I'm running OpenCL on Intel's HD4000 GPU.
It seems that this GPU has no vector calculations.
According to the query:
The vector width is 1. Which means that float4 calculations are executed as 4 float scalars.
But what about the size of the loaded vector from global memory ?
Can HD4000 load 4 floats (or even 8) at one clock ?
Can you send a document describing those features ?
I did not find the relevant document in google.
IVB GPU can actually execute 8 SP floats / cycle / EU. With MAD that's 2X so peak is 16 SP float ops / cycle / EU. I will try to find the public doc (if one is available) and post the link here.