internal/optimized/depthwiseconv_float.h

/*
 * Copyright (C) 2017 The Android Open Source Project
 *
 * 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 ANDROID_ML_NN_COMMON_OPERATIONS_INTERNAL_OPTIMIZED_DEPTHWISECONV_FLOAT_H_
#define ANDROID_ML_NN_COMMON_OPERATIONS_INTERNAL_OPTIMIZED_DEPTHWISECONV_FLOAT_H_

#include "gemmlowp.h"
#include "../common.h"
#include "../types.h"

namespace android {
namespace nn {
namespace optimized_ops {

// Implementation of float DepthwiseConv

template <bool kAllowStrided, int kFixedInputDepth, int kFixedDepthMultiplier>
struct FloatDepthwiseConvKernel {};

#ifdef USE_NEON

template <>
struct FloatDepthwiseConvKernel<false, 8, 1> {
  static void Run(int num_output_pixels, int input_depth, int depth_multiplier,
                  const float* input_ptr, int input_ptr_increment,
                  const float* filter_ptr, float* acc_buffer_ptr) {
    // Load the filters
    float32x4_t filter[2];
    for (int i = 0; i < 2; i++) {
      filter[i] = vld1q_f32(filter_ptr + 4 * i);
    }
    int outp = 0;
    // Handle 2 output pixels at a time.
    for (; outp <= num_output_pixels - 2; outp += 2) {
      // Load the inputs
      float32x4_t input[4];
      for (int i = 0; i < 4; i++) {
        input[i] = vld1q_f32(input_ptr + 4 * i);
      }
      input_ptr += 16;
      // Load the accumulators from acc_buffer
      float32x4_t acc[4];
      for (int i = 0; i < 4; i++) {
        acc[i] = vld1q_f32(acc_buffer_ptr + 4 * i);
      }
      // Multiply-accumulate
      acc[0] = vmlaq_f32(acc[0], input[0], filter[0]);
      acc[1] = vmlaq_f32(acc[1], input[1], filter[1]);
      acc[2] = vmlaq_f32(acc[2], input[2], filter[0]);
      acc[3] = vmlaq_f32(acc[3], input[3], filter[1]);
      // Store the accumulators back to acc_buffer
      for (int i = 0; i < 4; i++) {
        vst1q_f32(acc_buffer_ptr + 4 * i, acc[i]);
      }
      acc_buffer_ptr += 16;
    }
    // Handle one output pixel at a time.
    for (; outp < num_output_pixels; outp++) {
      // Load the inputs
      float32x4_t input[2];
      for (int i = 0; i < 2; i++) {
        input[i] = vld1q_f32(input_ptr + 4 * i);
      }
      input_ptr += 8;
      // Load the accumulators from acc_buffer
      float32x4_t acc[2];
      for (int i = 0; i < 2; i++) {
        acc[i] = vld1q_f32(acc_buffer_ptr + 4 * i);
      }
      // Multiply-accumulate
      for (int i = 0; i < 2; i++) {
        acc[i] = vmlaq_f32(acc[i], input[i], filter[i]);
      }
      // Store the accumulators back to acc_buffer
      for (int i = 0; i < 2; i++) {
        vst1q_f32(acc_buffer_ptr + 4 * i, acc[i]);
      }
      acc_buffer_ptr += 8;
    }
  }
};

