, axis_dim, X, Y);
}

template <class DeviceContext>
using enable_if_CPU = typename std::enable_if<
    std::is_same<DeviceContext, CPUContext>::value>::type;

template <typename DeviceContext, typename T>
class SoftmaxFunctor<DeviceContext, T, enable_if_CPU<DeviceContext>> {
 public:
  void operator()(const DeviceContext& dev_ctx,
                  const int axis_dim,
                  const DenseTensor* X,
                  DenseTensor* Y) {
    const auto& in_dims = X->dims();
    constexpr int64_t kBatchDim = 0;
    constexpr int64_t kClassDim = 1;

    const int64_t num_classes = in_dims[kClassDim];
    const int64_t batch_size = in_dims[kBatchDim];
    const int64_t num_remain = num_classes / axis_dim;

    if (num_remain == 1 &&
        phi::backends::cpu::MayIUse(phi::backends::cpu::avx)) {
      const T* in_data = X->data<T>();
      T* out_data = Y->data<T>();
      for (int64_t bs = 0; bs < batch_size; ++bs) {
        T max_val = *std::max_element(in_data, in_data + num_classes);
        max_val *= static_cast<T>(-1);
        vec_add_bias<T, phi::backends::cpu::avx>(
            num_classes, max_val, in_data, out_data);
        vec_clip<T, phi::backends::cpu::avx>(
            num_classes, static_cast<T>(-64), out_data, out_data);
        vec_exp<T>(num_classes, out_data, out_data);

        T sum = 0;
        vec_sum<T, phi::backends::cpu::avx>(num_classes, out_data, &sum);
        sum = static_cast<T>(1) / sum;
        vec_scal<T, phi::backends::cpu::avx>(
            num_classes, sum, out_data, out_data);

        in_data += num_classes;
        out_data += num_classes;
      }
    } else {
      SoftmaxEigen<DeviceContext, T>()(dev_ctx, axis_dim, X, Y);
    }
  }
};

template <typename DeviceContext, typename T>
class SoftmaxGradEigen {
 public:
  void operator()(const DeviceContext& dev_ctx,
                  const int axis_dim,
                  const DenseTensor* y,
                  const DenseTensor* y_grad,
                  DenseTensor* x_grad) {
    auto softmax = EigenMatrix<T>::From(*y);
    auto softmax_grad = EigenMatrix<T>::From(*y_grad);
    auto logits_grad = EigenMatrix<T>::From(*x_grad);

    constexpr int64_t kBatchDim = 0;
    constexpr int64_t kClassDim = 1;

    const int64_t batch_size = softmax.dimension(kBatchDim);
    const int64_t num_classes = softmax.dimension(kClassDim);
    const int64_t num_remain = num_classes / axis_dim;

    Eigen::DSizes<int64_t, 1> along_class(kClassDim);
    Eigen::DSizes<int64_t, 2> batch_by_one(batch_size, 1);
    Eigen::DSizes<int64_t, 2> one_by_class(1, num_classes);
    Eigen::DSizes<int64_t, 3> batch_axis_remain(
        batch_size, axis_dim, num_remain);
    Eigen::DSizes<int64_t, 2> one_axis(1, axis_dim);

    auto dot = (softmax * softmax_grad)
                   .reshape(batch_axis_remain)
                   .sum(along_class)
                   .eval()
                   .broadcast(one_axis);
    logits_grad.device(*dev_ctx.eigen_device()) =
        (softmax_grad - dot) * softmax;
  }
};

template <typename DeviceContext>
class SoftmaxGradEigen<DeviceContext, phi::float16> {
 public:
  void operator()(const DeviceContext& dev_ctx,
                  const int axis_dim,
                  const DenseTensor* y,
                  const DenseTensor* y_grad,
                  DenseTensor* x_grad) {
    auto softmax = EigenMatrix<phi::float16>::From(*y);
    auto softmax_grad = EigenMatrix<phi::float16>::From(*y_grad);
    auto logits_grad = EigenMatrix<phi::float16>::From(*x_grad);

    constexpr int64_t kBatchDim = 0;
    constexpr int64_t kClassDim = 1;

    const int64_t batch_size = softmax.dimension(kBatchDim);
    const int64_t num_classes = softmax.dimension(kClassDim);
    const int64_t num_remain = num_classes / axis_dim;

