Expand description
Numeric types used across the CUDA stack.
These are thin, #[repr(transparent)] / #[repr(C)] wrappers chosen to
match the layout NVIDIA’s headers use for __half, __nv_bfloat16,
cuFloatComplex, and cuDoubleComplex. All conversion methods return
the same bit patterns the CUDA runtime itself would produce for typical
inputs; exact agreement with NVIDIA’s rounding on edge cases is tested
in the integration suite against half and CUDA itself.
If you already depend on half / num-complex, enable the half-crate
/ num-complex-crate features for zero-cost From/Into adapters.
Structs§
- BFloat16
- Brain Floating Point 16 (
__nv_bfloat16in CUDA). The top 16 bits of an IEEE 754f32. - Complex32
- Single-precision complex number (
cuFloatComplex, layout-compatible withfloat2). - Complex64
- Double-precision complex number (
cuDoubleComplex, layout-compatible withdouble2). - Half
- IEEE 754 binary16 (“half-precision”,
__halfin CUDA).