amstradcpc
Member
You posted earlier hinting at the same thing, I believe:
Have your prior thoughts been confirmed or is this something different? (Or is this just reiteration of the same thoughts?)
-----
Did he say that?
That seems to conflict with earlier information in the thread. At least from Proelite.
So confusing.
The thing is i don´t see asset compression/decompression related to SIMD management but as something related to increase the effective bandwith: you pack the bit packages after being computed and before sending them through the bus so more of them can travel across, and more can fit into memory. Something like texture compression but more general. This would have some sense to make the effective bandwidth to the memory pools greater.
If there is hardware in to improve SIMD usage it must be something related to prevent the ALUs stalls when there are instructions dependencies in a wavefront that in GCN forces to process a instruction of another wavefront to prevent the ALU from waiting for the dependent instructions to be executed. This could be solve introducing some kind of out-of-order schedule that allowed to jump from one instruction to another inside the same wavefront.