Hi Dave,
Can you rebase your series on top of master because it doesn't apply.
I'm fine with the general idea. Note that I have a shader which spends
5s in copy_propagate()... That's huge! Looks like this series might
improve the situation here, but needs more testing.
Thanks!
On 05/30/2017
On 30 May 2017 at 22:31, Ilia Mirkin wrote:
> Is this a reimplementation of the same algorithm in a more efficient
> manner, or an implementation of a different algorithm which happens to
> also run faster? The reason I ask is whether I should be seeing any
> differences in shader-db with these at
Is this a reimplementation of the same algorithm in a more efficient
manner, or an implementation of a different algorithm which happens to
also run faster? The reason I ask is whether I should be seeing any
differences in shader-db with these at all or not.
Cheers,
-ilia
On Tue, May 30, 2017
While looking at the fp64 emulation code, we were spending minutes
in the glsl->tgsi passes as some tests were producing shaders with
> 32000 temporaries. Now it might be possible to reduce these earlier
with some GLSL passes, but this code is pretty bad as-is.
This reduces one test execution time
