Kernel fusion across slopes

Compile once; parameterize angles per-shot via Guppy comptime constants. Cuts G3 wall-clock from minutes to seconds.

Results · real Selene shots

host_fusion_verified_3.47x · in_kernel_variant_tracked_upstream

Host-level reuse of one built runner is verified at 3.47× faster than per-slope re-compilation, with shot statistics indistinguishable from the baseline. The in-kernel parametric-entry-point variant is blocked upstream (guppylang 0.21.16 EntrypointArgsError) and tracked there — outside this repo's surface area.

speedup (A per-slope) / (B fused-host)3.47×

JSONsrc/data/demos/nadarasa_g7.json

Open full Nadarasa page →

Why this matters

G3 today regenerates and re-imports a .py file per slope. That round-trip dominates wall-clock. Compile once, sweep many — re-use the built Selene runner across slopes — is the engineering claim the card makes.

What the repo already has

Host-fused runner in nadarasa_g7.py: one compile, one builder, N×run_shots — verified to match per-slope shot statistics exactly.

What's missing

True in-kernel parametric fusion (`angles: array[float, n]` as entry-point arg) requires Guppy entry-point support that 0.21.16 does not have; tracked upstream.

Smallest experiment

Build it or kill it

Qubits

Same as G3 (n + 1).

Ancilla pattern

Same as G3.

Shots

6 slopes × 200 shots = 1200, same as current G3.

Predicted outcome

Host-fused wall-clock drops by >1.5×; shot statistics indistinguishable from per-slope baseline.

Refutation criterion

If host fusion gives <1.5× speedup or alters shot statistics, the card is refuted.

Kernel sketch

Untested — sketch only

# host-side fusion (verified · G7)
compiled = program.compile()           # once
runner   = build(compiled)             # once
for angles in slope_grid:
    shots = runner.run_shots(Quest(), n_qubits=n+1, n_shots=200, angles=angles)
# wall_clock = compile_once + sum(run_shots) vs. per-slope re-compile

Host pipeline

Driver compiles once; per-slope = one runner.run_shots call with new angles. Verified speedup A/B = 3.47×.

Files in this repo

· quantum/nadarasa_g7.py
· src/data/demos/nadarasa_g7.json
· quantum/nadarasa_g3.py
· quantum/nadarasa_g3_lib.py

References

· Upstream tracker — guppylang 0.21.16 EntrypointArgsError ↗

More in Compilation

2 cards

Verified

Parameter sweep as a first-class object