Brutex
/
cavis
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
cavis/libnd4j/include/ops/ops.h

4453 lines
118 KiB
C
Raw Normal View History

Eclipse Migration Initial Commit
2019-06-06 14:21:15 +02:00
/*******************************************************************************
* Copyright (c) 2015-2018 Skymind, Inc.
*
* This program and the accompanying materials are made available under the
* terms of the Apache License, Version 2.0 which is available at
* https://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.
*
* SPDX-License-Identifier: Apache-2.0
******************************************************************************/
#pragma once
#ifndef OPS_H_
#define OPS_H_
#include <op_boilerplate.h>
#include <array/DataTypeUtils.h>
#include <helpers/shape.h>
#include <vector>
#include <Environment.h>
#include <loops/summarystatsreduce.h>
#include <loops/ReduceType.h>
#define MIN_V 1e-12
#define MAX_FLOAT 1e37
#define MIN_FLOAT 1e-37
#define MAX_INT 2147483647
#define MIN_CUTFOFF -3.79297773665f
#define FLOAT_MIN_NORMAL 1.17549435e-38
#define EPS 1e-5
#define AFFINITY close
#define DOUBLE_PI_T T(2.0 * 3.14159265358979323846)
#define DOUBLE_PI_X X(2.0 * 3.14159265358979323846)
#define no_op_exec_special_any static const bool requiresSpecial = false; static void execSpecial(X *dx, Nd4jLong *xShapeBuffer, Z *result, Nd4jLong *resultShapeBuffer, X *extraParams, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets) {}
#define no_op_exec_special_bool static const bool requiresSpecial = false; static void execSpecial(X *dx, Nd4jLong *xShapeBuffer, Z *result, Nd4jLong *resultShapeBuffer, X *extraParams, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets) {}
#define no_op_exec_special_same static const bool requiresSpecial = false; static void execSpecial(X *dx, Nd4jLong *xShapeBuffer, X *result, Nd4jLong *resultShapeBuffer, X *extraParams, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets) {}
#define no_op_exec_special static const bool requiresSpecial = false; static void execSpecial(X *dx, Nd4jLong *xShapeBuffer, Z *result, Nd4jLong *resultShapeBuffer, Z *extraParams, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets) {}
#define no_op_exec_special_accumulation static const bool requiresSpecialAccumulation = false; static void execSpecial(X *x, Nd4jLong *xShapeInfo, Z *extraParams, Z *result, Nd4jLong *resultShapeInfoBuffer, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffset){}
#define no_op_exec_special_accumulation_long static const bool requiresSpecialAccumulation = false; static void execSpecial(X *x, Nd4jLong *xShapeInfo, X *extraParams, Z *result, Nd4jLong *resultShapeInfoBuffer, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffset){}
#define no_op_exec_special_accumulation_same static const bool requiresSpecialAccumulation = false; static void execSpecial(X *x, Nd4jLong *xShapeInfo, X *extraParams, X *result, Nd4jLong *resultShapeInfoBuffer, int *dimension, int dimensionLength, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffset){}
#ifdef __CUDACC__
#define no_op_exec_special_any_cuda static __device__ void execSpecialCuda(X *dx, Nd4jLong *xShapeBuffer, Z *result, Nd4jLong *resultShapeBuffer, X *extraParams, int *allocationPointer, Z *reductionPointer, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets) {}
#define no_op_exec_special_bool_cuda static __device__ void execSpecialCuda(X *dx, Nd4jLong *xShapeBuffer, Z *result, Nd4jLong *resultShapeBuffer, X *extraParams, int *allocationPointer, Z *reductionPointer, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets) {}
#define no_op_exec_special_same_cuda static __device__ void execSpecialCuda(X *dx, Nd4jLong *xShapeBuffer, X *result, Nd4jLong *resultShapeBuffer, X *extraParams, int *allocationPointer, X *reductionPointer, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets) {}
#define no_op_exec_special_cuda static __device__ void execSpecialCuda(X *dx, Nd4jLong *xShapeBuffer,Z *result, Nd4jLong *resultShapeBuffer,Z *extraParams, int *allocationPointer, Z *reductionPointer, Nd4jLong *tadShapeInfo, Nd4jLong *tadOffsets) {}
#define no_op_exec_special_accumulation_same_cuda static inline __device__ void execSpecialCuda(X *dx, Nd4jLong *xShapeInfo, X *extraParams, X *result, Nd4jLong *resultShapeInfo, int *dimension, int dimensionLength, X *reductionBuffer, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets) {}
#define no_op_exec_special_accumulation_long_cuda static inline __device__ void execSpecialCuda(X *dx, Nd4jLong *xShapeInfo, X *extraParams, Z *result, Nd4jLong *resultShapeInfo, int *dimension, int dimensionLength, Z *reductionBuffer, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets) {}
#define no_op_exec_special_accumulation_cuda static inline __device__ void execSpecialCuda(X *dx, Nd4jLong *xShapeInfo, Z *extraParams, Z *result, Nd4jLong *resultShapeInfo, int *dimension, int dimensionLength, Z *reductionBuffer, Nd4jLong *tadOnlyShapeInfo, Nd4jLong *tadOffsets) {}
#else
// hacky fix for isnan/being being out of scope
//#ifdef IOS
//#define isinf(x) 0 // this isn't right. But std::isinf fails
//#define isnan(x) 0
//#else
//#define isnan std::isnan
//#define isinf std::isinf
//#endif
#define no_op_exec_special_cuda
#define no_op_exec_special_accumulation_cuda
#define no_op_exec_special_accumulation_same_cuda
#define no_op_exec_special_accumulation_long_cuda
#define no_op_exec_special_any_cuda
#define no_op_exec_special_bool_cuda
#define no_op_exec_special_same_cuda
#define no_op_exec_special_accumulation_same_cuda
#endif
#define SELU_ALPHA 1.6732632423543772848170429916717
#define SELU_LAMBDA 1.0507009873554804934193349852946
#ifdef _OPENMP
#pragma omp declare reduction(maxTF : float,double,float16,bfloat16 : \
omp_out = nd4j::math::nd4j_max(omp_in, omp_out) )\
initializer (omp_priv=-MAX_FLOAT)
#pragma omp declare reduction(minTF : float,double,float16,bfloat16 : \
omp_out = nd4j::math::nd4j_min(omp_in, omp_out) )\
initializer (omp_priv=MAX_FLOAT)
#pragma omp declare reduction(maxT : float,double,float16,bfloat16,int,Nd4jLong,Nd4jULong,int8_t,uint8_t,bool,int16_t,uint16_t,uint32_t : \
omp_out = nd4j::math::nd4j_max(omp_in, omp_out) )\
initializer (omp_priv=0)
#pragma omp declare reduction(minT : float,double,float16,bfloat16,int,Nd4jLong,Nd4jULong,int8_t,uint8_t,bool,int16_t,uint16_t,uint32_t : \
omp_out = nd4j::math::nd4j_min(omp_in, omp_out) )\
initializer (omp_priv=0)
#pragma omp declare reduction(amaxT : float,double,float16,bfloat16,int,Nd4jLong,Nd4jULong,int8_t,uint8_t,bool,int16_t,uint16_t,uint32_t : \
omp_out = nd4j::math::nd4j_max(nd4j::math::nd4j_abs(omp_in), nd4j::math::nd4j_abs(omp_out)) )
#pragma omp declare reduction(aminT : float,double,float16,bfloat16,int,Nd4jLong,Nd4jULong,int8_t,uint8_t,bool,int16_t,uint16_t,uint32_t : \
omp_out = nd4j::math::nd4j_min(nd4j::math::nd4j_abs(omp_in), nd4j::math::nd4j_abs(omp_out)) )
#pragma omp declare reduction(asumT : float,double,float16,bfloat16,int,Nd4jLong,Nd4jULong,int8_t,uint8_t,bool,int16_t,uint16_t,uint32_t : \
omp_out = nd4j::math::nd4j_abs(omp_in) + nd4j::math::nd4j_abs(omp_out))\
initializer (omp_priv=0)
#pragma omp declare reduction(sumT : float,double,float16,bfloat16,int,Nd4jLong,Nd4jULong,int8_t,uint8_t,bool,int16_t,uint16_t,uint32_t : \
omp_out = omp_in + omp_out)\
initializer (omp_priv=0)
#pragma omp declare reduction(prodT : float,double,float16,bfloat16,int,Nd4jLong,Nd4jULong,int8_t,uint8_t,bool,int16_t,uint16_t,uint32_t : \
omp_out = omp_in * omp_out)\
initializer (omp_priv=1)
#endif
namespace functions {
namespace indexreduce {
template <typename T>
struct IndexValue {
T value;
Nd4jLong index;
_CUDA_HD IndexValue() = default;
_CUDA_HD IndexValue(const T val, const Nd4jLong ind): index(ind), value(val) {}
};
}
namespace summarystats {
template <typename T>
class SummaryStatsData;
}
}
namespace simdOps {
template <typename X, typename Y, typename Z>
class Add {
public:
op_def static Z op(X d1, Y d2) {
return static_cast<Z>(d1 + d2);
}
op_def static Z op(X d1, Y d2, Z *params) {
return static_cast<Z>(d1 + d2);
}
op_def static Z op(X d1) {
return static_cast<Z>(d1);
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
return static_cast<Z>(d1 + params[0]);
}
op_def static X startingValue() {
return static_cast<X>(0.f);
}
};
template <typename X, typename Y>
class NewAdd {
public:
op_def static X op(X d1, Y d2, X *params) {
return d1 + d2;
}
};
template <typename X, typename Y, typename Z>
class Subtract {
public:
op_def static Z op(X d1, Y d2) {
return static_cast<Z>(d1 - d2);
}
op_def static Z op(X d1, Y d2, Z *params) {
return static_cast<Z>(d1 - d2);
}
op_def static Z op(X d1) {
return static_cast<Z>(d1);
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
return static_cast<Z>(d1 - params[0]);
}
};
template <typename X, typename Y, typename Z>
class SquaredSubtract {
public:
op_def static Z op(X d1, Y d2) {
auto d = static_cast<Z>(d1 - d2);
return d * d;
}
op_def static Z op(X d1, Y d2, Z *params) {
auto d = static_cast<Z>(d1 - d2);
return d * d;
}
op_def static Z op(X d1) {
return d1;
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
auto d = static_cast<Z>(d1 - params[0]);
return d * d;
}
};
template <typename X, typename Y, typename Z>
class SquaredReverseSubtract {
public:
op_def static Z op(X d1, Y d2) {
auto d = static_cast<Z>(d2 - d1);
return d * d;
}
op_def static Z op(X d1, Y d2, Z *params) {
auto d = static_cast<Z>(d2 - d1);
return d * d;
}
op_def static Z op(X d1) {
return d1;
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
auto d = static_cast<Z>(params[0] - d1);
return d * d;
}
};
template <typename X, typename Y, typename Z>
class ReverseSubtract {
public:
op_def static Z op(X d1, Y d2) {
return static_cast<Z>(d2 - d1);
}
op_def static Z op(X d1, Y d2, Z *params) {
return static_cast<Z>(d2 - d1);
}
op_def static Z op(X d1) {
return d1;
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
return static_cast<Z>(params[0] - d1);
}
};
template <typename X, typename Y, typename Z>
class LogPoissonLossFull {
public:
op_def static Z op(X z, Y c) {
auto zz = static_cast<Z>(z);
auto zc = static_cast<Z>(c);
return (nd4j::math::nd4j_exp<Y, Z>(c) - zz * zc + (zz * nd4j::math::nd4j_log<X, Z>(z) - zz + static_cast<Z>(0.5f) * nd4j::math::nd4j_log<Z, Z>(static_cast<Z>(DOUBLE_PI_X) * zz)));
}
op_def static Z op(X z, Y c, Z *params) {
auto zz = static_cast<Z>(z);
auto zc = static_cast<Z>(c);
return (nd4j::math::nd4j_exp<Y, Z>(c) - zz * zc + (zz * nd4j::math::nd4j_log<X, Z>(z) - zz + static_cast<Z>(0.5f) * nd4j::math::nd4j_log<Z, Z>(static_cast<Z>(DOUBLE_PI_X) * zz)));
}
op_def static Z op(X z) {
auto zz = static_cast<Z>(z);
return (zz * nd4j::math::nd4j_log<Y, Z>(z) - zz + static_cast<Z>(0.5f) * nd4j::math::nd4j_log<Z, Z>(static_cast<Z>(DOUBLE_PI_X) * zz));
}
// op for MetaOps
op_def static X op(X z, Y *params) {
return (nd4j::math::nd4j_exp<X, X>(params[0]) - z * params[0] + (z * nd4j::math::nd4j_log<X, Z>(z) - z + static_cast<X>(0.5f) * nd4j::math::nd4j_log<X, Z>(DOUBLE_PI_X * z)));
}
};
template <typename X, typename Y, typename Z>
class LogPoissonLoss {
public:
op_def static Z op(X z, Y c) {
auto zz = static_cast<Z>(z);
auto zc = static_cast<Z>(c);
return (nd4j::math::nd4j_exp<Y, Z>(c) - zz * zc);
}
op_def static Z op(X z, Y c, Z *params) {
auto zz = static_cast<Z>(z);
auto zc = static_cast<Z>(c);
return (nd4j::math::nd4j_exp<Y, Z>(c) - zz * zc);
}
op_def static Z op(X z) {
return static_cast<Z>(z);
}
// op for MetaOps
op_def static Z op(X z, Y *params) {
return (nd4j::math::nd4j_exp<Y, Z>(params[0]) - static_cast<Z>(z) * static_cast<Z>(params[0]));
}
};
template <typename X, typename Y, typename Z>
class Multiply {
public:
op_def static Z op(X d1, Y d2) {
return static_cast<Z>(d1 * d2);
}
op_def static Z op(X d1, Y d2, Z *params) {
return static_cast<Z>(d1 * d2);
}
op_def static Z op(X d1) {
return static_cast<Z>(d1);
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
return static_cast<Z>(d1 * params[0]);
}
op_def static X startingValue() {
return static_cast<X>(1.f);
}
};
template <typename X, typename Y, typename Z>
class Divide {
public:
op_def static Z op(X d1, Y d2) {
return static_cast<Z>(d1 / d2);
}
op_def static Z op(X d1, Y d2, Z *params) {
return static_cast<Z>(d1 / d2);
}
op_def static Z op(X d1) {
return static_cast<Z>(d1);
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
return static_cast<Z>(d1 / params[0]);
}
op_def static X startingValue() {
return static_cast<X>(1);
}
};
template <typename X, typename Y, typename Z>
class SafeDivide {
public:
op_def static Z op(X d1, Y d2) {
if(d2 == static_cast<Y>(0))
return static_cast<Z>(0);
return static_cast<Z>(d1 / d2);
}
op_def static Z op(X d1, Y d2, Z *params) {
if(d2 == static_cast<Y>(0))
return static_cast<Z>(0);
return static_cast<Z>(d1 / d2);
}
op_def static Z op(X d1) {
return static_cast<Z>(d1);
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
if(params[0] == static_cast<Y>(0))
return static_cast<Z>(0);
return static_cast<Z>(d1 / params[0]);
}
};
template <typename X, typename Y, typename Z>
class FloorDiv {
public:
op_def static Z op(X d1, Y d2) {
return nd4j::math::nd4j_floor<Z,Z>(static_cast<Z>(d1 / d2));
}
op_def static Z op(X d1, Y d2, Z *params) {
return nd4j::math::nd4j_floor<Z,Z>(static_cast<Z>(d1 / d2));
}
op_def static Z op(X d1) {
return nd4j::math::nd4j_floor<Z,Z>(static_cast<Z>(d1));
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
return nd4j::math::nd4j_floor<Z,Z>(static_cast<Z>(d1 / params[0]));
}
};
template <typename X, typename Y, typename Z>
class TruncateDiv {
public:
op_def static Z op(X d1, Y d2) {
auto i1 = static_cast<int>(d1);
auto i2 = static_cast<int>(d2);
return static_cast<Z>(i1 / i2);
}
op_def static Z op(X d1, Y d2, Z *params) {
auto i1 = static_cast<int>(d1);
auto i2 = static_cast<int>(d2);
return static_cast<Z>(i1 / i2);
}
op_def static Z op(X d1) {
return d1;
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
auto i1 = static_cast<int>(d1);
auto i2 = static_cast<int>(params[0]);
return static_cast<Z>(i1 / i2);
}
};
template <typename X, typename Y, typename Z>
class TruncateMod {
public:
op_def static Z op(X d1, Y d2) {
auto i1 = static_cast<int>(d1);
auto i2 = static_cast<int>(d2);
return static_cast<Z>(i1 % i2);
}
op_def static Z op(X d1, Y d2, Z *params) {
auto i1 = static_cast<int>(d1);
auto i2 = static_cast<int>(d2);
return static_cast<Z>(i1 % i2);
}
op_def static Z op(X d1) {
return static_cast<Z>(d1);
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
auto i1 = static_cast<int>(d1);
auto i2 = static_cast<int>(params[0]);
return static_cast<Z>(i1 % i2);
}
};
template<typename X, typename Y, typename Z>
class Remainder {
public:
op_def static Z op(X d1, Y d2) {
return nd4j::math::nd4j_remainder<X, Y, Z>(d1, d2);
}
op_def static Z op(X d1, Y d2, Z *params) {
return nd4j::math::nd4j_remainder<X, Y, Z>(d1, d2);
}
op_def static Z op(X d1) {
return d1;
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
return nd4j::math::nd4j_remainder<X, Y, Z>(d1, params[0]);
}
};
template <typename X, typename Y, typename Z>
class FMod {
public:
op_def static Z op(X d1, Y d2) {
return nd4j::math::nd4j_fmod<X, Y, Z>(d1, d2);
}
op_def static Z op(X d1, Y d2, Z *params) {
return nd4j::math::nd4j_fmod<X, Y, Z>(d1, d2);
}
op_def static Z op(X d1) {
return d1;
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
return nd4j::math::nd4j_fmod<X, Y, Z>(d1, params[0]);
}
};
template <typename X, typename Y, typename Z>
class FloorMod {
public:
op_def static Z op(X d1, Y d2) {
auto m = nd4j::math::nd4j_fmod<X, Y, Z>(d1, d2);
