tesseract/arch/intsimdmatrix.cpp

///////////////////////////////////////////////////////////////////////
// File:        intsimdmatrix.cpp
// Description: Base class for 8-bit int SIMD matrix multipliers.
// Author:      Ray Smith
// Created:     Tue Aug 15 08:01:32 PST 2017
//
// (C) Copyright 2017, Google Inc.
// 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 "intsimdmatrix.h"
#include "intsimdmatrixavx2.h"
#include "intsimdmatrixsse.h"
#include "simddetect.h"

namespace tesseract {

// Factory makes and returns an IntSimdMatrix (sub)class of the best
// available type for the current architecture.
/* static */
IntSimdMatrix* IntSimdMatrix::GetFastestMultiplier() {
  IntSimdMatrix* multiplier = nullptr;
  if (SIMDDetect::IsAVX2Available()) {
    multiplier = new IntSimdMatrixAVX2();
  } else if (SIMDDetect::IsSSEAvailable()) {
    multiplier = new IntSimdMatrixSSE();
  } else {
    // Default c++ implementation.
    multiplier = new IntSimdMatrix();
  }
  return multiplier;
}

// Computes a reshaped copy of the weight matrix w. If there are no
// partial_funcs_, it does nothing.
void IntSimdMatrix::Init(const GENERIC_2D_ARRAY<int8_t>& w) {
  if (partial_funcs_.empty()) return;
  int num_out = w.dim1();
  int num_in = w.dim2() - 1;
  // The rounded-up sizes of the reshaped weight matrix, excluding biases.
  int rounded_num_in = Roundup(num_in, num_inputs_per_group_);
  int rounded_num_out = RoundOutputs(num_out);
  // Add the bias and compute the required size.
  shaped_w_.resize((rounded_num_in + 1) * rounded_num_out, 0);
  int shaped_index = 0;
  int output = 0;
  // Each number of registers needs a different format! Iterates over the
  // different numbers of registers (each a power of 2).
  for (int num_registers = max_output_registers_; num_registers >= 1;
       num_registers /= 2) {
    // The number of outputs that we will generate with this many registers.
    int num_outputs_per_register_set =
        num_registers * num_outputs_per_register_;
    // Use the max number of registers until we have to go fewer.
    while (output + num_outputs_per_register_set <= rounded_num_out) {
      // Accumulating outputs in registers saves iterating over the inputs, so
      // we only have to do it once per output register set.
      for (int input = 0; input < num_in; input += num_inputs_per_group_) {
        // Iterate over the number of outputs in a register set.
        for (int j = 0; j < num_outputs_per_register_set; ++j) {
          // Inner-most loop corresponds to the number of inputs in an input
          // group.
          for (int i = 0; i < num_inputs_per_group_; ++i) {
            int8_t weight = 0;
            if (output + j < num_out && input + i < num_in)
              weight = w(output + j, input + i);
            shaped_w_[shaped_index++] = weight;
          }
        }
      }
      // Append the bias weights for the register set.
      for (int j = 0; j < num_outputs_per_register_set; ++j) {
        int8_t weight = 0;
        if (output + j < num_out) weight = w(output + j, num_in);
        shaped_w_[shaped_index++] = weight;
      }
      output += num_outputs_per_register_set;
    }
  }
}

// Computes matrix.vector v = Wu.
// u is of size W.dim2() - 1 and the output v is of size W.dim1().
// u is imagined to have an extra element at the end with value 1, to
// implement the bias, but it doesn't actually have it.
void IntSimdMatrix::MatrixDotVector(const GENERIC_2D_ARRAY<int8_t>& w,
                                    const GenericVector<double>& scales,
                                    const int8_t* u, double* v) const {
  int num_out = w.dim1();
  int num_in = w.dim2() - 1;
  if (partial_funcs_.empty()) {
    // Base implementation.
    for (int i = 0; i < num_out; ++i) {
      const int8_t* wi = w[i];
      int total = 0;
      for (int j = 0; j < num_in; ++j) total += wi[j] * u[j];
      // Add in the bias and correct for integer values.
      v[i] = (static_cast<double>(total) / INT8_MAX + wi[num_in]) * scales[i];
    }
  } else {
    const int8_t* w_data = shaped_w_.data();
    const double* scales_data = &scales[0];
    // Each call to a partial_func_ produces group_size outputs, except the
    // last one, which can produce less.
    int group_size = num_outputs_per_register_ * max_output_registers_;
    int rounded_num_in = Roundup(num_in, num_inputs_per_group_);
    int rounded_num_out = RoundOutputs(num_out);
    int output = 0;
    for (auto fn : partial_funcs_) {
      // The amount of w_data consumed by each call to fn.
      int w_step = (rounded_num_in + 1) * group_size;
      // Run with this group size, until it would produce too much output, then
      // switch to a smaller size.
      for (; output + group_size <= rounded_num_out; output += group_size) {
        (*fn)(w_data, scales_data, u, rounded_num_in, num_out - output, v);
        w_data += w_step;
        scales_data += group_size;
        v += group_size;
      }
      group_size /= 2;
    }
  }
}

}  // namespace tesseract