307 lines
10 KiB
C++
307 lines
10 KiB
C++
/* ******************************************************************************
|
|
*
|
|
*
|
|
* 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.
|
|
*
|
|
* See the NOTICE file distributed with this work for additional
|
|
* information regarding copyright ownership.
|
|
* 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
|
|
******************************************************************************/
|
|
|
|
//
|
|
// @author raver119@protonmail.com
|
|
//
|
|
|
|
#ifndef LIBND4J_GRAPH_RNG_H
|
|
#define LIBND4J_GRAPH_RNG_H
|
|
|
|
#include <types/u64.h>
|
|
#include <types/u32.h>
|
|
#include <system/pointercast.h>
|
|
#include <system/op_boilerplate.h>
|
|
#include <system/dll.h>
|
|
#include <chrono>
|
|
#include <array/DataTypeUtils.h>
|
|
#include <helpers/logger.h>
|
|
#include <stdexcept>
|
|
#include <math/templatemath.h>
|
|
|
|
#ifdef __CUDACC__
|
|
#include <cuda.h>
|
|
#include <cuda_runtime.h>
|
|
#endif
|
|
|
|
namespace sd {
|
|
namespace graph {
|
|
#ifdef __CUDACC__
|
|
class ND4J_EXPORT CudaManagedRandomGenerator {
|
|
private:
|
|
|
|
protected:
|
|
void *devHolder;
|
|
|
|
public:
|
|
void *operator new(size_t len) {
|
|
void *ptr;
|
|
auto res = cudaHostAlloc(&ptr, len, cudaHostAllocDefault);
|
|
if (res != 0)
|
|
throw std::runtime_error("CudaManagedRandomGenerator: failed to allocate memory");
|
|
|
|
return ptr;
|
|
}
|
|
|
|
void operator delete(void *ptr) {
|
|
cudaFreeHost(ptr);
|
|
}
|
|
};
|
|
|
|
class ND4J_EXPORT RandomGenerator : public CudaManagedRandomGenerator {
|
|
#else
|
|
class ND4J_EXPORT RandomGenerator {
|
|
#endif
|
|
private:
|
|
#ifndef __CUDACC__
|
|
void *placeHolder;
|
|
#endif
|
|
// GRAPH-LEVEL STATE
|
|
u64 _rootState;
|
|
|
|
// NODE-LEVEL STATE
|
|
u64 _nodeState;
|
|
|
|
/**
|
|
* Utility method, returns number of milliseconds since 1970
|
|
* Leave this static if possible to avoid problems in constructor
|
|
*/
|
|
static FORCEINLINE Nd4jLong currentMilliseconds();
|
|
|
|
public:
|
|
FORCEINLINE _CUDA_HD uint32_t xoroshiro32(uint64_t index);
|
|
FORCEINLINE _CUDA_HD uint64_t xoroshiro64(uint64_t index);
|
|
|
|
/**
|
|
* This method returns integer value between 0 and MAX_UINT
|
|
*/
|
|
//uint32_t relativeUInt32(Nd4jLong index);
|
|
|
|
public:
|
|
FORCEINLINE RandomGenerator(Nd4jLong rootSeed = 0, Nd4jLong nodeSeed = 0);
|
|
|
|
/**
|
|
* This method allows to change graph-level state in runtime.
|
|
* PLEASE NOTE: this method will change state of node as well.