template <>
struct FloatDepthwiseConvKernel<false, 2, 1> {
  static void Run(int num_output_pixels, int input_depth, int depth_multiplier,
                  const float* input_ptr, int input_ptr_increment,
                  const float* filter_ptr, float* acc_buffer_ptr) {
    const float32x2_t filters = vld1_f32(filter_ptr);
    const float32x4_t filters_dup2 = vcombine_f32(filters, filters);
    int outp = 0;
    // Handle 8 output pixels at a time.
    for (; outp <= num_output_pixels - 8; outp += 8) {
      // Load the inputs
      float32x4_t input[4];
      for (int i = 0; i < 4; i++) {
        input[i] = vld1q_f32(input_ptr + 4 * i);
      }
      input_ptr += 16;
      // Load the accumulators from acc_buffer
      float32x4_t acc[4];
      for (int i = 0; i < 4; i++) {
        acc[i] = vld1q_f32(acc_buffer_ptr + 4 * i);
      }
      // Multiply-accumulate
      for (int i = 0; i < 4; i++) {
        acc[i] = vmlaq_f32(acc[i], input[i], filters_dup2);
      }
      // Store the accumulators back to acc_buffer
      for (int i = 0; i < 4; i++) {
        vst1q_f32(acc_buffer_ptr + 4 * i, acc[i]);
      }
      acc_buffer_ptr += 16;
    }
    // Handle 4 output pixels at a time.
    for (; outp <= num_output_pixels - 4; outp += 4) {
      // Load the inputs
      float32x4_t input[2];
      for (int i = 0; i < 2; i++) {
        input[i] = vld1q_f32(input_ptr + 4 * i);
      }
      input_ptr += 8;
      // Load the accumulators from acc_buffer
      float32x4_t acc[2];
      for (int i = 0; i < 2; i++) {
        acc[i] = vld1q_f32(acc_buffer_ptr + 4 * i);
      }
      // Multiply-accumulate
      for (int i = 0; i < 2; i++) {
        acc[i] = vmlaq_f32(acc[i], input[i], filters_dup2);
      }
      // Store the accumulators back to acc_buffer
      for (int i = 0; i < 2; i++) {
        vst1q_f32(acc_buffer_ptr + 4 * i, acc[i]);
      }
      acc_buffer_ptr += 8;
    }
    // Handle 2 output pixels at a time.
    for (; outp <= num_output_pixels - 2; outp += 2) {
      // Load the inputs
      const float32x4_t input = vld1q_f32(input_ptr);
      input_ptr += 4;
      // Load the accumulators from acc_buffer
      float32x4_t acc = vld1q_f32(acc_buffer_ptr);
      // Multiply-accumulate
      acc = vmlaq_f32(acc, input, filters_dup2);
      // Store the accumulators back to acc_buffer
      vst1q_f32(acc_buffer_ptr, acc);
      acc_buffer_ptr += 4;
    }
    // Handle 1 output pixel at a time
    for (; outp < num_output_pixels; outp++) {
      // Load the inputs
      const float32x2_t input = vld1_f32(input_ptr);
      input_ptr += 2;
      // Load the accumulators from acc_buffer
      float32x2_t acc = vld1_f32(acc_buffer_ptr);
      // Multiply-accumulate
      acc = vmla_f32(acc, input, filters);
      // Store the accumulators back to acc_buffer
      vst1_f32(acc_buffer_ptr, acc);
      acc_buffer_ptr += 2;
    }
  }
};

template <>
struct FloatDepthwiseConvKernel<true, 0, 1> {
  static void Run(int num_output_pixels, int input_depth, int depth_multiplier,
                  const float* input_ptr, int input_ptr_increment,
                  const float* filter_ptr, float* acc_buffer_ptr) {
    // Handle one output pixel at a time.
    for (int outp = 0; outp < num_output_pixels; outp++) {
      const float* local_filter_ptr = filter_ptr;
      const float* local_input_ptr = input_ptr;
      int ic = 0;
      // Handle 16 input channels at a time.
      for (; ic <= input_depth - 16; ic += 16) {
        // Load the filters
        float32x4_t filter[4];
        for (int i = 0; i < 4; i++) {
          filter[i] = vld1q_f32(local_filter_ptr + 4 * i);
        }
        local_filter_ptr += 16;
        // Load the inputs
        float32x4_t input[4];
        for (int i = 0; i < 4; i++) {
          input[i] = vld1q_f32(local_input_ptr + 4 * i);
        }
        local_input_ptr += 16;
        // Load the accumulators from acc_buffer
        float32x4_t acc[4];
        for (int i = 0; i < 4; i++) {
          acc[i] = vld1q_f32(acc_buffer_ptr + 4 * i);
        }
        // Multiply-accumulate
        for (int i = 0; i < 4; i++) {
          acc[i] = vmlaq_f32(acc[i], input[i], filter[i]);
        }
        // Store the accumulators back to acc_buffer
        for (int i = 0; i < 4; i++) {
          vst1q_f32(acc_buffer_ptr + 4 * i, acc[i]);
        }
        acc_buffer_ptr += 16;
      }
      // Handle 4 input channels at a time.
      for (; ic <= input_depth - 4; ic += 4) {
        // Load the filters
        float32x4_t filter;
        filter = vld1q_f32(local_filter_ptr);
        local_filter_ptr += 4;
        // Load the inputs
        float32x4_t input;
        input = vld1q_f32(local_input_ptr);
        local_input_ptr += 4;
        // Load the accumulators from acc_buffer
        float32x4_t acc;
        acc = vld1q_f32(acc_buffer_ptr);
        // Multiply-accumulate
        acc = vmlaq_f32(acc, input, filter);
        // Store the accumulators back to acc_buffer
        vst1q_f32(acc_buffer_ptr, acc);
        acc_buffer_ptr += 4;
      }
      // Handle one input channel at a time.
      for (; ic < input_depth; ic++) {
        const float input_val = *local_input_ptr++;
        const float filter_val = *local_filter_ptr++;
        *acc_buffer_ptr++ += filter_val * input_val;
      }
      input_ptr += input_ptr_increment;
    }
  }
};

