linear_combination_relu.h

Go to the documentation of this file.

/***************************************************************************************************
 * Copyright (c) 2017-2019, NVIDIA CORPORATION.  All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without modification, are permitted
 * provided that the following conditions are met:
 *     * Redistributions of source code must retain the above copyright notice, this list of
 *       conditions and the following disclaimer.
 *     * Redistributions in binary form must reproduce the above copyright notice, this list of
 *       conditions and the following disclaimer in the documentation and/or other materials
 *       provided with the distribution.
 *     * Neither the name of the NVIDIA CORPORATION nor the names of its contributors may be used
 *       to endorse or promote products derived from this software without specific prior written
 *       permission.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND ANY EXPRESS OR
 * IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND
 * FITNESS FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL NVIDIA CORPORATION BE LIABLE
 * FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
 * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE, DATA, OR PROFITS;
 * OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT,
 * STRICT LIABILITY, OR TOR (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 *
 **************************************************************************************************/
#pragma once

#include "cutlass/cutlass.h"
#include "cutlass/numeric_types.h"
#include "cutlass/array.h"
#include "cutlass/functional.h"
#include "cutlass/numeric_conversion.h"


namespace cutlass {
namespace epilogue {
namespace thread {


template <
  typename ElementOutput_,                             
  int Count,                                           
  typename ElementAccumulator_ = ElementOutput_,       
  typename ElementCompute_ = ElementOutput_,           
  FloatRoundStyle Round = FloatRoundStyle::round_to_nearest
>
class LinearCombinationRelu {
public:

  using ElementOutput = ElementOutput_;
  using ElementAccumulator = ElementAccumulator_;
  using ElementCompute = ElementCompute_;

  static int const kCount = Count;

  using FragmentOutput = Array<ElementOutput, kCount>;
  using FragmentAccumulator = Array<ElementAccumulator, kCount>;
  using ComputeFragment = Array<ElementCompute, kCount>;

  static FloatRoundStyle const kRound = Round;

  struct Params {

    ElementCompute alpha;                  
    ElementCompute beta;                   
    ElementCompute threshold;              
    ElementCompute const *alpha_ptr;       
    ElementCompute const *beta_ptr;        

    //
    // Methods
    //

    CUTLASS_HOST_DEVICE
    Params(): 
      alpha(ElementCompute(1)), 
      beta(ElementCompute(0)),
      threshold(ElementCompute(0)),
      alpha_ptr(nullptr), 
      beta_ptr(nullptr) { }

    CUTLASS_HOST_DEVICE
    Params(
      ElementCompute alpha,
      ElementCompute beta,
      ElementCompute threshold  = ElementCompute(0)
    ): alpha(alpha), beta(beta), threshold(threshold), alpha_ptr(nullptr), beta_ptr(nullptr) {

    }

    CUTLASS_HOST_DEVICE
    Params(
      ElementCompute const *alpha_ptr,
      ElementCompute const *beta_ptr,
      ElementCompute threshold = ElementCompute(0)
    ): alpha(0), beta(0), threshold(threshold), alpha_ptr(alpha_ptr), beta_ptr(beta_ptr) {

    }
  };

private:

  //
  // Data members
  //

  ElementCompute alpha_;
  ElementCompute beta_;
  ElementCompute threshold_;

public:

  CUTLASS_HOST_DEVICE
  LinearCombinationRelu(Params const &params) {

    alpha_ = (params.alpha_ptr ? *params.alpha_ptr : params.alpha);
    beta_ = (params.beta_ptr ? *params.beta_ptr : params.beta);
    threshold_ = params.threshold;
  }

  CUTLASS_HOST_DEVICE
  bool is_source_needed() const {
    return beta_ != ElementCompute(0);
  }

  CUTLASS_HOST_DEVICE
  void set_k_partition(int k_partition) {
    if (k_partition) {
      beta_ = ElementCompute(1);
    }
  }

  CUTLASS_HOST_DEVICE
  FragmentOutput operator()(
    FragmentAccumulator const &accumulator, 
    FragmentOutput const &source,
    ElementCompute uniform = ElementCompute(0)) const {

    // Convert source to interal compute numeric type
    NumericArrayConverter<ElementCompute, ElementOutput, kCount, Round> source_converter;
    NumericArrayConverter<ElementCompute, ElementAccumulator, kCount, Round> accumulator_converter;

    ComputeFragment converted_source = source_converter(source);
    ComputeFragment converted_accumulator = accumulator_converter(accumulator);

