SYRK / SYR2K Sample¶

Demonstrates symmetric rank-k and rank-2k updates using the LIBXS dispatch-and-call model. The sample validates correctness against a plain Fortran reference and reports performance.

Operations¶

SYRK: C := alpha * A * A^T + beta * C (lower triangle) SYR2K: C := alpha * (A * B^T + B * A^T) + beta * C (upper triangle)

Only the specified triangle of C is written; the other triangle is left untouched.

Build¶

make

Requires a Fortran compiler (gfortran, ifort, or ifx). The Makefile picks up the compiler from the top-level Makefile.inc. MKL or BLAS linkage is enabled (BLAS=1) so that dispatch can use JIT-compiled kernels when available.

Run¶

./syrkf.x [N [K [nrepeat]]]

Arguments (all optional, positional):

N        Matrix dimension of C (N x N).  Default: 64
K        Inner dimension (columns of A).  Default: N
nrepeat  Number of timed repetitions.     Default: 100

Example Output¶

syrk(F): N=64 K=64 nrepeat=100

--- libxs_syrk (lower) ---
  max error (lower): 0.00000E+00

--- libxs_syr2k (upper) ---
  max error (upper): 0.00000E+00

--- SYRK performance ---
  time:      0.002 s (100 calls)
  perf:      28.4 GFLOPS/s

Notes¶

• The dispatch step (libxs_syrk_dispatch / libxs_syr2k_dispatch) returns a pointer to a registry-owned config. This pointer remains valid until libxs_finalize or the registry is destroyed. There is no need to release it manually.

• Internally, SYRK/SYR2K decompose into GEMM tiles on the diagonal and off-diagonal blocks. The dispatched GEMM kernel (MKL JIT, LIBXSMM, or fallback BLAS) handles the inner loop.

• Scratch memory for the temporary full-panel product is managed via a thread-local buffer that grows on demand and is freed at finalization.