return (d1 < static_cast<X>(0)) == (d2 < static_cast<Y>(0)) ? m : nd4j::math::nd4j_fmod<Z, Y, Z>(m + static_cast<Z>(d2), d2);
}
op_def static Z op(X d1, Y d2, Z *params) {
auto m = nd4j::math::nd4j_fmod<X, Y, Z>(d1, d2);
return (d1 < static_cast<X>(0.0f)) == (d2 < static_cast<Y>(0)) ? m : nd4j::math::nd4j_fmod<Z, Y, Z>(m + static_cast<Z>(d2), d2);
}
op_def static Z op(X d1) {
return d1;
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
return op(d1, params[0]);
}
};
template <typename X, typename Y, typename Z>
class ReverseDivide {
public:
op_def static Z op(X d1, Y d2) {
return static_cast<Z>(d2 / d1);
}
op_def static Z op(X d1, Y d2, Z *params) {
return static_cast<Z>(d2 / d1);
}
op_def static Z op(X d1) {
return static_cast<Z>(d1);
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
return static_cast<Z>(params[0] / d1);
}
};
template <typename X, typename Y, typename Z>
class CopyPws {
public:
op_def static Z op(X d1, Y d2) {
return static_cast<Z>(d2);
}
op_def static Z op(X d1, Y d2, Z *params) {
return static_cast<Z>(d2);
}
op_def static Z op(X d1) {
return static_cast<Z>(d1);
}
op_def static Z op(X d1, Y *params) {
return static_cast<Z>(d1);
}
};
template <typename X>
class Copy {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return d1;
}
};
template <typename X, typename Y, typename Z>
class Copy2 {
public:
op_def static Z op(X d1, Y d2) {
return static_cast<Z>(d2);
}
op_def static Z op(X d1, Y d2, Z *params) {
return static_cast<Z>(d2);
}
op_def static Z op(X d1) {
return static_cast<Z>(d1);
}
op_def static Z op(X d1, Y *params) {
return static_cast<Z>(d1);
}
};
template <typename X, typename Y, typename Z>
class Axpy {
public:
op_def static Z op(X d1, Y d2) {
return static_cast<Z>(d2 + d1);
}
op_def static Z op(X d1, Y d2, Z *params) {
auto alpha = params[0];
return alpha * static_cast<Z>(d1) + static_cast<Z>(d2);
}
op_def static Z op(X d1) {
return static_cast<Z>(d1);
}
};
template <typename X, typename Z>
class Assign {
public:
no_op_exec_special_any
no_op_exec_special_any_cuda
op_def static Z op(X d1, X *params) {
return static_cast<Z>(d1);
}
};
template <typename X, typename Z>
class And {
public:
no_op_exec_special_bool
no_op_exec_special_bool_cuda
op_def static Z op(X d1, X d2) {
return d2 + d1;
}
op_def static Z op(X d1, X d2, X *params) {
if (params != nullptr) {
auto comp = params[0];
return d1 != comp && d2 != comp ? static_cast<Z>(1) : static_cast<Z>(0);
} else {
auto b1 = static_cast<bool>(d1);
auto b2 = static_cast<bool>(d2);
return (b1 && b2) ? static_cast<Z>(1) : static_cast<Z>(0);
}
}
op_def static Z op(X d1) {
return d1;
}
// op for MetaOps
op_def static Z op(X d1, X *params) {
return static_cast<Z>(119);
}
};
template <typename X, typename Z>
class Or {
public:
no_op_exec_special_bool
no_op_exec_special_bool_cuda
op_def static Z op(X d1, X d2) {
return d2 + d1;
}
op_def static Z op(X d1, X d2, X *params) {
if (params != nullptr) {
auto comp = params[0];
return d1 != comp || d2 != comp ? static_cast<Z>(1) : static_cast<Z>(0);
} else {
auto b1 = static_cast<bool>(d1);
auto b2 = static_cast<bool>(d2);
return b1 || b2 ? static_cast<Z>(1) : static_cast<Z>(0);
}
}
op_def static Z op(X d1) {
return d1;
}
// op for MetaOps
op_def static Z op(X d1, X *params) {
return static_cast<Z>(119);
}
};
template <typename X, typename Z>
class Xor {
public:
no_op_exec_special_bool
no_op_exec_special_bool_cuda
op_def static Z op(X d1, X d2) {
return d2 + d1;
}
op_def static Z op(X d1, X d2, X *params) {
if (params != nullptr) {
auto comp = params[0];
return ((d1 == comp && d2 != comp) || (d1 != comp && d2 == comp)) ? static_cast<Z>(1) : static_cast<Z>(0);
} else {
auto b1 = static_cast<bool>(d1);
auto b2 = static_cast<bool>(d2);
return (!b1 && b2 )||(b1 && !b2) ? static_cast<Z>(1) : static_cast<Z>(0);
}
}
op_def static Z op(X d1) {
return d1;
}
};
template <typename X, typename Z>
class Not {
public:
no_op_exec_special_bool
no_op_exec_special_bool_cuda
op_def static Z op(X d1, X d2) {
return static_cast<Z>(0);
}
op_def static Z op(X d1, X d2, X *params) {
return d1 != d2 ? static_cast<Z>(1) : static_cast<Z>(0);
}
// this transform op should run only on boolean input
op_def static Z op(X d1, X *params) {
auto b1 = static_cast<bool>(d1);
return !b1;
}
};
template <typename X, typename Y, typename Z>
class LogicalNot {
public:
op_def static Z op(X d1, Y d2) {
return !((int) d1 && (int) d2);
}
op_def static Z op(X d1, Y d2, Z *params) {
return static_cast<X>(!(static_cast<int>(d1) && static_cast<int>(d2)));
}
op_def static Z op(X d1) {
return d1;
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
return static_cast<X>(119);
}
};
template <typename X, typename Y, typename Z>
class LogicalXor {
public:
op_def static Z op(X d1, Y d2) {
auto i1 = static_cast<int>(d1);
auto i2 = static_cast<int>(d2);
return (i1 | i2) &~ (i1 & i2);
}
op_def static Z op(X d1, Y d2, Z *params) {
return op(d1, d2);
}
op_def static Z op(X d1) {
return d1;
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
return static_cast<Z>(119);
}
};
template <typename X, typename Y, typename Z>
class LogicalAnd {
public:
op_def static Z op(X d1, Y d2) {
return static_cast<int>(d1) & static_cast<int>(d2);
}
op_def static Z op(X d1, Y d2, Z *params) {
return op(d1, d2);
}
op_def static Z op(Y d1) {
return d1;
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
return static_cast<Z>(119);
}
};
template <typename X, typename Y, typename Z>
class LogicalOr {
public:
op_def static Z op(X d1, Y d2) {
return static_cast<int>(d1) | static_cast<int>(d2);
}
op_def static Z op(X d1, Y d2, Z *params) {
return op(d1, d2);
}
op_def static Z op(X d1) {
return d1;
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
return static_cast<X>(119);
}
};
template <typename X, typename Y, typename Z>
class Mod {
public:
/*
// just a optional note, feel free to remove later
op_def static half op(half d1, half d2, half *params) {
return __float2half(simdOps::Mod<float>::op(__half2float(d1), __half2float(d2), nullptr));
}
*/
op_def static Z op(X d1, Y d2) {
return static_cast<int>(d1) % static_cast<int>(d2);
}
op_def static Z op(X d1, Y d2, Z *params) {
return op(d1, d2);
}
// op for MetaOp
op_def static Z op(X d1, Y *params) {
return op(d1, params[0]);
}
};
template <typename X, typename Y, typename Z>
class ReverseMod {
public:
op_def static Z op(X d1, Y d2) {
return static_cast<int>(d2) % static_cast<int>(d1);
}
op_def static Z op(X d1, Y d2, Z *params) {
return op(d1, d2);
}
// op for MetaOp
op_def static Z op(X d1, Y *params) {
return op(d1, params[0]);
}
};
/**
* Whether 2 elements in an array
* are epsilion equal
*/
template <typename X, typename Z>
class Epsilon {
public:
op_def static Z op(X d1, X d2) {
X diff = d1 - d2;
X absDiff = nd4j::math::nd4j_abs<X>(diff);
if (absDiff <= static_cast<X>(MIN_V))
return static_cast<Z>(1);
return static_cast<Z>(0);
}
op_def static Z op(X d1, X d2, X *params) {
return op(d1, d2);
}
op_def static Z op(X d1, X *params) {
return d1;
}
};
template <typename X, typename Z>
class EqualTo {
public:
op_def static Z op(X d1, X d2) {
return d1 == d2;
}
op_def static Z op(X d1, X d2, X *params) {
return op(d1, d2);
}
op_def static Z op(X d1, X *params) {
return d1;
}
};
template <typename X, typename Z>
class NotEqualTo {
public:
op_def static Z op(X d1, X d2) {
return d1 != d2;
}
op_def static Z op(X d1, X d2, X *params) {
return op(d1, d2);
}
op_def static Z op(X d1, X *params) {
return d1;
}
};
template <typename X, typename Z>
class GreaterThanOrEqual {
public:
op_def static Z op(X d1, X d2) {
return d1 >= d2;
}
op_def static Z op(X d1, X d2, X *params) {
return op(d1, d2);
}
// FIXME: this signature clashes with MetaOp stuff
op_def static Z op(X d1, X *params) {
return d1;
}
};
template <typename X, typename Z>
class GreaterThan {
public:
op_def static Z op(X d1, X d2) {
return d1 > d2;
}
op_def static Z op(X d1, X d2, X *params) {
return op(d1, d2);
}
// FIXME: this signature clashes with MetaOp stuff
op_def static Z op(X d1, X *params) {
return d1;
}
};
template <typename X, typename Z>
class LessThan {
public:
op_def static Z op(X d1, X d2) {
return d1 < d2;
}
op_def static Z op(X d1, X d2, X *params) {
return op(d1, d2);
}
op_def static Z op(X d1, X *params) {
return d1;
}
};
template <typename X, typename Z>
class LessThanOrEqual {
public:
op_def static Z op(X d1, X d2) {
return d1 <= d2;
}
op_def static Z op(X d1, X d2, X *params) {
return op(d1, d2);
}
op_def static Z op(X d1, X *params) {
return d1;
}
};
template <typename X>
class Abs {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_abs<X>(d1);
}
};
template <typename X>
class Ceiling {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_ceil<X,X>(d1);
}
};
template <typename X>
class Cosine {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_cos<X,X>(d1);
}
};
template <typename X>
class Exp {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_exp<X, X>(d1);
}
};
template <typename X>
class HardTanhDerivative {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return ((d1 >= static_cast<X>(-1.f) && d1 <= static_cast<X>(1.f)) ? static_cast<X>(1.f) : static_cast<X>(0.f));
}
};
template <typename X>
class HardTanh {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
if (d1 < static_cast<X>(-1))
return static_cast<X>(-1);
else if (d1 > static_cast<X>(1))
return static_cast<X>(1);
else
return d1;
}
};
template <typename X>
class Floor {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_floor<X,X>(d1);
}
};
template <typename X>
class Log {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_log<X, X>(d1);
}
};
template <typename X>
class Log1p {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_log<X, X>(1 + d1);
}
};
template <typename X, typename Y, typename Z>
class LogX {
public:
op_def static Z op(X d1, Y d2, Z *params) {
return nd4j::math::nd4j_log<X, Z>(d1) / nd4j::math::nd4j_log<Y, Z>(d2) ;
}
};
template <typename X>
class StabilizeFP16 {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
if (d1 <= static_cast<X>(0))
return static_cast<X>(nd4j::DataTypeUtils::min<float16>());
else return d1;
}
};
template <typename X>
class StabilizeX {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
if (d1 <= static_cast<X>(0))
return nd4j::DataTypeUtils::min<X>();
else return d1;
}
};
template <typename X>
class SpecialDerivative {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return d1 * (static_cast<X>(1.f) - d1);
}
};
template <typename X>
class Neg {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return -d1;
}
};
template <typename X>
class Erf {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_erf<X,X>(d1);
}
};
template <typename X>
class Erfc {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_erfc<X,X>(d1);
}
};
template <typename X>
class Reciprocal {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
// op_def static T op(T d1) {
// return (T(1.0f) / d1);
// }
// op for MetaOps
op_def static X op(X d1, X *params) {
return (static_cast<X>(1) / d1);
}
};
template <typename X, typename Z>
class Sqr {
public:
no_op_exec_special
no_op_exec_special_cuda
op_def static Z op(X d1, Z *params) {
return nd4j::math::nd4j_pow<X, X, Z>(d1, static_cast<X>(2));
}
op_def static Z op(X d1) {
return nd4j::math::nd4j_pow<X, X, Z>(d1, static_cast<X>(2));
}
};
template <typename X, typename Y, typename Z>
class RelativeError {
public:
no_op_exec_special
no_op_exec_special_cuda
op_def static Z op(X d1, Y d2) {
return nd4j::math::nd4j_re<X>(d1, d2);
}
op_def static Z op(X d1, Y d2, Z *params) {
return op(d1, d2);
}
op_def static Z op(X d1) {
return static_cast<Z>(0);
}
};
template <typename X, typename Y, typename Z>
class BinaryRelativeError {
public:
no_op_exec_special
no_op_exec_special_cuda
op_def static Z op(X d1, Y d2, Z *params) {
X threshold = params[0];
return nd4j::math::nd4j_re<X>(d1, d2) > threshold ? static_cast<Z>(1) : static_cast<Z>(0);
}
op_def static Z op(X d1) {
return static_cast<Z>(0);
}
};
template <typename X, typename Y, typename Z>
class BinaryMinimumAbsoluteRelativeError {
public:
no_op_exec_special
no_op_exec_special_cuda
op_def static Z op(X d1, X *params) {
X d2 = params[0];
X thresholdRelative = params[1];
X thresholdAbsolute = params[2];
return nd4j::math::nd4j_re<X>(d1, d2) > thresholdRelative ? (nd4j::math::nd4j_abs<X>(d1 - static_cast<X>(d2)) < thresholdAbsolute ? static_cast<Z>(0) : static_cast<Z>(1)) : static_cast<Z>(0);
}
op_def static Z op(X d1, Y d2, Z *params) {
X thresholdRelative = params[0];
X thresholdAbsolute = params[1];
return nd4j::math::nd4j_re<X>(d1, d2) > thresholdRelative ? (nd4j::math::nd4j_abs<X>(d1 - static_cast<X>(d2)) < thresholdAbsolute ? static_cast<Z>(0) : static_cast<Z>(1)) : static_cast<Z>(0);
}
op_def static Z op(X d1) {
return static_cast<Z>(0);
}
};
template <typename X, typename Y, typename Z>
class ReversePow {
public:
no_op_exec_special
no_op_exec_special_cuda
op_def static Z op(X d1, Z *params) {
return nd4j::math::nd4j_pow<X, X, Z>(params[0], d1);
}
op_def static Z op(X d1, Y d2) {
return nd4j::math::nd4j_pow<X, Y, Z>(d2, d1);
}
op_def static Z op(X d1, Y d2, Z *params) {
return nd4j::math::nd4j_pow<X, Y, Z>(d2, d1);
}
op_def static Z op(X d1) {
return d1;
}
};
template <typename X, typename Y, typename Z>
class Pow {
public:
no_op_exec_special
no_op_exec_special_cuda
op_def static Z op(X d1, Z *params) {
return nd4j::math::nd4j_pow<X, X, Z>(d1, params[0]);
}
op_def static Z op(X d1, Y d2) {
return nd4j::math::nd4j_pow<X, Y, Z>(d1, d2);
}
op_def static Z op(X d1, Y d2, Z *params) {
return nd4j::math::nd4j_pow<X, Y, Z>(d1, d2);
}
op_def static Z op(X d1) {
return d1;
}
};
template <typename X, typename Y, typename Z>
class PowDerivative {
public:
no_op_exec_special
no_op_exec_special_cuda
op_def static Z op(X d1, Z *params) {
return params[0] * nd4j::math::nd4j_pow<X, Z, Z>(d1, static_cast<Z>(params[0]) - static_cast<Z>(1.f));
}
op_def static Z op(X d1, Y d2) {
return static_cast<Z>(d2) * nd4j::math::nd4j_pow<X, Z, Z>(d1, static_cast<Z>(d2) - static_cast<Z>(1.f));
}
op_def static Z op(X d1, Y d2, Z *params) {
return static_cast<Z>(d2) * nd4j::math::nd4j_pow<X, Z, Z>(d1, static_cast<Z>(d2) - static_cast<Z>(1.f));
}
op_def static Z op(X d1) {
return d1;
}
};
template <typename X>
class Round {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_round<X,X>(d1);
}
};
template <typename X, typename Z>
class IsNan {
public:
no_op_exec_special_bool
no_op_exec_special_bool_cuda
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
op_def static Z op(X d1, X *params) {
return nd4j::math::nd4j_isnan(d1) ? static_cast<X>(1) : static_cast<X>(0);
}
op_def static X startingValue(const X *input) {
return static_cast<X>(0);
}
op_def static Z merge(X old, X opOutput, X *extraParams) {
return opOutput + old;
}
op_def static Z update(X old, X opOutput, X *extraParams) {
return opOutput + old;
}
op_def static Z postProcess(X reduction, Nd4jLong n, X *extraParams) {
return reduction;
}
};
template <typename X>
class Expm1 {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_exp<X, X>(d1) - static_cast<X>(1);
}
};
template <typename X, typename Z>
class IsPositive {
public:
no_op_exec_special_bool
no_op_exec_special_bool_cuda
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
op_def static Z op(X d1, X *params) {
return d1 > (X)0.f;
}
op_def static X startingValue(const X *input) {
return static_cast<X>(0);
}
op_def static Z merge(X old, X opOutput, X *extraParams) {
return opOutput + old;
}
op_def static Z update(X old, X opOutput, X *extraParams) {
return opOutput + old;
}
op_def static Z postProcess(X reduction, Nd4jLong n, X *extraParams) {
