[Libclc-dev] [PATCH 2/2] r600: Cleanup barrier implementation.

[Aaron Watry via Libclc-dev]

On Fri, Sep 1, 2017, 7:37 PM Jan Vesely via Libclc-dev <
libclc-dev at lists.llvm.org> wrote:

> On Tue, 2017-08-15 at 18:23 -0400, Jan Vesely wrote:
> > We don't have memory fences for r600 so just call group barrier directly
> >
> > Signed-off-by: Jan Vesely <jan.vesely at rutgers.edu>
> > ---
> >  amdgpu/lib/SOURCES                       |  1 -
> >  amdgpu/lib/synchronization/barrier.cl    | 10 ----------
> >  r600/lib/synchronization/barrier_impl.ll | 31
> +++++--------------------------
> >  3 files changed, 5 insertions(+), 37 deletions(-)
> >  delete mode 100644 amdgpu/lib/synchronization/barrier.cl
> >
> > diff --git a/amdgpu/lib/SOURCES b/amdgpu/lib/SOURCES
> > index 9379ad7..f817538 100644
> > --- a/amdgpu/lib/SOURCES
> > +++ b/amdgpu/lib/SOURCES
> > @@ -1,7 +1,6 @@
> >  atomic/atomic.cl
> >  math/nextafter.cl
> >  math/sqrt.cl
> > -synchronization/barrier.cl
> >  image/get_image_width.cl
> >  image/get_image_height.cl
> >  image/get_image_depth.cl
> > diff --git a/amdgpu/lib/synchronization/barrier.cl
> b/amdgpu/lib/synchronization/barrier.cl
> > deleted file mode 100644
> > index 6f2900b..0000000
> > --- a/amdgpu/lib/synchronization/barrier.cl
> > +++ /dev/null
> > @@ -1,10 +0,0 @@
> > -
> > -#include <clc/clc.h>
> > -
> > -_CLC_DEF int __clc_clk_local_mem_fence() {
> > -  return CLK_LOCAL_MEM_FENCE;
> > -}
> > -
> > -_CLC_DEF int __clc_clk_global_mem_fence() {
> > -  return CLK_GLOBAL_MEM_FENCE;
> > -}
> > diff --git a/r600/lib/synchronization/barrier_impl.ll
> b/r600/lib/synchronization/barrier_impl.ll
> > index 9b8fefb..777001a 100644
> > --- a/r600/lib/synchronization/barrier_impl.ll
> > +++ b/r600/lib/synchronization/barrier_impl.ll
> > @@ -1,32 +1,11 @@
> > -declare i32 @__clc_clk_local_mem_fence() #1
> > -declare i32 @__clc_clk_global_mem_fence() #1
> >  declare void @llvm.r600.group.barrier() #0
> >
> > -define void @barrier(i32 %flags) #2 {
> > -barrier_local_test:
> > -  %CLK_LOCAL_MEM_FENCE = call i32 @__clc_clk_local_mem_fence()
> > -  %0 = and i32 %flags, %CLK_LOCAL_MEM_FENCE
> > -  %1 = icmp ne i32 %0, 0
> > -  br i1 %1, label %barrier_local, label %barrier_global_test
> > -
> > -barrier_local:
> > -  call void @llvm.r600.group.barrier()
> > -  br label %barrier_global_test
> > -
> > -barrier_global_test:
> > -  %CLK_GLOBAL_MEM_FENCE = call i32 @__clc_clk_global_mem_fence()
> > -  %2 = and i32 %flags, %CLK_GLOBAL_MEM_FENCE
> > -  %3 = icmp ne i32 %2, 0
> > -  br i1 %3, label %barrier_global, label %done
> > -
> > -barrier_global:
> > -  call void @llvm.r600.group.barrier()
> > -  br label %done
> > -
> > -done:
> > +define void @barrier(i32 %flags) #1 {
> > +entry:
> > +  ; We should call mem_fence here, but that is not implemented for r600
> yet
> > +  tail call void @llvm.r600.group.barrier()
> >    ret void
> >  }
> >
> >  attributes #0 = { nounwind convergent }
> > -attributes #1 = { nounwind alwaysinline }
> > -attributes #2 = { nounwind convergent alwaysinline }
> > +attributes #1 = { nounwind convergent alwaysinline }
>
> ping. these are minor cleanups
>

Sorry, this looks fine to me, and I think it fixes a bug where no barrier
would get triggered if flags was 0.

--Aaroon

>
> Jan_______________________________________________
> Libclc-dev mailing list
> Libclc-dev at lists.llvm.org
> http://lists.llvm.org/cgi-bin/mailman/listinfo/libclc-dev
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/libclc-dev/attachments/20170902/4ac1fc64/attachment.html>