Mojo function
get_accum_type
get_accum_type[type: DType, *, preferred_accum_type: DType = float32]() -> DType
Returns the recommended type for accumulation operations.
Half precision and float8 types can introduce numerical error if they are used in reduction/accumulation operations. This method returns a higher precision type to use for accumulation if a half precision types is provided, otherwise it returns the original type.
The rules are as follows: - If the type is a float8 type, return a float16 type. - If the type is a bfloat16 precision type, return a float32 type. - If the type is a float16 precision type, return a float32 type if the preferred_accum_type is float32, otherwise return a float16 type. - Otherwise, return the original type.
Parameters:
- type (
DType): The type of some accumulation operation. - preferred_accum_type (
DType): The preferred type for accumulation.
Returns:
DType.float32 if type is a half-precision float, type otherwise.
Was this page helpful?
Thank you! We'll create more content like this.
Thank you for helping us improve!