return reduction;
}
};
template <typename X, typename Z>
class IsInf {
public:
no_op_exec_special_bool
no_op_exec_special_bool_cuda
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
op_def static Z op(X d1, X *params) {
return nd4j::math::nd4j_isinf<X>(d1) ? static_cast<Z>(1) : static_cast<Z>(0);
}
op_def static X startingValue(const X *input) {
return static_cast<X>(0);
}
op_def static Z merge(X old, X opOutput, X *extraParams) {
return opOutput + old;
}
op_def static Z update(X old, X opOutput, X *extraParams) {
return opOutput + old;
}
op_def static Z postProcess(X reduction, Nd4jLong n, X *extraParams) {
return reduction;
}
};
template <typename X, typename Z>
class IsInfOrNan{
public:
no_op_exec_special_bool
no_op_exec_special_bool_cuda
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
op_def static Z op(X d1, X *params) {
return nd4j::math::nd4j_isfin<X>(d1) ? static_cast<Z>(0) : static_cast<Z>(1);
}
op_def static X startingValue(const X *input) {
return static_cast<X>(0);
}
op_def static Z merge(X old, X opOutput, X *extraParams) {
return opOutput + old;
}
op_def static Z update(X old, X opOutput, X *extraParams) {
return opOutput + old;
}
op_def static Z postProcess(X reduction, Nd4jLong n, X *extraParams) {
return reduction;
}
};
template <typename X, typename Z>
class IsFinite {
public:
no_op_exec_special_bool
no_op_exec_special_bool_cuda
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
op_def static Z op(X d1, X *params) {
return nd4j::math::nd4j_isfin<X>(d1) ? static_cast<Z>(1) : static_cast<Z>(0);
}
op_def static X startingValue(const X *input) {
return static_cast<X>(0);
}
op_def static Z merge(X old, X opOutput, X *extraParams) {
return opOutput + old;
}
op_def static Z update(X old, X opOutput, X *extraParams) {
return opOutput + old;
}
op_def static Z postProcess(X reduction, Nd4jLong n, X *extraParams) {
return reduction;
}
};
template <typename X>
class ClipByValue {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
if (d1 > params[1])
return params[1];
if (d1 < params[0])
return params[0];
return d1;
}
};
template <typename X, typename Y, typename Z>
class LstmClip {
public:
no_op_exec_special
no_op_exec_special_cuda
op_def static Z op(X d1, Y d2, Z *params) {
X _v = (X) d2;
if (d1 > _v)
return _v;
else if (d1 < -_v)
return -_v;
else return d1;
}
};
template <typename X>
class Swish {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return d1 * nd4j::math::nd4j_sigmoid<X,X>(d1);
}
};
template <typename X>
class GELU {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return d1 * nd4j::math::nd4j_sigmoid<X,X>(static_cast<X>(1.702f) * d1);
}
};
template <typename X>
class PreciseGELU {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
auto sp = nd4j::math::nd4j_sqrt<X, X>(static_cast<X>(2) / static_cast<X>(M_PI));
auto xp = d1 + nd4j::math::nd4j_pow<X, X, X>(static_cast<X>(0.044715) * d1, static_cast<X>(3));
return (d1 / static_cast<X>(2)) * (static_cast<X>(1) + nd4j::math::nd4j_tanh<X, X>(sp * xp));
}
};
template <typename X>
class GELUDerivative {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
auto x17 = static_cast<X>(1.702f) * d1;
auto ep = nd4j::math::nd4j_pow<X,X,X>(static_cast<X>(M_E), x17);
// (E^(1.702 x) (1. + E^(1.702 x) + 1.702 x))/(1. + E^(1.702 x))^2
return (ep * (static_cast<X>(1.f) + ep + x17)) / nd4j::math::nd4j_pow<X, int, X>((static_cast<X>(1.f) + ep), 2);
}
};
template <typename X>
class PreciseGELUDerivative {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
auto x79 = static_cast<X>(0.797885) * d1;
auto x03 = nd4j::math::nd4j_pow<X, int, X>(static_cast<X>(0.0356774) * d1, 3);
auto x39 = static_cast<X>(0.398942) * d1;
auto x05 = nd4j::math::nd4j_pow<X, int, X>(static_cast<X>(0.0535161) * d1, 3);
auto scz = nd4j::math::nd4j_sech<X, X>(x79 + x03);
// 0.5 + (0.398942 x + 0.0535161 x^3) Sech[0.797885 x + 0.0356774 x^3]^2 + 0.5 Tanh[0.797885 x + 0.0356774 x^3]
return static_cast<X>(0.5) + (x39 + x05) * (scz * scz) + static_cast<X>(0.5) * nd4j::math::nd4j_tanh<X, X>(x79 + x03);
}
};
template <typename X>
class SwishDerivative {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
X ex = nd4j::math::nd4j_pow<X, X, X>(static_cast<X>(M_E), d1);
return (ex * (d1 + ex + static_cast<X>(1.f))) / nd4j::math::nd4j_pow<X, X, X>((ex + static_cast<X>(1.f)) , static_cast<X>(2.f));
}
};
template <typename X>
class LogSigmoid {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_log<X, X>(nd4j::math::nd4j_sigmoid<X, X>(d1));
}
};
template <typename X>
class LogSigmoidDerivative {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
X ex = nd4j::math::nd4j_pow<X, X, X>(M_E, d1);
return static_cast<X>(1.f) / (ex + static_cast<X>(1.f));
}
};
template <typename X>
class Sigmoid {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_sigmoid<X, X>(d1);
}
};
template <typename X>
class SigmoidDerivative {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_sigmoidderivative<X, X>(d1);
}
};
template <typename X>
class HardSigmoid {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_min<X>(static_cast<X>(1), nd4j::math::nd4j_max<X>(static_cast<X>(0), (static_cast<X>(0.2f)) * d1 + static_cast<X>(0.5f)));
}
};
template <typename X>
class HardSigmoidDerivative {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return d1 < static_cast<X>(-2.5f) || d1 > static_cast<X>(2.5f) ? static_cast<X>(0.f) : static_cast<X>(0.2f);
}
};
/**
* Scale to be between a min and max
*/
template <typename X>
class SetRange {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
auto min = params[0];
auto max = params[1];
if (static_cast<X>(d1) >= min && static_cast<X>(d1) <= max)
return d1;
if (min == static_cast<X>(0) && max == static_cast<X>(1)) {
auto val = static_cast<X>(1) / (static_cast<X>(1) + nd4j::math::nd4j_exp<X, X>(-d1));
return (nd4j::math::nd4j_floor<X,X>(val * (max - min)) + min);
}
return (nd4j::math::nd4j_floor<X,X>(d1 * (max - min)) + min);
}
};
template <typename X>
class Sin {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_sin<X,X>(d1);
}
};
template <typename X>
class Square {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return d1 * d1;
}
};
template <typename X, typename Z>
class Sqrt {
public:
no_op_exec_special
no_op_exec_special_cuda
op_def static Z op(X d1, Z *params) {
return nd4j::math::nd4j_sqrt<X, Z>(d1);
}
};
template <typename X, typename Z>
class RSqrt {
public:
no_op_exec_special
no_op_exec_special_cuda
op_def static Z op(X d1, Z *params) {
return static_cast<Z>(1) / nd4j::math::nd4j_sqrt<X, Z>(d1);
}
};
template <typename X>
class Rint {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_rint<X,X>(d1);
}
};
template <typename X>
class SoftPlus {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::softplus<X, X>(d1);
}
};
template <typename X>
class Sign {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return (d1 > static_cast<X>(0)) - (d1 < static_cast<X>(0));
}
};
template <typename X>
class TimesOneMinus {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return d1 * (static_cast<X>(1) - d1);
}
};
template <typename X>
class RationalTanh {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
// keep 2/3 as runtime variable, to match precision
auto dis = (static_cast<X>(2) / static_cast<X>(3)) * d1;
auto tanh = nd4j::math::nd4j_sgn<X,X>(dis) * (static_cast<X>(1) - (static_cast<X>(1) / (static_cast<X>(1) + static_cast<X>(nd4j::math::nd4j_abs<X>(dis)) + nd4j::math::nd4j_pow<X, X, X>(dis, static_cast<X>(2)) + static_cast<X>(1.41645f) * nd4j::math::nd4j_pow<X, X, X>(dis, static_cast<X>(4)) )));
return static_cast<X>(1.7159f) * tanh;
}
};
template <typename X>
class RationalTanhDerivative {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
auto dis = (static_cast<X>(2.f) / static_cast<X>(3.f)) * d1;
auto a = static_cast<X>(1.f) + nd4j::math::nd4j_abs<X>(dis) + nd4j::math::nd4j_pow<X, X, X>(dis, static_cast<X>(2.f)) + static_cast<X>(1.41645f) * nd4j::math::nd4j_pow<X, X, X>(dis, static_cast<X>(4));
auto tDeriv = (static_cast<X>(1.f) + nd4j::math::nd4j_sign<X,X>(dis) * (static_cast<X>(2.f) * dis + static_cast<X>(4.f) * static_cast<X>(1.41645f) * nd4j::math::nd4j_pow<X, X, X>(dis, static_cast<X>(3)))) / (a * a);
return static_cast<X>(1.7159f) * (static_cast<X>(2.f) / static_cast<X>(3.f)) * tDeriv;
}
};
template <typename X>
class Tanh {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_tanh<X, X>(d1);
}
};
template <typename X>
class RectifiedTanh {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_max<X>(static_cast<X>(0), nd4j::math::nd4j_tanh<X,X>(d1));
}
};
template <typename X>
class RectifiedTanhDerivative {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return d1 > static_cast<X>(0.f) ? nd4j::math::nd4j_tanhderivative<X,X>(d1) : static_cast<X>(0.f);
}
};
template <typename X>
class ATanh {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_atanh<X,X>(d1);
}
};
template <typename X>
class TanhDerivative {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_tanhderivative<X,X>(d1);
}
};
template <typename X>
class Cube {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return d1 * d1 * d1;
}
};
template <typename X>
class CubeDerivative {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return static_cast<X>(3) * d1 * d1;
}
};
template <typename X>
class ACos {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_acos<X, X>(d1);
}
};
template <typename X>
class ASinh {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_asinh<X, X>(d1);
}
};
template <typename X>
class ASinhDerivative {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return static_cast<X>(1.f) / (nd4j::math::nd4j_sqrt<X, X>(nd4j::math::nd4j_pow<X, X, X>(d1, static_cast<X>(2.f)) + static_cast<X>(1.f)));
}
};
template <typename X>
class ACosh {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_acosh<X, X>(d1);
}
};
template <typename X>
class ACoshDerivative {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return static_cast<X>(1.f) / (nd4j::math::nd4j_sqrt<X, X>(d1 - static_cast<X>(1.f)) * nd4j::math::nd4j_sqrt<X, X>(d1 + static_cast<X>(1.f)));
}
};
template <typename X>
class Ones {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return static_cast<X>(1.0f);
}
};
template <typename X>
class SoftSign {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_softsign<X, X>(d1);
}
};
template <typename X>
class SoftSignDerivative {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_softsignderivative<X,X>(d1);
}
};
template <typename X, typename Z>
class MatchConditionBool {
public:
no_op_exec_special_bool
no_op_exec_special_bool_cuda
// this op return 1.0 if condition met, 0.0 otherwise
op_def static Z op(X d1, X *extraParams) {
X compare = extraParams[0];
X eps = extraParams[1];
auto mode = static_cast<int>(extraParams[2]);
//nd4j_printf("value: %f; comp: %f; eps: %f; mode: %i;\n", d1, compare, eps, mode);
switch (mode) {
case 0: // equals
return nd4j::math::nd4j_abs<X>(d1 - compare) <= eps ? true : false;
case 1: // not equals
return nd4j::math::nd4j_abs<X>(d1 - compare) > eps ? true : false;
case 2: // less_than
return d1 < compare ? true : false;
case 3: // greater_than
return d1 > compare ? true : false;
case 4: // less_or_equals_than
return d1 <= compare ? true : false;
case 5: // greater_or_equals_than
return d1 >= compare ? true : false;
case 6: // abs_less_than
return nd4j::math::nd4j_abs<X>(d1) < compare ? true : false;
case 7: // abs_greater_than
return nd4j::math::nd4j_abs<X>(d1) > compare ? true : false;
case 8: // is inf
return nd4j::math::nd4j_isinf(d1) ? true : false;
case 9: // is nan
return nd4j::math::nd4j_isnan(d1) ? true : false;
case 10:
return (d1 == compare) ? true : false;
case 11:
return (d1 != compare) ? true : false;
case 12: // abs_greater_or_equals_than
return nd4j::math::nd4j_abs<X>(d1) >= compare ? true : false;
case 13: // abs_less_or_equals_than
return nd4j::math::nd4j_abs<X>(d1) <= compare ? true : false;
case 14:
// isFinite
return !(nd4j::math::nd4j_isinf(d1) || nd4j::math::nd4j_isnan(d1));
case 15:
// isInfinite
return nd4j::math::nd4j_isinf(d1) || nd4j::math::nd4j_isnan(d1);
default:
printf("Undefined match condition: [%i]\n", mode);
}
return d1;
}
};
template <typename X, typename Z>
class MatchCondition {
public:
no_op_exec_special
no_op_exec_special_cuda
no_op_exec_special_accumulation_long
no_op_exec_special_accumulation_cuda
op_def static Z startingValue(const X *input) {
return static_cast<Z>(0);
}
op_def static Z merge(Z old, Z opOutput, X *extraParams) {
return old + opOutput;
}
op_def static Z update(Z old, Z opOutput, X *extraParams) {
return old + opOutput;
}
// this op return 1.0 if condition met, 0.0 otherwise
op_def static Z op(X d1, X *extraParams) {
X compare = extraParams[0];
X eps = extraParams[1];
auto mode = static_cast<int>(extraParams[2]);
//printf("value: %f; comp: %f; eps: %f; mode: %i;\n", (float) d1, (float) compare, (float) eps, mode);
switch (mode) {
case 0: // equals
return nd4j::math::nd4j_abs<X>(d1 - compare) <= eps ? 1 : 0;
case 1: // not equals
return nd4j::math::nd4j_abs<X>(d1 - compare) > eps ? 1 : 0;
case 2: // less_than
return d1 < compare ? 1 : 0;
case 3: // greater_than
return d1 > compare ? 1 : 0;
case 4: // less_or_equals_than
return d1 <= compare ? 1 : 0;
case 5: // greater_or_equals_than
return d1 >= compare ? 1 : 0;
case 6: // abs_less_than
return nd4j::math::nd4j_abs<X>(d1) < compare ? 1 : 0;
case 7: // abs_greater_than
return nd4j::math::nd4j_abs<X>(d1) > compare ? 1 : 0;
case 8: // is inf
return nd4j::math::nd4j_isinf(d1) ? 1 : 0;
case 9: // is nan
return nd4j::math::nd4j_isnan(d1) ? 1 : 0;
case 10:
return (d1 == compare) ? 1 : 0;
case 11:
return (d1 != compare) ? 1 : 0;
case 12: // abs_greater_or_equals_than
return nd4j::math::nd4j_abs<X>(d1) >= compare ? 1 : 0;
case 13: // abs_less_or_equals_than
return nd4j::math::nd4j_abs<X>(d1) <= compare ? 1 : 0;
case 14:
// isFinite
return !(nd4j::math::nd4j_isinf(d1) || nd4j::math::nd4j_isnan(d1)) ? 1 : 0;
case 15:
// isInfinite
return nd4j::math::nd4j_isinf(d1) || nd4j::math::nd4j_isnan(d1) ? 1 : 0;
default:
printf("Undefined match condition: [%i]\n", mode);
}
return d1;
}
op_def static Z postProcess(Z reduction, Nd4jLong n, X *extraParams) {
return reduction;
}
};
template <typename X>
class ELU {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_elu<X,X>(d1);
}
};
template <typename X>
class ELUDerivative {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_eluderivative<X,X>(d1);
}
};
template <typename X, typename Y, typename Z>
class RELU {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static Z op(X d1, Y d2, Z *params) {
auto xt = static_cast<Z>(d1);
auto xf = static_cast<Z>(d2);
return xt < xf ? xf : xt;
}
};
template <typename X, typename Y, typename Z>
class SXELogitsSmoother {
public:
op_def static Z op(X d1, Y d2, Z *params) {
return d1 * ((X)1.f - (X) d2) + (X)(0.5f) * (X) d2;
}
};
template <typename X, typename Y, typename Z>
class RELU6 {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static Z op(X d1, Y d2, Z *params) {
auto relu = simdOps::RELU<X,Y,Z>::op(d1, d2, params);
return relu < static_cast<Z>(6) ? relu : static_cast<Z>(6);
}
};
template <typename X, typename Y, typename Z>
class LeakyRELU {
public:
no_op_exec_special
no_op_exec_special_cuda
op_def static Z op(X d1, Y d2, Z *params) {
auto val = static_cast<Z>(d1);
auto alpha = static_cast<Z>(d2);
return val < 0.0f ? alpha * val : val;
}
};
template <typename X>
class SELU {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return d1 > static_cast<X>(0.0f) ? static_cast<X>(SELU_LAMBDA) * static_cast<X>(d1) : static_cast<X>(SELU_LAMBDA) * (static_cast<X>(SELU_ALPHA) * nd4j::math::nd4j_exp<X, X>(d1) - static_cast<X>(SELU_ALPHA));
