vload_half3_* and vstore_half3_* are not supported about clspv HOT 2 CLOSED

Here's an example of the complexity I'm concerned about:

global float2* OUT;
float3 value = ...
vstore_half3(value, 1, (half*)OUT);

vstore_half3 says:

The floatn value given by data is converted
to a halfn value using the appropriate
rounding mode. The halfn value is then
written to address computed as (p + (offset * n)). The address computed as (p + (offset * n)) must be 16-bit aligned.
vstore_halfn uses the default rounding
mode. The default rounding mode is round
to nearest even.

In this case n = 3. So we should be writing at least 48 bits of data (possibly 64 bits) starting at 48bits into the OUT array. Each OUT element is 64 bits wide, so that means we have to update two adjacent elements:

• the last 16 bits of element 0 of OUT
• the first 32 bits (possibly 48 bits) of element 1 of OUT

from clspv.

I think this is fixed now with the ThreeElementVector pass

from clspv.