template <>
struct FloatDepthwiseConvKernel<true, 0, 8> {
  static void Run(int num_output_pixels, int input_depth, int depth_multiplier,
                  const float* input_ptr, int input_ptr_increment,
                  const float* filter_ptr, float* acc_buffer_ptr) {
    // Handle one output pixel at a time.
    for (int outp = 0; outp < num_output_pixels; outp++) {
      const float* local_filter_ptr = filter_ptr;
      const float* local_input_ptr = input_ptr;
      int ic = 0;
      // Handle 2 input channels at a time.
      for (; ic <= input_depth - 2; ic += 2) {
        // Load the filters
        float32x4_t filter[4];
        for (int i = 0; i < 4; i++) {
          filter[i] = vld1q_f32(local_filter_ptr + 4 * i);
        }
        local_filter_ptr += 16;
        // Load the inputs
        const float32x2_t input = vld1_f32(local_input_ptr);
        local_input_ptr += 2;
        // Load the accumulators from acc_buffer
        float32x4_t acc[4];
        for (int i = 0; i < 4; i++) {
          acc[i] = vld1q_f32(acc_buffer_ptr + 4 * i);
        }
        // Multiply-accumulate
        acc[0] = vmlaq_lane_f32(acc[0], filter[0], input, 0);
        acc[1] = vmlaq_lane_f32(acc[1], filter[1], input, 0);
        acc[2] = vmlaq_lane_f32(acc[2], filter[2], input, 1);
        acc[3] = vmlaq_lane_f32(acc[3], filter[3], input, 1);
        // Store the accumulators back to acc_buffer
        for (int i = 0; i < 4; i++) {
          vst1q_f32(acc_buffer_ptr + 4 * i, acc[i]);
        }
        acc_buffer_ptr += 16;
      }
      // Handle one input channel at a time.
      for (; ic < input_depth; ic++) {
        // Load the filters
        float32x4_t filter[2];
        for (int i = 0; i < 2; i++) {
          filter[i] = vld1q_f32(local_filter_ptr + 4 * i);
        }
        local_filter_ptr += 8;
        // Load the inputs
        const float input_val = *local_input_ptr++;
        // Load the accumulators from acc_buffer
        float32x4_t acc[2];
        for (int i = 0; i < 2; i++) {
          acc[i] = vld1q_f32(acc_buffer_ptr + 4 * i);
        }
        // Multiply-accumulate
        for (int i = 0; i < 2; i++) {
          acc[i] = vmlaq_n_f32(acc[i], filter[i], input_val);
        }
        // Store the accumulators back to acc_buffer
        for (int i = 0; i < 2; i++) {
          vst1q_f32(acc_buffer_ptr + 4 * i, acc[i]);
        }
        acc_buffer_ptr += 8;
      }
      input_ptr += input_ptr_increment;
    }
  }
};

