cuRAND随机数生成库

Host API

要使用cuRAND host API ，需要引用头文件<curand.h>，步骤如下

1. Create a new generator of the desired type with curandCreateGenerator().
2. Set the generator options ;for example, use curandSetPseudoRandomGeneratorSeed() to set the seed.
3. Allocate memory on the device with cudaMalloc().
4. Generate random numbers with curandGenerate() or another generation function.
5. Clean up with curandDestroyGenerator().

如果想在CPU生成随机数，在step1调用curandCreateGeneratorHost()，并在step3申请host 内存接受结果，其他步骤不变。

在同一时间创建多个generators是合法的，每个generators生成的随机数序列跟设置参数有关，设置参数相同，生成的随机数序列也会相同，无论是在GPU还是CPU。

不能向运行在GPU上的generator传递host memory pointer，也不能向运行在CPU上的generator传递device memory pointer.

Generator Types

curandStatus_t curandCreateGenerator(curandGenerator_t* generator , curandRngType_t rng_type); 

• generator

  Pointer to generator

• rng_type

  Type of generator to create

通过传递一个类型来创建 Random number generators。

伪随机数生成器(pseudorandom number generators)：

• CURAND_RNG_PSEUDO_XORWOW,
• CURAND_RNG_PSEUDO_MRG32K3A,
• CURAND_RNG_PSEUDO_MTGP32,
• CURAND_RNG_PSEUDO_PHILOX4_32_10
• CURAND_RNG_PSEUDO_MT19937
  只能用于Host API，sm_35架构以上

拟随机数生成器(quasi random number generator)：

• CURAND_RNG_QUASI_SOBOL32,
• CURAND_RNG_QUASI_SCRAMBLED_SOBOL32,
• CURAND_RNG_QUASI_SOBOL64
• CURAND_RNG_QUASI_SCRAMBLED_SOBOL64

Generator Options

curandStatus_t curandSetPseudoRandomGeneratorSeed(curandGenerator_t generator, unsigned long long seed);
curandStatus_t  curandSetQuasiRandomGeneratorDimensions(curandGenerator_t generator, unsigned int  num_dimensions );
curandStatus_t curandSetGeneratorOffset(curandGenerator_t generator, unsigned long long offset);
curandStatus_t curandSetGeneratorOrdering(curandGenerator_t generator,curandOrdering_t order);
curandStatus_t curandSetStream(curandGenerator_t generator,cudaStream_t stream);

Random number generators一旦创建好了，就可以用general options seed, offset, and order来定义generators。

• Seed
  是一个64-bit integar，用来初始化伪随机数生成器的起始状态，相同的seed经常生成相同的随机数序列
• Offset
  offset参数用来跳过随机数序列的开头，第一个随机数从序列的第offset个开始取，这使得多次运行同一程序，从相同随机数序列生成的随机数不重叠。is not available for the CURAND_RNG_PSEUDO_MTGP32 and CURAND_RNG_PSEUDO_MT19937 generators
• Order
  用来选择结果如何在全局内存中排序
  • CURAND_ORDERING_PSEUDO_DEFAULT
  • CURAND_ORDERING_PSEUDO_LEGACY
  • CURAND_ORDERING_PSEUDO_BEST
  • CURAND_ORDERING_PSEUDO_SEEDED
  • CURAND_ORDERING_QUASI_DEFAULT
    用于拟随机数的选项

Return Values

All cuRAND host library calls have a return value of curandStatus_t. Calls that succeed without errors return CURAND_STATUS_SUCCESS.

Generation Functions

curandStatus_t curandGenerate(curandGenerator_t generator,unsigned int *outputPtr, size_t num);
    
curandStatus_t curandGenerateLongLong(curandGenerator_t generator,unsigned long long *outputPtr, size_t num);

curandGenerate() 用于生成32位unsigned int型 pseudo- or quasi-random for XORWOW, MRG32k3a, MTGP32, MT19937, Philox_4x32_10 and SOBOL32 generators，其中每一位都是随机的。

curandGenerateLongLong()用于生成64位unsigned long long类型的随机数with the SOBOL64 generators.

curandStatus_t curandGenerateUniform(curandGenerator_t generator, float *outputPtr, size_t num);

用于生成$(0.0,1.0]$的均匀分布float型随机数，0是取不到的。

curandStatus_t curandGenerateNormal(curandGenerator_t generator, float *outputPtr, size_t n, float mean, float stddev);

