squall/cutlass
1
0
Fork 0
Code Issues Pull Requests Actions 1 Packages Projects Releases Wiki Activity

add leaky relu (#542)

Browse Source 
Authored-by: Haicheng Wu <haichengw@nvidia.com>
This commit is contained in:
Haicheng Wu 2022-06-26 10:07:50 -04:00 committed by GitHub
parent e5d4669f16
commit 9ab9110168
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23
2 changed files with 236 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

26
include/cutlass/epilogue/thread/activation.h
View File

@ -98,6 +98,32 @@ struct ReLu<Array<T, N>> {
} }
}; };
// Leaky Relu operator
template <typename T>
struct LeakyReLU {
CUTLASS_HOST_DEVICE
T operator()(T const &value, T const & alpha_recip) const {
T res = value > T(0) ? value : value * alpha_recip;
return res;
}
};
template <typename T, int N>
struct LeakyReLU<Array<T, N> > {
CUTLASS_HOST_DEVICE
Array<T, N> operator()(Array<T, N> const &rhs, T const & alpha_recip) const {
Array<T, N> y;
LeakyReLU<T> leaky_op;
CUTLASS_PRAGMA_UNROLL
for (int i = 0; i < int(rhs.size()); ++i) {
y[i] = leaky_op(rhs[i], alpha_recip);
}
return y;
}
};
// Sigmoid operator // Sigmoid operator
template <typename T> template <typename T>
struct Sigmoid { struct Sigmoid {

210
include/cutlass/epilogue/thread/linear_combination_leaky_relu.h Normal file
View File

@ -0,0 +1,210 @@
/***************************************************************************************************
* Copyright (c) 2017 - 2022 NVIDIA CORPORATION & AFFILIATES. All rights reserved.
* SPDX-License-Identifier: BSD-3-Clause
*
* Redistribution and use in source and binary forms, with or without
* modification, are permitted provided that the following conditions are met:
*
* 1. Redistributions of source code must retain the above copyright notice, this
* list of conditions and the following disclaimer.
*
* 2. 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.
*
* 3. Neither the name of the copyright holder 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 THE COPYRIGHT HOLDER OR CONTRIBUTORS 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 TORT (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"
#include "cutlass/epilogue/thread/activation.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
FloatRoundStyle Round = FloatRoundStyle::round_to_nearest
>
class LinearCombinationLeakyRelu {
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_bias; ///< scales bias tensor
ElementCompute leaky_alpha; ///< leaky_alpha
//
// Methods
//
CUTLASS_HOST_DEVICE
Params():
alpha(ElementCompute(1)),
beta_bias(ElementCompute(0)),
leaky_alpha(ElementCompute(1))
{ }
CUTLASS_HOST_DEVICE
Params(
ElementCompute alpha,
ElementCompute beta_bias,
ElementCompute leaky_alpha = ElementCompute(1)
): alpha(alpha), beta_bias(beta_bias), leaky_alpha(leaky_alpha) {
}
};
private:
//
// Data members
//
ElementCompute alpha_;
ElementCompute beta_bias_;
ElementCompute leaky_alpha_recip_;
public:
/// Constructs the function object, possibly loading from pointers in host memory
CUTLASS_HOST_DEVICE
LinearCombinationLeakyRelu(Params const &params) {
alpha_ = (params.alpha);
beta_bias_ = (params.beta_bias);
leaky_alpha_recip_ = (ElementCompute(params.leaky_alpha));
}
/// Returns true if source is needed
CUTLASS_HOST_DEVICE
bool is_source_needed() const {
return beta_bias_ != ElementCompute(0);
}
/// Functionally required for serial reduction in the epilogue
CUTLASS_HOST_DEVICE
void set_k_partition(int k_partition) {
if (k_partition) {
beta_bias_ = ElementCompute(1);
}
}
CUTLASS_HOST_DEVICE
void set_k_partition(int k_partition, int k_partition_count) {
if (k_partition) {
beta_bias_ = 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;
LeakyReLU<ComputeFragment> leakyrelu;
intermediate = mul_add_source(beta_bias_, converted_source); // X = beta * C + uniform
intermediate = mul_add_accumulator(alpha_, converted_accumulator, intermediate); // D = alpha * Accum + X
// Compute threshold optionally
intermediate = leakyrelu(intermediate, leaky_alpha_recip_);
// 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;
LeakyReLU<ComputeFragment> leakyrelu;
//printf("in doing with bias");
intermediate = mul_accumulator(alpha_, converted_accumulator); // D = alpha * Accum
// Compute threshold optionally
intermediate = leakyrelu(intermediate, leaky_alpha_recip_);
// Convert to destination numeric type
NumericArrayConverter<ElementOutput, ElementCompute, kCount, Round> destination_converter;
return destination_converter(intermediate);
}
};
/////////////////////////////////////////////////////////////////////////////////////////////////
} // namespace thread
} // namespace epilogue
} // namespace cutlass
/////////////////////////////////////////////////////////////////////////////////////////////////