template <>
struct FloatDepthwiseConvKernel<true, 0, 2> {
  static void Run(int num_output_pixels, int input_depth, int depth_multiplier,
                  const float* input_ptr, int input_ptr_increment,
                  const float* filter_ptr, float* acc_buffer_ptr) {
    // Handle one output pixel at a time.
    for (int outp = 0; outp < num_output_pixels; outp++) {
      const float* local_filter_ptr = filter_ptr;
      const float* local_input_ptr = input_ptr;
      int ic = 0;
      // Handle 8 input channels at a time.
      for (; ic <= input_depth - 8; ic += 8) {
        // Load the filters
        float32x4_t filter[4];
        for (int i = 0; i < 4; i++) {
          filter[i] = vld1q_f32(local_filter_ptr + 4 * i);
        }
        local_filter_ptr += 16;
        // Load the inputs
        float32x4x2_t input_dup2[2];
        for (int i = 0; i < 2; i++) {
          const float32x4_t input = vld1q_f32(local_input_ptr + 4 * i);
          input_dup2[i] = vzipq_f32(input, input);
        }
        local_input_ptr += 8;
        // Load the accumulators from acc_buffer
        float32x4_t acc[4];
        for (int i = 0; i < 4; i++) {
          acc[i] = vld1q_f32(acc_buffer_ptr + 4 * i);
        }
        // Multiply-accumulate
        acc[0] = vmlaq_f32(acc[0], filter[0], input_dup2[0].val[0]);
        acc[1] = vmlaq_f32(acc[1], filter[1], input_dup2[0].val[1]);
        acc[2] = vmlaq_f32(acc[2], filter[2], input_dup2[1].val[0]);
        acc[3] = vmlaq_f32(acc[3], filter[3], input_dup2[1].val[1]);
        // Store the accumulators back to acc_buffer
        for (int i = 0; i < 4; i++) {
          vst1q_f32(acc_buffer_ptr + 4 * i, acc[i]);
        }
        acc_buffer_ptr += 16;
      }
      // Handle 4 input channels at a time.
      for (; ic <= input_depth - 4; ic += 4) {
        // Load the filters
        float32x2_t filter[4];
        for (int i = 0; i < 4; i++) {
          filter[i] = vld1_f32(local_filter_ptr + 2 * i);
        }
        local_filter_ptr += 8;
        // Load the inputs
        const float32x4_t input = vld1q_f32(local_input_ptr);
        local_input_ptr += 4;
        // Load the accumulators from acc_buffer
        float32x2_t acc[4];
        for (int i = 0; i < 4; i++) {
          acc[i] = vld1_f32(acc_buffer_ptr + 2 * i);
        }
        // Multiply-accumulate
        acc[0] = vmla_lane_f32(acc[0], filter[0], vget_low_f32(input), 0);
        acc[1] = vmla_lane_f32(acc[1], filter[1], vget_low_f32(input), 1);
        acc[2] = vmla_lane_f32(acc[2], filter[2], vget_high_f32(input), 0);
        acc[3] = vmla_lane_f32(acc[3], filter[3], vget_high_f32(input), 1);
        // Store the accumulators back to acc_buffer
        for (int i = 0; i < 4; i++) {
          vst1_f32(acc_buffer_ptr + 2 * i, acc[i]);
        }
        acc_buffer_ptr += 8;
      }
      // Handle 2 input channels at a time.
      for (; ic <= input_depth - 2; ic += 2) {
        // Load the filters
        const float32x4_t filter = vld1q_f32(local_filter_ptr);
        local_filter_ptr += 4;
        // Load the inputs
        const float32x2_t input = vld1_f32(local_input_ptr);
        local_input_ptr += 2;
        // Load the accumulators from acc_buffer
        float32x2_t acc[2];
        for (int i = 0; i < 2; i++) {
          acc[i] = vld1_f32(acc_buffer_ptr + 2 * i);
        }
        // Multiply-accumulate
        acc[0] = vmla_lane_f32(acc[0], vget_low_f32(filter), input, 0);
        acc[1] = vmla_lane_f32(acc[1], vget_high_f32(filter), input, 1);
        // Store the accumulators back to acc_buffer
        for (int i = 0; i < 2; i++) {
          vst1_f32(acc_buffer_ptr + 2 * i, acc[i]);
        }
        acc_buffer_ptr += 4;
      }
      // Handle one input channel at a time.
      for (; ic < input_depth; ic++) {
        // Load the inputs
        const float input_val = *local_input_ptr++;
        // Multiply-accumulate
        for (int i = 0; i < 2; i++) {
          acc_buffer_ptr[i] += local_filter_ptr[i] * input_val;
        }
        local_filter_ptr += 2;
        acc_buffer_ptr += 2;
      }
      input_ptr += input_ptr_increment;
    }
  }
};

