Data type support#

The following table shows the supported input and output datatypes.

Table 1 Supported Input/Output Types#

Input/Output Types

Library Data Type

Support

int8

int8_t

⚠️

float8

Not Supported

bfloat8

Not Supported

int16

int16_t

float16

rocprim::half

⚠️

bfloat16

rocprim::bfloat16

⚠️

int32

int

tensorfloat32

Not Supported

float32

float

float64

double

The ⚠️ means that the data type is mostly supported, but there are some API tests, that do not work.

  • The block_histogram test fails with int8.

  • The device_histogram and device_reduce_by_key doesn’t work with rocprim::half and rocprim::bfloat16.

  • The device_run_length_encode, warp_exchange and warp_load doesn’t work with rocprim::half.