Casting and Datatype Conversion - 2023.2 English

AI Engine Kernel and Graph Programming Guide (UG1079)

Document ID
Release Date
2023.2 English

Casting intrinsic functions (as_[Type]()) allow casting between vector types or scalar types of the same size. The casting can work on accumulator vector types too. Generally, using the smallest data type possible will reduce register spillage and improve performance. For example, if a 48-bit accumulator (acc48) can meet the design requirements then it is preferable to use that instead of a larger 80-bit accumulator (acc80).

Note: The acc80 vector data type occupies two neighboring 48-bit lanes.

Standard C casts can be also used and works identically in almost all cases as shown in the following example.

aie::vector<int16,8> iv;
aie::vector<cint16,4> cv=as_v4cint16(iv); 
aie::vector<cint16,4> cv2=*(aie::vector<cint16,4>*)&iv; 
aie::accum<acc80,8> cas_iv;
aie::accum<acc48,8> cas_cv=as_v8cacc48(cas_iv);

There is hardware support built-in for floating-point to fixed-point (float2fix()) and fixed-point to floating-point (fix2float()) conversions. For example, the fixed-point square root, inverse square root, and inverse are implemented with floating point precision and the fix2float() and float2fix() conversions are used before and after the function.

Note: The AI Engine floating-point is not completely compliant with the IEEE standards. For more information about the exceptions, see Versal Adaptive SoC AI Engine Architecture Manual (AM009).

The scalar engine is used in this example because the square root and inverse functions are not vectorizable. This can be verified by looking at the function prototype's input data types:

float _sqrtf(float a) //scalar float operation
int sqrt(int a,...) //scalar integer operation

Note that the input data types are scalar types (int) and not vector types (vint).

The conversion functions (fix2float, float2fix) can be handled by either the vector or scalar engines depending on the function called. Note the difference in data return type and data argument types:

float fix2float(int n,...) //Uses the scalar engine
v8float fix2float(aie::vector<int32,8> ivec,...) //Uses the vector engine
Note: For float2fix, there are two types of implementations, float2fix_safe (default) and float2fix_fast with the float2fix_safe implementation offering a more strict data type check. You can define the macro FLOAT2FIX_FAST to make float2fix choose the float2fix_fast implementation.