template <>
struct FloatDepthwiseConvKernel<true, 1, 8> {
  static void Run(int num_output_pixels, int input_depth, int depth_multiplier,
                  const float* input_ptr, int input_ptr_increment,
                  const float* filter_ptr, float* acc_buffer_ptr) {
    // Handle one output pixel at a time.
    for (int outp = 0; outp < num_output_pixels; outp++) {
      // Load the filters
      float32x4_t filter[2];
      for (int i = 0; i < 2; i++) {
        filter[i] = vld1q_f32(filter_ptr + 4 * i);
      }
      // Load the inputs
      const float input_val = *input_ptr;
      input_ptr += input_ptr_increment;
      // Load the accumulators from acc_buffer
      float32x4_t acc[2];
      for (int i = 0; i < 2; i++) {
        acc[i] = vld1q_f32(acc_buffer_ptr + 4 * i);
      }
      // Multiply-accumulate
      for (int i = 0; i < 2; i++) {
        acc[i] = vmlaq_n_f32(acc[i], filter[i], input_val);
      }
      // Store the accumulators back to acc_buffer
      for (int i = 0; i < 2; i++) {
        vst1q_f32(acc_buffer_ptr + 4 * i, acc[i]);
      }
      acc_buffer_ptr += 8;
    }
  }
};

template <>
struct FloatDepthwiseConvKernel<true, 0, 16> {
  static void Run(int num_output_pixels, int input_depth, int depth_multiplier,
                  const float* input_ptr, int input_ptr_increment,
                  const float* filter_ptr, float* acc_buffer_ptr) {
    // Handle one output pixel at a time.
    for (int outp = 0; outp < num_output_pixels; outp++) {
      const float* local_filter_ptr = filter_ptr;
      const float* local_input_ptr = input_ptr;
      for (int ic = 0; ic < input_depth; ic++) {
        // Load the filters
        float32x4_t filter[4];
        for (int i = 0; i < 4; i++) {
          filter[i] = vld1q_f32(local_filter_ptr + 4 * i);
        }
        local_filter_ptr += 16;
        // Load the inputs
        const float input_val = *local_input_ptr++;
        // Load the accumulators from acc_buffer
        float32x4_t acc[4];
        for (int i = 0; i < 4; i++) {
          acc[i] = vld1q_f32(acc_buffer_ptr + 4 * i);
        }
        // Multiply-accumulate
        for (int i = 0; i < 4; i++) {
          acc[i] = vmlaq_n_f32(acc[i], filter[i], input_val);
        }
        // Store the accumulators back to acc_buffer
        for (int i = 0; i < 4; i++) {
          vst1q_f32(acc_buffer_ptr + 4 * i, acc[i]);
        }
        acc_buffer_ptr += 16;
      }
      input_ptr += input_ptr_increment;
    }
  }
};
#endif

// Accumulates the effect of one row of the filter, on a segment of one row
// of the output, accessing the corresponding one row of the input.
template <bool kAllowStrided, int kFixedInputDepth, int kFixedDepthMultiplier>
void FloatDepthwiseConvAccumRow(int stride, int input_depth, int input_width,
                                const float* input_data, int pad_width,
                                int depth_multiplier, int filter_width,
                                const float* filter_data,
                                int out_x_buffer_start, int out_x_buffer_end,
                                int output_depth, float* acc_buffer) {
#ifdef GEMMLOWP_PROFILING
  gemmlowp::ScopedProfilingLabel label(__PRETTY_FUNCTION__);
#endif
  // Sanity check parameters. This is important in particular to ensure
  // that we keep the number of template instantiations minimal, so we don't
  // increase binary size unnecessarily.
  static_assert(kFixedDepthMultiplier || !kFixedInputDepth, "");
  static_assert(kFixedInputDepth || kAllowStrided, "");
  DCHECK(stride == 1 || kAllowStrided);
  if (kFixedInputDepth) {
    DCHECK_EQ(input_depth, kFixedInputDepth);
  }
  if (kFixedDepthMultiplier) {
    DCHECK_EQ(depth_multiplier, kFixedDepthMultiplier);
  }
  DCHECK_EQ(output_depth, input_depth * depth_multiplier);
  const int input_ptr_increment = stride * input_depth;
  const float* filter_base_ptr = filter_data;
  for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
    // For the current (filter_x, filter_y) point in the filter,
    // compute the boundaries of the corresponding output row segment.
    int out_x_loop_start_unclampled = 0;
    int out_x_loop_end_unclampled = 0;
    if (kAllowStrided) {
      if (stride == 2) {
        out_x_loop_start_unclampled = (pad_width - filter_x + 1) / 2;
        out_x_loop_end_unclampled =
            (pad_width + input_width - filter_x + 1) / 2;
      } else if (stride == 4) {
        out_x_loop_start_unclampled = (pad_width - filter_x + 3) / 4;
        out_x_loop_end_unclampled =
            (pad_width + input_width - filter_x + 3) / 4;
      } else {
        out_x_loop_start_unclampled =
            (pad_width - filter_x + stride - 1) / stride;
        out_x_loop_end_unclampled =
            (pad_width + input_width - filter_x + stride - 1) / stride;
      }
    } else {
      out_x_loop_start_unclampled = pad_width - filter_x;
      out_x_loop_end_unclampled = pad_width + input_width - filter_x;
    }
    // The kernel will have to iterate on the segment of the
    // output row that starts at out_x_loop_start and out_x_loop_end.
    const int out_x_loop_start =
        std::max(out_x_buffer_start, out_x_loop_start_unclampled);
    const int out_x_loop_end =
        std::min(out_x_buffer_end, out_x_loop_end_unclampled);

