Set the Default Data Types

Set the default data types (see default_dtypes()) for a scope: local_default_dtypes() until the calling frame exits, with_default_dtypes() for the duration of code. Both write the anvl.default_dtypes option for one backend, and change only the categories they name.

Usage

local_default_dtypes(dtypes, backend = NULL, envir = parent.frame())

with_default_dtypes(dtypes, code, backend = NULL)

Arguments

dtypes

(named character() | named list())
A mapping of the data type categories (float and int) to data types, e.g. c(float = "f64", int = "i32"). Each may be a string or a DataType. Can also be a partial override, such as c(float = "f64"), in which case the category it does not name is left as it is.

backend

(NULL | character(1))
The backend whose defaults to set. Uses active_backend() by default.

envir

(environment)
The environment to scope the change to.

code

An expression to evaluate with the given defaults.

Value

local_default_dtypes() returns the previous values of the options it set, invisibly. with_default_dtypes() returns the result of evaluating code.

Details

Inside a jit()ted body the defaults the program was keyed on are the baseline and an override applies to its scope, so one program can use different precisions in different parts of itself. Only that baseline is part of the compilation cache key, so an override in a body must not change between calls: write it out literally rather than reading it from a variable. with_default_dtypes(c(float = prec), ...) with a prec that later changes keeps serving the program traced at the first value, exactly as a changing dtype argument would – and just as silently. Nothing checks this for you.

See also

default_dtypes(), local_backend(), with_backend()

Examples

with_default_dtypes(c(float = "f64"), dtype(nv_array(1.5)))
#> <f64>
# A value that meets a typed array still takes that array's data type
with_default_dtypes(c(float = "f64"), dtype(nv_array(1, dtype = "f32") + 1.5))
#> <f32>
# untyped values in one program can materialize at different precisions: one at
# whatever the default is, one at the `f64` the override asks for
jit(function() {
  list(
    at_default = nv_fill(0, 2),
    forced_f64 = with_default_dtypes(c(float = "f64"), nv_fill(0, 2))
  )
})()
#> $at_default
#> AnvlArray
#>  0
#>  0
#> [ CPUf32{2} ] 
#> 
#> $forced_f64
#> AnvlArray
#>  0
#>  0
#> [ CPUf64{2} ] 
#>