Interface Description

Handle class interfaces

krl_create_distance_handle

Initializes and constructs a KRLDistanceHandle instance for dense distance computation.

krl_create_reorder_handle

Initializes and constructs a KRLDistanceHandle instance for reordering computation.

krl_clean_distance_handle

Destroys the KRLDistanceHandle instance and releases memory space.

krl_create_LUT8b_handle

Initializes and constructs a KRLLUT8bHandle instance for accumulation based on 8-bit table lookup.

krl_clean_LUT8b_handle

Destroys the KRLLUT8bHandle instance and releases memory space.

krl_get_idx_pointer

Obtains the IDs of to-be-computed base vectors stored in the KRLLUT8bHandle instance.

krl_get_dist_pointer

Obtains the start address of the distance array stored in the KRLLUT8bHandle instance. The address contains a random number before a table lookup-based accumulation operator is called, and holds the computed distance value after the operator is called.

Distance computation interfaces

krl_L2sqr

Performs one-to-one Euclidean distance computation on vectors represented in float precision.

krl_L2sqr_f16f32

Performs one-to-one Euclidean distance computation on vectors represented in fp16 precision.

krl_L2sqr_u8u32

Performs one-to-one Euclidean distance computation on vectors represented as 8-bit unsigned integers (uint8).

krl_ipdis

Performs one-to-one distance computation using inner product on vectors represented in float precision.

krl_negative_ipdis_f16f32

Performs one-to-one distance computation using inner product on vectors represented in fp16 precision.

krl_negative_ipdis_s8s32

Performs one-to-one distance computation using inner product on vectors represented as 8-bit unsigned integers (uint8).

krl_L2sqr_by_idx

Performs one-to-many Euclidean distance computation on vectors represented in float precision.

krl_L2sqr_by_idx_f16f32

Performs one-to-many Euclidean distance computation on vectors represented in fp16 precision.

krl_L2sqr_by_idx_u8f32

Performs one-to-many Euclidean distance computation on vectors represented as 8-bit unsigned integers (uint8).

krl_inner_product_by_idx

Performs one-to-many distance computation using inner product on vectors represented in float precision.

krl_inner_product_by_idx_f16f32

Performs one-to-many distance computation using inner product on vectors represented in fp16 precision.

krl_negative_inner_product_by_idx_f16f32

Performs one-to-many distance computation using inner product on vectors represented in fp16 precision, and negates the computation results.

krl_inner_product_by_idx_s8f32

Performs one-to-many distance computation using inner product on vectors represented as 8-bit unsigned integers (uint8).

krl_L2sqr_ny

Performs one-to-many Euclidean distance computation on vectors represented in float precision.

krl_L2sqr_ny_f16f32

Performs one-to-many Euclidean distance computation on vectors represented in fp16 precision.

krl_L2sqr_ny_u8f32

Performs one-to-many Euclidean distance computation on vectors represented as 8-bit unsigned integers (uint8).

krl_L2sqr_ny_with_handle

Performs one-to-many Euclidean distance computation on vectors represented in float precision. The base vectors and dimensions are stored in the handle.

krl_inner_product_ny

Performs one-to-many distance computation using inner product on vectors represented in float precision.

krl_inner_product_ny_f16f32

Performs one-to-many distance computation using inner product on vectors represented in fp16 precision.

krl_inner_product_ny_s8f32

Performs one-to-many distance computation using inner product on vectors represented as 8-bit unsigned integers (uint8).

krl_inner_product_ny_with_handle

Performs one-to-many distance computation using inner product on vectors represented in float precision. The base vectors and dimensions are stored in the handle.

