[Mesa-dev] [PATCH] i965/fs: Don't try 16-wide if 8-wide uses more than half the registers.

[Kenneth Graunke]

On 08/21/2012 10:49 PM, Eric Anholt wrote:
> I don't like the idea of losing 16-wide on apps where it might have
> succeeded if we just tried.  Note that a chunk of register space gets
> eaten by things that don't scale with number of pixels, like attribute
> setup, push constants, and the MRF hack.

Nor do I, which is why I was trying to be conservative.  I figured the
attribute setup could be accounted for by a large enough fudge factor,
but I forgot about the MRF hack region.

> I could be convinced otherwise with some shader-db stats, though.

Good call.

total instructions in shared programs: 507477 -> 507477 (0.00%)
instructions in affected programs:     0 -> 0

However, the margin is more slim than I thought: Unigine Sanctuary
starts losing out on 16-wide with the cut-off at 70 (my patch was 75).
Perhaps I should refine the heuristic to account for the non-scaling
sections.