[tesch1]

Addresses the copy part of #6146.

Before: on VORTEXM4 the B-side GEMM copies were the generic C copies, and the A-side copies of the SME2 kernels used NEON. At n <= 512 these copies took most of the level-3 run time (about 70-80% of SSYRK and STRMM at n = 256, one thread). NEON versions of the B-side copies were tried first: they gain little, because the copies are limited by the bandwidth that one core gets from L2.

After: the copies run in streaming mode. The SME unit moves the same data 2-4x faster (fp32, one thread, data in L2: oncopy 50 -> 165 GB/s, otcopy 75 -> 220 GB/s at m = n = 256).

How:

• sme2_gemm_copy.h: s2c_rows() (contiguous rows) uses predicated streaming vector loads and stores; s2c_cols() (contiguous columns) loads columns into horizontal ZA slices and stores rows from vertical slices. Pieces smaller than 512 elements, and CPUs without SME2 (s2_usable()), keep the NEON code.
• New gemm_[nt]copy_b_sme2.c for S/DGEMMONCOPY and S/DGEMMOTCOPY: the same helpers, with the layout of generic/gemm_[nt]copy_16.c (UNROLL_N columns, then 8/4/2/1), in one streaming-mode entry per call.
• TRMM/TRSM A-side copies: the diagonal block goes through the panel copy, then each row gets its zero side and its unit or inverted diagonal (no per-element tests).

Only VORTEXM4 files change. GCC builds keep the generic copies (the KERNEL.VORTEXM4 section for non-GCC compilers).

Not in this PR: the OUTER TRMM/TRSM/SYMM copies (right-side B panels) are fixed to the generic files in kernel/Makefile.L3; changing them needs a shared-file change.

Results

M4 Pro, one thread, GFLOPS, develop 791b02a -> this PR:

fp32	n = 64	n = 256	n = 1024
SSYMM (L, L)	181 -> 215	764 -> 1124	1285 -> 1648
SSYRK (L, N)	91 -> 102	423 -> 518	1016 -> 1144
SSYRK (L, T)	76 -> 100	355 -> 508	915 -> 1157
STRMM (L, L, N)	50 -> 85	285 -> 421	805 -> 1010
STRMM (R, L, T)	78 -> 90	314 -> 374	812 -> 967
STRSM (L, L, N)	42 -> 68	221 -> 321	677 -> 837
STRSM (R, U, N)	63 -> 71	273 -> 329	810 -> 837

fp64	n = 64	n = 256	n = 1024
DSYMM	97 -> 129	255 -> 381	340 -> 423
DSYRK (L, T)	41 -> 60	149 -> 229	285 -> 369
DTRMM (L, L, N)	32 -> 44	119 -> 171	257 -> 321
DTRSM (L, L, N)	29 -> 38	104 -> 146	242 -> 296

SGEMM/DGEMM do not change. Default threads: also faster, no case slower (SSYMM n = 1024 2540 -> 2955, STRMM 1753 -> 1967, at n = 4096 5-11%).

Tests

• make: BLAS/CBLAS tests, utest 144/144 and 1564/1564.
• make lapack-test: 0 numerical errors.
• Randomized GEMM/SYMM/SYRK/SYR2K/TRMM/TRSM check against naive references, s and d, 1000 trials each: 0 failures.
• New copies against the generic ones for all m, n <= 70: identical output, with and without SME2 (NEON path).
• CMake (-DTARGET=VORTEXM4 -DNOFORTRAN=1 -DBUILD_TESTING=ON, ctest 14/14) and DYNAMIC_ARCH=1 NO_FORTRAN=1 builds.

AI assistance

AI coding tools (Claude Code) helped to prepare this PR. The changes were reviewed, tested and benchmarked as described above.