|
template<typename T > |
__global__ void | mma_sub1_kernel (T *__restrict__ xlow, T *__restrict__ xupp, const T *__restrict__ x, const T *__restrict__ xmin, const T *__restrict__ xmax, const T asyinit, const int n) |
|
template<typename T > |
__global__ void | mma_sub2_kernel (T *__restrict__ low, T *__restrict__ upp, const T *__restrict__ x, const T *__restrict__ xold1, const T *__restrict__ xold2, const T *__restrict__ xmin, const T *__restrict__ xmax, const T asydecr, const T asyincr, const int n) |
|
template<typename T > |
__global__ void | mma_sub3_kernel (const T *__restrict__ x, const T *__restrict__ df0dx, const T *__restrict__ dfdx, T *__restrict__ low, T *__restrict__ upp, const T *__restrict__ xmin, const T *__restrict__ xmax, T *__restrict__ alpha, T *__restrict__ beta, T *__restrict__ p0j, T *__restrict__ q0j, T *__restrict__ pij, T *__restrict__ qij, const int n, const int m) |
|
template<typename T > |
__global__ void | mma_sub4_kernel (const T *__restrict__ x, T *__restrict__ low, T *__restrict__ upp, T *__restrict__ pij, T *__restrict__ qij, T *__restrict__ temp, const int n, const int m) |
|
template<typename T > |
__global__ void | mma_max2_kernel (T *__restrict__ xsi, const T *__restrict__ x, T *__restrict__ alpha, const int n) |
|
template<typename T > |
__global__ void | relambda_kernel (T *__restrict__ temp, const T *__restrict__ x, const T *__restrict__ xupp, const T *__restrict__ xlow, const T *__restrict__ pij, const T *__restrict__ qij, const int n, const int m) |
|
template<typename T > |
__global__ void | sub2cons2_kernel (T *__restrict__ a, const T *__restrict__ b, const T *__restrict__ c, const T *__restrict__ d, const T e, const int n) |
|
template<typename T > |
__inline__ __device__ T | max_reduce_warp (T val) |
|
template<typename T > |
__global__ void | maxval_kernel (const T *__restrict__ a, T *temp, const int n) |
|
template<typename T > |
__global__ void | max_reduce_kernel (T *__restrict__ bufred, const int n) |
|
template<typename T > |
__global__ void | delx_kernel (T *__restrict__ delx, const T *__restrict__ x, const T *__restrict__ xlow, const T *__restrict__ xupp, const T *__restrict__ pij, const T *__restrict__ qij, const T *__restrict__ p0j, const T *__restrict__ q0j, const T *__restrict__ alpha, const T *__restrict__ beta, const T *__restrict__ lambda, const T epsi, const int n, const int m) |
|
template<typename T > |
__global__ void | GG_kernel (T *__restrict__ GG, const T *__restrict__ x, const T *__restrict__ xlow, const T *__restrict__ xupp, const T *__restrict__ pij, const T *__restrict__ qij, const int n, const int m) |
|
template<typename T > |
__global__ void | diagx_kernel (T *__restrict__ diagx, const T *__restrict__ x, const T *__restrict__ xsi, const T *__restrict__ xlow, const T *__restrict__ xupp, const T *__restrict__ p0j, const T *__restrict__ q0j, const T *__restrict__ pij, const T *__restrict__ qij, const T *alpha, const T *beta, const T *eta, const T *lambda, const int n, const int m) |
|
template<typename T > |
__inline__ __device__ T | reduce_warp (T val) |
|
template<typename T > |
__global__ void | mmareduce_kernel (T *__restrict__ bufred, const int n) |
|
template<typename T > |
__global__ void | mmasum_kernel (const T *__restrict__ a, T *__restrict__ buf_h, const int n, const int m, const int k) |
|
template<typename T > |
__global__ void | mmasumbb_kernel (const T *__restrict__ GG, const T *__restrict__ delx, const T *__restrict__ diagx, T *__restrict__ buf_h, const int n, const int m, const int k) |
|
template<typename T > |
__global__ void | mmasumAA_kernel (const T *__restrict__ GG, const T *__restrict__ diagx, T *__restrict__ buf_h, const int n, const int m, const int k0, const int k1) |
|
template<typename T > |
__global__ void | mma_copy_kernel (T *__restrict__ a, const T *__restrict__ b, const int n, const int m) |
|
template<typename T > |
