Experiment: the A-prep OMP loop heap-allocated ready_row (plus signed_row and full_hadamard_row on the Hadamard path) once per row -- up to M*(1-3) malloc/free per k-segment per matmul. Move to per-thread pools sized omp_get_max_threads(); resize() reuses the buffer after the first row. Byte-identical (consumers read only the fully-written prefix). pp300/512/700 35.8/41.6/40.6 -> 36.5/42.1/42.4 (+2-4%); tg128 2.93. Cumulative vs baseline: pp300 +19%, tg +14%.
mfritsche
3705e84765