}
};
template <typename X>
class SELUDerivative {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return d1 > static_cast<X>(0.f) ? static_cast<X>(SELU_LAMBDA) : static_cast<X>(SELU_ALPHA) * static_cast<X>(SELU_LAMBDA) * nd4j::math::nd4j_exp<X, X>(d1);
}
};
template <typename X, typename Y, typename Z>
class LeakyRELUDerivative {
public:
no_op_exec_special
no_op_exec_special_cuda
op_def static Z op(X d1, Y d2, Z *params) {
if (d1 >= static_cast<X>(0))
return static_cast<Z>(1);
else
return static_cast<Z>(d2);
}
};
template <typename X>
class ASin {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_asin<X,X>(d1);
}
};
template <typename X>
class Sinh {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_sinh<X,X>(d1);
}
};
template <typename X>
class SinhDerivative {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_cosh<X, X>(d1);
}
};
template <typename X>
class Cosh {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_cosh<X,X>(d1);
}
};
template <typename X>
class Tan {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_tan<X,X>(d1);
}
};
template <typename X>
class TanDerivative {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return static_cast<X>(1.f) / nd4j::math::nd4j_pow<X, X, X>(nd4j::math::nd4j_cos<X, X>(d1), static_cast<X>(2.0f));
}
};
template <typename X>
class ATan {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return nd4j::math::nd4j_atan<X, X>(d1);
}
};
template <typename X, typename Y, typename Z>
class Atan2 {
public:
no_op_exec_special
no_op_exec_special_cuda
op_def static Z op(X d1, Y d2) {
return nd4j::math::nd4j_atan2<X, Z>(d2, d1);
}
op_def static Z op(X d1, Y d2, Z *params) {
return op(d1, d2);
}
// op for MetaOps
op_def static Z op(X d1, Y *params) {
return op(d1, params[0]);
}
};
template <typename X>
class Identity {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return d1;
}
};
template <typename X>
class Stabilize {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
X k = params[0];
if (d1 * k > static_cast<X>(- MIN_CUTFOFF))
return static_cast<X>(- MIN_CUTFOFF) / k;
else if (d1 * k < static_cast<X>(MIN_CUTFOFF))
return static_cast<X>(MIN_CUTFOFF) / k;
return d1;
}
};
template <typename X, typename Y, typename Z>
class Step {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static Z op(X d1, Y d2, Z *params) {
return (d1 > static_cast<X>(d2) ? static_cast<Z>(1) : static_cast<Z>(0));
}
};
template <typename X>
class OneMinus {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
op_def static X op(X d1, X *params) {
return static_cast<X>(1) - d1;
}
};
template <typename X>
class Sum {
public:
no_op_exec_special_accumulation_same
no_op_exec_special_accumulation_same_cuda
op_def static X startingValue(const X *input) {
return static_cast<X>(0.0f);
}
op_def static X merge(X old, X opOutput, X *extraParams) {
return opOutput + old;
}
op_def static X update(X old, X opOutput, X *extraParams) {
return opOutput + old;
}
op_def static X op(X d1, X *extraParams) {
return d1;
}
op_def static X postProcess(X reduction, Nd4jLong n, X *extraParams) {
return reduction;
}
};
template <typename X>
class ReduceSameBenchmarkOp {
public:
no_op_exec_special_accumulation_same
no_op_exec_special_accumulation_same_cuda
const static functions::ReduceType reduceType = functions::ReduceType::SUM;
op_def static X startingValue(const X *input) {
return static_cast<X>(0.0f);
}
op_def static X merge(X old, X opOutput, X *extraParams) {
return opOutput + old;
}
op_def static X update(X old, X opOutput, X *extraParams) {
return opOutput + old;
}
op_def static X op(X d1, X *extraParams) {
auto f1 = static_cast<float>(d1);
return static_cast<X>(nd4j::math::nd4j_pow<float,float,float>(f1, 3)
+ nd4j::math::nd4j_log<float,float>(f1) * nd4j::math::nd4j_sin<float,float>(f1)
/ nd4j::math::nd4j_tanh<float,float>(static_cast<float>(M_E) * static_cast<float>(M_PI) * f1)
* nd4j::math::nd4j_sqrt<float,float>(static_cast<float>(M_PI) / f1)
- nd4j::math::nd4j_atan<float,float>(static_cast<float>(M_E) / f1));
}
op_def static X postProcess(X reduction, Nd4jLong n, X *extraParams) {
return reduction;
}
};
template <typename X, typename Z>
class ShannonEntropy {
public:
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
const static functions::ReduceType reduceType = functions::ReduceType::SUM;
op_def static X startingValue(const X *input) {
return static_cast<X>(0);
}
op_def static Z merge(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z update(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z op(X d1, Z *extraParams) {
auto p = d1 * d1;
return static_cast<Z>(p) * nd4j::math::nd4j_log<X, Z>(p);
}
op_def static Z postProcess(Z reduction, Nd4jLong n, Z *extraParams) {
return -reduction;
}
};
template <typename X, typename Z>
class LogEntropy {
public:
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
const static functions::ReduceType reduceType = functions::ReduceType::SUM;
op_def static X startingValue(const X *input) {
return static_cast<X>(0);
}
op_def static Z merge(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z update(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z op(X d1, Z *extraParams) {
return static_cast<Z>(d1) * nd4j::math::nd4j_log<X, Z>(d1);
}
op_def static Z postProcess(Z reduction, Nd4jLong n, Z *extraParams) {
//entropy is -sum(p(x) * log(p(x))); log entropy is log of this
return nd4j::math::nd4j_log<Z, Z>(-reduction);
}
};
template <typename X, typename Z>
class Entropy {
public:
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
const static functions::ReduceType reduceType = functions::ReduceType::SUM;
op_def static X startingValue(const X *input) {
return static_cast<X>(0);
}
op_def static Z merge(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z update(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z op(X d1, Z *extraParams) {
return static_cast<Z>(d1) * nd4j::math::nd4j_log<X, Z>(d1);
}
op_def static Z postProcess(Z reduction, Nd4jLong n, Z *extraParams) {
return static_cast<Z>(-reduction); //entropy is -sum(p(x) * log(p(x)))
}
};
template <typename X>
class ASum {
public:
no_op_exec_special_accumulation_same
no_op_exec_special_accumulation_same_cuda
const static functions::ReduceType reduceType = functions::ReduceType::ASUM;
op_def static X startingValue(const X *input) {
return static_cast<X>(0);
}
op_def static X merge(X old, X opOutput, X *extraParams) {
return nd4j::math::nd4j_abs<X>(opOutput) + nd4j::math::nd4j_abs<X>(old);
}
op_def static X update(X old, X opOutput, X *extraParams) {
return nd4j::math::nd4j_abs<X>(opOutput) + nd4j::math::nd4j_abs<X>(old);
}
op_def static X op(X d1, X *extraParams) {
return nd4j::math::nd4j_abs<X>(d1);
}
op_def static X postProcess(X reduction, Nd4jLong n, X *extraParams) {
return nd4j::math::nd4j_abs<X>(reduction);
}
};
template <typename X, typename Z>
class CountNonZero {
public:
no_op_exec_special_accumulation_long
no_op_exec_special_accumulation_cuda
const static functions::ReduceType reduceType = functions::ReduceType::ASUM;
op_def static Z startingValue(const X *input) {
return static_cast<Z>(0);
}
op_def static Z merge(Z old, Z opOutput, X *extraParams) {
return opOutput + old;
}
op_def static Z update(Z old, Z opOutput, X *extraParams) {
return opOutput + old;
}
op_def static Z op(X d1, X *extraParams) {
return d1 == static_cast<X>(0.0f) ? static_cast<Z>(0.0f) : static_cast<Z>(1.0f);
}
op_def static Z postProcess(Z reduction, Nd4jLong n, X *extraParams) {
return reduction;
}
};
template <typename X, typename Z>
class CountZero {
public:
no_op_exec_special_accumulation_long
no_op_exec_special_accumulation_cuda
const static functions::ReduceType reduceType = functions::ReduceType::SUM;
op_def static Z startingValue(const X *input) {
return static_cast<Z>(0.0f);
}
op_def static Z merge(Z old, Z opOutput, X *extraParams) {
return opOutput + old;
}
op_def static Z update(Z old, Z opOutput, X *extraParams) {
return opOutput + old;
}
op_def static Z op(X d1, X *extraParams) {
return d1 == static_cast<X>(0) ? static_cast<X>(1) : static_cast<X>(0);
}
op_def static Z postProcess(X reduction, Nd4jLong n, X *extraParams) {
return static_cast<Z>(reduction);
}
};
template <typename X>
class Prod {
public:
no_op_exec_special_accumulation_same
no_op_exec_special_accumulation_same_cuda
const static functions::ReduceType reduceType = functions::ReduceType::PRODUCT;
op_def static X startingValue(const X *input) {
return static_cast<X>(1);
}
op_def static X merge(X old, X opOutput, X *extraParams) {
return opOutput * old;
}
op_def static X update(X old, X opOutput, X *extraParams) {
return opOutput * old;
}
op_def static X op(X d1, X *extraParams) {
return d1;
}
op_def static X postProcess(X reduction, Nd4jLong n, X *extraParams) {
return reduction;
}
};
template <typename X, typename Z>
class Any {
public:
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
const static functions::ReduceType reduceType = functions::ReduceType::SUM;
op_def static X startingValue(const X *input) {
return static_cast<X>(0.0f);
}
op_def static Z merge(X old, X opOutput, X *extraParams) {
return opOutput + old;
}
op_def static Z update(X old, X opOutput, X *extraParams) {
return opOutput + old;
}
op_def static Z op(X d1, X *extraParams) {
return d1;
}
op_def static Z postProcess(X reduction, Nd4jLong n, X *extraParams) {
return reduction > static_cast<X>(0) ? static_cast<Z>(1) : static_cast<Z>(0) ;
}
};
template <typename X, typename Z>
class All {
public:
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
const static functions::ReduceType reduceType = functions::ReduceType::PRODUCT;
op_def static X startingValue(const X *input) {
return static_cast<X>(1);
}
op_def static Z merge(X old, X opOutput, X *extraParams) {
return opOutput * old;
}
op_def static Z update(X old, X opOutput, X *extraParams) {
return opOutput * old;
}
op_def static Z op(X d1, X *extraParams) {
return d1;
}
op_def static Z postProcess(X reduction, Nd4jLong n, X *extraParams) {
return reduction > static_cast<X>(0) ? static_cast<Z>(1) : static_cast<Z>(0);
}
};
template <typename X, typename Z>
class Mean {
public:
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
const static functions::ReduceType reduceType = functions::ReduceType::SUM;
op_def static X startingValue(const X *input) {
return static_cast<X>(0);
}
op_def static Z merge(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z update(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z op(X d1, Z *extraParams) {
return d1;
}
op_def static Z postProcess(Z reduction, Nd4jLong n, Z *extraParams) {
return reduction / (Z) n;
}
};
template <typename X, typename Z>
class ReduceFloatBenchmarkOp {
public:
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
const static functions::ReduceType reduceType = functions::ReduceType::SUM;
op_def static X startingValue(const X *input) {
return static_cast<X>(0);
}
op_def static Z merge(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z update(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z op(X d1, Z *extraParams) {
auto f1 = static_cast<float>(d1);
return static_cast<Z>(nd4j::math::nd4j_pow<float,float,float>(f1, 3)
+ nd4j::math::nd4j_log<float,float>(f1) * nd4j::math::nd4j_sin<float,float>(f1)
/ nd4j::math::nd4j_tanh<float,float>(static_cast<float>(M_E) * static_cast<float>(M_PI) * f1)
* nd4j::math::nd4j_sqrt<float,float>(static_cast<float>(M_PI) / f1)
- nd4j::math::nd4j_atan<float,float>(static_cast<float>(M_E) / f1));
}
op_def static Z postProcess(Z reduction, Nd4jLong n, Z *extraParams) {
return (Z) reduction / (Z) n;
}
};
template <typename X, typename Z>
class AMean {
public:
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
const static functions::ReduceType reduceType = functions::ReduceType::SUM;
op_def static X startingValue(const X *input) {
return static_cast<X>(0);
}
op_def static Z merge(Z old, Z opOutput, Z *extraParams) {
return nd4j::math::nd4j_abs<X>(opOutput) + nd4j::math::nd4j_abs<X>(old);
}
op_def static Z update(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z op(X d1, Z *extraParams) {
return nd4j::math::nd4j_abs<X>(d1);
}
op_def static Z postProcess(Z reduction, Nd4jLong n, Z *extraParams) {
return nd4j::math::nd4j_abs<Z>(reduction) / static_cast<Z>(n);
}
};
template <typename X>
class Max {
public:
no_op_exec_special_accumulation_same
no_op_exec_special_accumulation_same_cuda
const static functions::ReduceType reduceType = functions::ReduceType::MAX;
op_def static X startingValue(const X *input) {
Dev branch merge: dev_20190606 (#7904) * correct logsoftmax looss (#2) * Small SameDiff listener fix (#4) * Various fixes (#6) * #7839 Fix for asXMatrix and tests * #7866 EmbeddingSequenceLayer dtype fix + test * #7856 SameDiff save/load stream methods * #7859 RegressionEvaluation rank 4 fix + tests + axis configuration * EvaluationBinary 3d/4d * More evaluation 3d/4d tests * #7847 Evaluation empty checks * Small test ifx * #7848 Fix median edge case * Improve DL4J samediff layer tests * [WIP] FastText wrapper implemented (#8) * FastText implemented * Some fixes * Fix shapes for wordsNearest * Validation of input vectors * Fixes * Fixed test * Thread tagged * Some tweaks * setContextClassLoader for DeallocatorServiceThread * Numpy format tests (#1) * Various fixes (#11) * #7852 SameDiff gather fix * #7892 SameDiff placeholder to constant conversion * #7890 validate input rank for MLN/CG init methods * Fix broken permute shape calculation * Permute and gather fixes * Tests * #7850 LogSumExp fix + test * Handful of test fixes * Empty arrays with non-scalar shapes (#10) * minor rearrangements for lambdas * empty tensors with non-scalar shapes * numpy empty tensors with non-scalar shapes * few more empty tweaks * Small fixes * conv3d signature update * micro fix in batchnorm mkldnn * Import fixes * Fix * MKL-DNN update * Small fill fix * fill with empty input + test * Fixes * Small error improvement * Fix * one special test * couple of fixes for lstm * Rewrite TFGraphMapper.getNDArrayFromTensor to be maintainable and less error prone * Fixes * FP16 * Unsigned * BFloat16 * Fill op - empty tweaks * - couple of fixes for empty arrays construction - stack updated * strided slice fix * one transform test * provide method for reducing shapeInfo in case of input array is empty * Fixed reduceAlongDimensions to use empty input properly. * couple of broadcast tests * couple of tests broadcast tests + tweak to make them pass * add check of non-empty to methods producing sub-arrays * Fixed reshapeC with zeros in shape. * complete empty check in reduce_... legacy ops * Concat and cumsum/prod * Tweak to empty shape inference on import * add empty check to the rest of reduce legacy ops * one more test * correct typo in evalReduceShapeInfoEmpty * Added tests for reduce_* ops to tests with zero shapes. * few more tests for empty reductions * Fixed strided_slice op with empty case and tests. * one more empty reduction test * Fixed strided_slice test. * add empty check to NDArray::reshapei * infOrMax * empty min/max with infinity tests * made unstack working correctly with empty arrays * few IndexReduce tests + tweaks for empty shapes * add test for empty concat * few tests fixed * Validation fix for reductions on empty shapes * Reverse fix * Reduction shape calc fixes * SameDiff.generateOutputVariable: don't use shape function to determine number of outputs * Range fix * - NDArray constructor updated for scalars/empty arrays - few tests fixed * More fixes * Empty creator fixes * concat fix * concat fix * TF import tests: allow 'both all NaN' and 'both all inf' to pass * Slice, zero fraction, and reshape fixes * transpose, gather * Zero fraction * scalar cast fix * Empty reduction axis support * few more tests fixed * Fixed input checks conforming with TF for concat op and tests. * few tests fixed * matmul scalar shape fix * Fixed checkout for data type and scalarity with concat to allow non-empty scalars with vector concats. * broadcast bool fix * few more tests * few more tests * correct evalReduceShapeInfoEmpty * argmax/argmin + tests * one more empty edge case + one more test * argmax/argmin/realdiv_bp tweaks * empty reshape test + fix * Helper fixes * Small fixes * Gather test fix * Gather test fix * Small fixes * reduce scalar zero values * scalar mean workaround * Remove debug code * along dim mean workaround * one more test * - equalsTo() tweak for empty arrays - one more test * broadcast tweaks
