cutlass/include/cutlass/epilogue/thread/linear_combination.h
2021-05-12 22:23:55 +08:00

223 lines
7.8 KiB
C++

/***************************************************************************************************
* Copyright (c) 2017-2021, 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.
*
**************************************************************************************************/
/*! \file
\brief Functor performing linear combination operations used by epilogues.
*/
#pragma once
#include "cutlass/cutlass.h"
#include "cutlass/numeric_types.h"
#include "cutlass/array.h"
#include "cutlass/functional.h"
#include "cutlass/numeric_conversion.h"
#include "cutlass/epilogue/thread/scale_type.h"
/////////////////////////////////////////////////////////////////////////////////////////////////
namespace cutlass {
namespace epilogue {
namespace thread {
/////////////////////////////////////////////////////////////////////////////////////////////////
/// Applies a linear combination operator to an array of elements.
///
/// D = alpha * accumulator + beta * source + uniform
///
template <
typename ElementOutput_, ///< Data type used to load and store tensors
int Count, ///< Number of elements computed per operation
typename ElementAccumulator_ = ElementOutput_, ///< Accumulator data type
typename ElementCompute_ = ElementOutput_, ///< Data type used to compute linear combination
ScaleType::Kind Scale = ScaleType::Default, ///< Control Alpha and Beta scaling
FloatRoundStyle Round = FloatRoundStyle::round_to_nearest
>
class LinearCombination {
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;
/// Host-constructable parameters structure
struct Params {
ElementCompute alpha; ///< scales accumulators
ElementCompute beta; ///< scales source tensor
ElementCompute const *alpha_ptr; ///< pointer to accumulator scalar - if not null, loads it from memory
ElementCompute const *beta_ptr; ///< pointer to source scalar - if not null, loads it from memory
//
// Methods
//
CUTLASS_HOST_DEVICE
Params():
alpha(ElementCompute(1)),
beta(ElementCompute(0)),
alpha_ptr(nullptr),
beta_ptr(nullptr) { }
CUTLASS_HOST_DEVICE
Params(
ElementCompute alpha,
ElementCompute beta
): alpha(alpha), beta(beta), alpha_ptr(nullptr), beta_ptr(nullptr) {
}
CUTLASS_HOST_DEVICE
Params(
ElementCompute alpha
): alpha(alpha), beta(0), alpha_ptr(nullptr), beta_ptr(nullptr) {
}
CUTLASS_HOST_DEVICE
Params(
ElementCompute const *alpha_ptr,
ElementCompute const *beta_ptr
): alpha(0), beta(0), alpha_ptr(alpha_ptr), beta_ptr(beta_ptr) {
}
CUTLASS_HOST_DEVICE
Params(
ElementCompute const *alpha_ptr
): alpha(0), beta(0), alpha_ptr(alpha_ptr), beta_ptr(nullptr) {
}
};
private:
//
// Data members
//
ElementCompute alpha_;
ElementCompute beta_;
public:
/// Constructs the function object, possibly loading from pointers in host memory
CUTLASS_HOST_DEVICE
LinearCombination(Params const &params) {
alpha_ = (params.alpha_ptr ? *params.alpha_ptr : params.alpha);
beta_ = (params.beta_ptr ? *params.beta_ptr : params.beta);
}
/// Returns true if source is needed
CUTLASS_HOST_DEVICE
bool is_source_needed() const {
if (Scale == ScaleType::NoBetaScaling) return true;
if (Scale == ScaleType::OnlyAlphaScaling) return false;
return beta_ != ElementCompute(0);
}
/// Functionally required for serial reduction in the epilogue
CUTLASS_HOST_DEVICE
void set_k_partition(int k_partition, int k_partition_count) {
if (k_partition) {
beta_ = ElementCompute(1);
}
}
/// Computes linear scaling: D = alpha * accumulator + beta * source
CUTLASS_HOST_DEVICE
FragmentOutput operator()(
FragmentAccumulator const &accumulator,
FragmentOutput const &source) 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;
if (Scale == ScaleType::NoBetaScaling)
intermediate = converted_source;
else
intermediate = mul_add_source(beta_, converted_source); // X = beta * C + uniform
intermediate = mul_add_accumulator(alpha_, converted_accumulator, intermediate); // D = alpha * Accum + X
// Convert to destination numeric type
NumericArrayConverter<ElementOutput, ElementCompute, kCount, Round> destination_converter;
return destination_converter(intermediate);
}
/// Computes linear scaling: D = alpha * accumulator
CUTLASS_HOST_DEVICE
FragmentOutput operator()(
FragmentAccumulator const &accumulator) const {
// Convert source to interal compute numeric type
NumericArrayConverter<ElementCompute, ElementAccumulator, kCount, Round> accumulator_converter;
ComputeFragment converted_accumulator = accumulator_converter(accumulator);
// Perform binary operations
ComputeFragment intermediate;
multiplies<ComputeFragment> mul_accumulator;
intermediate = mul_accumulator(alpha_, converted_accumulator); // D = alpha * Accum
// Convert to destination numeric type
NumericArrayConverter<ElementOutput, ElementCompute, kCount, Round> destination_converter;
return destination_converter(intermediate);
}
};
/////////////////////////////////////////////////////////////////////////////////////////////////
} // namespace thread
} // namespace epilogue
} // namespace cutlass
/////////////////////////////////////////////////////////////////////////////////////////////////