The default data types for the active backend. They decide the data type an R value is materialized at when it cannot be inferred from another operand.
This includes array creation via (nv_array(1)) or passing R values to unary functions
(prim_exp(1)).
default_dtypes() reports both categories at once; default_float() and
default_int() report one each.
Each backend registers its own – f32 / i32 for "pjrt", f64 / i32
for "quickr" – and the anvl.default_dtypes option overrides them.
local_default_dtypes() and with_default_dtypes() set the option for
a scope.
Below, we configure any backend to use the default f64 for floats and i64 for integers:
You can also only change the float dtype, leaving the backend's default integer dtype unchanged:
It is also possible to specify the defaults per-backend:
options(anvl.default_dtypes = list(
pjrt = list(float = "f64", int = "i64"),
quickr = list(int = "i32")
))An entry that names a backend wins over the categories beside it.
When the option is not set, the ANVL_DEFAULT_DTYPES environment variable
(read once, when anvl is loaded) is used instead, written as category=dtype pairs that apply to every
backend, e.g. ANVL_DEFAULT_DTYPES="float=f64,int=i64".
The defaults decide only what a value becomes when nothing else does: an R
value that meets a typed array of its own category still takes that array's
data type, whatever the default (vignette("type-promotion")). The data
type you name is taken on trust, so one that does not fit is an error where
the data is allocated or the program compiled rather than where it is set.
Which ones fit is the backend's own business – see the Supported data
types section of AnvlBackendPjrt() and of AnvlBackendQuickr(),
which has only f64, i32 and bool, so both "f32" and "i64" are
errors there. A compiled program is keyed on the defaults it was compiled
under, so changing them never serves a stale program.
Value
default_dtypes() returns a named list with elements float and
int, each a DataType
Examples
with_backend("quickr", default_dtypes())
#> $float
#> <f64>
#>
#> $int
#> <i32>
#>
with_backend("pjrt", default_dtypes())
#> $float
#> <f32>
#>
#> $int
#> <i32>
#>
default_dtypes()
#> $float
#> <f32>
#>
#> $int
#> <i32>
#>
default_float()
#> <f32>
default_int()
#> <i32>
dtype(nv_array(1.5))
#> <f32>