2019-06-15 13:34:34 +02:00
return -nd4j::DataTypeUtils::infOrMax<X>();
Eclipse Migration Initial Commit
2019-06-06 14:21:15 +02:00
}
op_def static X merge(X old, X opOutput, X *extraParams) {
return nd4j::math::nd4j_max<X>(old, opOutput);
}
op_def static X update(X old, X opOutput, X *extraParams) {
return nd4j::math::nd4j_max<X>(opOutput, old);
}
op_def static X op(X d1, X d2, X *params) {
return nd4j::math::nd4j_max<X>(d1, d2);
}
op_def static X op(X d1, X d2) {
return nd4j::math::nd4j_max<X>(d1, d2);
}
// FIXME: this signature overlaps with MetaOp
op_def static X op(X d1, X *extraParams) {
return d1;
}
op_def static X postProcess(X reduction, Nd4jLong n, X *extraParams) {
return reduction;
}
};
template <typename X, typename Y, typename Z>
class AMaxPairwise {
public:
op_def static Z op(X d1, Y d2, Z *params) {
return op(d1, d2);
}
op_def static Z op(X d1, Y d2) {
auto z1 = static_cast<Z>(d1);
auto z2 = static_cast<Z>(d2);
if (nd4j::math::nd4j_abs<Z>(z1) > nd4j::math::nd4j_abs<Z>(z2))
return z1;
else
return z2;
}
};
template <typename X, typename Y, typename Z>
class AMinPairwise {
public:
op_def static Z op(X d1, Y d2, Z *params) {
return op(d1, d2);
}
op_def static Z op(X d1, Y d2) {
auto z1 = static_cast<Z>(d1);
auto z2 = static_cast<Z>(d2);
if (nd4j::math::nd4j_abs<Z>(z1) < nd4j::math::nd4j_abs<Z>(z2))
return z1;
else
return z2;
}
};
template <typename X, typename Y, typename Z>
class MaxPairwise {
public:
op_def static Z op(X d1, Y d2, Z *params) {
return nd4j::math::nd4j_max<Z>(static_cast<Z>(d1), static_cast<Z>(d2));
}
op_def static Z op(X d1, Y d2) {
return nd4j::math::nd4j_max<Z>(static_cast<Z>(d1), static_cast<Z>(d2));
}
};
template <typename X, typename Y, typename Z>
class MinPairwise {
public:
op_def static Z op(X d1, Y d2, Z *params) {
return nd4j::math::nd4j_min<Z>(static_cast<Z>(d1), static_cast<Z>(d2));
}
op_def static Z op(X d1, Y d2) {
return nd4j::math::nd4j_min<Z>(static_cast<Z>(d1), static_cast<Z>(d2));
}
};
template <typename X>
class AMax {
public:
no_op_exec_special_accumulation_same
no_op_exec_special_accumulation_same_cuda
const static functions::ReduceType reduceType = functions::ReduceType::AMAX;
op_def static X startingValue(const X *input) {
return input[0];
}
op_def static X merge(X old, X opOutput, X *extraParams) {
return nd4j::math::nd4j_max<X>(nd4j::math::nd4j_abs<X>(old), nd4j::math::nd4j_abs<X>(opOutput));
}
op_def static X update(X old, X opOutput, X *extraParams) {
return nd4j::math::nd4j_max<X>(nd4j::math::nd4j_abs<X>(opOutput), nd4j::math::nd4j_abs<X>(old));
}
op_def static X op(X d1, X d2, X *params) {
return nd4j::math::nd4j_max<X>(nd4j::math::nd4j_abs<X>(d1), nd4j::math::nd4j_abs<X>(d2));
}
op_def static X op(X d1, X d2) {
return nd4j::math::nd4j_abs<X>(d1) > nd4j::math::nd4j_abs<X>(d2) ? d1 : d2;
}
// FIXME: this signature overlaps with MetaOp
op_def static X op(X d1, X *extraParams) {
return nd4j::math::nd4j_abs<X>(d1);
}
op_def static X postProcess(X reduction, Nd4jLong n, X *extraParams) {
return nd4j::math::nd4j_abs<X>(reduction);
}
};
template <typename X>
class AMin {
public:
no_op_exec_special_accumulation_same
no_op_exec_special_accumulation_same_cuda
const static functions::ReduceType reduceType = functions::ReduceType::AMIN;
op_def static X startingValue(const X *input) {
return input[0];
}
op_def static X merge(X old, X opOutput, X *extraParams) {
return nd4j::math::nd4j_min<X>(nd4j::math::nd4j_abs<X>(old), nd4j::math::nd4j_abs<X>(opOutput));
}
op_def static X update(X old, X opOutput, X *extraParams) {
return nd4j::math::nd4j_min<X>(nd4j::math::nd4j_abs<X>(opOutput), nd4j::math::nd4j_abs<X>(old));
}
op_def static X op(X d1, X d2, X *params) {
return nd4j::math::nd4j_min<X>(nd4j::math::nd4j_abs<X>(d1), nd4j::math::nd4j_abs<X>(d2));
}
op_def static X op(X d1, X d2) {
return nd4j::math::nd4j_min<X>(nd4j::math::nd4j_abs<X>(d1), nd4j::math::nd4j_abs<X>(d2));
}
// FIXME: this signature overlaps with MetaOp
op_def static X op(X d1, X *extraParams) {
return nd4j::math::nd4j_abs<X>(d1);
}
op_def static X postProcess(X reduction, Nd4jLong n, X *extraParams) {
return nd4j::math::nd4j_abs<X>(reduction);
}
};
template <typename X>
class Min {
public:
no_op_exec_special_accumulation_same
no_op_exec_special_accumulation_same_cuda
const static functions::ReduceType reduceType = functions::ReduceType::MIN;
op_def static X startingValue(const X *input) {
Dev branch merge: dev_20190606 (#7904) * correct logsoftmax looss (#2) * Small SameDiff listener fix (#4) * Various fixes (#6) * #7839 Fix for asXMatrix and tests * #7866 EmbeddingSequenceLayer dtype fix + test * #7856 SameDiff save/load stream methods * #7859 RegressionEvaluation rank 4 fix + tests + axis configuration * EvaluationBinary 3d/4d * More evaluation 3d/4d tests * #7847 Evaluation empty checks * Small test ifx * #7848 Fix median edge case * Improve DL4J samediff layer tests * [WIP] FastText wrapper implemented (#8) * FastText implemented * Some fixes * Fix shapes for wordsNearest * Validation of input vectors * Fixes * Fixed test * Thread tagged * Some tweaks * setContextClassLoader for DeallocatorServiceThread * Numpy format tests (#1) * Various fixes (#11) * #7852 SameDiff gather fix * #7892 SameDiff placeholder to constant conversion * #7890 validate input rank for MLN/CG init methods * Fix broken permute shape calculation * Permute and gather fixes * Tests * #7850 LogSumExp fix + test * Handful of test fixes * Empty arrays with non-scalar shapes (#10) * minor rearrangements for lambdas * empty tensors with non-scalar shapes * numpy empty tensors with non-scalar shapes * few more empty tweaks * Small fixes * conv3d signature update * micro fix in batchnorm mkldnn * Import fixes * Fix * MKL-DNN update * Small fill fix * fill with empty input + test * Fixes * Small error improvement * Fix * one special test * couple of fixes for lstm * Rewrite TFGraphMapper.getNDArrayFromTensor to be maintainable and less error prone * Fixes * FP16 * Unsigned * BFloat16 * Fill op - empty tweaks * - couple of fixes for empty arrays construction - stack updated * strided slice fix * one transform test * provide method for reducing shapeInfo in case of input array is empty * Fixed reduceAlongDimensions to use empty input properly. * couple of broadcast tests * couple of tests broadcast tests + tweak to make them pass * add check of non-empty to methods producing sub-arrays * Fixed reshapeC with zeros in shape. * complete empty check in reduce_... legacy ops * Concat and cumsum/prod * Tweak to empty shape inference on import * add empty check to the rest of reduce legacy ops * one more test * correct typo in evalReduceShapeInfoEmpty * Added tests for reduce_* ops to tests with zero shapes. * few more tests for empty reductions * Fixed strided_slice op with empty case and tests. * one more empty reduction test * Fixed strided_slice test. * add empty check to NDArray::reshapei * infOrMax * empty min/max with infinity tests * made unstack working correctly with empty arrays * few IndexReduce tests + tweaks for empty shapes * add test for empty concat * few tests fixed * Validation fix for reductions on empty shapes * Reverse fix * Reduction shape calc fixes * SameDiff.generateOutputVariable: don't use shape function to determine number of outputs * Range fix * - NDArray constructor updated for scalars/empty arrays - few tests fixed * More fixes * Empty creator fixes * concat fix * concat fix * TF import tests: allow 'both all NaN' and 'both all inf' to pass * Slice, zero fraction, and reshape fixes * transpose, gather * Zero fraction * scalar cast fix * Empty reduction axis support * few more tests fixed * Fixed input checks conforming with TF for concat op and tests. * few tests fixed * matmul scalar shape fix * Fixed checkout for data type and scalarity with concat to allow non-empty scalars with vector concats. * broadcast bool fix * few more tests * few more tests * correct evalReduceShapeInfoEmpty * argmax/argmin + tests * one more empty edge case + one more test * argmax/argmin/realdiv_bp tweaks * empty reshape test + fix * Helper fixes * Small fixes * Gather test fix * Gather test fix * Small fixes * reduce scalar zero values * scalar mean workaround * Remove debug code * along dim mean workaround * one more test * - equalsTo() tweak for empty arrays - one more test * broadcast tweaks
2019-06-15 13:34:34 +02:00
return nd4j::DataTypeUtils::infOrMax<X>();
Eclipse Migration Initial Commit
2019-06-06 14:21:15 +02:00
}
op_def static X merge(X old, X opOutput, X *extraParams) {
return nd4j::math::nd4j_min<X>(old, opOutput);
}
op_def static X update(X old, X opOutput, X *extraParams) {
return nd4j::math::nd4j_min<X>(opOutput, old);
}
op_def static X op(X d1, X d2, X *params) {
return nd4j::math::nd4j_min<X>(d1, d2);
}
op_def static X op(X d1, X d2) {
return nd4j::math::nd4j_min<X>(d1, d2);
}
// FIXME: this signature overlaps with MetaOp
op_def static X op(X d1, X *extraParams) {
return d1;
}
op_def static X postProcess(X reduction, Nd4jLong n, X *extraParams) {
return reduction;
}
};
template <typename X, typename Z>
class Norm1 {
public:
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
const static functions::ReduceType reduceType = functions::ReduceType::SUM;
op_def static X startingValue(const X *input) {
return static_cast<X>(0);
}
op_def static Z merge(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z update(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z op(X d1, Z *extraParams) {
return static_cast<Z>(nd4j::math::nd4j_abs<X>(d1));
}
op_def static Z postProcess(Z reduction, Nd4jLong n, Z *extraParams) {
return reduction;
}
};
template <typename X, typename Z>
class Norm2 {
public:
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
const static functions::ReduceType reduceType = functions::ReduceType::SUM;
op_def static X startingValue(const X *input) {
return static_cast<X>(0);
}
op_def static Z merge(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z update(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z postProcess(Z reduction, Nd4jLong n, Z *extraParams) {
return nd4j::math::nd4j_sqrt<Z, Z>(reduction);
}
op_def static Z op(X d1, Z *extraParams) {
return static_cast<Z>(d1 * d1);
}
};
template <typename X, typename Z>
class SquaredNorm {
public:
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
const static functions::ReduceType reduceType = functions::ReduceType::SUM;
op_def static X startingValue(const X *input) {
return static_cast<X>(0);
}
op_def static Z merge(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z update(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z op(X d1, Z *extraParams) {
return static_cast<Z>(d1 * d1);
}
op_def static Z postProcess(Z reduction, Nd4jLong n, Z *extraParams) {
return reduction;
}
};
template <typename X, typename Z>
class NormFrobenius {
public:
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
const static functions::ReduceType reduceType = functions::ReduceType::SUM;
op_def static X startingValue(const X *input) {
return static_cast<X>(0);
}
op_def static Z merge(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z update(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z op(X d1, Z *extraParams) {
X v = nd4j::math::nd4j_abs<X>(d1);
return static_cast<Z>(v * v);
}
op_def static Z postProcess(Z reduction, Nd4jLong n, Z *extraParams) {
return nd4j::math::nd4j_sqrt<Z, Z>(reduction);
}
};
template <typename X, typename Z>
class NormP {
public:
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
const static functions::ReduceType reduceType = functions::ReduceType::SUM;
op_def static X startingValue(const X *input) {
return static_cast<X>(0);
}
op_def static Z merge(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z update(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z op(X d1, Z *extraParams) {
return nd4j::math::nd4j_pow<X, Z, Z>(nd4j::math::nd4j_abs<X>(d1), extraParams[0]);
}
op_def static Z postProcess(Z reduction, Nd4jLong n, Z *extraParams) {
return nd4j::math::nd4j_pow<Z, Z, Z>(reduction, static_cast<Z>(1.0f) / extraParams[0]);
}
};
template <typename X, typename Z>
class NormMax {
public:
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
const static functions::ReduceType reduceType = functions::ReduceType::SUM;
op_def static X startingValue(const X *input) {
return static_cast<X>(0);
}
op_def static Z merge(Z old, Z opOutput, Z *extraParams) {
return opOutput + old;
}
op_def static Z update(Z old, Z opOutput, Z *extraParams) {
return nd4j::math::nd4j_max<Z>(nd4j::math::nd4j_abs<Z>(old),
nd4j::math::nd4j_abs<Z>(opOutput));
}
op_def static Z op(X d1, Z *extraParams) {
return static_cast<Z>(d1);
}
op_def static Z postProcess(Z reduction, Nd4jLong n, Z *extraParams) {
return nd4j::math::nd4j_max<Z>(nd4j::math::nd4j_abs<Z>(reduction), nd4j::math::nd4j_abs<Z>(reduction));
}
};
template <typename X, typename Z>
class Variance {
public:
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
const static functions::ReduceType reduceType = functions::ReduceType::SUM;
op_def static X startingValue(const X *input) {
return static_cast<X>(0.0f);
}
op_def static Z merge(X old, X opOutput, Z *extraParams) {
return old + opOutput;
}
op_def static Z update(X old, X opOutput, Z *extraParams) {
return old + opOutput;
}
op_def static X op(X d1, Z *extraParams) {
X mean = static_cast<X>(extraParams[0]);
X ret = d1 - mean;
return ret * ret;
}
op_def static Z postProcess(X reduction, Nd4jLong n, Z *extraParams) {
// T bias = extraParams[1];
// return (reduction - (nd4j::math::nd4j_pow<T>(bias, static_cast<T>(2.0f)) / static_cast<T>(n))) / (n - 1)
return static_cast<Z>(reduction) / static_cast<Z>(n - 1);
}
};
/**
* Standard deviation of a buffer
*/
template <typename X, typename Z>
class StandardDeviation {
public:
no_op_exec_special_accumulation
no_op_exec_special_accumulation_cuda
const static functions::ReduceType reduceType = functions::ReduceType::SUM;
op_def static X startingValue(const X *input) {
return static_cast<X>(0.0f);
}
op_def static Z merge(X old, X opOutput, Z *extraParams) {
return old + opOutput;
}
op_def static Z update(X old, X opOutput, Z *extraParams) {
return old + opOutput;
}
op_def static Z op(X d1, Z *extraParams) {
X mean = extraParams[0];
X ret = d1 - mean;
return ret * ret;
}
op_def static Z postProcess(X reduction, Nd4jLong n, Z *extraParams) {
Z ret = Variance<X,Z>::postProcess(reduction, n, extraParams);
Z sqrtRet = nd4j::math::nd4j_sqrt<X, Z>(ret);
return sqrtRet;
}
};
template <typename X, typename Y>
class CosineSimilarity {
public:
static const int extraParamsLen = 2;
op_def static X *generateExtraParams() {
//T *extraParams = new T[2];
return nullptr;