    float* acc_buffer_ptr =
        acc_buffer + (out_x_loop_start - out_x_buffer_start) * output_depth;
    const int in_x_origin = (out_x_loop_start * stride) - pad_width + filter_x;
    const float* input_ptr = input_data + in_x_origin * input_depth;
    const int num_output_pixels = out_x_loop_end - out_x_loop_start;
    FloatDepthwiseConvKernel<kAllowStrided, kFixedInputDepth,
                             kFixedDepthMultiplier>::Run(num_output_pixels,
                                                         input_depth,
                                                         depth_multiplier,
                                                         input_ptr,
                                                         input_ptr_increment,
                                                         filter_base_ptr,
                                                         acc_buffer_ptr);
    filter_base_ptr += output_depth;
  }
}

// generic fallback of FloatDepthwiseConvAccumRow, portable, non-templatized.
inline void FloatDepthwiseConvAccumRowGeneric(
    int stride, int input_depth, int input_width, const float* input_data,
    int pad_width, int depth_multiplier, int filter_width,
    const float* filter_data, int out_x_buffer_start, int out_x_buffer_end,
    int output_depth, float* acc_buffer) {
  gemmlowp::ScopedProfilingLabel label("DepthwiseConvAccumRowGeneric (slow)");
  const float* filter_base_ptr = filter_data;
  for (int filter_x = 0; filter_x < filter_width; ++filter_x) {
    const int out_x_loop_start = std::max(
        out_x_buffer_start, (pad_width - filter_x + stride - 1) / stride);
    const int out_x_loop_end =
        std::min(out_x_buffer_end,
                 (pad_width + input_width - filter_x + stride - 1) / stride);

    float* acc_buffer_ptr =
        acc_buffer + (out_x_loop_start - out_x_buffer_start) * output_depth;
    const int in_x_origin = (out_x_loop_start * stride) - pad_width + filter_x;
    const float* input_ptr = input_data + in_x_origin * input_depth;
    const int input_ptr_increment = (stride - 1) * input_depth;
    for (int out_x = out_x_loop_start; out_x < out_x_loop_end; out_x++) {
      const float* filter_ptr = filter_base_ptr;
      for (int ic = 0; ic < input_depth; ++ic) {
        const float input_val = *input_ptr++;
        for (int m = 0; m < depth_multiplier; m++) {
          const float filter_val = *filter_ptr++;
          *acc_buffer_ptr++ += filter_val * input_val;
        }
      }
      input_ptr += input_ptr_increment;
    }
    filter_base_ptr += output_depth;
  }
}

// Initializes the accumulator buffer with bias values.
inline void DepthwiseConvInitAccBuffer(int num_output_pixels, int output_depth,
                                       const float* bias_data,
                                       float* acc_buffer) {
  for (int i = 0; i < num_output_pixels; i++) {
    memcpy(acc_buffer + i * output_depth, bias_data,
           sizeof(acc_buffer[0]) * output_depth);
  }
}

