simd(riscv64): 6x8 RVV GEMM micro-kernel

[tannevaled]

What

gemmMicro6x8RVV, generated by go-asmgen (riscv64/gen.go), with no WORDs. NR is fixed by the packing, while VLEN varies between RISC-V cores. Since V guarantees VLEN ≥ 128, a group of four registers (LMUL=4) always holds 8 float64, so VSETIVLI $8, E64, M4 gives vl = 8 on every core. The tile uses six accumulator groups. Each step loads the B row once and fuses each A value in with vfmacc.vf (Go: VFMACCVF Vb, Fa, Vacc computes Vacc += Fa·Vb, decoded from Go's testdata).

gemm_riscv64.go: MR=6, NR=8, gated on useRVV. Without V, the same tile runs in Go (gemmMicroGo). gemm_generic.go now excludes riscv64.

Tests on real hardware

• cfarm95 (SpacemiT X60, V): GEMM, MatMul, fence and stress tests all pass, plus the full root and kernels suites.
• cfarm94 (SiFive U74, no V): the same binary takes the Go 6×8 tile and passes.
• Sabotage, caught on cfarm95: dropping the FMA of tile row 5 (MatMul m=50 k=70 n=90 [450]: 0 != -94).
• Coverage 100% with -race; go vet on 7 architectures.

Measured (X60, one core) with the fleet's Go tools, no ad-hoc scripts

hostload -pick 1 (go-fleettools/fleettools#60) picked an idle CPU on a host that was 64% busy with others' niced jobs. benchstat compared 8 interleaved runs each against v0.7.0, with p = 0.000 everywhere:

	64²	128²	256²	512²	1024²
MatMul time	−31%	−32%	−45%	−43%	−44% (1.79 s → 1.00 s)

Inner −44%, small rectangular products −22% to −38%, geomean −36%. Absolute speed is about 2.1 GFLOP/s per core. With VLEN = 256, LMUL=4 uses half of each group; that is the price of one tile portable to every VLEN, and it is documented.

Version: v0.7.1 (performance, no API change).

🤖 Generated with Claude Code