Module policy 

CPU-vs-GPU backend selection.

The decision compares estimated end-to-end times built from measured throughputs: a one-time GPU micro-benchmark (GpuCalibration, cached per context) and CPU throughputs supplied by the caller (CpuCosts, measured by the flashraw facade the same way). Hard rules come first: software adapters are never auto-selected, tiny frames stay on the CPU, and outputs that need no processing never go through the GPU.

Structs

CpuCosts

CPU throughputs (pixels per ms with all threads) for the same operations.

Decision

Outcome of decide.

GpuCalibration

Measured GPU characteristics.

Workload

Description of one decode for the cost model.

Enums

WorkOutput

What the caller wants produced.

Constants

AHD_CPU_FACTOR

The same for the AHD demosaic: on the CPU 2.3x (x86-64, 45 MP Bayer, develop to RGBA8 on one thread), on the GPU 5.2x (Apple M1 Ultra, 24 MP Bayer, kernel time 26 ms against 5 ms).

AHD_GPU_FACTOR

EDGE_CPU_FACTOR

Cost of the edge-aware demosaic relative to the bilinear one, for the develop step alone: on the CPU (measured 4.4x to 9.2x: Apple M1 Ultra, X-Trans and Bayer; 3.7x to 5.5x on x86-64) and on the GPU (measured 6.2x to 6.3x on an Apple M1 Ultra, kernel time). Rounded; only the CPU/GPU ratio matters.

EDGE_GPU_FACTOR

GPU_MARGIN

The GPU must win by this factor (estimates are noisy; CPU is the safer default).

MIN_GPU_SAMPLES

Frames below this size stay on the CPU (fixed GPU overheads dominate).

Functions

decide

Estimate CPU and GPU times and choose.