Performance

OpenQP's performance knobs are ordinary input keys (no environment variables) bundled behind a single opt-in preset, [input] perf, that acts as one accuracy↔speed dial.

[input]
perf = 1        # 0, 1, 2, or 3 (default 1; perf=-1 disables the preset)

The four levels

perf	Use it for	What it enables	Accuracy
0	strict reference / reproducibility	every accelerator off, every cutoff tightest	bit-reference (fixed thread count)
1	recommended production	exact, proven-helpful only: MRSF response cutoff 1e-8, z-vector warm-start (+ always-on Fock digestion)	≈ reference (≤ µEh)
2	faster, tiny degradation	perf=1 + coarse-to-fine XC grid + gradient Schwarz cutoff 1e-8	gradients within ~5×10⁻⁷ a.u.; SCF exact at convergence
3	aggressive, degradation allowed	looser cutoffs traded for speed: gradient 1e-7, response 1e-6	small, controlled (~few µeV on excitation energies)

perf=1 is the default (recommended production, exact). Set perf=-1 to disable the preset entirely (every knob at its control default). The levels were calibrated on a CPU benchmark (MKL and Apple Accelerate) over HF/DFT/TDDFT/MRSF energies and MRSF gradients.

Why some accelerators are not in any preset

On the CPU builds benchmarked, the XC Φ-cache, IncDFT, MRSF FP32 and progressive screening (pscreen) were performance-neutral to negative (FP32 was ~2× slower under MKL). No preset enables them; they remain available as explicit input keys (xc_phi_cache, xc_incdft, fp32, and pscreen/pscreen_cap) for regimes the CPU benchmark does not cover, e.g. GPU XC. Start at perf=1 and raise it only if your own timing shows a gain.

Individual input keys

Every knob is also a direct input key. Most keys below default to auto (defer to the preset); an explicit value overrides the preset. Progressive screening is the exception: pscreen is an existing boolean key (False by default; use on/off or true/false), and pscreen_cap is numeric (1.0e-8 by default). No current preset enables pscreen, so set it explicitly only when you want to test progressive screening.

Section	Key	Meaning
[scf]	xc_c2f	coarse-to-fine XC grid during the SCF descent (on/off/auto)
[scf]	xc_phi_cache	cache collocation Φ across SCF iterations — exact, opt-in (on/off/auto)
[scf]	xc_incdft	incremental DFT — experimental, opt-in (on/off/auto)
[scf]	pscreen, pscreen_cap	progressive integral/grid screening (pscreen is on/off; pscreen_cap is numeric)
[scf]	grad_cutoff	Schwarz cutoff for the 2e-derivative gradient build (number or auto)
[tdhf]	resp_cutoff	2e cutoff for the MRSF response build (number or auto)
[tdhf]	fp32	single-precision MRSF response digestion — opt-in (on/off/auto)
[tdhf]	zv_warmstart	reuse the previous step's z-vector as the CPHF seed — exact (on/off/auto)

Precedence (low → high): control default → perf preset → explicit input key. For example, production speed but keep the full XC grid and a tight response on one job:

[input]
perf = 2
[scf]
xc_c2f = off
[tdhf]
resp_cutoff = 5e-11

The resolved settings (and any warnings) are printed near the top of the run log:

   Performance settings (perf = 2)
     scf.xc_c2f        = off       (input)
     scf.grad_cutoff   = 1.0d-8    (preset)
     tdhf.resp_cutoff  = 5e-11     (input)
     tdhf.zv_warmstart = on        (preset)
     ...

Recommendations

• HF / DFT / TDDFT / MRSF energies: perf=1 for exact production; perf=2 to add the coarse-to-fine grid (helps large pure-functional DFT). Use perf=3 when a few µeV on excitation energies is acceptable for ~10–20% more speed.
• Gradients, geometry optimization, MD: perf=1 (z-vector warm-start is most effective across many nearby geometries) or perf=2. Avoid perf=3 for MD through conical intersections.
• Reference numbers and tight 2nd-order properties (IR/Raman/NMR): perf=0 or perf=1.

Reproducibility

perf=0 is the bitwise reference for a fixed thread count. The threaded Fock build uses a dynamic-schedule reduction, so results are not bit-identical across thread counts at any level — pin OMP_NUM_THREADS for strict reproducibility.