quantize_kernel.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 PADDLE_MOBILE_CPU

#include "operators/kernel/quantize_kernel.h"
#include <cmath>
#include <limits>

#if defined(__ARM_NEON__) || defined(__ARM_NEON)
#include <arm_neon.h>

#ifndef __aarch64__
float32_t vmaxvq_f32(float32x4_t r) {
  float32x2_t v = vmax_f32(vget_high_f32(r), vget_low_f32(r));
  return vget_lane_f32(vpmax_f32(v, v), 0);
}
#endif

int32x4_t vrnd_towards_zero(float32x4_t r) {
  return vcvtq_s32_f32(r);
}

int32x4_t vrnd_away_zero(float32x4_t r) {
  float32x4_t plus  = vdupq_n_f32(0.5);
  float32x4_t minus = vdupq_n_f32(-0.5);
  float32x4_t zero  = vdupq_n_f32(0);
  uint32x4_t more_than_zero = vcgtq_f32(r, zero);
  float32x4_t temp = vbslq_f32(more_than_zero, plus, minus);
  temp = vaddq_f32(r, temp);
  int32x4_t ret = vcvtq_s32_f32(temp);
  return ret;
}

int32x4_t vrnd_to_even(float32x4_t r) {
  int32x4_t ret;
  float value[4];
  vst1q_f32(value, r);
  for (int i = 0; i < 4; ++i) {
    float v = round(value[i]);
    int32_t q = (int32_t)v;
    if (abs(abs(v - value[i]) - 0.5) > 0) {
      ret[i] = q;
    } else {
      if (abs(q) % 2 == 0) {
        ret[i] = q;
      } else {
        ret[i] = q + (q > 0) ? -1 : 1;
      }
    }
  }
  return ret;
}
#endif

namespace paddle_mobile {
namespace operators {

static float find_abs_max(const Tensor *input) {
  float max_abs = float(0);
  const float *x = input->data<const float>();
  size_t size = input->numel();
#if defined(__ARM_NEON__) || defined(__ARM_NEON)
  size_t loop = size >> 4;
  size_t remain = size & 0xF;
  for (size_t i = 0; i < loop; ++i) {
    float32x4_t max;
    float32x4_t r0 = vld1q_f32(x);
    float32x4_t r1 = vld1q_f32(x + 4);
    float32x4_t r2 = vld1q_f32(x + 8);
    float32x4_t r3 = vld1q_f32(x + 12);
    r0 = vabsq_f32(r0);
    r1 = vabsq_f32(r1);
    r2 = vabsq_f32(r2);
    r3 = vabsq_f32(r3);
    max[0] = vmaxvq_f32(r0);
    max[1] = vmaxvq_f32(r1);
    max[2] = vmaxvq_f32(r2);
    max[3] = vmaxvq_f32(r3);
    max[0] = vmaxvq_f32(max);
    if (max[0] > max_abs) {
      max_abs = max[0];
    }
    x += 16;
  }
  size = remain;
#endif
  for (size_t i = 0; i < size; ++i) {
    float value = std::abs(x[i]);
    if (value > max_abs) {
      max_abs = value;
    }
  }
  return max_abs;
}

static void quantize_round_to_even(const Tensor *input,
                                   const float scale,
                                   Tensor *output) {
  const float *x = input->data<const float>();
  int8_t *y = output->data<int8_t>();
  size_t size = input->numel();
#if defined(__ARM_NEON__) || defined(__ARM_NEON)
  size_t loop = size >> 4;
  size_t remain = size & 0xF;
  for (size_t i = 0; i < loop; ++i) {
    float32x4_t r0 = vld1q_f32(x);
    float32x4_t r1 = vld1q_f32(x + 4);
    float32x4_t r2 = vld1q_f32(x + 8);
    float32x4_t r3 = vld1q_f32(x + 12);
    r0 = vmulq_n_f32(r0, scale);
    r1 = vmulq_n_f32(r1, scale);
    r2 = vmulq_n_f32(r2, scale);
    r3 = vmulq_n_f32(r3, scale);
    int32x4_t q0 = vrnd_to_even(r0);
    int32x4_t q1 = vrnd_to_even(r1);
    int32x4_t q2 = vrnd_to_even(r2);
    int32x4_t q3 = vrnd_to_even(r3);
    int16x4_t d0 = vmovn_s32(q0);
    int16x4_t d1 = vmovn_s32(q1);
    int16x4_t d2 = vmovn_s32(q2);
    int16x4_t d3 = vmovn_s32(q3);
    int16x8_t q5 = vcombine_s16(d1, d0);
    int16x8_t q6 = vcombine_s16(d3, d2);
    int8x8_t d5 = vmovn_s16(q5);
    int8x8_t d6 = vmovn_s16(q6);
    vst1_s8(y, d5);
    vst1_s8(y + 8, d6);
    x += 16;
    y += 16;
  }
  size = remain;
#endif
  for (size_t i = 0; i < size; ++i) {
    float value = x[i] * scale;
    float v = round(value);
    int32_t q = (int32_t)v;
    if (abs(abs(q - value) - 0.5) > 0) {
      y[i] = q;
    } else {
      if (abs(q) % 2 == 0) {
        y[i] = q;
      } else {
        y[i] = q + (q > 0) ? -1 : 1;
      }
    }
  }
}

