Add RVV distance kernels for QT_8bit_direct_signed

[Xlawy]

RVV QT_8bit_direct_signed distances currently fall through to the generic DCTemplate. This adds RVV specializations for L2 and inner product, covering both floating-point query-to-code and code-to-code distances.

The kernel reconstructs code[i] - 128 in float, accumulates in vector registers, and performs a single reduction after the loop. Tail-undisturbed accumulation preserves earlier lanes when the final iteration is short. Query values remain floating point, including fractional queries — the query is not truncated to the integer byte domain.

Relationship to other RVV scalar-quantizer PRs

This is an independent change based on main. It overlaps the direct-signed distance area of #5539, and #5535 also works on RVV scalar-quantizer primitives. #5539 also specializes Quantizer8bitDirectSigned, so the difference is worth stating explicitly: #5539's set_query truncates the query via int(x[i]) + 128, which changes results for fractional queries. This PR keeps the floating-point query path, matching the current RVV baseline (which falls back to the SIMDLevel::NONE DCTemplate, whose set_query stores the const float* directly) and SimilarityL2<SIMDLevel::NONE>::add_component(float).

For the vectors used in this PR's test, the two approaches disagree:

Case	This PR (float)	Integer-truncated	Equal
d=1, code=0, q=-127.75, L2	0.0625	1	no
d=1, code=128, q=-0.5, L2	0.25	0	no

On x86, sq-dispatch.h routes this qtype to DistanceComputerByteSigned<AVX512_SPR> (integer domain, query re-biased by +128) only when d % 64 == 0, and otherwise falls back to DCTemplate — so both semantics are reachable upstream depending on dimension. This PR keeps the DCTemplate entry point, i.e. the floating-point / all-dimensions path.

The benchmark below compares with the stated official baseline and does not compare against either PR's head.

Performance

Measured on a native SG2044 RISC-V host (VLEN=128), GCC 15.1, Release -O3, dynamic dispatch (FAISS_OPT_LEVEL=dd), rv64gcv_zvfhmin/lp64d, one thread pinned to CPU 2. The baseline is official commit 80a16564f86530dbf0bfaf96c2b71feffeb5093f; the candidate is that same baseline plus only this kernel change. These measurements were not collected on the newer PR base 2ed4c106e9fb9686e7727e5daf8ad6ad1e164109. The affected scalar-quantizer source files are unchanged between those bases, and the submitted kernel differs from the measured one only in comments/formatting.

Public path	Dimension	Baseline ns/element	Candidate ns/element	Paired speedup	95% interval
IP query-to-code	16	3.341	2.162	1.545x	[1.537, 1.550]
IP query-to-code	32	2.963	1.236	2.400x	[2.145, 2.406]
IP query-to-code	128	2.745	0.543	5.058x	[4.887, 5.084]
IP query-to-code	768	2.628	0.360	7.261x	[7.213, 7.371]
IP code-to-code	16	3.524	2.419	1.457x	[1.456, 1.458]
IP code-to-code	32	3.060	1.462	2.095x	[2.092, 2.096]
IP code-to-code	128	2.769	0.749	3.700x	[3.693, 3.701]
IP code-to-code	768	2.635	0.563	4.701x	[4.685, 4.706]
L2 query-to-code	16	3.410	2.248	1.515x	[1.512, 1.521]
L2 query-to-code	32	3.002	1.321	2.269x	[2.251, 2.279]
L2 query-to-code	128	2.754	0.611	4.493x	[4.457, 4.512]
L2 query-to-code	768	2.630	0.428	6.187x	[6.103, 6.230]
L2 code-to-code	16	3.592	2.461	1.460x	[1.459, 1.462]
L2 code-to-code	32	3.100	1.499	2.069x	[2.066, 2.070]
L2 code-to-code	128	2.778	0.777	3.573x	[3.568, 3.581]
L2 code-to-code	768	2.635	0.586	4.497x	[4.470, 4.529]

Geometric mean of the dimension-specific speedups at d=32/128/768: IP query-to-code 4.450x, IP code-to-code 3.315x, L2 query-to-code 3.981x, L2 code-to-code 3.215x.

There are three consecutive sessions, each with four alternating ABBA/BAAB blocks per dimension/path: 192 paired blocks for this candidate. A is the baseline and B is the candidate. Each call is calibrated to at least 0.1 s (the shortest formal call in the full campaign was 0.192 s), with three warm-up batches and n = max(32, floor(32768/d)). Each block uses the ratio of the two-call geometric mean times. Reported speedup is the median of the three session medians; intervals use 5,000 hierarchical bootstrap resamples, resampling sessions and then blocks within each ABBA/BAAB order stratum. The timing columns are separate medians, so their quotient need not equal the paired speedup. All 192 paired blocks favored this candidate.

The timed operations are public SQDistanceComputer::query_to_code and symmetric_dis calls over randomly generated codes (fixed seed 718), with fractional floating-point queries.

These results describe public encoding/distance throughput on one non-exclusive host, not end-to-end ANN search speedup. The intervals describe these sessions only, with no multiple-comparison correction; they do not establish portability across machines or vector lengths.

Validation

• Native public-path oracle: 13,152 distance/index checks; L2/IP, query-to-code, code-to-code, top-k, fractional queries, NaN/Inf, and tails. Float results are checked against a double reference with a dimension-dependent error bound; integer code-to-code sums are required to be exact when the sum of absolute terms is at most 2^24.
• Original benchmark baseline full C++ suite: 273 passed, 7 skipped, 0 failed.
• Submission version on 2ed4c106e9fb9686e7727e5daf8ad6ad1e164109: all three independent candidate builds succeeded; this PR's focused C++ suite (NONE: 11 passed, 5 skipped, 0 failed; RISCV_RVV: 12 passed, 4 skipped, 0 failed) and the public-path oracle above pass on native RISC-V. The newly added RVV regression test is executed and passed in the RVV run, and skipped when RVV is disabled in the NONE run.
• All touched C++ files pass clang-format 21.1.8; git diff --check clean.

The added regression test ScalarQuantizer.RVVDirectSignedDistances covers byte boundaries (0, 127, 128, 255), fractional queries, both metrics, symmetric distances in both directions, and dimensions around vector tails (1, 15, 16, 17, 31, 32, 33, 65). It skips when RVV is unavailable.

Notes

• Only faiss/impl/scalar_quantizer/sq-rvv.cpp and tests/test_scalar_quantizer.cpp are touched.
• The implementation is VLEN-agnostic: it uses e8m1 loads and an e32m4 accumulator with vsetvl-driven loop bounds, so the vector length adapts at runtime rather than assuming VLEN=128.
• Correctness relies on tail-undisturbed (_tu) accumulation, so the final short iteration does not overwrite earlier lanes.

Co-authored-by: ihb2032 hebome@foxmail.com
Co-authored-by: lyd1992 liuyudong@iscas.ac.cn
Co-authored-by: Yuansheng yuansheng@isrc.iscas.ac.cn

[meta-cla]

Hi @Xlawy!

Thank you for your pull request and welcome to our community.

Action Required

In order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you.

Process

In order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA.

Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with CLA signed. The tagging process may take up to 1 hour after signing. Please give it that time before contacting us about it.

If you have received this in error or have any questions, please contact us at cla@meta.com. Thanks!