|
template<int SrcLaneIdx> |
__device__ void | ck::dpp8::inline_v_dot2c_dpp8_instr (const half2_t &a, const half2_t &b, float &c) |
|
template<> |
__device__ void | ck::dpp8::inline_v_dot2c_dpp8_instr< 0 > (const half2_t &a, const half2_t &b, float &c) |
|
template<> |
__device__ void | ck::dpp8::inline_v_dot2c_dpp8_instr< 1 > (const half2_t &a, const half2_t &b, float &c) |
|
template<> |
__device__ void | ck::dpp8::inline_v_dot2c_dpp8_instr< 2 > (const half2_t &a, const half2_t &b, float &c) |
|
template<> |
__device__ void | ck::dpp8::inline_v_dot2c_dpp8_instr< 3 > (const half2_t &a, const half2_t &b, float &c) |
|
template<> |
__device__ void | ck::dpp8::inline_v_dot2c_dpp8_instr< 4 > (const half2_t &a, const half2_t &b, float &c) |
|
template<> |
__device__ void | ck::dpp8::inline_v_dot2c_dpp8_instr< 5 > (const half2_t &a, const half2_t &b, float &c) |
|
template<> |
__device__ void | ck::dpp8::inline_v_dot2c_dpp8_instr< 6 > (const half2_t &a, const half2_t &b, float &c) |
|
template<> |
__device__ void | ck::dpp8::inline_v_dot2c_dpp8_instr< 7 > (const half2_t &a, const half2_t &b, float &c) |
|
template<int SrcLaneIdx, bool ShareA> |
__device__ void | ck::dpp8::inline_v_dot2c_dpp8 (const half2_t &a, const half2_t &b, float &c) |
|
template<int SrcLaneIdx> |
constexpr int | ck::dpp8::get_dpp_sel_mask_broadcast () |
|
template<int SrcLaneIdx> |
__device__ void | ck::dpp8::intrinsic_fdot2_impl (const half2_t &a, const half2_t &b, float &c) |
|
template<int SrcLaneIdx, bool ShareA> |
__device__ void | ck::dpp8::intrinsic_fdot2 (const half2_t &a, const half2_t &b, float &c) |
|
template<typename TA , typename TB , typename TC , int SrcLaneIdx, bool ShareA> |
__device__ void | ck::dpp8::inner_product_dpp (const TA &a, const TB &b, TC &c) |
|