Quicker ADC is an implementation of fast distance computation techniques for nearest neighbor search in large-scale databases of high-dimensional vectors. It is based upon Quick ADC but provides (i) AVX512 support, (ii) new optimized product quantizers, (iii) full integration with FAISS from Facebook AI Research. It is thus compatible with Inverted Multi-Indexes and HNSW-based IVFs.
Quicker ADC achieves excellent performance outperforming polysemous codes in numerous configurations. We evaluated its performance for exhaustive search in the SIFT1M dataset with both 64-bit and 128-bit codes.
We also evaluated its performance for index-based (i.e., non-exhaustive) search on the SIFT1000M dataset and Deep1B dataset. For this evaluation, we used three different type of indexes (standard IVF, Inverted Multi-Index, and HNSW-based IVFs). For each combination of a product quantizer implementation with an index, we plot the curve that gives the achievable recall (R@1 and R@100) given a query time budget. Quicker ADC is particularly efficient for low query times, R@100 and Deep1B dataset.
Results for SIFT1000M on AWS m5 instances
Results for Deep1B on AWS m5 instances
The algorithms implemented and complete commented results are described in
Quicker ADC : Unlocking the hidden potential of Product Quantization with SIMD
Fabien André, Anne-Marie Kermarrec and Nicolas Le Scouarnec
arXiv 1812.09162
December 2018
https://arxiv.org/abs/1812.09162
For more details on FAISS see the main repository and the faiss's README.md.
Modifications related the introduction of Quicker ADC (Vectorized Product Quantization) are licensed by Thomson Licensing, SAS under the Clear BSD license found in the LICENSE file in the root directory of this source tree.
The rest of the source code is licensed by Facebook under the BSD license found in the LICENSE file in the root directory of this source tree
If you use or compare to this work, please cite
Quicker ADC : Unlocking the hidden potential of Product Quantization with SIMD
Fabien André, Anne-Marie Kermarrec and Nicolas Le Scouarnec
arXiv 1812.09162
December 2018
https://arxiv.org/abs/1812.09162
Accelerated Nearest Neighbor Search with Quick ADC
Fabien André, Anne-Marie Kermarrec and Nicolas Le Scouarnec
ICMR '17 Proceedings of the 2017 ACM on International Conference on Multimedia Retrieval
June 2017
https://doi.org/10.1145/3078971.3078992
./configure
make
make demo/demo_ivfvpq_indexing
The configure script detects the architecture of the CPU and enable SIMD instruction sets accordingly.
make py
PYTHONPATH=. python benchs/bench_vpq_sift1m.py
make install
make -C python install
More advanced compilation and installation instructions are detailled into INSTALL.md. Note that Quicker ADC is not pushed to Conda and must be compiled from source.
Note that only L2 distance is supported in Quicker ADC's implementation.
#include <faiss/VPQ_Impl.h>
...
// Declare and IVF based on Quicker ADC (similar to IndexIVFPQ)
// The suffix describes the variant desired, as specified in VPQ_Impl.h
// In this case it is (16x{4,4} based on SSE)
faiss::IndexIVFVPQ_2v4 index (&coarse_quantizer, d, ncentroids);
// Declare and structure for exhaustive search based on Quicker ADC (similar to IndexPQ)
// The suffix describes the variant desired, as specified in VPQ_Impl.h
// In this case it is (16x{4,4} based on SSE)
faiss::IndexIVFVPQ_2v4 index (d);
// You can use them similarly to IndexPQ and IndexIVFPQ. Note that they do not support OnDiskInvertedList.
Check VPQ_Impl.h for a complete list of available quantizers, or if you want to add a new quantizer.
The Quicker ADC implementation is mostly contained into IndexVPQ.h, IndexIVFVPQ.h and VecProductQuantizer.h. The first contains a flat-database relying on Quicker ADC and the second is used for database with an IVF and relying on Quicker ADC. The last contains the actual Product Quantizer implementation and is referenced by the two first. The implementation relies on templates so that numerous variants share the same code that is specialized at compile time thanks to compiler knowledge template parameter values. This increase compile time but allow numerous variants to be implemented in a concise and maintanable way. To use a particular variant, it should be declared in VPQ_Impl.h: this enables auto-tuning (index construction from CLI) and serialization specific routines. If you need to have a switch/case that depends on Index type, you should use EXPAND_VPQTEST macro; use it as depicted in this example for serialization
template<class T_IVPQ>
bool write_indexVec(const AbstractIndexVPQ* idx, IOWriter *f){
if(const T_IVPQ* ivpq = dynamic_cast<const T_IVPQ*>(idx)){
std::string a = fourcc_vpq(ivpq);
uint32_t h = fourcc(a.c_str());
WRITE1 (h);
write_index_header (ivpq, f);
write_VecProductQuantizer<typename T_IVPQ::VPQ_t>(&ivpq->pq, f);
WRITEVECTOR (ivpq->codes);
return true;
}else{
return false;
}
}
...