}
op_def static void finalizeExtraParams(X *extraParams) {
//delete[] extraParams;
}
op_def static Y startingValue(const X *input) {
return static_cast<Y>(0.0f);
}
op_def static Y postProcess(Y reduction, Nd4jLong n, Y *extraParams) {
return reduction / (nd4j::math::nd4j_sqrt<Y, Y>(extraParams[0]) * nd4j::math::nd4j_sqrt<Y, Y>(extraParams[1]));
}
op_def static Y op(X d1, X d2, Y *extraParams) {
extraParams[0] += static_cast<Y>(d1 * d1);
extraParams[1] += static_cast<Y>(d2 * d2);
return static_cast<Y>(d1 * d2);
}
op_def static void aggregateExtraParams(Y *extraParamsTotal, Y *extraParamsLocal) {
extraParamsTotal[0] += extraParamsLocal[0];
extraParamsTotal[1] += extraParamsLocal[1];
}
#ifdef __CUDACC__
static _CUDA_D inline Y opAtomic(X d1, X d2, Y *extraParams) {
nd4j::math::atomics::nd4j_atomicAdd(&extraParams[0],static_cast<Y>(d1 * d1));
nd4j::math::atomics::nd4j_atomicAdd(&extraParams[1],static_cast<Y>(d2 * d2));
return static_cast<Y>(d1 * d2);
}
#endif
op_def static Y update(Y old, Y opOutput, Y *extraParams) {
return old + opOutput;
}
op_def static Y merge(Y old, Y opOutput, Y *extraParams) {
return update(old, opOutput, extraParams);
}
};
template <typename X, typename Y>
class JaccardDistance {
public:
static const int extraParamsLen = 2;
op_def static X *generateExtraParams() {
//T *extraParams = new T[2];
return nullptr;
}
op_def static void finalizeExtraParams(X *extraParams) {
//delete[] extraParams;
}
op_def static Y startingValue(const X *input) {
return static_cast<X>(0.0f);
}
op_def static Y postProcess(Y reduction, Nd4jLong n, Y *extraParams) {
// num / denom
return (static_cast<Y>(1.0f)) - (extraParams[0] / extraParams[1]);
}
op_def static Y num(X d1, X d2) {
return nd4j::math::nd4j_min<X>(d1, d2);
}
op_def static Y denom(X d1, X d2) {
return nd4j::math::nd4j_max<X>(d1, d2);
}
op_def static Y op(X d1, X d2, Y *extraParams) {
extraParams[0] += static_cast<Y>(num(d1, d2));
extraParams[1] += static_cast<Y>(denom(d1, d2));
return static_cast<Y>(0.0f);
}
op_def static void aggregateExtraParams(Y *extraParamsTotal, Y *extraParamsLocal) {
extraParamsTotal[0] += extraParamsLocal[0];
extraParamsTotal[1] += extraParamsLocal[1];
}
#ifdef __CUDACC__
__device__
static inline Y opAtomic(X d1, X d2, Y *extraParams) {
nd4j::math::atomics::nd4j_atomicAdd(&extraParams[0],num(d1, d2));
nd4j::math::atomics::nd4j_atomicAdd(&extraParams[1], denom(d1, d2));
return static_cast<Y>(0.0f);
}
#endif
op_def static Y update(Y old, Y opOutput, Y *extraParams) {
return old + opOutput;
}
op_def static Y merge(Y old, Y opOutput, Y *extraParams) {
return update(old, opOutput, extraParams);
}
};
template <typename X, typename Y>
class SimpleHammingDistance {
public:
static const int extraParamsLen = 0;
op_def static X *generateExtraParams() {
//T *extraParams = new T[2];
return nullptr;
}
op_def static void finalizeExtraParams(X *extraParams) {
//delete[] extraParams;
}
op_def static Y startingValue(const X *input) {
return static_cast<Y>(0.0f);
}
op_def static Y postProcess(Y reduction, Nd4jLong n, Y *extraParams) {
return static_cast<Y>(reduction / n);
}
op_def static Y op(X d1, X d2, Y *extraParams) {
return (d1 == d2) ? static_cast<Y>(0.0f) : static_cast<Y>(1.0f);
}
op_def static void aggregateExtraParams(Y *extraParamsTotal, Y *extraParamsLocal) {
}
#ifdef __CUDACC__
__device__
static inline Y opAtomic(X d1, X d2, Y *extraParams) {
return op(d1, d2, extraParams);
}
#endif
op_def static Y update(Y old, Y opOutput, Y *extraParams) {
return old + opOutput;
}
op_def static Y merge(Y old, Y opOutput, Y *extraParams) {
return update(old, opOutput, extraParams);
}
};
template <typename X, typename Y>
class CosineDistance {
public:
static const int extraParamsLen = 2;
op_def static X *generateExtraParams() {
//T *extraParams = new T[2];
return nullptr;
}
op_def static void finalizeExtraParams(X *extraParams) {
//delete[] extraParams;
}
op_def static Y startingValue(const X *input) {
return static_cast<Y>(0.0f);
}
op_def static Y postProcess(Y reduction, Nd4jLong n, Y *extraParams) {
return (static_cast<Y>(1.0f)) - (reduction / (nd4j::math::nd4j_sqrt<Y, Y>(extraParams[0]) * nd4j::math::nd4j_sqrt<Y, Y>(extraParams[1])));
}
op_def static Y op(X d1, X d2, Y *extraParams) {
extraParams[0] += static_cast<Y>(nd4j::math::nd4j_abs<X>(d1) * nd4j::math::nd4j_abs<X>(d1));
extraParams[1] += static_cast<Y>(nd4j::math::nd4j_abs<X>(d2) * nd4j::math::nd4j_abs<X>(d2));
return (d1 * d2);
}
op_def static void aggregateExtraParams(Y *extraParamsTotal, Y *extraParamsLocal) {
extraParamsTotal[0] += extraParamsLocal[0];
extraParamsTotal[1] += extraParamsLocal[1];
}
#ifdef __CUDACC__
static _CUDA_D inline Y opAtomic(X d1, X d2, Y *extraParams) {
nd4j::math::atomics::nd4j_atomicAdd(&extraParams[0], nd4j::math::nd4j_abs<Y>(d1) * nd4j::math::nd4j_abs<Y>(d1));
nd4j::math::atomics::nd4j_atomicAdd(&extraParams[1], nd4j::math::nd4j_abs<Y>(d2) * nd4j::math::nd4j_abs<Y>(d2));
return (d1 * d2);
}
#endif
op_def static Y update(Y old, Y opOutput, Y *extraParams) {
return old + opOutput;
}
op_def static Y merge(Y old, Y opOutput, Y *extraParams) {
return update(old, opOutput, extraParams);
}
};
/**
* Dot product between 2 arrays
*/
template <typename X, typename Y>
class Dot {
public:
static const int extraParamsLen = 0;
op_def static X * generateExtraParams() {
return nullptr;
}
op_def static void finalizeExtraParams(X *extraParamsRef) {
//no-op
//delete[] * extraParamsRef;
}
op_def static Y startingValue(const X *input) {
return static_cast<Y>(0.0f);
}
op_def static Y postProcess(Y reduction, Nd4jLong n, Y *extraParamsRef) {
return reduction;
}
op_def static Y op(X d1, X d2, Y *extraParamsRef) {
return static_cast<Y>(d1 * d2);
}
#ifdef __CUDACC__
__device__
static inline Y opAtomic(X d1, X d2, Y *extraParamsRef) {
return op(d1, d2, extraParamsRef);
}
#endif
op_def static Y update(Y old, Y opOutput, Y *extraParamsRef) {
return opOutput + old;
}
op_def static Y merge(Y old, Y opOutput, Y *extraParamsRef) {
return update(old, opOutput, extraParamsRef);
}
op_def static void aggregateExtraParams(Y *extraParamsTotal, Y *extraParamsLocal) {}
};
/**
* Op to check equality within arrays
*/
template <typename X, typename Z>
class EqualsWithEps {
public:
static const int extraParamsLen = 0;
op_def static X * generateExtraParams() {
return nullptr;
}
op_def static void finalizeExtraParams(X *extraParamsRef) {
//no-op
}
op_def static Z startingValue(const X *input) {
return static_cast<Z>(0.0f);
}
op_def static Z postProcess(Z reduction, Nd4jLong n, Z *extraParamsRef) {
return reduction;
}
op_def static Z op(X d1, X d2, Z *extraParamsRef) {
double eps = nd4j::math::nd4j_abs<double>(extraParamsRef[2]);
return static_cast<Z>(!nd4j::math::nd4j_eq<X>(d1, d2, eps));
}
#ifdef __CUDACC__
__device__
static inline Z opAtomic(X d1, X d2, Z *extraParamsRef) {
return op(d1, d2, extraParamsRef);
}
#endif
op_def static Z update(Z old, Z opOutput, Z *extraParamsRef) {
return opOutput + old;
}
op_def static Z merge(X old, Z opOutput, Z *extraParamsRef) {
return update(old, opOutput, extraParamsRef);
}
op_def static void aggregateExtraParams(Z *extraParamsTotal, Z *extraParamsLocal) {}
};
template <typename X, typename Y>
class EuclideanDistance {
public:
static const int extraParamsLen = 0;
op_def static X * generateExtraParams() {
return nullptr;
}
op_def static void finalizeExtraParams(X *extraParamsRef) {
//no-op
}
op_def static Y startingValue(const X *input) {
return static_cast<Y>(0.0f);
}
op_def static Y postProcess(Y reduction, Nd4jLong n, Y *extraParamsRef) {
return nd4j::math::nd4j_sqrt<Y, Y>(reduction);
}
op_def static Y op(X d1, X d2, Y *extraParamsRef) {
X ret = d1 - d2;
return static_cast<Y>(ret * ret);
}
#ifdef __CUDACC__
__device__
static inline Y opAtomic(X d1, X d2, Y *extraParamsRef) {
return op(d1, d2, extraParamsRef);
}
#endif
op_def static Y update(Y old, Y opOutput, Y *extraParamsRef) {
return opOutput + old;
}
op_def static Y merge(Y old, Y opOutput, Y *extraParamsRef) {
return update(old, opOutput, extraParamsRef);
}
op_def static void aggregateExtraParams(Y *extraParamsTotal, Y *extraParamsLocal) {}
};
template <typename X, typename Y>
class ManhattanDistance {
public:
static const int extraParamsLen = 0;
op_def static X * generateExtraParams() {
return nullptr;
}
op_def static void finalizeExtraParams(X *extraParamsRef) {
//no-op
}
op_def static Y startingValue(const X *input) {
return static_cast<Y>(0.0f);
}
op_def static Y postProcess(Y reduction, Nd4jLong n, Y *extraParamsRef) {
return reduction;
}
op_def static Y op(X d1, X d2, Y *extraParamsRef) {
return nd4j::math::nd4j_abs<X>(d1 - d2);
}
op_def static Y update(Y old, Y opOutput, Y *extraParamsRef) {
return old + opOutput;
}
op_def static void aggregateExtraParams(Y *extraParamsTotal, Y *extraParamsLocal) {
}
#ifdef __CUDACC__
__device__
static inline Y opAtomic(X d1, X d2, Y *extraParamsRef) {
return op(d1, d2, extraParamsRef);
}
#endif
#ifndef __clang__
#pragma omp declare simd uniform(extraParamsRef)
#endif
op_def static Y merge(X old, X opOutput, X *extraParamsRef) {
return update(old, opOutput, extraParamsRef);
}
};
template <typename X>
class IndexAbsoluteMax {
public:
static _CUDA_HD inline functions::indexreduce::IndexValue<X> op(functions::indexreduce::IndexValue<X> val, X *extraParams) {
return nd4j::math::nd4j_abs<X>(val);
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> update(functions::indexreduce::IndexValue<X> &old, functions::indexreduce::IndexValue<X> &opOutput, X *extraParams) {
opOutput.value = nd4j::math::nd4j_abs<X>(opOutput.value);
old.value = nd4j::math::nd4j_abs<X>(old.value);
if (opOutput.value > old.value)
return opOutput;
#ifdef __CUDACC__
// workaround for cuda race condition at merge phase
else if (opOutput.value == old.value && opOutput.index < old.index)
return opOutput;
#elif defined(__GNUC__)
#endif
return old;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> merge(
functions::indexreduce::IndexValue<X> f1,
functions::indexreduce::IndexValue<X> f2, X *extraParams) {
if (nd4j::math::nd4j_abs<X>(f1.value) > nd4j::math::nd4j_abs<X>(f2.value))
return f2;
return f1;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> postProcess(
functions::indexreduce::IndexValue<X> reduction, int n, int xOffset,
X *dx, int incx, X *extraParams, X *result) {
return reduction;
}
static _CUDA_HD inline X startingValue(const X *input) {
return 0;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> startingIndexValue(X *input) {
functions::indexreduce::IndexValue<X> local;
local.value = startingValue(input);
local.index = 0;
return local;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> op(functions::indexreduce::IndexValue<X> d1,
functions::indexreduce::IndexValue<X> d2, X *extraParams) {
return d1;
}
};
template <typename X>
class FirstIndex {
public:
static _CUDA_HD inline functions::indexreduce::IndexValue<X> op(functions::indexreduce::IndexValue<X> val, X *extraParams) {
return val;
}
static _CUDA_HD functions::indexreduce::IndexValue<X> update(functions::indexreduce::IndexValue<X> &old, functions::indexreduce::IndexValue<X> &opOutput, X *extraParams) {
#ifdef __CUDACC__
if (opOutput.index < 0)
return old;
#endif
auto res = simdOps::MatchCondition<X,X>::op(opOutput.value, extraParams);
//printf("res: %f; oldIdx: %i; newIdx: %i\n", res, old.index, opOutput.index);
if (res == static_cast<X>(0))
return old;
if (old.index < 0)
return opOutput;
if (old.index > opOutput.index)
return opOutput;
return old;
}
static _CUDA_HD inline X startingValue(const X *input) {
Dev branch merge: dev_20190606 (#7904) * correct logsoftmax looss (#2) * Small SameDiff listener fix (#4) * Various fixes (#6) * #7839 Fix for asXMatrix and tests * #7866 EmbeddingSequenceLayer dtype fix + test * #7856 SameDiff save/load stream methods * #7859 RegressionEvaluation rank 4 fix + tests + axis configuration * EvaluationBinary 3d/4d * More evaluation 3d/4d tests * #7847 Evaluation empty checks * Small test ifx * #7848 Fix median edge case * Improve DL4J samediff layer tests * [WIP] FastText wrapper implemented (#8) * FastText implemented * Some fixes * Fix shapes for wordsNearest * Validation of input vectors * Fixes * Fixed test * Thread tagged * Some tweaks * setContextClassLoader for DeallocatorServiceThread * Numpy format tests (#1) * Various fixes (#11) * #7852 SameDiff gather fix * #7892 SameDiff placeholder to constant conversion * #7890 validate input rank for MLN/CG init methods * Fix broken permute shape calculation * Permute and gather fixes * Tests * #7850 LogSumExp fix + test * Handful of test fixes * Empty arrays with non-scalar shapes (#10) * minor rearrangements for lambdas * empty tensors with non-scalar shapes * numpy empty tensors with non-scalar shapes * few more empty tweaks * Small fixes * conv3d signature update * micro fix in batchnorm mkldnn * Import fixes * Fix * MKL-DNN update * Small fill fix * fill with empty input + test * Fixes * Small error improvement * Fix * one special test * couple of fixes for lstm * Rewrite TFGraphMapper.getNDArrayFromTensor to be maintainable and less error prone * Fixes * FP16 * Unsigned * BFloat16 * Fill op - empty tweaks * - couple of fixes for empty arrays construction - stack updated * strided slice fix * one transform test * provide method for reducing shapeInfo in case of input array is empty * Fixed reduceAlongDimensions to use empty input properly. * couple of broadcast tests * couple of tests broadcast tests + tweak to make them pass * add check of non-empty to methods producing sub-arrays * Fixed reshapeC with zeros in shape. * complete empty check in reduce_... legacy ops * Concat and cumsum/prod * Tweak to empty shape inference on import * add empty check to the rest of reduce legacy ops * one more test * correct typo in evalReduceShapeInfoEmpty * Added tests for reduce_* ops to tests with zero shapes. * few more tests for empty reductions * Fixed strided_slice op with empty case and tests. * one more empty reduction test * Fixed strided_slice test. * add empty check to NDArray::reshapei * infOrMax * empty min/max with infinity tests * made unstack working correctly with empty arrays * few IndexReduce tests + tweaks for empty shapes * add test for empty concat * few tests fixed * Validation fix for reductions on empty shapes * Reverse fix * Reduction shape calc fixes * SameDiff.generateOutputVariable: don't use shape function to determine number of outputs * Range fix * - NDArray constructor updated for scalars/empty arrays - few tests fixed * More fixes * Empty creator fixes * concat fix * concat fix * TF import tests: allow 'both all NaN' and 'both all inf' to pass * Slice, zero fraction, and reshape fixes * transpose, gather * Zero fraction * scalar cast fix * Empty reduction axis support * few more tests fixed * Fixed input checks conforming with TF for concat op and tests. * few tests fixed * matmul scalar shape fix * Fixed checkout for data type and scalarity with concat to allow non-empty scalars with vector concats. * broadcast bool fix * few more tests * few more tests * correct evalReduceShapeInfoEmpty * argmax/argmin + tests * one more empty edge case + one more test * argmax/argmin/realdiv_bp tweaks * empty reshape test + fix * Helper fixes * Small fixes * Gather test fix * Gather test fix * Small fixes * reduce scalar zero values * scalar mean workaround * Remove debug code * along dim mean workaround * one more test * - equalsTo() tweak for empty arrays - one more test * broadcast tweaks
2019-06-15 13:34:34 +02:00
return -nd4j::DataTypeUtils::infOrMax<X>();
Eclipse Migration Initial Commit