|
|
*/
|
|
FORCEINLINE _CUDA_H void setStates(Nd4jLong rootSeed, Nd4jLong nodeState = 0);
|
|
|
|
|
|
|
|
/**
|
|
* This method returns T value between from and to
|
|
*/
|
|
template <typename T>
|
|
FORCEINLINE _CUDA_HD T relativeT(Nd4jLong index, T from, T to);
|
|
|
|
/**
|
|
* This method returns T value between 0 and MAX_T
|
|
*/
|
|
template <typename T>
|
|
FORCEINLINE _CUDA_HD T relativeT(Nd4jLong index);
|
|
|
|
/**
|
|
* These two methods are made for JVM
|
|
* @param index
|
|
* @return
|
|
*/
|
|
FORCEINLINE _CUDA_HD int relativeInt(Nd4jLong index);
|
|
FORCEINLINE _CUDA_HD Nd4jLong relativeLong(Nd4jLong index);
|
|
|
|
FORCEINLINE _CUDA_HD void rewindH(uint64_t steps);
|
|
|
|
/**
|
|
* These methods set up only node states, with non-changed root ones
|
|
*/
|
|
FORCEINLINE _CUDA_H void setSeed(int seed) {
|
|
_nodeState._ulong = static_cast<uint64_t>(seed);
|
|
}
|
|
|
|
FORCEINLINE _CUDA_H void setSeed(uint64_t seed) {
|
|
_nodeState._ulong = seed;
|
|
}
|
|
|
|
FORCEINLINE _CUDA_HD Nd4jLong rootState() {
|
|
return _rootState._long;
|
|
}
|
|
|
|
FORCEINLINE _CUDA_HD Nd4jLong nodeState() {
|
|
return _nodeState._long;
|
|
}
|
|
};
|
|
|
|
|
|
FORCEINLINE RandomGenerator::RandomGenerator(Nd4jLong rootSeed, Nd4jLong nodeSeed) {
|
|
// this seed is used graph-level state
|
|
if (rootSeed == 0)
|
|
rootSeed = currentMilliseconds();
|
|
|
|
// graph-level state is just first seed
|
|
_rootState._long = rootSeed;
|
|
|
|
// used to build second, node state
|
|
_nodeState._long = (nodeSeed != 0 ? nodeSeed: 1298567341LL);
|
|
}
|
|
|
|
FORCEINLINE void RandomGenerator::setStates(Nd4jLong rootSeed, Nd4jLong nodeSeed) {
|
|
// this seed is used graph-level state
|
|
if (rootSeed == 0)
|
|
rootSeed = currentMilliseconds();
|
|
|
|
// graph-level state is just first seed
|
|
_rootState._long = rootSeed;
|
|
|
|
// used to build second, node state
|
|
_nodeState._long = (nodeSeed != 0 ? nodeSeed: 1298567341LL);
|
|
}
|
|
|
|
FORCEINLINE Nd4jLong RandomGenerator::currentMilliseconds() {
|
|
auto s = std::chrono::system_clock::now().time_since_epoch();
|
|
auto v = std::chrono::duration_cast<std::chrono::milliseconds>(s).count();
|
|
return v;
|
|
}
|
|
|
|
template <>
|
|
_CUDA_HD FORCEINLINE float RandomGenerator::relativeT<float>(Nd4jLong index) {
|
|
u32 u;
|
|
u._u32 = (0x3f800000 | (this->xoroshiro32(index) >> 9));
|
|
return u._f32 - 1.0f;
|
|
}
|
|
|
|
template <>
|
|
_CUDA_HD FORCEINLINE double RandomGenerator::relativeT<double>(Nd4jLong index) {
|
|
#ifdef __DOUBLE_RNG__
|
|
u64 u;
|
|
u._ulong = ((UINT64_C(0x3FF) << 52) | (this->xoroshiro64(index) >> 12));
|
|
return u._double - 1.0;
|
|
#else
|
|
return (double) relativeT<float>(index);
|
|
#endif
|
|
}
|
|
|
|
template <>
|
|
_CUDA_HD FORCEINLINE uint64_t RandomGenerator::relativeT<uint64_t>(Nd4jLong index) {
|
|
return this->xoroshiro64(index);
|
|
}
|
|
|
|
template <>
|
|
_CUDA_HD FORCEINLINE uint32_t RandomGenerator::relativeT<uint32_t>(Nd4jLong index) {
|
|
return this->xoroshiro32(index);
|
|
}
|
|
|
|
template <>
|
|
_CUDA_HD FORCEINLINE int RandomGenerator::relativeT<int>(Nd4jLong index) {
|
|
auto r = relativeT<uint32_t>(index);