template <FusedActivationFunctionType Ac>
void DepthwiseConv(const float* input_data, const Dims<4>& input_dims,
                   const float* filter_data, const Dims<4>& filter_dims,
                   const float* bias_data, const Dims<4>& bias_dims,
                   int stride_width, int stride_height,
                   int pad_width, int pad_height, int depth_multiplier,
                   float* output_data, const Dims<4>& output_dims) {
  gemmlowp::ScopedProfilingLabel label("DepthwiseConv");
  static_assert(Ac == FusedActivationFunctionType::kNone ||
                    Ac == FusedActivationFunctionType::kRelu ||
                    Ac == FusedActivationFunctionType::kRelu6 ||
                    Ac == FusedActivationFunctionType::kRelu1,
                "");
  const int batches = MatchingArraySize(input_dims, 3, output_dims, 3);
  const int output_depth = MatchingArraySize(filter_dims, 0, output_dims, 0);
  const int input_height = ArraySize(input_dims, 2);
  const int input_width = ArraySize(input_dims, 1);
  const int input_depth = ArraySize(input_dims, 0);
  const int filter_height = ArraySize(filter_dims, 2);
  const int filter_width = ArraySize(filter_dims, 1);
  const int output_height = ArraySize(output_dims, 2);
  const int output_width = ArraySize(output_dims, 1);
  DCHECK(output_depth == input_depth * depth_multiplier);

  static const int kAccBufferMaxSize = 1024;
  float acc_buffer[kAccBufferMaxSize];
  DCHECK_GE(kAccBufferMaxSize, output_depth);
  const int kOutputPixelsInAccBuffer = kAccBufferMaxSize / output_depth;
  const int kAccBufferActualSize = kOutputPixelsInAccBuffer * output_depth;
  DCHECK_LE(kOutputPixelsInAccBuffer * output_depth, kAccBufferActualSize);
  DCHECK_LE(kAccBufferActualSize, kAccBufferMaxSize);
  DCHECK_GE(kOutputPixelsInAccBuffer, 1);

  // row_accum_func will point to the core accumulation function to be used
  // for this DepthwiseConv op.
  auto* row_accum_func = FloatDepthwiseConvAccumRowGeneric;

  const int kMaxFixedDepthMultiplier = 16;
  int fixed_depth_multiplier = 0;
  if (depth_multiplier <= kMaxFixedDepthMultiplier) {
    fixed_depth_multiplier = depth_multiplier;
  }
  // kMaxUnrolling is the max number of output values that we aim to handle
  // in one unrolled iteration of the inner loop. For practical performance
  // reasons, it is limited by the number of available registers. We could
  // fine-tune it depending on the architecture, but that's not worth doing
  // since this whole code is not very optimized to begin with. The
  // present value reflects what's realistic on ARM 32bit NEON with 16 128-bit
  // vector registers.
  const int kMaxUnrolling = 8;
  int fixed_input_depth = 0;
  if (fixed_depth_multiplier &&
      input_depth * fixed_depth_multiplier <= kMaxUnrolling) {
    fixed_input_depth = input_depth;
  }
#define TFMINI_USE_DEPTHWISECONV_KERNEL(ALLOW_STRIDED, FIXED_INPUT_DEPTH, \
                                        FIXED_DEPTH_MULTIPLIER)           \
  if ((stride_width == 1 || ALLOW_STRIDED) &&                             \
      fixed_input_depth == FIXED_INPUT_DEPTH &&                           \
      fixed_depth_multiplier == FIXED_DEPTH_MULTIPLIER) {                 \
    row_accum_func =                                                      \
        FloatDepthwiseConvAccumRow<ALLOW_STRIDED, FIXED_INPUT_DEPTH,      \
                                   FIXED_DEPTH_MULTIPLIER>;               \
  }

#ifdef USE_NEON
  TFMINI_USE_DEPTHWISECONV_KERNEL(true, 0, 1)
  TFMINI_USE_DEPTHWISECONV_KERNEL(true, 0, 8)
  TFMINI_USE_DEPTHWISECONV_KERNEL(true, 0, 2)
  TFMINI_USE_DEPTHWISECONV_KERNEL(false, 8, 1)
  TFMINI_USE_DEPTHWISECONV_KERNEL(false, 2, 1)
  TFMINI_USE_DEPTHWISECONV_KERNEL(true, 0, 16)
  TFMINI_USE_DEPTHWISECONV_KERNEL(true, 1, 8)
#endif  // USE_NEON

