Re: [Mesa-dev] [PATCH] ac: Use mov_dpp for derivatives.

2017-06-15 Thread Connor Abbott
On Jun 15, 2017 3:31 AM, "Nicolai Hähnle" wrote: On 15.06.2017 00:36, Connor Abbott wrote: > I was looking into WQM stuff today, and I realized that LLVM will no > longer mark this instruction as needing WQM, which seems like a > problem. Seems like we need a patch to LLVM. Other uses of DPP (e.

Re: [Mesa-dev] [PATCH] ac: Use mov_dpp for derivatives.

2017-06-15 Thread Nicolai Hähnle
On 15.06.2017 00:36, Connor Abbott wrote: I was looking into WQM stuff today, and I realized that LLVM will no longer mark this instruction as needing WQM, which seems like a problem. Seems like we need a patch to LLVM. Other uses of DPP (e.g. for the subgroup reduction stuff) won't want WQM, so

Re: [Mesa-dev] [PATCH] ac: Use mov_dpp for derivatives.

2017-06-14 Thread Connor Abbott
I was looking into WQM stuff today, and I realized that LLVM will no longer mark this instruction as needing WQM, which seems like a problem. Seems like we need a patch to LLVM. Other uses of DPP (e.g. for the subgroup reduction stuff) won't want WQM, so I'm not sure what's the best approach there.

Re: [Mesa-dev] [PATCH] ac: Use mov_dpp for derivatives.

2017-06-12 Thread Nicolai Hähnle
On 10.06.2017 22:05, Bas Nieuwenhuizen wrote: Slightly faster than bpermute, and seems supported since at least LLVM 3.9. v2: Since this supersedes bpermute, remove the bpermute code. Signed-off-by: Bas Nieuwenhuizen This version looks better. R-b as well. --- src/amd/common/ac_llvm_buil

Re: [Mesa-dev] [PATCH] ac: Use mov_dpp for derivatives.

2017-06-12 Thread Nicolai Hähnle
On 10.06.2017 21:52, Bas Nieuwenhuizen wrote: Slightly faster than bpermute, and seems supported since at least LLVM 3.9. Signed-off-by: Bas Nieuwenhuizen --- src/amd/common/ac_llvm_build.c | 78 +- 1 file changed, 54 insertions(+), 24 deletions(-) di

Re: [Mesa-dev] [PATCH] ac: Use mov_dpp for derivatives.

2017-06-11 Thread Marek Olšák
Hi Bas, Have you tested piglit on radeonsi with this? Marek On Sat, Jun 10, 2017 at 10:05 PM, Bas Nieuwenhuizen wrote: > Slightly faster than bpermute, and seems supported since at least > LLVM 3.9. > > v2: Since this supersedes bpermute, remove the bpermute code. > Signed-off-by: Bas Nieuwenhu

[Mesa-dev] [PATCH] ac: Use mov_dpp for derivatives.

2017-06-10 Thread Bas Nieuwenhuizen
Slightly faster than bpermute, and seems supported since at least LLVM 3.9. v2: Since this supersedes bpermute, remove the bpermute code. Signed-off-by: Bas Nieuwenhuizen --- src/amd/common/ac_llvm_build.c | 47 src/amd/common/ac_llvm_build.h

[Mesa-dev] [PATCH] ac: Use mov_dpp for derivatives.

2017-06-10 Thread Bas Nieuwenhuizen
Slightly faster than bpermute, and seems supported since at least LLVM 3.9. Signed-off-by: Bas Nieuwenhuizen --- src/amd/common/ac_llvm_build.c | 78 +- 1 file changed, 54 insertions(+), 24 deletions(-) diff --git a/src/amd/common/ac_llvm_build.c b/src/am