2019-06-06 14:21:15 +02:00
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> startingIndexValue(X *input) {
functions::indexreduce::IndexValue<X> local;
local.value = startingValue(input);
local.index = -1;
return local;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> op(functions::indexreduce::IndexValue<X> d1,
functions::indexreduce::IndexValue<X> d2, X *extraParams) {
return d1;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> merge(
functions::indexreduce::IndexValue<X> f1,
functions::indexreduce::IndexValue<X> f2, X *extraParams) {
if (f1.index > f2.index)
return f2;
return f1;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> postProcess(
functions::indexreduce::IndexValue<X> reduction, int n, int xOffset,
X *dx, int incx, X *extraParams, X *result) {
return reduction;
}
};
template <typename X>
class LastIndex {
public:
static _CUDA_HD inline functions::indexreduce::IndexValue<X> op(functions::indexreduce::IndexValue<X> val, X *extraParams) {
return val;
}
static _CUDA_HD functions::indexreduce::IndexValue<X> update(functions::indexreduce::IndexValue<X> &old, functions::indexreduce::IndexValue<X> &opOutput, X *extraParams) {
#ifdef __CUDACC__
if (opOutput.index < 0)
return old;
#endif
auto res = simdOps::MatchCondition<X,X>::op(opOutput.value, extraParams);
if (res == static_cast<X>(0))
return old;
if (old.index < 0)
return opOutput;
if (old.index < opOutput.index)
return opOutput;
return old;
}
static _CUDA_HD inline X startingValue(const X *input) {
Dev branch merge: dev_20190606 (#7904) * correct logsoftmax looss (#2) * Small SameDiff listener fix (#4) * Various fixes (#6) * #7839 Fix for asXMatrix and tests * #7866 EmbeddingSequenceLayer dtype fix + test * #7856 SameDiff save/load stream methods * #7859 RegressionEvaluation rank 4 fix + tests + axis configuration * EvaluationBinary 3d/4d * More evaluation 3d/4d tests * #7847 Evaluation empty checks * Small test ifx * #7848 Fix median edge case * Improve DL4J samediff layer tests * [WIP] FastText wrapper implemented (#8) * FastText implemented * Some fixes * Fix shapes for wordsNearest * Validation of input vectors * Fixes * Fixed test * Thread tagged * Some tweaks * setContextClassLoader for DeallocatorServiceThread * Numpy format tests (#1) * Various fixes (#11) * #7852 SameDiff gather fix * #7892 SameDiff placeholder to constant conversion * #7890 validate input rank for MLN/CG init methods * Fix broken permute shape calculation * Permute and gather fixes * Tests * #7850 LogSumExp fix + test * Handful of test fixes * Empty arrays with non-scalar shapes (#10) * minor rearrangements for lambdas * empty tensors with non-scalar shapes * numpy empty tensors with non-scalar shapes * few more empty tweaks * Small fixes * conv3d signature update * micro fix in batchnorm mkldnn * Import fixes * Fix * MKL-DNN update * Small fill fix * fill with empty input + test * Fixes * Small error improvement * Fix * one special test * couple of fixes for lstm * Rewrite TFGraphMapper.getNDArrayFromTensor to be maintainable and less error prone * Fixes * FP16 * Unsigned * BFloat16 * Fill op - empty tweaks * - couple of fixes for empty arrays construction - stack updated * strided slice fix * one transform test * provide method for reducing shapeInfo in case of input array is empty * Fixed reduceAlongDimensions to use empty input properly. * couple of broadcast tests * couple of tests broadcast tests + tweak to make them pass * add check of non-empty to methods producing sub-arrays * Fixed reshapeC with zeros in shape. * complete empty check in reduce_... legacy ops * Concat and cumsum/prod * Tweak to empty shape inference on import * add empty check to the rest of reduce legacy ops * one more test * correct typo in evalReduceShapeInfoEmpty * Added tests for reduce_* ops to tests with zero shapes. * few more tests for empty reductions * Fixed strided_slice op with empty case and tests. * one more empty reduction test * Fixed strided_slice test. * add empty check to NDArray::reshapei * infOrMax * empty min/max with infinity tests * made unstack working correctly with empty arrays * few IndexReduce tests + tweaks for empty shapes * add test for empty concat * few tests fixed * Validation fix for reductions on empty shapes * Reverse fix * Reduction shape calc fixes * SameDiff.generateOutputVariable: don't use shape function to determine number of outputs * Range fix * - NDArray constructor updated for scalars/empty arrays - few tests fixed * More fixes * Empty creator fixes * concat fix * concat fix * TF import tests: allow 'both all NaN' and 'both all inf' to pass * Slice, zero fraction, and reshape fixes * transpose, gather * Zero fraction * scalar cast fix * Empty reduction axis support * few more tests fixed * Fixed input checks conforming with TF for concat op and tests. * few tests fixed * matmul scalar shape fix * Fixed checkout for data type and scalarity with concat to allow non-empty scalars with vector concats. * broadcast bool fix * few more tests * few more tests * correct evalReduceShapeInfoEmpty * argmax/argmin + tests * one more empty edge case + one more test * argmax/argmin/realdiv_bp tweaks * empty reshape test + fix * Helper fixes * Small fixes * Gather test fix * Gather test fix * Small fixes * reduce scalar zero values * scalar mean workaround * Remove debug code * along dim mean workaround * one more test * - equalsTo() tweak for empty arrays - one more test * broadcast tweaks
2019-06-15 13:34:34 +02:00
return -nd4j::DataTypeUtils::infOrMax<X>();
Eclipse Migration Initial Commit
2019-06-06 14:21:15 +02:00
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> startingIndexValue(X *input) {
functions::indexreduce::IndexValue<X> local;
local.value = startingValue(input);
local.index = -1;
return local;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> op(functions::indexreduce::IndexValue<X> d1,
functions::indexreduce::IndexValue<X> d2, X *extraParams) {
return d1;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> merge(
functions::indexreduce::IndexValue<X> f1,
functions::indexreduce::IndexValue<X> f2, X *extraParams) {
if (f1.index < f2.index)
return f2;
return f1;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> postProcess(
functions::indexreduce::IndexValue<X> reduction, int n, int xOffset,
X *dx, int incx, X *extraParams, X *result) {
return reduction;
}
};
template <typename X>
class IndexMax {
public:
static _CUDA_HD inline functions::indexreduce::IndexValue<X> op(functions::indexreduce::IndexValue<X> val, X *extraParams) {
return val;
}
static _CUDA_HD functions::indexreduce::IndexValue<X> update(functions::indexreduce::IndexValue<X> &old, functions::indexreduce::IndexValue<X> &opOutput, X *extraParams) {
if (opOutput.value > old.value) {
return opOutput;
}
#ifdef __CUDACC__
// workaround for cuda race condition at merge phase
else if (opOutput.value == old.value && opOutput.index < old.index)
return opOutput;
#elif defined(__GNUC__)
#endif
return old;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> merge(
functions::indexreduce::IndexValue<X> f1,
functions::indexreduce::IndexValue<X> f2, X *extraParams) {
if (f1.value > f2.value)
return f2;
return f1;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> postProcess(
functions::indexreduce::IndexValue<X> reduction, int n, int xOffset,
X *dx, int incx, X *extraParams, X *result) {
return reduction;
}
static _CUDA_HD inline X startingValue(const X *input) {
Dev branch merge: dev_20190606 (#7904) * correct logsoftmax looss (#2) * Small SameDiff listener fix (#4) * Various fixes (#6) * #7839 Fix for asXMatrix and tests * #7866 EmbeddingSequenceLayer dtype fix + test * #7856 SameDiff save/load stream methods * #7859 RegressionEvaluation rank 4 fix + tests + axis configuration * EvaluationBinary 3d/4d * More evaluation 3d/4d tests * #7847 Evaluation empty checks * Small test ifx * #7848 Fix median edge case * Improve DL4J samediff layer tests * [WIP] FastText wrapper implemented (#8) * FastText implemented * Some fixes * Fix shapes for wordsNearest * Validation of input vectors * Fixes * Fixed test * Thread tagged * Some tweaks * setContextClassLoader for DeallocatorServiceThread * Numpy format tests (#1) * Various fixes (#11) * #7852 SameDiff gather fix * #7892 SameDiff placeholder to constant conversion * #7890 validate input rank for MLN/CG init methods * Fix broken permute shape calculation * Permute and gather fixes * Tests * #7850 LogSumExp fix + test * Handful of test fixes * Empty arrays with non-scalar shapes (#10) * minor rearrangements for lambdas * empty tensors with non-scalar shapes * numpy empty tensors with non-scalar shapes * few more empty tweaks * Small fixes * conv3d signature update * micro fix in batchnorm mkldnn * Import fixes * Fix * MKL-DNN update * Small fill fix * fill with empty input + test * Fixes * Small error improvement * Fix * one special test * couple of fixes for lstm * Rewrite TFGraphMapper.getNDArrayFromTensor to be maintainable and less error prone * Fixes * FP16 * Unsigned * BFloat16 * Fill op - empty tweaks * - couple of fixes for empty arrays construction - stack updated * strided slice fix * one transform test * provide method for reducing shapeInfo in case of input array is empty * Fixed reduceAlongDimensions to use empty input properly. * couple of broadcast tests * couple of tests broadcast tests + tweak to make them pass * add check of non-empty to methods producing sub-arrays * Fixed reshapeC with zeros in shape. * complete empty check in reduce_... legacy ops * Concat and cumsum/prod * Tweak to empty shape inference on import * add empty check to the rest of reduce legacy ops * one more test * correct typo in evalReduceShapeInfoEmpty * Added tests for reduce_* ops to tests with zero shapes. * few more tests for empty reductions * Fixed strided_slice op with empty case and tests. * one more empty reduction test * Fixed strided_slice test. * add empty check to NDArray::reshapei * infOrMax * empty min/max with infinity tests * made unstack working correctly with empty arrays * few IndexReduce tests + tweaks for empty shapes * add test for empty concat * few tests fixed * Validation fix for reductions on empty shapes * Reverse fix * Reduction shape calc fixes * SameDiff.generateOutputVariable: don't use shape function to determine number of outputs * Range fix * - NDArray constructor updated for scalars/empty arrays - few tests fixed * More fixes * Empty creator fixes * concat fix * concat fix * TF import tests: allow 'both all NaN' and 'both all inf' to pass * Slice, zero fraction, and reshape fixes * transpose, gather * Zero fraction * scalar cast fix * Empty reduction axis support * few more tests fixed * Fixed input checks conforming with TF for concat op and tests. * few tests fixed * matmul scalar shape fix * Fixed checkout for data type and scalarity with concat to allow non-empty scalars with vector concats. * broadcast bool fix * few more tests * few more tests * correct evalReduceShapeInfoEmpty * argmax/argmin + tests * one more empty edge case + one more test * argmax/argmin/realdiv_bp tweaks * empty reshape test + fix * Helper fixes * Small fixes * Gather test fix * Gather test fix * Small fixes * reduce scalar zero values * scalar mean workaround * Remove debug code * along dim mean workaround * one more test * - equalsTo() tweak for empty arrays - one more test * broadcast tweaks
2019-06-15 13:34:34 +02:00
return -nd4j::DataTypeUtils::infOrMax<X>();
Eclipse Migration Initial Commit
2019-06-06 14:21:15 +02:00
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> startingIndexValue(X *input) {
functions::indexreduce::IndexValue<X> local;
local.value = startingValue(input);
local.index = 0;
return local;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> op(functions::indexreduce::IndexValue<X> d1,
functions::indexreduce::IndexValue<X> d2, X *extraParams) {
return d1;
}
};
template <typename X>
class IndexAbsoluteMin {
public:
static _CUDA_HD inline functions::indexreduce::IndexValue<X> op(
functions::indexreduce::IndexValue<X> val, X *extraParams) {
return val;
}
static _CUDA_HD inline X startingValue(const X *input) {
Dev branch merge: dev_20190606 (#7904) * correct logsoftmax looss (#2) * Small SameDiff listener fix (#4) * Various fixes (#6) * #7839 Fix for asXMatrix and tests * #7866 EmbeddingSequenceLayer dtype fix + test * #7856 SameDiff save/load stream methods * #7859 RegressionEvaluation rank 4 fix + tests + axis configuration * EvaluationBinary 3d/4d * More evaluation 3d/4d tests * #7847 Evaluation empty checks * Small test ifx * #7848 Fix median edge case * Improve DL4J samediff layer tests * [WIP] FastText wrapper implemented (#8) * FastText implemented * Some fixes * Fix shapes for wordsNearest * Validation of input vectors * Fixes * Fixed test * Thread tagged * Some tweaks * setContextClassLoader for DeallocatorServiceThread * Numpy format tests (#1) * Various fixes (#11) * #7852 SameDiff gather fix * #7892 SameDiff placeholder to constant conversion * #7890 validate input rank for MLN/CG init methods * Fix broken permute shape calculation * Permute and gather fixes * Tests * #7850 LogSumExp fix + test * Handful of test fixes * Empty arrays with non-scalar shapes (#10) * minor rearrangements for lambdas * empty tensors with non-scalar shapes * numpy empty tensors with non-scalar shapes * few more empty tweaks * Small fixes * conv3d signature update * micro fix in batchnorm mkldnn * Import fixes * Fix * MKL-DNN update * Small fill fix * fill with empty input + test * Fixes * Small error improvement * Fix * one special test * couple of fixes for lstm * Rewrite TFGraphMapper.getNDArrayFromTensor to be maintainable and less error prone * Fixes * FP16 * Unsigned * BFloat16 * Fill op - empty tweaks * - couple of fixes for empty arrays construction - stack updated * strided slice fix * one transform test * provide method for reducing shapeInfo in case of input array is empty * Fixed reduceAlongDimensions to use empty input properly. * couple of broadcast tests * couple of tests broadcast tests + tweak to make them pass * add check of non-empty to methods producing sub-arrays * Fixed reshapeC with zeros in shape. * complete empty check in reduce_... legacy ops * Concat and cumsum/prod * Tweak to empty shape inference on import * add empty check to the rest of reduce legacy ops * one more test * correct typo in evalReduceShapeInfoEmpty * Added tests for reduce_* ops to tests with zero shapes. * few more tests for empty reductions * Fixed strided_slice op with empty case and tests. * one more empty reduction test * Fixed strided_slice test. * add empty check to NDArray::reshapei * infOrMax * empty min/max with infinity tests * made unstack working correctly with empty arrays * few IndexReduce tests + tweaks for empty shapes * add test for empty concat * few tests fixed * Validation fix for reductions on empty shapes * Reverse fix * Reduction shape calc fixes * SameDiff.generateOutputVariable: don't use shape function to determine number of outputs * Range fix * - NDArray constructor updated for scalars/empty arrays - few tests fixed * More fixes * Empty creator fixes * concat fix * concat fix * TF import tests: allow 'both all NaN' and 'both all inf' to pass * Slice, zero fraction, and reshape fixes * transpose, gather * Zero fraction * scalar cast fix * Empty reduction axis support * few more tests fixed * Fixed input checks conforming with TF for concat op and tests. * few tests fixed * matmul scalar shape fix * Fixed checkout for data type and scalarity with concat to allow non-empty scalars with vector concats. * broadcast bool fix * few more tests * few more tests * correct evalReduceShapeInfoEmpty * argmax/argmin + tests * one more empty edge case + one more test * argmax/argmin/realdiv_bp tweaks * empty reshape test + fix * Helper fixes * Small fixes * Gather test fix * Gather test fix * Small fixes * reduce scalar zero values * scalar mean workaround * Remove debug code * along dim mean workaround * one more test * - equalsTo() tweak for empty arrays - one more test * broadcast tweaks