    // Perform binary operations

    ComputeFragment intermediate;

    multiplies<ComputeFragment> mul_add_source;
    multiply_add<ComputeFragment> mul_add_accumulator;
    
    maximum<ComputeFragment> max_accumulator;

    intermediate = mul_add_source(beta_, converted_source);                             // X =  beta * C + uniform
    intermediate = mul_add_accumulator(alpha_, converted_accumulator, intermediate);    // D = alpha * Accum + X
    
    intermediate = max_accumulator(intermediate, threshold_);

    // Convert to destination numeric type
    NumericArrayConverter<ElementOutput, ElementCompute, kCount, Round> destination_converter;

    return destination_converter(intermediate);
  }
};



template <
  typename ElementOutput_,                             
  int Count,                                           
  FloatRoundStyle Round
>
class LinearCombinationRelu<ElementOutput_, Count, int, float, Round> {
public:

  using ElementOutput = ElementOutput_;
  using ElementAccumulator = int;
  using ElementCompute = float;

  static int const kCount = Count;

  using FragmentOutput = Array<ElementOutput, kCount>;
  using FragmentAccumulator = Array<ElementAccumulator, kCount>;
  using ComputeFragment = Array<ElementCompute, kCount>;

  static FloatRoundStyle const kRound = Round;

  struct Params {

    ElementCompute alpha;                  
    ElementCompute beta;                   
    ElementCompute threshold;              
    ElementCompute const *alpha_ptr;       
    ElementCompute const *beta_ptr;        

    //
    // Methods
    //

    CUTLASS_HOST_DEVICE
    Params(): 
      alpha(ElementCompute(1)), 
      beta(ElementCompute(0)),
      threshold(ElementCompute(0)),
      alpha_ptr(nullptr), 
      beta_ptr(nullptr) { }

    CUTLASS_HOST_DEVICE
    Params(
      ElementCompute alpha,
      ElementCompute beta,
      ElementCompute threshold  = ElementCompute(0)
    ): alpha(alpha), beta(beta), threshold(threshold), alpha_ptr(nullptr), beta_ptr(nullptr) {

    }

    CUTLASS_HOST_DEVICE
    Params(
      ElementCompute const *alpha_ptr,
      ElementCompute const *beta_ptr,
      ElementCompute threshold = ElementCompute(0)
    ): alpha(0), beta(0), threshold(threshold), alpha_ptr(alpha_ptr), beta_ptr(beta_ptr) {

    }
  };

private:

  //
  // Data members
  //

  ElementCompute alpha_;
  ElementCompute beta_;
  ElementCompute threshold_;

public:

  CUTLASS_HOST_DEVICE
  LinearCombinationRelu(Params const &params) {

    alpha_ = (params.alpha_ptr ? *params.alpha_ptr : params.alpha);
    beta_ = (params.beta_ptr ? *params.beta_ptr : params.beta);
    threshold_ = params.threshold;
  }

  CUTLASS_HOST_DEVICE
  bool is_source_needed() const {
    return beta_ != ElementCompute(0);
  }

  CUTLASS_HOST_DEVICE
  void set_k_partition(int k_partition) {
    if (k_partition) {
      beta_ = ElementCompute(1);
    }
  }

  CUTLASS_HOST_DEVICE
  FragmentOutput operator()(
    FragmentAccumulator const &accumulator, 
    FragmentOutput const &source,
    ElementCompute uniform = ElementCompute(0)) const {

    // Convert source to interal compute numeric type
    NumericArrayConverter<ElementCompute, ElementOutput, kCount, Round> source_converter;
    NumericArrayConverter<ElementCompute, ElementAccumulator, kCount, Round> accumulator_converter;

    ComputeFragment converted_source = source_converter(source);
    ComputeFragment converted_accumulator = accumulator_converter(accumulator);

    // Perform binary operations

    ComputeFragment intermediate;

    multiplies<ComputeFragment> mul_add_source;
    multiply_add<ComputeFragment> mul_add_accumulator;
    
    maximum<ComputeFragment> max_accumulator;

    intermediate = mul_add_source(beta_, converted_source);                             // X =  beta * C + uniform
    intermediate = mul_add_accumulator(alpha_, converted_accumulator, intermediate);    // D = alpha * Accum + X
    
    // Clamp to theshold
    intermediate = max_accumulator(intermediate, threshold_);

    // Convert back to accumulator data type
    FragmentAccumulator scaled_accumulator;

    CUTLASS_PRAGMA_UNROLL
    for (int i = 0; i < kCount; ++i) {
      scaled_accumulator[i] = static_cast<int>(intermediate[i]);
    }

    // Convert to destination numeric type and pack
    NumericArrayConverter<ElementOutput, ElementAccumulator, kCount, Round> destination_converter;

    return destination_converter(scaled_accumulator);
  }
};


} // namespace thread
} // namespace epilogue
} // namespace cutlass