static void quantize_round_to_zero(const Tensor *input,
                            const float scale,
                            Tensor *output) {
  const float *x = input->data<const float>();
  int8_t *y = output->data<int8_t>();
  size_t size = input->numel();
#ifdef defined(__ARM_NEON__) || defined(__ARM_NEON)
  size_t loop = size >> 4;
  size_t remain = size & 0xF;
  for (size_t i = 0; i < loop; ++i) {
    float32x4_t r0 = vld1q_f32(x);
    float32x4_t r1 = vld1q_f32(x + 4);
    float32x4_t r2 = vld1q_f32(x + 8);
    float32x4_t r3 = vld1q_f32(x + 12);
    r0 = vmulq_n_f32(r0, scale);
    r1 = vmulq_n_f32(r1, scale);
    r2 = vmulq_n_f32(r2, scale);
    r3 = vmulq_n_f32(r3, scale);
    int32x4_t q0 = vrnd_towards_zero(r0);
    int32x4_t q1 = vrnd_towards_zero(r1);
    int32x4_t q2 = vrnd_towards_zero(r2);
    int32x4_t q3 = vrnd_towards_zero(r3);
    int16x4_t d0 = vmovn_s32(q0);
    int16x4_t d1 = vmovn_s32(q1);
    int16x4_t d2 = vmovn_s32(q2);
    int16x4_t d3 = vmovn_s32(q3);
    int16x8_t q5 = vcombine_s16(d1, d0);
    int16x8_t q6 = vcombine_s16(d3, d2);
    int8x8_t d5 = vmovn_s16(q5);
    int8x8_t d6 = vmovn_s16(q6);
    vst1_s8(y, d5);
    vst1_s8(y + 8, d6);
    x += 16;
    y += 16;
  }
  size = remain;
#endif
  for (size_t i = 0; i < size; ++i) {
    y[i] = trunc(x[i] * scale);
  }
}

static void quantize_round_to_nearest(const Tensor *input,
                               const float scale,
                               Tensor *output) {
  const float *x = input->data<const float>();
  int8_t *y = output->data<int8_t>();
  size_t size = input->numel();
#ifdef defined(__ARM_NEON__) || defined(__ARM_NEON)
  size_t loop = size >> 4;
  size_t remain = size & 0xF;
  for (size_t i = 0; i < loop; ++i) {
    float32x4_t r0 = vld1q_f32(x);
    float32x4_t r1 = vld1q_f32(x + 4);
    float32x4_t r2 = vld1q_f32(x + 8);
    float32x4_t r3 = vld1q_f32(x + 12);
    r0 = vmulq_n_f32(r0, scale);
    r1 = vmulq_n_f32(r1, scale);
    r2 = vmulq_n_f32(r2, scale);
    r3 = vmulq_n_f32(r3, scale);
    int32x4_t q0 = vrnd_away_zero(r0);
    int32x4_t q1 = vrnd_away_zero(r1);
    int32x4_t q2 = vrnd_away_zero(r2);
    int32x4_t q3 = vrnd_away_zero(r3);
    int16x4_t d0 = vmovn_s32(q0);
    int16x4_t d1 = vmovn_s32(q1);
    int16x4_t d2 = vmovn_s32(q2);
    int16x4_t d3 = vmovn_s32(q3);
    int16x8_t q5 = vcombine_s16(d1, d0);
    int16x8_t q6 = vcombine_s16(d3, d2);
    int8x8_t d5 = vmovn_s16(q5);
    int8x8_t d6 = vmovn_s16(q6);
    vst1_s8(y, d5);
    vst1_s8(y + 8, d6);
    x += 16;
    y += 16;
  }
  size = remain;
#endif
  for (size_t i = 0; i < size; ++i) {
    y[i] = trunc(x[i] * scale);
  }
}

template<>
bool QuantizeKernel<CPU, float>::Init(QuantizeParam<CPU> *param) {
  return true;
}

template<>
void QuantizeKernel<CPU, float>::Compute(
    const QuantizeParam<CPU> &param) const {
  // TODO
  float max_abs = 0.f;
  const Tensor *input = param.input_;
  Tensor *output = param.out_;
  Tensor *output_scale = param.online_scale_;
  if (param.is_static_) {
    max_abs = param.static_scale_;
  } else {
    max_abs = find_abs_max(input);
  }
  if (max_abs < std::numeric_limits<float>::min()) {
    max_abs = std::numeric_limits<float>::min();
  }
  // only support int8 currently
  float online_scale = 127 / max_abs;
  param.online_scale_->mutable_data<float>()[0] = online_scale;
  switch (param.round_type_) {
    case ROUND_NEAREST_TO_EVEN:
      quantize_round_to_even(input, online_scale, output);
      break;
    case ROUND_NEAREST_TOWARDS_ZERO:
      quantize_round_to_zero(input, online_scale, output);
      break;
    case ROUND_NEAREST_AWAY_ZERO:
      quantize_round_to_nearest(input, online_scale, output);
    default:
      LOG(kLOG_ERROR) << "round type is not supported.";
      break;
  }
}

}  // namespace paddle_mobile
}  // namespace operators

#endif