Convert BFloat16sr to Float32 by padding trailing zeros.

The BFloat16 + stochastic rounding type.

Convert to BFloat16sr from Float32 via round-to-nearest and tie to even. Identical to BFloat16(::Float32).

Float16 + stochastic rounding type.

The Float32 + stochastic rounding type.

Chance that x::Float32 is round up when converted to BFloat16sr.

Stochastically round x::Float32 to BFloat16 with distance-proportional probabilities.

Float16_chance_roundup(x::Float32)

Chance that x::Float32 is round up when converted to Float16sr.

Float16_stochastic_round(x::Float32) -> Float16sr

Stochastically round x to Float16 with distance-proportional probabilities.

Float32_chance_roundup(x::Float64)

Chance that x is round up when converted to Float32sr.

Float32_stochastic_round(x::Float64) -> Float32sr

Stochastically round x::Float64 to Float32 with distance-proportional probabilities.

Reseed the PRNG randomly by recalling.

rand_subnormal(rbits::UInt32) -> Float32

Create a random perturbation for the Float16 subnormals for stochastic rounding of Float32 -> Float16. This function samples uniformly from [-2.980232f-8,2.9802319f-8]. This function is algorithmically similar to randfloat from RandomNumbers.jl

rand_subnormal(rbits::UInt64) -> Float64

Create a random perturbation for the Float16 subnormals for stochastic rounding of Float32 -> Float16. This function samples uniformly from [-7.0064923216240846e-46,7.006492321624084e-46]. This function is algorithmically similar to randfloat from RandomNumbers.jl

Seed the PRNG with any integer >0.

Mask for both sign and exponent bits. Equiv to ~significand_mask(Float64).