} else if(const AbstractIndexVPQ * ivpq =
dynamic_cast<const AbstractIndexVPQ *> (idx)) {
#define WRITE_IVPQ(t) write_indexVec<IndexVPQ<t>>(ivpq,f)
bool success= EXPAND_VPQTEST(WRITE_IVPQ);
FAISS_THROW_IF_NOT_MSG(success,"Unsupported Serialization for Vectorized PQ Index type");
}Alternative usage examples are given in AutoTune.cpp; in that case note that macro EXPAND_VPQTEST is inside the if test and that the action block is empty (action is performed within the if, leveraging lazy evaluation). It also shows the usage of the string-based representation of the Quantizer (e.g., VPQ_12x6.6.4_AVX512) declared within VPQ_Impl.h.
Note that each variant comes in multiple form useful for evaluation and debugging:
- non-vectorized non-quantized
- non-vectorized quantized but not relying on quantized lookup tables
- non-vectorized quantized and relying on quantized lookup tables
- vectorized using SSE
- vectorized using AVX2
- vectorized using AVX512
- and all for multiple types of quantization (using either signed or unsigned arithmetic)
For example for 16x{4,4} product quantizer, the following classes are available.
// Unsigned arithmetic
VecProductQuantizer_NoVecNoQuant<16,2,4,4,0,0,uint8_t,__m128i,__m128i>
VecProductQuantizer_NoVec<16,2,4,4,0,0,uint8_t,__m128i,__m128i>
VecProductQuantizer_NoVecTable<16,2,4,4,0,0,uint8_t,__m128i,__m128i>
VecProductQuantizer_NoVecTable<16,2,4,4,0,0,uint8_t,__m128i,__m256i> // With AVX2 compatible distance table layout
VecProductQuantizer_4_SSE128_unsigned<16>
VecProductQuantizer_4_AVX512_unsigned<16>
VecProductQuantizer_4_AVX256_unsigned<16>
// Signed arithmetic
VecProductQuantizer_NoVecNoQuant<16,2,4,4,0,0,int8_t,__m128i,__m128i>
VecProductQuantizer_NoVec<16,2,4,4,0,0,int8_t,__m128i,__m128i>
VecProductQuantizer_NoVecTable<16,2,4,4,0,0,int8_t,__m128i,__m128i>
VecProductQuantizer_NoVecTable<16,2,4,4,0,0,int8_t,__m128i,__m256i> // With AVX2 compatible distance table layout
VecProductQuantizer_4_SSE128<16>
VecProductQuantizer_4_AVX512<16>
VecProductQuantizer_4_AVX256<16>Check VPQ_Impl.h for a complete list of available quantizers (mx{6,6,4}, mx{6,5,5}, mx{5,5,5}, mx{8,8}, mx{8}), or if you want to add a new quantizer.
The technical report Quicker ADC : Unlocking the hidden potential of Product Quantization with SIMD relies on code commit d11c5af. To reproduce results, you can use the following commands:
export PYTHON_PATH=.
