but uhm, how would that work if you assert(atom->subOp ==
NV50_IR_SUBOP_ATOM_ADD); inside handleSharedATOMGM107? I thought
that's only needed for fadd, not for all atoms
On Wed, Dec 5, 2018 at 3:17 PM Ilia Mirkin <imir...@alum.mit.edu> wrote:
>
> On Wed, Dec 5, 2018 at 4:59 AM Karol Herbst <kher...@redhat.com> wrote:
> >
> > On Wed, Dec 5, 2018 at 6:30 AM Ilia Mirkin <imir...@alum.mit.edu> wrote:
> > >
> > > Signed-off-by: Ilia Mirkin <imir...@alum.mit.edu>
> > > ---
> > > .../nouveau/codegen/nv50_ir_lowering_nvc0.cpp | 49 +++++++++++++++++++
> > > .../nouveau/codegen/nv50_ir_lowering_nvc0.h | 1 +
> > > 2 files changed, 50 insertions(+)
> > >
> > > diff --git
> > > a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
> > > b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
> > > index 295497be2f9..44c62820342 100644
> > > --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
> > > +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp
> > > @@ -1347,6 +1347,53 @@ NVC0LoweringPass::handleBUFQ(Instruction *bufq)
> > > return true;
> > > }
> > >
> > > +void
> > > +NVC0LoweringPass::handleSharedATOMGM107(Instruction *atom)
> > > +{
> > > + if (atom->dType != TYPE_F32)
> > > + return;
> > > +
> > > + assert(atom->subOp == NV50_IR_SUBOP_ATOM_ADD);
> > > + assert(atom->src(0).getFile() == FILE_MEMORY_SHARED);
> > > +
> > > + BasicBlock *currBB = atom->bb;
> > > + BasicBlock *addAndCasBB = atom->bb->splitBefore(atom, false);
> > > + BasicBlock *joinBB = atom->bb->splitAfter(atom);
> > > +
> > > + bld.setPosition(currBB, true);
> > > +
> > > + Value *load = atom->getDef(0), *newval = bld.getSSA();
> > > + // TODO: Use "U" subop?
> > > + bld.mkLoad(TYPE_U32, load, atom->getSrc(0)->asSym(),
> > > atom->getIndirect(0, 0));
> > > + assert(!currBB->joinAt);
> > > + currBB->joinAt = bld.mkFlow(OP_JOINAT, joinBB, CC_ALWAYS, NULL);
> > > +
> > > + bld.mkFlow(OP_BRA, addAndCasBB, CC_ALWAYS, NULL);
> > > + currBB->cfg.attach(&addAndCasBB->cfg, Graph::Edge::TREE);
> > > +
> > > + bld.setPosition(addAndCasBB, true);
> > > + bld.remove(atom);
> > > +
> > > + bld.mkOp2(OP_ADD, TYPE_F32, newval, load, atom->getSrc(1));
> > > +
> > > + // Try to do a compare-and-swap. If the old value doesn't match the
> > > loaded
> > > + // value, repeat.
> > > + Value *old = bld.getSSA();
> > > + Instruction *cas =
> > > + bld.mkOp3(OP_ATOM, TYPE_U32, old, atom->getSrc(0), load, newval);
> > > + cas->setIndirect(0, 0, atom->getIndirect(0, 0));
> > > + cas->subOp = NV50_IR_SUBOP_ATOM_CAS;
> > > + Value *pred = bld.getSSA(1, FILE_PREDICATE);
> > > + bld.mkCmp(OP_SET, CC_EQ, TYPE_U32, pred, TYPE_U32, old, load);
> > > + bld.mkMov(load, old);
> > > + bld.mkFlow(OP_BRA, addAndCasBB, CC_NOT_P, pred);
> > > + bld.mkFlow(OP_BRA, joinBB, CC_ALWAYS, NULL);
> > > + addAndCasBB->cfg.attach(&addAndCasBB->cfg, Graph::Edge::BACK);
> > > +
> > > + bld.setPosition(joinBB, false);
> > > + bld.mkFlow(OP_JOIN, NULL, CC_ALWAYS, NULL)->fixed = 1;
> > > +}
> > > +
> > > void
> > > NVC0LoweringPass::handleSharedATOMNVE4(Instruction *atom)
> > > {
> > > @@ -1559,6 +1606,8 @@ NVC0LoweringPass::handleATOM(Instruction *atom)
> > > handleSharedATOM(atom);
> > > else if (targ->getChipset() < NVISA_GM107_CHIPSET)
> > > handleSharedATOMNVE4(atom);
> > > + else
> > > + handleSharedATOMGM107(atom);
> >
> > but doesn't this makes all shared ATOM operations get lowered now?
>
> All shared ATOM operations (gm107+) call this function, yes.
_______________________________________________
mesa-dev mailing list
mesa-dev@lists.freedesktop.org
https://lists.freedesktop.org/mailman/listinfo/mesa-dev