#undef TFMINI_USE_DEPTHWISECONV_KERNEL

  // Now that we have determined row_accum_func, we can start work.
  float* output_ptr = output_data;
  for (int b = 0; b < batches; ++b) {
    for (int out_y = 0; out_y < output_height; ++out_y) {
      const int in_y_origin = (out_y * stride_height) - pad_height;
      const int filter_y_start = std::max(0, -in_y_origin);
      const int filter_y_end =
          std::min(filter_height, input_height - in_y_origin);
      for (int out_x_buffer_start = 0; out_x_buffer_start < output_width;
           out_x_buffer_start += kOutputPixelsInAccBuffer) {
        const int out_x_buffer_end = std::min(
            output_width, out_x_buffer_start + kOutputPixelsInAccBuffer);
        // We call a 'pixel' a group of activation that share all but the
        // 'depth'/'channel' coordinate. num_output_pixels is the number of
        // output pixels that we will accumulate in this loop iteration.
        const int num_output_pixels = out_x_buffer_end - out_x_buffer_start;
        // Initialize our local accumulator with the bias values, so we don't
        // have to add them later.
        DepthwiseConvInitAccBuffer(num_output_pixels, output_depth, bias_data,
                                   acc_buffer);
        // Accumulation loop. Most of the time should be spent in here.
        for (int filter_y = filter_y_start; filter_y < filter_y_end;
             ++filter_y) {
          const int in_y = in_y_origin + filter_y;
          row_accum_func(stride_width, input_depth, input_width,
                         input_data + in_y * input_dims.strides[2] +
                             b * input_dims.strides[3],
                         pad_width, depth_multiplier, filter_width,
                         filter_data + filter_y * filter_dims.strides[2],
                         out_x_buffer_start, out_x_buffer_end, output_depth,
                         acc_buffer);
        }
        // Finished accumulating. Now store to destination.
        const int num_output_values = output_depth * num_output_pixels;
        int i = 0;
#ifdef USE_NEON
        // Handle 16 values at a time
        for (; i <= num_output_values - 16; i += 16) {
          float32x4_t acc[4];
          for (int k = 0; k < 4; k++) {
            acc[k] = vld1q_f32(acc_buffer + i + 4 * k);
          }
          if (Ac == FusedActivationFunctionType::kRelu) {
            for (int k = 0; k < 4; k++) {
              acc[k] = vmaxq_f32(vdupq_n_f32(0.f), acc[k]);
            }
          } else if (Ac == FusedActivationFunctionType::kRelu6) {
            for (int k = 0; k < 4; k++) {
              acc[k] = vmaxq_f32(vdupq_n_f32(0.f),
                                 vminq_f32(vdupq_n_f32(6.f), acc[k]));
            }
          } else if (Ac == FusedActivationFunctionType::kRelu1) {
            for (int k = 0; k < 4; k++) {
              acc[k] = vmaxq_f32(vdupq_n_f32(-1.f),
                                 vminq_f32(vdupq_n_f32(1.f), acc[k]));
            }
          }
          for (int k = 0; k < 4; k++) {
            vst1q_f32(output_ptr + 4 * k, acc[k]);
          }
          output_ptr += 16;
        }
        // Handle 4 values at a time
        for (; i <= num_output_values - 4; i += 4) {
          float32x4_t acc = vld1q_f32(acc_buffer + i);
          if (Ac == FusedActivationFunctionType::kRelu) {
            acc = vmaxq_f32(vdupq_n_f32(0.f), acc);
          } else if (Ac == FusedActivationFunctionType::kRelu6) {
            acc = vmaxq_f32(vdupq_n_f32(0.f), vminq_f32(vdupq_n_f32(6.f), acc));
          } else if (Ac == FusedActivationFunctionType::kRelu1) {
            acc =
                vmaxq_f32(vdupq_n_f32(-1.f), vminq_f32(vdupq_n_f32(1.f), acc));
          }
          vst1q_f32(output_ptr, acc);
          output_ptr += 4;
        }
#endif
        // Handle leftover values, one by one. This is very slow.
        for (; i < num_output_values; i++) {
          float acc = acc_buffer[i];
          if (Ac == FusedActivationFunctionType::kRelu) {
            acc = std::max(0.f, acc);
          } else if (Ac == FusedActivationFunctionType::kRelu6) {
            acc = std::max(0.f, std::min(6.f, acc));
          } else if (Ac == FusedActivationFunctionType::kRelu1) {
            acc = std::max(-1.f, std::min(1.f, acc));
          }
          *output_ptr++ = acc;
        }
      }
    }
  }
}

}  // namespace optimized_ops
}  // namespace nn
}  // namespace android


#endif  // ANDROID_ML_NN_COMMON_OPERATIONS_INTERNAL_OPTIMIZED_DEPTHWISECONV_FLOAT_H_