## For exhaustive search (results are displayed on stdout) on SIFT1M (dataset must be in sift1M folder)
pyton benchs/bench_vpq_sift1000m.py
## For non-exhaustive search (commands outputs result files in /tmp for R@1 and R@100)
# For SIFT1000M (dataset must be in bigann folder)
python benchs/bench_vpq_1bn.py SIFT1000M IMI2x12,VPQ_12x6.6.4_AVX512 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IMI2x12,VPQ_24x6.6.4_AVX512 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IMI2x12,VPQ_32x4.4_AVX2 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IMI2x12,VPQ_16x4.4_AVX2 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IMI2x12,VPQ_8x8.8_AVX512 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IMI2x12,VPQ_16x8.8_AVX512 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IMI2x12,PQ8 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IMI2x12,PQ16 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IMI2x12,PQ8np autotuneNP
python benchs/bench_vpq_1bn.py SIFT1000M IMI2x12,PQ16np autotuneNP
python benchs/bench_vpq_1bn.py SIFT1000M IVF262144_HNSW32,VPQ_12x6.6.4_AVX512 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IVF262144_HNSW32,VPQ_24x6.6.4_AVX512 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IVF262144_HNSW32,VPQ_32x4.4_AVX2 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IVF262144_HNSW32,VPQ_16x4.4_AVX2 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IVF262144_HNSW32,VPQ_8x8.8_AVX512 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IVF262144_HNSW32,VPQ_16x8.8_AVX512 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IVF262144_HNSW32,PQ8 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IVF262144_HNSW32,PQ16 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IVF262144_HNSW32,PQ8np autotuneNP
python benchs/bench_vpq_1bn.py SIFT1000M IVF262144_HNSW32,PQ16np autotuneNP
python benchs/bench_vpq_1bn.py SIFT1000M IVF65536,VPQ_12x6.6.4_AVX512 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IVF65536,VPQ_24x6.6.4_AVX512 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IVF65536,VPQ_32x4.4_AVX2 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IVF65536,VPQ_16x4.4_AVX2 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IVF65536,VPQ_8x8.8_AVX512 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IVF65536,VPQ_16x8.8_AVX512 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IVF65536,PQ8 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IVF65536,PQ16 autotune
python benchs/bench_vpq_1bn.py SIFT1000M IVF65536,PQ8np autotuneNP
python benchs/bench_vpq_1bn.py SIFT1000M IVF65536,PQ16np autotuneNP
# For Deep1B (dataset must be in deep1b folder)
python benchs/bench_vpq_1bn.py Deep1B IMI2x12,VPQ_12x6.6.4_AVX512 autotune
python benchs/bench_vpq_1bn.py Deep1B IMI2x12,VPQ_24x6.6.4_AVX512 autotune
python benchs/bench_vpq_1bn.py Deep1B IMI2x12,VPQ_32x4.4_AVX2 autotune
python benchs/bench_vpq_1bn.py Deep1B IMI2x12,VPQ_16x4.4_AVX2 autotune
python benchs/bench_vpq_1bn.py Deep1B IMI2x12,VPQ_8x8.8_AVX512 autotune
python benchs/bench_vpq_1bn.py Deep1B IMI2x12,VPQ_16x8.8_AVX512 autotune
python benchs/bench_vpq_1bn.py Deep1B IMI2x12,PQ8 autotune
python benchs/bench_vpq_1bn.py Deep1B IMI2x12,PQ16 autotune
python benchs/bench_vpq_1bn.py Deep1B IMI2x12,PQ8np autotuneNP
python benchs/bench_vpq_1bn.py Deep1B IMI2x12,PQ16np autotuneNP
python benchs/bench_vpq_1bn.py Deep1B IVF262144_HNSW32,VPQ_12x6.6.4_AVX512 autotune
python benchs/bench_vpq_1bn.py Deep1B IVF262144_HNSW32,VPQ_24x6.6.4_AVX512 autotune
python benchs/bench_vpq_1bn.py Deep1B IVF262144_HNSW32,VPQ_32x4.4_AVX2 autotune
python benchs/bench_vpq_1bn.py Deep1B IVF262144_HNSW32,VPQ_16x4.4_AVX2 autotune
python benchs/bench_vpq_1bn.py Deep1B IVF262144_HNSW32,VPQ_8x8.8_AVX512 autotune
python benchs/bench_vpq_1bn.py Deep1B IVF262144_HNSW32,VPQ_16x8.8_AVX512 autotune
python benchs/bench_vpq_1bn.py Deep1B IVF262144_HNSW32,PQ8 autotune
python benchs/bench_vpq_1bn.py Deep1B IVF262144_HNSW32,PQ16 autotune
python benchs/bench_vpq_1bn.py Deep1B IVF262144_HNSW32,PQ8np autotuneNP
python benchs/bench_vpq_1bn.py Deep1B IVF262144_HNSW32,PQ16np autotuneNP
python benchs/bench_vpq_1bn.py Deep1B IVF65536,VPQ_12x6.6.4_AVX512 autotune
python benchs/bench_vpq_1bn.py Deep1B IVF65536,VPQ_24x6.6.4_AVX512 autotune
python benchs/bench_vpq_1bn.py Deep1B IVF65536,VPQ_32x4.4_AVX2 autotune
python benchs/bench_vpq_1bn.py Deep1B IVF65536,VPQ_16x4.4_AVX2 autotune
python benchs/bench_vpq_1bn.py Deep1B IVF65536,VPQ_8x8.8_AVX512 autotune
python benchs/bench_vpq_1bn.py Deep1B IVF65536,VPQ_16x8.8_AVX512 autotune
python benchs/bench_vpq_1bn.py Deep1B IVF65536,PQ8 autotune
python benchs/bench_vpq_1bn.py Deep1B IVF65536,PQ16 autotune
python benchs/bench_vpq_1bn.py Deep1B IVF65536,PQ8np autotuneNP
python benchs/bench_vpq_1bn.py Deep1B IVF65536,PQ16np autotuneNP