    Eigen::DSizes<int64_t, 1> along_class(kClassDim);
    Eigen::DSizes<int64_t, 2> batch_by_one(batch_size, 1);
    Eigen::DSizes<int64_t, 2> one_by_class(1, num_classes);
    Eigen::DSizes<int64_t, 3> batch_axis_remain(
        batch_size, axis_dim, num_remain);
    Eigen::DSizes<int64_t, 2> one_axis(1, axis_dim);

    auto dot = (softmax * softmax_grad)
                   .reshape(batch_axis_remain)
                   .sum(along_class)
                   .broadcast(one_axis);
    logits_grad.device(*dev_ctx.eigen_device()) =
        (softmax_grad - dot) * softmax;
  }
};

template <typename DeviceContext>
class SoftmaxGradEigen<DeviceContext, phi::bfloat16> {
 public:
  void operator()(const DeviceContext& dev_ctx,
                  const int axis_dim,
                  const DenseTensor* y,
                  const DenseTensor* y_grad,
                  DenseTensor* x_grad) {
    auto softmax = EigenMatrix<phi::bfloat16>::From(*y);
    auto softmax_grad = EigenMatrix<phi::bfloat16>::From(*y_grad);
    auto logits_grad = EigenMatrix<phi::bfloat16>::From(*x_grad);

    constexpr int64_t kBatchDim = 0;
    constexpr int64_t kClassDim = 1;

    const int64_t batch_size = softmax.dimension(kBatchDim);
    const int64_t num_classes = softmax.dimension(kClassDim);
    const int64_t num_remain = num_classes / axis_dim;

    Eigen::DSizes<int64_t, 1> along_class(kClassDim);
    Eigen::DSizes<int64_t, 2> batch_by_one(batch_size, 1);
    Eigen::DSizes<int64_t, 2> one_by_class(1, num_classes);
    Eigen::DSizes<int64_t, 3> batch_axis_remain(
        batch_size, axis_dim, num_remain);
    Eigen::DSizes<int64_t, 2> one_axis(1, axis_dim);

    auto dot = (softmax * softmax_grad)
                   .reshape(batch_axis_remain)
                   .sum(along_class)
                   .broadcast(one_axis);
    logits_grad.device(*dev_ctx.eigen_device()) =
        (softmax_grad - dot) * softmax;
  }
};

template <typename DeviceContext, typename T, typename Enable>
void SoftmaxGradFunctor<DeviceContext, T, Enable>::operator()(
    const DeviceContext& dev_ctx,
    const int axis_dim,
    const DenseTensor* y,
    const DenseTensor* y_grad,
    DenseTensor* x_grad) {
  SoftmaxGradEigen<DeviceContext, T>()(dev_ctx, axis_dim, y, y_grad, x_grad);
}

template <typename DeviceContext, typename T>
class SoftmaxGradFunctor<DeviceContext, T, enable_if_CPU<DeviceContext>> {
 public:
  void operator()(const DeviceContext& dev_ctx,
                  const int axis_dim,
                  const DenseTensor* y,
                  const DenseTensor* y_grad,
                  DenseTensor* x_grad) {
    const auto& out_dims = y->dims();
    constexpr int64_t kBatchDim = 0;
    constexpr int64_t kClassDim = 1;
    const int64_t num_classes = out_dims[kClassDim];
    const int64_t batch_size = out_dims[kBatchDim];
    const int64_t num_remain = num_classes / axis_dim;

    if (num_remain == 1 &&
        phi::backends::cpu::MayIUse(phi::backends::cpu::avx)) {
      const T* out_data = y->data<T>();
      const T* out_grad = y_grad->data<T>();
      T* in_grad = x_grad->data<T>();
      for (int64_t bs = 0; bs < batch_size; ++bs) {
        T scalar;
        vec_mul_reduce<T, phi::backends::cpu::avx>(
            num_classes, out_grad, out_data, &scalar);
        scalar *= static_cast<T>(-1);
        vec_add_bias<T, phi::backends::cpu::avx>(
            num_classes, scalar, out_grad, in_grad);
        vec_mul<T, phi::backends::cpu::avx>(
            num_classes, out_data, in_grad, in_grad);
        out_data += num_classes;
        out_grad += num_classes;
        in_grad += num_classes;
      }
    } else {
      SoftmaxGradEigen<DeviceContext, T>()(
          dev_ctx, axis_dim, y, y_grad, x_grad);
    }
  }
};

}  // namespace funcs
}  // namespace phi