[michalharakal]

What

skainet_ternary_f32_gemv (the BitNet b1.58 / BITNET_B1_58 projection) called the vendored NeoGPU hs_ml_ternary_f32_proj directly. That function creates and joins 4 pthreads on every call once N >= 512, which is about 210 calls per BitNet 2B4T decode step. It's the create/join pattern skainet_row_threads.c documents as the slow one (#1195).

The adapter now drives the vendored kernel from the shared row pool (skainet_run_rows) in slices of < 512 rows, so each slice stays below NeoGPU's THREAD_THRESHOLD and runs inline on a pool thread.

• Vendored file untouched: still byte-identical, "Local modifications: none", as the vendor README requires. The README's constraints section is updated to describe the adapter's threading.
• Bit-identical results: per-row math doesn't depend on the slicing. threaded_regime_matches_single_row_calls_exactly (outputDim 2048, partial slice included) covers it; its comment is updated.
• lm_head stage 1 unchanged: hs_ml_lmhead_stage1 always threads and has no inline entry point. Pooling it as well needs an upstream NeoGPU change; it was worth about +6 % more in my experiment below.

Measurements: Arduino Ventuno Q, 4× Cortex-A78C, taskset -c 0-3

Microbenchmark of the vendored kernel at the real 2B4T shapes, weights streamed from DRAM (harness):

op	create/join (before)	row pool (after)
q/o 2560×2560	654 µs	309 µs
k/v 640×2560	251 µs	84 µs
gate/up 6912×2560	1371 µs	785 µs
down 2560×6912	1430 µs	831 µs
trunk per token (210 GEMVs)	185.9 ms	93.9 ms

End to end: skainet-cli, JVM (Temurin 21), microsoft/bitnet-b1.58-2B-4T-gguf I2_S, 128 greedy tokens:

build	decode tok/s	greedy output
before	1.90	md5 52f57d1d…
this PR	2.26 (+19 %)	md5 52f57d1d… (byte-identical)
(experiment: also pooling lm_head by editing the vendored file, not proposed)	2.40	identical

Tests

• :skainet-backends:skainet-backend-native-cpu:jvmTest --tests '*Ternary*' on linux-x86_64 passes 13/13: NativeTernaryF32GemvKernelTest 10, TernaryPlanesFfmTest 2, TernaryF32FfmPackTest 1.
• I cross-built libskainet_kernels.so for linux-arm64 from this branch and ran it on the board (numbers above).
• The JNI-cpu CMake already compiles skainet_row_threads.c, so the Android AAR links unchanged.

Refs #1357

🤖 Generated with Claude Code