|
|
return r <= DataTypeUtils::max<int>() ? r : r % DataTypeUtils::max<int>();
|
|
}
|
|
|
|
template <>
|
|
_CUDA_HD FORCEINLINE Nd4jLong RandomGenerator::relativeT<Nd4jLong>(Nd4jLong index) {
|
|
auto r = relativeT<uint64_t>(index);
|
|
return r <= DataTypeUtils::max<Nd4jLong>() ? r : r % DataTypeUtils::max<Nd4jLong>();
|
|
}
|
|
|
|
template <typename T>
|
|
_CUDA_HD FORCEINLINE T RandomGenerator::relativeT(Nd4jLong index, T from, T to) {
|
|
auto t = this->relativeT<T>(index);
|
|
auto z = from + T(t * (to - from));
|
|
return z;
|
|
}
|
|
|
|
template <>
|
|
_CUDA_HD FORCEINLINE Nd4jLong RandomGenerator::relativeT(Nd4jLong index, Nd4jLong from, Nd4jLong to) {
|
|
auto t = this->relativeT<double>(index);
|
|
auto z = from + Nd4jLong(t * (to - from));
|
|
return z;
|
|
}
|
|
|
|
template <>
|
|
_CUDA_HD FORCEINLINE int RandomGenerator::relativeT(Nd4jLong index, int from, int to) {
|
|
auto t = this->relativeT<float>(index);
|
|
auto z = from + float(t * (to - from));
|
|
return z;
|
|
}
|
|
|
|
template <typename T>
|
|
_CUDA_HD FORCEINLINE T RandomGenerator::relativeT(Nd4jLong index) {
|
|
// This is default implementation for floating point types
|
|
return static_cast<T>(relativeT<float>(index));
|
|
}
|
|
|
|
|
|
_CUDA_HD FORCEINLINE int RandomGenerator::relativeInt(Nd4jLong index) {
|
|
auto r = relativeT<uint32_t>(index);
|
|
return r <= DataTypeUtils::max<int>() ? r : r % DataTypeUtils::max<int>();
|
|
}
|
|
|
|
_CUDA_HD FORCEINLINE Nd4jLong RandomGenerator::relativeLong(Nd4jLong index) {
|
|
auto r = relativeT<uint64_t>(index);
|
|
return r <= DataTypeUtils::max<Nd4jLong>() ? r : r % DataTypeUtils::max<Nd4jLong>();
|
|
}
|
|
|
|
//////
|
|
static FORCEINLINE _CUDA_HD uint32_t rotl(const uint32_t x, int k) {
|
|
return (x << k) | (x >> (32 - k));
|
|
}
|
|
|
|
static FORCEINLINE _CUDA_HD uint64_t rotl(const uint64_t x, int k) {
|
|
return (x << k) | (x >> (64 - k));
|
|
}
|
|
|
|
static FORCEINLINE _CUDA_HD uint32_t next(uint32_t s0, uint32_t s1, uint32_t s2, uint32_t s3) {
|
|
const uint32_t result = rotl(s0 + s3, 7) + s0;
|
|
return result;
|
|
}
|
|
|
|
_CUDA_HD FORCEINLINE uint32_t RandomGenerator::xoroshiro32(uint64_t index) {
|
|
auto s0 = _rootState._ulong;
|
|
auto s1 = _nodeState._ulong;
|
|
|
|
// xor by idx
|
|
s0 |= ((index + 2) * (s1 + 24243287));
|
|
s1 ^= ((index + 2) * (s0 + 723829));
|
|
|
|
unsigned long val = 0;
|
|
val = s1 ^ s0;
|
|
int* pHalf = reinterpret_cast<int*>(&val);
|
|
|
|
return rotl(*pHalf * 0x9E3779BB, 5) * 5;
|
|
}
|
|
|
|
_CUDA_HD FORCEINLINE uint64_t RandomGenerator::xoroshiro64(uint64_t index) {
|
|
uint64_t upper = ((uint64_t) xoroshiro32(index)) << 32;
|
|
uint32_t lower = xoroshiro32(sd::math::nd4j_rotl<uint64_t>(index, 32));
|
|
return upper + lower;
|
|
}
|
|
|
|
_CUDA_HD FORCEINLINE void RandomGenerator::rewindH(uint64_t steps) {
|
|
// we only update node state, if any
|
|
auto s0 = _nodeState._du32._v0;
|
|
auto s1 = _nodeState._du32._v1;
|
|
|
|
s1 ^= s0;
|
|
_nodeState._du32._v0 = rotl(s0, 26) ^ s1 ^ (s1 << 9); // a, b
|
|
_nodeState._du32._v1 = rotl(s1, 13); // c
|
|
|
|
_nodeState._long ^= (steps ^ 0xdeadbeef);
|
|
}
|
|
}
|
|
}
|
|
|
|
#endif
|