2019-06-15 13:34:34 +02:00
return nd4j::DataTypeUtils::infOrMax<X>();
Eclipse Migration Initial Commit
2019-06-06 14:21:15 +02:00
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> startingIndexValue(X *input) {
functions::indexreduce::IndexValue<X> local;
local.value = startingValue(input);
local.index = 0;
return local;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> update(functions::indexreduce::IndexValue<X> &old, functions::indexreduce::IndexValue<X> &opOutput, X *extraParams) {
opOutput.value = nd4j::math::nd4j_abs<X>(opOutput.value);
old.value = nd4j::math::nd4j_abs<X>(old.value);
if (opOutput.value < old.value)
return opOutput;
#ifdef __CUDACC__
// workaround for cuda race condition at merge phase
else if (opOutput.value == old.value && opOutput.index < old.index)
return opOutput;
#elif defined(__GNUC__)
#endif
return old;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> merge(
functions::indexreduce::IndexValue<X> f1,
functions::indexreduce::IndexValue<X> f2, X *extraParams) {
if (nd4j::math::nd4j_abs<X>(f1.value) < nd4j::math::nd4j_abs<X>(f2.value))
return f2;
return f1;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> postProcess(
functions::indexreduce::IndexValue<X> reduction, int n, int xOffset,
X *dx, int incx, X *extraParams, X *result) {
return reduction;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> op(functions::indexreduce::IndexValue<X> d1,
functions::indexreduce::IndexValue<X> d2, X *extraParams) {
return d1;
}
};
template <typename X>
class IndexMin {
public:
static _CUDA_HD inline functions::indexreduce::IndexValue<X> op(
functions::indexreduce::IndexValue<X> val, X *extraParams) {
return val;
}
static _CUDA_HD inline X startingValue(const X *input) {
Dev branch merge: dev_20190606 (#7904) * correct logsoftmax looss (#2) * Small SameDiff listener fix (#4) * Various fixes (#6) * #7839 Fix for asXMatrix and tests * #7866 EmbeddingSequenceLayer dtype fix + test * #7856 SameDiff save/load stream methods * #7859 RegressionEvaluation rank 4 fix + tests + axis configuration * EvaluationBinary 3d/4d * More evaluation 3d/4d tests * #7847 Evaluation empty checks * Small test ifx * #7848 Fix median edge case * Improve DL4J samediff layer tests * [WIP] FastText wrapper implemented (#8) * FastText implemented * Some fixes * Fix shapes for wordsNearest * Validation of input vectors * Fixes * Fixed test * Thread tagged * Some tweaks * setContextClassLoader for DeallocatorServiceThread * Numpy format tests (#1) * Various fixes (#11) * #7852 SameDiff gather fix * #7892 SameDiff placeholder to constant conversion * #7890 validate input rank for MLN/CG init methods * Fix broken permute shape calculation * Permute and gather fixes * Tests * #7850 LogSumExp fix + test * Handful of test fixes * Empty arrays with non-scalar shapes (#10) * minor rearrangements for lambdas * empty tensors with non-scalar shapes * numpy empty tensors with non-scalar shapes * few more empty tweaks * Small fixes * conv3d signature update * micro fix in batchnorm mkldnn * Import fixes * Fix * MKL-DNN update * Small fill fix * fill with empty input + test * Fixes * Small error improvement * Fix * one special test * couple of fixes for lstm * Rewrite TFGraphMapper.getNDArrayFromTensor to be maintainable and less error prone * Fixes * FP16 * Unsigned * BFloat16 * Fill op - empty tweaks * - couple of fixes for empty arrays construction - stack updated * strided slice fix * one transform test * provide method for reducing shapeInfo in case of input array is empty * Fixed reduceAlongDimensions to use empty input properly. * couple of broadcast tests * couple of tests broadcast tests + tweak to make them pass * add check of non-empty to methods producing sub-arrays * Fixed reshapeC with zeros in shape. * complete empty check in reduce_... legacy ops * Concat and cumsum/prod * Tweak to empty shape inference on import * add empty check to the rest of reduce legacy ops * one more test * correct typo in evalReduceShapeInfoEmpty * Added tests for reduce_* ops to tests with zero shapes. * few more tests for empty reductions * Fixed strided_slice op with empty case and tests. * one more empty reduction test * Fixed strided_slice test. * add empty check to NDArray::reshapei * infOrMax * empty min/max with infinity tests * made unstack working correctly with empty arrays * few IndexReduce tests + tweaks for empty shapes * add test for empty concat * few tests fixed * Validation fix for reductions on empty shapes * Reverse fix * Reduction shape calc fixes * SameDiff.generateOutputVariable: don't use shape function to determine number of outputs * Range fix * - NDArray constructor updated for scalars/empty arrays - few tests fixed * More fixes * Empty creator fixes * concat fix * concat fix * TF import tests: allow 'both all NaN' and 'both all inf' to pass * Slice, zero fraction, and reshape fixes * transpose, gather * Zero fraction * scalar cast fix * Empty reduction axis support * few more tests fixed * Fixed input checks conforming with TF for concat op and tests. * few tests fixed * matmul scalar shape fix * Fixed checkout for data type and scalarity with concat to allow non-empty scalars with vector concats. * broadcast bool fix * few more tests * few more tests * correct evalReduceShapeInfoEmpty * argmax/argmin + tests * one more empty edge case + one more test * argmax/argmin/realdiv_bp tweaks * empty reshape test + fix * Helper fixes * Small fixes * Gather test fix * Gather test fix * Small fixes * reduce scalar zero values * scalar mean workaround * Remove debug code * along dim mean workaround * one more test * - equalsTo() tweak for empty arrays - one more test * broadcast tweaks
2019-06-15 13:34:34 +02:00
return nd4j::DataTypeUtils::infOrMax<X>();
Eclipse Migration Initial Commit
2019-06-06 14:21:15 +02:00
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> startingIndexValue(X *input) {
functions::indexreduce::IndexValue<X> local;
local.value = startingValue(input);
local.index = 0;
return local;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> update(functions::indexreduce::IndexValue<X> &old, functions::indexreduce::IndexValue<X> &opOutput, X *extraParams) {
if (opOutput.value < old.value)
return opOutput;
#ifdef __CUDACC__
// workaround for cuda race condition at merge phase
else if (opOutput.value == old.value && opOutput.index < old.index)
return opOutput;
#elif defined(__GNUC__)
#endif
return old;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> merge(
functions::indexreduce::IndexValue<X> f1,
functions::indexreduce::IndexValue<X> f2, X *extraParams) {
if (f1.value < f2.value)
return f2;
return f1;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> postProcess(
functions::indexreduce::IndexValue<X> reduction, int n, int xOffset,
X *dx, int incx, X *extraParams, X *result) {
return reduction;
}
static _CUDA_HD inline functions::indexreduce::IndexValue<X> op(functions::indexreduce::IndexValue<X> d1,
functions::indexreduce::IndexValue<X> d2, X *extraParams) {
return d1;
}
};
template <typename X, typename Z>
class SummaryStatsVariance {
public:
static _CUDA_HD inline Z getValue(const bool biasCorrected, functions::summarystats::SummaryStatsData<X> val) {
if (biasCorrected) {
Z ret = static_cast<Z>(val.varianceBiasCorrected());
if (ret < static_cast<Z>(0.0f))
return static_cast<Z>(val.variance());
return ret;
}
return static_cast<Z>(val.variance());
}
static _CUDA_HD inline functions::summarystats::SummaryStatsData<X> op(functions::summarystats::SummaryStatsData<X> d1, Z *extraParams) {
return d1;
}
};
template <typename X, typename Z>
class SummaryStatsStandardDeviation {
public:
static _CUDA_HD inline Z getValue(const bool biasCorrected, functions::summarystats::SummaryStatsData<X> val) {
if (biasCorrected) {
auto ret = static_cast<Z>(val.varianceBiasCorrected());
if (ret < static_cast<Z>(0.0f))
return nd4j::math::nd4j_sqrt<double, Z>(val.variance());
else
return nd4j::math::nd4j_sqrt<double, Z>(ret);
}
return nd4j::math::nd4j_sqrt<double, Z>(val.variance());
}
static _CUDA_HD inline functions::summarystats::SummaryStatsData<X> op(functions::summarystats::SummaryStatsData<X> d1, Z *extraParams) {
return d1;
}
};
template <typename X>
class DropOut {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
inline _CUDA_D static X op(X d1, X *params) {
X prob = params[0];
#ifdef __CUDACC__
X length = params[1];
X tid = blockIdx.x * blockDim.x + threadIdx.x;
X rnd = nd4j::math::nd4j_abs<X>(nd4j::math::nd4j_cos<X>(static_cast<X>(clock64()) * static_cast<X>(tid) + static_cast<X>(length) * static_cast<X>(tid)));
#else
X rnd = static_cast<X>(rand() / RAND_MAX);
#endif
return rnd >= prob ? static_cast<X>(0.0f) : d1;
}
};
template <typename X, typename Y, typename Z>
class DropOutInverted {
public:
no_op_exec_special
no_op_exec_special_cuda
#ifdef __CUDACC__
__device__
#endif
inline static Z op(X d1, Y d2, Z *params) {
Y prob = d2;
#ifdef __CUDACC__
X length = params[1];
X tid = blockIdx.x * blockDim.x + threadIdx.x;
X rnd = nd4j::math::nd4j_abs<X>(nd4j::math::nd4j_cos<X>(static_cast<X>(clock64()) * static_cast<X>(tid) + static_cast<X>(length) * static_cast<X>(tid)));
#else
X rnd = static_cast<X>(rand() / RAND_MAX);
#endif
return rnd >= static_cast<X>(prob) ? static_cast<Z>(0.0f) : reinterpret_cast<Z>(d1 / static_cast<X>(prob));
}
};
template <typename X, typename Y, typename Z>
class ReplaceNans {
public:
no_op_exec_special
no_op_exec_special_cuda
op_def static Z op(X d1, Y d2, Z *params) {
return nd4j::math::nd4j_isnan(d1) ? static_cast<Z>(d2) : static_cast<Z>(d1) ;
}
};
// this op is used for conditional pairwise transforms only
template <typename X, typename Y, typename Z>
class CompareAndReplace{
public:
// op definition for PairWise Transform
op_def static Z op(X d1, Y d2, Z *params) {
auto zd1 = static_cast<Z>(d1);
auto zd2 = static_cast<Z>(d2);
auto compare = params[0];
auto eps = params[2];
int mode = (int) params[3];
if (mode == 0) // equals
if (nd4j::math::nd4j_abs<Z>(zd1 - compare) <= eps)
return zd2;
else
return zd1;
else if (mode == 1) // not equals eps
if (nd4j::math::nd4j_abs<Z>(zd1 - compare) > eps)
return zd2;
else
return zd1;
else if (mode == 2) // less_than eps
if (zd1 < compare)
return zd2;
else
return zd1;
else if (mode ==3) // greater_than
if (zd1 > compare)
return zd2;
else
return zd1;
else if (mode == 4) // less_or_equals_than
if (zd1 <= compare)
return zd2;
else
return zd1;
else if (mode == 5) // greater_or_equals_than
if (zd1 >= compare)
return zd2;
else
return zd1;
else if (mode == 6) // abs_less_than
if (nd4j::math::nd4j_abs<Z>(zd1) < compare)
return zd2;
else
return zd1;
else if (mode == 7) // abs_greater_than
if (nd4j::math::nd4j_abs<Z>(zd1) > compare)
return zd2;
else
return zd1;
else if (mode == 8) // is inf
if (nd4j::math::nd4j_isinf(zd1))
return zd2;
else
return zd1;
else if (mode == 9) // is nan
if (nd4j::math::nd4j_isnan(zd1))
return zd2;
else
return zd1;
else if (mode == 10)
if (zd1 == compare)
return zd2;
else
return zd1;
else if (mode == 11)
if (zd1 != compare)
return zd2;
else
return zd1;
else if (mode == 12) // abs_greater_or_equals_than
if (nd4j::math::nd4j_abs<Z>(zd1) >= compare)
return zd2;
else
return zd1;
else if (mode == 13) // abs_less_or_equals_than
if (nd4j::math::nd4j_abs<Z>(zd1) <= compare)
return zd2;
else
return zd1;
else
printf("Undefined boolean operation: [%i]\n", mode);
return zd1;
}
};
template <typename X, typename Y, typename Z>
class CompareAndSet {
public:
// op definition for PairWise Transform
op_def static Z op(X dX, Y dY, Z *params) {
auto d1 = static_cast<Z>(dX);
auto d2 = static_cast<Z>(dY);
auto compare = params[0];
auto eps = params[2];
auto mode = static_cast<int>(params[3]);
if (mode == 0) // equals
if (nd4j::math::nd4j_abs<Z>(d2 - compare) <= eps)
return d2;
else
return d1;
else if (mode == 1) // not equals
if (nd4j::math::nd4j_abs<Z>(d2 - compare) > eps)
return d2;
else
return d1;
else if (mode == 2) // less_than
if (d2 < compare)
return d2;
else
return d1;
else if (mode ==3) // greater_than
if (d2 > compare)
return d2;
else
return d1;
else if (mode == 4) // less_or_equals_than
if (d2 <= compare)
return d2;
else
return d1;
else if (mode == 5) // greater_or_equals_than
if (d2 >= compare)
return d2;
else
return d1;
else if (mode == 6) // abs_less_than
if (nd4j::math::nd4j_abs<Z>(d2) < compare)
return d2;
else
return d1;
else if (mode == 7) // abs_greater_than
if (nd4j::math::nd4j_abs<Z>(d2) > compare)
return d2;
else
return d1;
else if (mode == 8) // is inf
if (nd4j::math::nd4j_isinf(d2))
return d2;
else
return d1;
else if (mode == 9) // is nan
if (nd4j::math::nd4j_isnan(d2))
return d2;
else
return d1;
else if (mode == 10)
if (d2 == compare)
return d2;
else
return d1;
else if (mode == 11)
if (d2 != compare)
return d2;
else
return d1;
else if (mode == 12) // abs_greater_or_equals_than
if (nd4j::math::nd4j_abs<Z>(d1) >= compare)
return d2;
else
return d1;
else if (mode == 13) // abs_less_or_equals_than
if (nd4j::math::nd4j_abs<Z>(d1) <= compare)
return d2;
else
return d1;
else
printf("Undefined boolean operation: [%i]\n", mode);
return d1;
}
};
template <typename X>
class CompareAndSetTransform {
public:
no_op_exec_special_same
no_op_exec_special_same_cuda
// op definition for Transform
op_def static X op(X d1, X *params) {
auto compare = params[0];
auto set = params[1];
auto eps = params[2];
// with mode == 0 we do set if d1 equals to compare, and with mode == 1 - we go otherwise
int mode = (int) params[3];
if (mode == 0) // equals
if (nd4j::math::nd4j_abs<X>(d1 - compare) <= eps)
return set;
else
return d1;
//return nd4j::math::nd4j_abs<T>(d1 - compare) <= eps ? set : d1;
else if (mode == 1) // not equals
if (nd4j::math::nd4j_abs<X>(d1 - compare) > eps)
return set;
else
return d1;
//return nd4j::math::nd4j_abs<T>(d1 - compare) > eps ? set : d1;
else if (mode == 2) // less_than
if (d1 < compare)
return set;
else
return d1;
else if (mode ==3) // greater_than
if (d1 > compare)
return set;
else
return d1;
else if (mode == 4) // less_or_equals_than
if (d1 <= compare)
return set;
else
return d1;
else if (mode == 5) // greater_or_equals_than
if (d1 >= compare)
return set;
else
return d1;
else if (mode == 6) // abs_less_than
if (nd4j::math::nd4j_abs<X>(d1) < compare)
return set;
else
return d1;
else if (mode == 7) // abs_greater_than
if (nd4j::math::nd4j_abs<X>(d1) > compare)
return set;
else
return d1;
else if (mode == 8) // is inf
if (nd4j::math::nd4j_isinf(d1))
return set;
else
return d1;
else if (mode == 9) // is nan
if (nd4j::math::nd4j_isnan(d1))
return set;
else
return d1;
else if (mode == 10)
if (d1 == compare)
return set;
else
return d1;
else if (mode == 11)
if (d1 != compare)
return set;
else
return d1;
else if (mode == 12) // abs_greater_or_equals_than
if (nd4j::math::nd4j_abs<X>(d1) >= compare)
return set;
else
return d1;
else if (mode == 13) // abs_less_or_equals_than
if (nd4j::math::nd4j_abs<X>(d1) <= compare)
return set;
else
return d1;
else
printf("Undefined boolean operation: [%i]\n", mode);
return d1;
}
};
}
#endif