What does UnsignedSaturate in SSE instruction mean?

Question

I am converting SIMD code into equivalent c code. I am stuck at one of SSE instructions

__m128i _mm_packus_epi16 (__m128i a, __m128i b)

it returns

r0 := UnsignedSaturate(a0)
r1 := UnsignedSaturate(a1)

...
r7 := UnsignedSaturate(a7)
r8 := UnsignedSaturate(b0)
r9 := UnsignedSaturate(b1)
...
r15 := UnsignedSaturate(b7)

What does UnsignedSaturate mean?

Answer 1

Basically, "saturation" means that values beyond some "max" value get set to "max", and values below a "min" value get set to "min". Usually, "min" and "max" are the values appropiate for some data type.

Thus, for example, if you take arithmetic on unsigned bytes, "128+128" would have to be "256" (which is hex 0x100), which doesn't fit into a byte. Normal integer arithmetic would create an overflow and discard the part that doesn't fit, which means "128+128 -> 0". With saturated arithmetic, "256 > 255" so the result is 255.

Another option would be scaling, which basically "compresses" the values to a smaller range. Saturation just cuts them off.

You can also use this to put larger types into smaller ones, like putting 16 bit values into 8 bit values. Your example most likely does exactly that, although you'll probably know better than I do what kind of types you are dealing with there.

"UnsignedSaturation" most likely has a min of "0" and a "max" of whatever the max of the result type is. Thus, negative inputs get turned into "0".

Answer 2

The instruction converts 16-bit signed integers to 8-bit unsigned integers. The problem is what to do when the value doesn't fit. In other words is less than 0 or larger than 255. Unsigned saturation specifies that the value is then clipped to its range. In other words, values less than 0 are converted to 0, more than 255 to 255.

Answer 3

Saturate simply means that the given variable keeps the maximum possible value instead of overflowing.

unsigned short v1 = 65535;
unsigned (saturate) short v2 = 65535;

v1++;
v2++;
printf("v1=%u v2=%u\n", v1, v2);

prints v1=0 v2=65535

The concept of saturates does not exist in standard C and must be provided by extensions.