用于生成均值为mean，标准差为stddev的正态分布float型随机数。

curandStatus_t curandGenerateNormal(curandGenerator_t generator, float *outputPtr, size_t n,float mean, float stddev);

用于生成均值为mean，标准差为stddev的对数正态分布float型随机数。

curandStatus_t curandGeneratePoisson(curandGenerator_t generator, unsigned int *outputPtr, size_t n,double lambda);

用于生成泊松分布随机数

curandStatus_t curandGenerateUniformDouble(curandGenerator_t generator, double *outputPtr, size_t num);

用于生成双精度均匀分布随机数

curandStatus_t curandGenerateNormalDouble(curandGenerator_t generator,double *outputPtr, size_t n,double mean, double stddev);

用于生成双精度正态分布随机数

curandStatus_t curandGenerateLogNormalDouble(curandGenerator_t generator,double *outputPtr, size_t n, double mean, double stddev);

用于生成双精度对数正态分布随机数

Device API

要使用cuRAND Device API ，需要引用头文件<curand_kernel.h>，包括 pseudorandom generationfor and quasirandom generation。

init

__device__ void curand_init(unsigned long long seed, unsigned long long sequence,unsigned long long offset, curandState_t *state);
__device__ unsigned int curand (curandState_t *state);

在调用curand_init()后，curand()返回一系列伪随机数。如果state相同，则curand()生成相同的随机数序列。curand_init()函数使用给定的种子、序列号和序列内的偏移量设置调用者分配的初始状态。不同的种子保证产生不同的起始状态和不同的序列。相同的种子总是产生相同的状态和相同的序列。
对于最高质量的并行伪随机数生成，每个实验都应该分配一个唯一的种子。在一个实验中，每个计算线程都应该分配一个唯一的序列号。

Distributions

__device__ float curand_uniform(curandState_t *state);
__device__ float curand_normal(curandState_t *state); 
__device__ float curand_log_normal(curandState_t *state, float mean, float stddev);
__device__ unsigned int curand_poisson(curandState_t *state, double lambda);
__device__ double curand_uniform_double(curandState_t *state);
__device__ double curand_log_normal_double(curandState_t *state, double mean, double stddev);

Quasirandom Sequences

state初始化

__device__ void curand_init(unsigned int *direction_vectors, unsigned int offset,curandStateSobol32_t *state);
__device__ void curand_init(unsigned int *direction_vectors,unsigned int scramble_c, unsigned int offset,curandStateScrambledSobol32_t *state);
__device__ void curand_init(unsigned long long *direction_vectors, unsigned long long offset,curandStateSobol64_t *state);
__device__ void curand_init(unsigned long long *direction_vectors,unsigned long long scramble_c, unsigned long long offset,curandStateScrambledSobol64_t *state);

curand_init()用来初始化quasirandom number generator 的state，没有seed参数，有一个额外参数scramble_c ，which is the initial value of the scrambled sequence。

对于curandStateSobol32_t 类型和 curandStateScrambledSobol32_t 类型 ，the direction vectors are an array of 32 unsigned integer values.
对于curandStateSobol64_t 类型和curandStateScrambledSobol64_t 类型， the direction vectors are an array of 64 unsigned long long values.

生成随机数

__device__ unsigned int curand(curandStateSobol32_t *state);
__device__ float curand_uniform(curandStateSobol32_t *state);
__device__ float curand_normal(curandStateSobol32_t *state);
__device__ float curand_log_normal(curandStateSobol32_t *state,float mean, float stddev);
__device__ unsigned int curand_poisson(curandStateSobol32_t *state, double lambda);
__device__ double curand_uniform_double (curandStateSobol32_t *state);
__device__ double curand_normal_double (curandStateSobol32_t *state);
__device__ double curand_log_normal_double (curandStateSobol32_t *state,double mean,double stddev);

Skip-Ahead

__device__ void
skipahead(unsigned long long n, curandState_t *state);
__device__ void
skipahead(unsigned int n, curandStateSobol32_t *state);

Using this function is equivalent to calling curand() n times without using the return value, but it is much faster.

__device__ void skipahead_sequence(unsigned long long n, curandState_t *state);

This function is the equivalent of calling curand() $n\cdot 2^{67}$ times without using the return value and is much faster.