softmax.cpp

/* 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. */

#ifdef SOFTMAX_OP

#include "operators/math/softmax.h"
#include "common/types.h"
#ifdef __ARM_NEON
#include <math.h>
#include <algorithm>
#include "operators/math/math_func_neon.h"
#endif

namespace paddle_mobile {
namespace operators {
namespace math {
using framework::DDim;
using framework::Tensor;
template <typename T>
class SoftmaxFuntor<CPU, T> {
#ifdef __ARM_NEON
  void sum(float *input, float *sumptr, int inner_size, int outter_size) {
    float32x4_t acc = vdupq_n_f32(0);
    float sum_ = 0;
    for (int i = 0; i < outter_size; ++i) {
      float *input_outer_ptr = input + i * inner_size;
      int nn = inner_size >> 2;
      int left = inner_size - (nn << 2);
      for (; nn > 0; nn--) {
        float32x4_t vec_input = vld1q_f32(input_outer_ptr);
        acc = vaddq_f32(acc, vec_input);
        input_outer_ptr += 4;
      }
      float32x2_t vsum_ = vadd_f32(vget_high_f32(acc), vget_low_f32(acc));
      sum_ = vget_lane_f32(vsum_, 0) + vget_lane_f32(vsum_, 1);
      for (; left > 0; left--) {
        sum_ += *input_outer_ptr;
        input_outer_ptr++;
      }
    }
    for (int j = 0; j < inner_size * outter_size; ++j) {
      sumptr[j] = sum_;
    }
  }

  void SoftmaxCacl(const Tensor *X, Tensor *Y) {
    const float *input = X->data<float>();
    const DDim &dDim = X->dims();
    int axis_index = 1;
    if (dDim.size() < 4) {
      axis_index = 0;
    }
    DDim outer_ddim =
        paddle_mobile::framework::slice_ddim(dDim, 0, axis_index + 1);
    DDim inner_ddim =
        paddle_mobile::framework::slice_ddim(dDim, axis_index + 1, dDim.size());
    int out_size = paddle_mobile::framework::product(outer_ddim);
    int inner_size = paddle_mobile::framework::product(inner_ddim);
    auto *max_ptr = new float[inner_size * out_size];
    // max
    for (int j = 0; j < out_size; ++j) {
      const float *input_outer_ptr = input + j * inner_size;
      float *max_outer_ptr = max_ptr + j * inner_size;
      float max_ = 0;
      for (int i = 0; i < inner_size; ++i) {
        const float *input_inner_ptr = input_outer_ptr + i;
        max_ = std::max(max_, input_inner_ptr[0]);
      }
      for (int k = 0; k < inner_size; ++k) {
        max_outer_ptr[k] = max_;
      }
    }
    // exp(value - max)
    float *exp_sub_max = new float[inner_size * out_size];
    float *exp_sub_max_ptr = &exp_sub_max[0];
    for (int l = 0; l < out_size; ++l) {
      const float *input_outer_ptr = input + l * inner_size;
      float *max_outer_ptr = max_ptr + l * inner_size;
      int nn = inner_size >> 2;
      int left = inner_size - (nn << 2);
      for (; nn > 0; nn--) {
        float32x4_t vec_input = vld1q_f32(input_outer_ptr);
        float32x4_t vec_max = vld1q_f32(max_outer_ptr);
        float32x4_t vec_sub = vsubq_f32(vec_input, vec_max);
        float32x4_t vec_exp = exp_ps(vec_sub);
        vst1q_f32(exp_sub_max_ptr, vec_exp);
        input_outer_ptr += 4;
        max_outer_ptr += 4;
        exp_sub_max_ptr += 4;
      }
      for (; left > 0; left--) {
        *exp_sub_max_ptr = expf(*input_outer_ptr - *max_outer_ptr);

        input_outer_ptr++;
        max_outer_ptr++;
        exp_sub_max_ptr++;
      }
    }
    float *sumptr = new float[inner_size * out_size];
    // sum exp
    sum(exp_sub_max, sumptr, inner_size, out_size);
    // div
    auto *out_ptr = Y->mutable_data<float>();
    for (int l = 0; l < out_size; ++l) {
      const float *input_outer_ptr = exp_sub_max + l * inner_size;
      float *output_outer_ptr = out_ptr + l * inner_size;
      float *sum_outer_ptr = sumptr + l * inner_size;
      int nn = inner_size >> 2;
      int left = inner_size - (nn << 2);
      for (; nn > 0; nn--) {
        float32x4_t vec_input = vld1q_f32(input_outer_ptr);
        float32x4_t vec_sum = vld1q_f32(sum_outer_ptr);
        float32x4_t vec_div = div_ps(vec_input, vec_sum);
        vst1q_f32(output_outer_ptr, vec_div);
        input_outer_ptr += 4;
        output_outer_ptr += 4;
        sum_outer_ptr += 4;
      }
      for (; left > 0; left--) {
        *output_outer_ptr = (*input_outer_ptr) / (*sum_outer_ptr);
        input_outer_ptr++;
        output_outer_ptr++;
        sum_outer_ptr++;
      }
    }
  }
#else
#endif  // ARM_NEON

 public:
  void operator()(const framework::Tensor *X, framework::Tensor *Y) {
    const DDim dDim = X->dims();
    for (int i = 0; i < dDim[0]; ++i) {
      framework::Tensor sub_X = X->Slice(i, i + 1);
      framework::Tensor sub_Y = Y->Slice(i, i + 1);

#ifdef __ARM_NEON
      SoftmaxCacl(&sub_X, &sub_Y);
#endif
    }
  }
};

template class SoftmaxFuntor<CPU, float>;

}  // namespace math
}  // namespace operators
}  // namespace paddle_mobile
#endif