Replace the free functions BraggIntegrate2D/ProfileIntegrate2D with the
BraggIntegrationEngine (CPU/GPU) as the live integrator.
- IndexAndRefine no longer holds the integrator: ProcessImage takes a
per-worker BraggIntegrateFn callback (ProcessImage is called concurrently by
the shared IndexAndRefine, so the stateful engine must not be a member).
- WithoutFPGA/jfjoch_process: owns a GPU engine when a GPU is present, else CPU,
and passes the GPU-resident preprocessed buffer so integration runs on-device.
- AfterFPGA: forces CPU and integrates straight off the assembled CompressedImage
via a templated per-pixel sampler - only the reflection-disk pixels are read,
no whole-image copy (the FPGA host runs up to 36 GB/s). Sampler maps type
min/max to INT32_MIN/INT32_MAX on read; special/saturation only, no +/-1 band.
- Remove BraggIntegrate2D/ProfileIntegrate2D and their test; keep IntegratorMode.
Prediction: buffer up to 20000 candidates but return the 10000 closest to the
Ewald sphere (deterministic partial_sort on |dist_ewald|, hkl tiebreak) instead
of the GPU atomic-fill order. Serialized output stays <=10000, so the frame
transport headroom and its CBOR guard are unchanged.
integration_model exposed via OpenAPI (bragg_integration_settings schema,
/config/bragg_integration PUT/GET, added to jfjoch_settings and jfjoch_statistics)
and the frontend (BraggIntegrationSettings dropdown). Regenerated C++/TS clients
and redoc.
Validated old-vs-new on all 18 /data/rotation_test crystals: indexing rate and
space group bit-identical; ISa/CC identical on 16/18 (one improved, EcwtAL500
ISa 0.0->6.7); new CompressedImage-vs-buffer and GPU-vs-CPU parity tests pass.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
leonarski_fandClaude Opus 4.8
4cda46c6b0