This project compares the run-time speed of the look-up table pre-computation procedure in Product Quantization and its several variants. Intel's MKL library (BLAS & SparseBLAS) is used for acceleration; a built-in implementation of matrix multipliation is also included for comparison.

We compare the following PQ-like methods:

• PQ (Product Quantization).
• OPQ (Optimized Product Quantization).
• CQ (Composite Quantization).
• SCQ (Sparse Composite Quantization).
• TQ (Tree Quantization).

Another two PQ-like methods, CKM (Cartesian K-means) and AQ (Additive Quantization), are not included, since their pre-computation procedure is the same as OPQ and CQ, respectively.

MKL library provides acceleration for matrix multiplication operations (both dense and sparse matrices). PQ, OPQ, and CQ can be accelerated by MKL's BLAS routine, and SCQ can be sped-up be MKL's SparseBLAS routine. TQ is a little bit special. Due to its structured sparsity, MKL's BLAS routine can still be adopted for better speed-up (compared against SparseBLAS).

We report the pre-computation time (in millisecond) of each query (average over 10k queries). Two input settings are evaluated, "single query" and "multiple queries" (50 queries in a batch). The time consumption is measured on a workstation, equiped with Intel Xeon E5-2640 CPU @ 2.50GHz (single-threaded).

Note: for query size X x Y, X is the number of feature dimensions, and Y is the number of queries in each input batch. The first time is for the built-in implementation, and the second time is for the MKL-based implementation. We choose SCQ's sparsity so that its complexity is the same as that of PQ.

• MKL's BLAS routine is much more powerful than the built-in implementation, and its advantage grows as the input batch size increases.
• MKL's SparseBLAS routine is also slightly better than the built-in version, but less efficient than the BLAS routine (comparing PQ and SCQ).
• TQ is more efficient than SCQ due to its structured sparsity.