D | multinomial_impl.cu | 21 __global__ void CheckZeroKernel(const size_t distributions, const size_t categories, const T *input… in CheckZeroKernel() argument 23 …for (size_t pos = blockIdx.x * blockDim.x + threadIdx.x; pos < (distributions); pos += blockDim.x … in CheckZeroKernel() 32 void CheckZero(const size_t distributions, const size_t categories, const T *input, T *output, in CheckZero() argument 34 …CheckZeroKernel<<<GET_BLOCKS(distributions), GET_THREADS, 0, cuda_stream>>>(distributions, categor… in CheckZero() 54 __global__ void NormInputKernel(T *input, const size_t distributions, const size_t categories) { in NormInputKernel() argument 55 size_t size = distributions * categories; in NormInputKernel() 66 void NormInput(T *input, const size_t distributions, const size_t categories, cudaStream_t cuda_str… in NormInput() argument 67 int count1 = distributions * categories; in NormInput() 68 …NormInputKernel<<<GET_BLOCKS(count1), GET_THREADS, 0, cuda_stream>>>(input, distributions, categor… in NormInput() 92 size_t distributions, size_t categories) { in MultinomialKernel() argument [all …]
|