|
|
template<typename T > |
| __global__ void | graphite::ops::axpy_kernel (size_t n, T *z, const T a, const T *x, T *y) |
| |
|
template<typename T > |
| void | graphite::ops::axpy_async (cudaStream_t stream, size_t n, T *z, const T a, const T *x, T *y) |
| |
|
template<typename T > |
| __global__ void | graphite::ops::damping_kernel (size_t n, T *z, const T damping_factor, const bool use_identity, const T *diag, const T *x) |
| |
|
template<typename T > |
| void | graphite::ops::damp_by_factor_async (cudaStream_t stream, size_t n, T *z, const T damping_factor, const bool use_identity, const T *diag, const T *x) |
| |
|
template<typename T > |
| __global__ void | graphite::ops::clamp_kernel (size_t n, T min_val, T max_val, T *x) |
| |
|
template<typename T > |
| void | graphite::ops::clamp_async (cudaStream_t stream, size_t n, T min_val, T max_val, T *x) |
| |
|
template<typename T > |
| __global__ void | graphite::ops::rescale_vec_kernel (size_t n, T *out, const T scale, const T *x) |
| |
|
template<typename T > |
| void | graphite::ops::rescale_vec_async (cudaStream_t stream, size_t n, T *out, const T scale, const T *x) |
| |
|
template<typename T > |
| __global__ void | graphite::ops::compute_adam_step (const size_t n, T *gradient, T *step, T *m, T *v, const T lr, const T beta1, const T beta2, const T epsilon, const size_t t) |
| |
|
template<typename T > |
| void | graphite::ops::compute_adam_step_async (cudaStream_t stream, const size_t n, T *gradient, T *step, T *m, T *v, const T lr, const T beta1, const T beta2, const T epsilon, const size_t t) |
| |