vibeproj.gpu_detect

GPU type detection for automatic precision selection.

Queries the NVIDIA driver to determine fp64:fp32 throughput ratio. Projection arithmetic remains fp64. The central registry can select bounded transcendental implementations for qualified Helmert, UTM, and projection domains on validated Ada consumer GPUs. Datacenter and unknown GPUs retain native fp64 until independently qualified.

Functions

get_fp64_ratio(→ float)	Return the fp64:fp32 throughput ratio for the current GPU.
favors_native_fp64(→ bool)	True if fp64 is fast enough to use directly (datacenter GPU or CPU).
select_compute_precision(→ str)	Select compute precision based on GPU type.
select_helmert_trig_mode(→ str)	Return the legacy name for the centrally resolved Helmert strategy.
select_tmerc_forward_mode(→ str)	Return the legacy name for the centrally resolved forward-UTM strategy.

Module Contents

vibeproj.gpu_detect.get_fp64_ratio() → float

Return the fp64:fp32 throughput ratio for the current GPU.

Returns 1.0 if no GPU is available (CPU mode — always use fp64). Returns ratio >= 0.25 for datacenter GPUs (use native fp64). Returns ratio < 0.25 for consumer GPUs (use compensated fp32).

vibeproj.gpu_detect.favors_native_fp64() → bool

True if fp64 is fast enough to use directly (datacenter GPU or CPU).

vibeproj.gpu_detect.select_compute_precision() → str

Select compute precision based on GPU type.

Always returns "fp64" for arithmetic and I/O. Individual fused kernels can still select any registry-qualified bounded transcendental strategy for the current device and exact operation domain without exposing a lower-precision public mode.

vibeproj.gpu_detect.select_helmert_trig_mode() → str

Return the legacy name for the centrally resolved Helmert strategy.

vibeproj.gpu_detect.select_tmerc_forward_mode() → str

Return the legacy name for the centrally resolved forward-UTM strategy.