perf: AMX BF16 dispatch in EuclideanDistances::computeABt

[napetrov]

AMX BF16 fast path for float32 Euclidean distance GEMM

Summary

Adds an opt-in AMX-BF16 fast path for the float32 Euclidean distance computation
used by KNN (brute-force) and DBSCAN. Internally, the A×B' matrix multiply in
EuclideanDistances::computeABt converts float32 inputs to BF16 and dispatches
to Intel AMX tile instructions. Output remains float32.

Activation (opt-in, default is unchanged strict / full float32):

ONEDAL_FLOAT32_MATMUL_PRECISION=ALLOW_BF16

Or programmatically:

daal_set_float32_matmul_precision(daal::internal::Float32MatmulPrecision::allow_bf16);

Design

Float32MatmulPrecision API (internal, daal::internal namespace, not installed in public headers):

Level	Behaviour
strict (default)	Always float32. IEEE-754, bit-exact.
allow_bf16	Permit BF16 kernels where oneDAL determines accuracy impact is bounded.

Hardware check — once at process init (g_hw_amx_bf16). No HW queries in getter/setter/hot path.

BF16GemmDispatcher<float, cpu> — partial specialization separates the dispatch struct from EuclideanDistances. The TU-init g_use_bf16_gemm bool combines HW capability and user hint; the hot path branches on a single static bool (branch predictor-friendly).

Eligibility gate: both GEMM dimensions (rows of A and rows of B) must be ≥ 64. Below this threshold, BF16 tiling overhead outweighs the benefit and accuracy impact grows.

Benchmark results

Hardware: Intel Xeon 6975P-C (Granite Rapids), KVM, AMX-BF16 enabled.
Method: scikit-learn_bench (sklbench), n_jobs set via --parameters.

DBSCAN euclidean (feat ≥ 64):

config	speedup range
8 vCPU, n_jobs 1–8	1.30–1.35x
16 vCPU, best case (f512, n_jobs=8)	1.60x
48 vCPU, n_jobs 1–8	1.21–1.25x

KNN classification inference (feat ≥ 128):

config	speedup range
8 vCPU	1.31–1.44x
16 vCPU	1.11–1.22x
48 vCPU	1.08–1.17x

No regressions observed: KNN training, KNN regressor (feat=10), DBSCAN (feat=3) all show ~1.00x.

Files changed

File	Change
cpp/daal/src/services/service_defines.h	Float32MatmulPrecision enum + API declarations (internal)
cpp/daal/src/services/compiler/generic/env_detect_precision.cpp	HW check, env parse, getter/setter impl
cpp/daal/src/algorithms/service_kernel_math.h	BF16GemmDispatcher, refactored computeABt
cpp/daal/src/services/cpu_type.h	CpuFeature::amx_bf16 flag

[napetrov]

/intelci: run

[Vika-F]

It would be better to extend the function we already have for CPU features detection, if needed:
https://github.com/uxlfoundation/oneDAL/blob/main/cpp/daal/src/services/compiler/generic/env_detect_features.cpp#L303

[Vika-F]

1. This kind of PRs are rather meaningless without performance & accuracy data.
   Need to ask an agent to run sklearn_bench on this.
   1.a) Probably the fp32->bf16 conversion done on every call to distances will "eat" all the potential performance.
2. The conversion code can be improved with _mm512_cvtneps_pbh, but this won't help much with the performance.
3. Using of unions in C++ can be harmful due to a strict aliasing rule (https://gist.github.com/shafik/848ae25ee209f698763cffee272a58f8). Need a union-free version of the code.
4. In DAAL we typically handle the case of specialized implementations with [partial] template specializations and not with the std::is_same. Template specializations allow to reduce the amount of branching and decrease the code complexity.