8-bit table lookup-based accumulation interfaces

krl_table_lookup_8b_f32

Obtains distance values from the float-type LUT using 8-bit indexes, accumulates the distance values, adds dis0 to the result after accumulation, and stores the final result in distance.

krl_table_lookup_8b_f32_by_idx

Obtains distance values from the float-type LUT using 8-bit indexes, accumulates the distance values, adds dis0 to the result after accumulation, and stores the final result in distance. Only the base vectors whose IDs are in the idx array are used for computation.

krl_table_lookup_8b_f32_with_handle

Obtains distance values from the float-type LUT using 8-bit indexes, accumulates the distance values, adds dis0 to the result after accumulation, and stores the final result in distance. The idx and distance arrays are contained in the KRLLUT8bHandle instance. Only the base vectors whose IDs are in the idx array are used for computation.

4-bit table lookup-based accumulation interfaces

krl_fast_table_lookup_step

An operator designed for 4-bit table lookup, accumulation, filtering, and compression. It is used for batch processing of float-type query vectors. This operator is used to compute distances between a maximum of 16 query vectors and 32 base vectors. After the distance computation, each result is compared against a threshold. If a distance value is less than the threshold, the bit in lt_mask of the corresponding base vector is set to 1. Otherwise, it is set to 0.

krl_L2_table_lookup_fast_scan_bs64

An operator designed for 4-bit table lookup, accumulation, filtering, and compression. It is used for processing float-type query vectors individually. Computes Euclidean distances between a query vector and 64 base vectors. Based on the distance comparison rule, if a distance value meets the comparison rule, the bit in lt_mask of the corresponding base vector is set to 1. Otherwise, it is set to 0.

krl_IP_table_lookup_fast_scan_bs64

An operator designed for 4-bit table lookup, accumulation, filtering, and compression. It is used for processing float-type query vectors individually. Computes distances between a query vector and 64 base vectors using inner product. Based on the distance comparison rule, if a distance value meets the comparison rule, the bit in lt_mask of the corresponding base vector is set to 1. Otherwise, it is set to 0.

krl_L2_table_lookup_fast_scan_bs96

An operator designed for 4-bit table lookup, accumulation, filtering, and compression. It is used for processing float-type query vectors individually. Computes Euclidean distances between a query vector and 96 base vectors. Based on the distance comparison rule, if a distance value meets the comparison rule, the bit in lt_mask of the corresponding base vector is set to 1. Otherwise, it is set to 0.

krl_IP_table_lookup_fast_scan_bs96

An operator designed for 4-bit table lookup, accumulation, filtering, and compression. It is used for processing float-type query vectors individually. Computes distances between a query vector and 96 base vectors using inner product. Based on the distance comparison rule, if a distance value meets the comparison rule, the bit in lt_mask of the corresponding base vector is set to 1. Otherwise, it is set to 0.

krl_table_lookup_4b_f16

An operator designed for 4-bit table lookup, accumulation, filtering, and compression. It is used for processing fp16 query vectors individually. Computes distances between a query vector and multiple base vectors using inner product. The initial distance is dis_f16. This interface does not perform filtering or compression (that is, comparison with the threshold).

krl_pack_codes_4b

An operator designed for 4-bit table lookup, accumulation, filtering, and compression. It is used for processing fp16 query vectors individually. Computes distances between a query vector and multiple base vectors using inner product. The initial distance is dis_f16. This interface does not perform filtering or compression (that is, comparison with the threshold).

Reranking interfaces

krl_reorder_2_vector

Computes high-accuracy distances between a single query vector and multiple non-contiguous base vectors, and returns the results sorted by distance.

krl_reorder_2_vector_continuous

Computes high-accuracy distances between a single query vector and multiple contiguous base vectors, and returns the results sorted by distance.

Store/Load interfaces

krl_store_LUT8Handle

Stores the handle of an 8-bit lookup table into a file.

krl_build_LUT8Handle_fromfile

Reads data of an 8-bit lookup table from a file and rebuilds a handle.

krl_store_distanceHandle

Stores the handle for distance computation into a file.

krl_build_distanceHandle_fromfile

Reads data of a distance computation handle from a file and rebuilds the handle.