__global__ void | AA_kernel (T *__restrict__ temp, const T *__restrict__ GG, const T *__restrict__ diagx, const int n, const int m) |
|
template<typename T > |
__global__ void | dx_kernel (T *__restrict__ dx, const T *__restrict__ delx, const T *__restrict__ diagx, const T *__restrict__ GG, const T *__restrict__ dlambda, const int n, const int m) |
|
template<typename T > |
__global__ void | dxsi_kernel (T *__restrict__ dxsi, const T *__restrict__ xsi, const T *__restrict__ dx, const T *__restrict__ x, const T *__restrict__ alpha, const T epsi, const int n) |
|
template<typename T > |
__global__ void | deta_kernel (T *__restrict__ deta, const T *__restrict__ eta, const T *__restrict__ dx, const T *__restrict__ x, const T *__restrict__ beta, const T epsi, const int n) |
|
template<typename T > |
__global__ void | RexCalculation_kernel (T *__restrict__ rex, const T *__restrict__ x, const T *__restrict__ xlow, const T *__restrict__ xupp, const T *__restrict__ pij, const T *__restrict__ p0j, const T *__restrict__ qij, const T *__restrict__ q0j, const T *__restrict__ lambda, const T *__restrict__ xsi, const T *__restrict__ eta, const int n, const int m) |
|
template<typename T > |
__global__ void | rey_calculation_kernel (T *__restrict__ rey, const T *__restrict__ c, const T *__restrict__ d, const T *__restrict__ y, const T *__restrict__ lambda, const T *__restrict__ mu, const int n) |
|
template<typename T > |
__global__ void | norm_kernel (const T *__restrict__ a, T *__restrict__ buf_h, const int n) |
|
template<typename T > |
__global__ void | sub2cons_kernel (T *__restrict__ a, const T *__restrict__ b, const T *__restrict__ c, const T d, const int n) |
|
template<typename T > |
__global__ void | dely_kernel (T *__restrict__ dely, const T *__restrict__ c, const T *__restrict__ d, const T *__restrict__ y, const T *__restrict__ lambda, const T epsi, const int n) |
|
template<typename T > |
__global__ void | maxval2_kernel (const T *__restrict__ a, const T *__restrict__ b, T *__restrict__ temp, const T cons, const int n) |
|
template<typename T > |
__global__ void | maxval3_kernel (const T *__restrict__ a, const T *__restrict__ b, const T *__restrict__ c, T *__restrict__ temp, const T cons, const int n) |
|
template<typename T > |
__global__ void | kkt_rex_kernel (T *__restrict__ rex, const T *__restrict__ df0dx, const T *__restrict__ dfdx, const T *__restrict__ xsi, const T *__restrict__ eta, const T *__restrict__ lambda, const int n, const int m) |
|
template<typename T > |
__global__ void | maxcons_kernel (T *__restrict__ a, const T b, const T c, const T *__restrict__ d, const int n) |
|
template<typename T > |
__global__ void | glsum_kernel (const T *a, T *buf_h, const int n) |
|
template<typename T > |
__global__ void | glsc2_kernel (const T *a, const T *b, T *buf_h, const int n) |
|
template<typename T > |
__global__ void | add2inv2_kernel (T *__restrict__ a, const T *__restrict__ b, const T c, const int n) |
|
template<typename T > |
__global__ void | max2_kernel (T *__restrict__ a, const T b, const T *__restrict__ c, const T d, const int n) |
|
template<typename T > |
__global__ void | updatebb_kernel (T *__restrict__ bb, const T *__restrict__ dellambda, const T *__restrict__ dely, const T *__restrict__ d, const T *__restrict__ mu, const T *__restrict__ y, const T delz, const int m) |
|
template<typename T > |
__global__ void | updateAA_kernel (T *__restrict__ AA, const T *__restrict__ globaltmp_mm, const T *__restrict__ s, const T *__restrict__ lambda, const T *__restrict__ d, const T *__restrict__ mu, const T *__restrict__ y, const T *__restrict__ a, const T zeta, const T z, const int m) |
|
template<typename T > |
__global__ void | dy_kernel (T *__restrict__ dy, const T *__restrict__ dely, const T *__restrict__ dlambda, const T *__restrict__ d, const T *__restrict__ mu, const T *__restrict__ y, const int n) |
|