recip_tile

template<ReciprocalDestAcc dest_acc = DST_ACCUM_MODE ? ReciprocalDestAcc::FP32 : ReciprocalDestAcc::BF16, ReciprocalApproxMode approximation = ReciprocalApproxMode::Default>
void ckernel::recip_tile_init()

Please refer to documentation for any_init. DEST precision defaults to DST_ACCUM_MODE. Explicit overrides must use ReciprocalDestAcc, e.g. recip_tile_init<ReciprocalDestAcc::FP32>(); use the same mode for recip_tile. approximation defaults to the kernel’s APPROX setting; override it for callers that need a more accurate reciprocal without changing the approximation of other operations in the kernel. The former legacy_compat Boolean template argument is no longer accepted.

template<ReciprocalDestAcc dest_acc = DST_ACCUM_MODE ? ReciprocalDestAcc::FP32 : ReciprocalDestAcc::BF16, ReciprocalApproxMode approximation = ReciprocalApproxMode::Default>
void ckernel::recip_tile(uint32_t idst, VectorMode vector_mode = VectorMode::RC)

Performs element-wise computation of the reciprocal on each element of a tile in DST register at index tile_index. The DST register buffer must be in acquired state via acquire_dst call. This call is blocking and is only available on the compute engine. Only works for Float32, Float16_b, Bfp8_b data formats for full accuracy. DEST precision defaults to DST_ACCUM_MODE and approximation defaults to APPROX. Explicit destination and approximation overrides must match recip_tile_init. Legacy Boolean template arguments are no longer accepted.

Return value: None

Argument Description

idst The index of the tile in DST register buffer to perform the computation on

vector_mode Specifies the vector mode for computation (e.g., Row, Column). (default: VectorMode::RC)