Like rounding, a VCOP store that includes saturation is translated as if were two operations: an explicit saturation operation, followed by the store. The saturation operation operates on 32-bit elements regardless of the data type of the store. VCOP supports several forms of saturation: SYMM, ASYMM, 4PARAM, and so on. Fundamentally, all of the forms operate as follows:

saturate(min, minset, max, maxset) = (x < min) ? minset 
                                   : (x > max) ? maxset : x; 

In the typical case when the saturation bounds are the same as the min/max set-to values, that is min == minset and max == maxset, the C7x translation is an efficient two instruction sequence:

VMINW        Vsrc,Vmax,Vdst 
VMAXW        Vdst,Vmin,Vdst 

If saturation bounds are to a power of 2 boundary, such as from 0 to 255, a single saturation instruction is used:

VGSATUW      Vsrc,Cwidth,Vdst 

If saturation bounds are different from the min/max set-to values, a less efficient 4-instruction sequence is required:

VCMPGTW      Vmin,Vsrc,Pred0 
VSEL         Pred0,Vminset,Vsrc 
VCMPGTW      Vmax,Vsrc,Pred1 
VSEL         Pred1,Vsrc,Vmaxset 

For unsigned vectors (see Section 1.5), unsigned forms of the compares are used.

C7x has a dedicated instruction for saturating to the range of a signed 16-bit value: VSATWH. Therefore the following statement:

__vptr_s16 out; 
out[Agen] = Vsrc.saturate(); // saturates to (-32768, 32767) 

translates to a single instruction.

VCC removes saturations that it determines to have no effect.