From 839a469166b9c0b9959620eda85a6481f9efa15f Mon Sep 17 00:00:00 2001 From: Samuel Pitoiset Date: Mon, 7 Mar 2016 18:56:21 +0100 Subject: [PATCH] nvc0/ir: do not lower shared+atomics on GM107+ For Maxwell, the ATOMS instruction can be used to perform atomic operations on shared memory instead of this load/store lowering pass. Signed-off-by: Samuel Pitoiset Reviewed-by: Ilia Mirkin --- .../nouveau/codegen/nv50_ir_lowering_nvc0.cpp | 17 ++++++++++------- 1 file changed, 10 insertions(+), 7 deletions(-) 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 fb2aec62144..ce83618d681 100644 --- a/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp +++ b/src/gallium/drivers/nouveau/codegen/nv50_ir_lowering_nvc0.cpp @@ -1288,11 +1288,12 @@ NVC0LoweringPass::handleATOM(Instruction *atom) sv = SV_LBASE; break; case FILE_MEMORY_SHARED: - if (targ->getChipset() >= NVISA_GK104_CHIPSET) { - handleSharedATOMNVE4(atom); - } else { + // For Fermi/Kepler, we have to use ld lock/st unlock to perform atomic + // operations on shared memory. For Maxwell, ATOMS is enough. + if (targ->getChipset() < NVISA_GK104_CHIPSET) handleSharedATOM(atom); - } + else if (targ->getChipset() < NVISA_GM107_CHIPSET) + handleSharedATOMNVE4(atom); return true; default: assert(atom->src(0).getFile() == FILE_MEMORY_GLOBAL); @@ -1320,9 +1321,11 @@ NVC0LoweringPass::handleATOM(Instruction *atom) bool NVC0LoweringPass::handleCasExch(Instruction *cas, bool needCctl) { - if (cas->src(0).getFile() == FILE_MEMORY_SHARED) { - // ATOM_CAS and ATOM_EXCH are handled in handleSharedATOM(). - return false; + if (targ->getChipset() < NVISA_GM107_CHIPSET) { + if (cas->src(0).getFile() == FILE_MEMORY_SHARED) { + // ATOM_CAS and ATOM_EXCH are handled in handleSharedATOM(). + return false; + } } if (cas->subOp != NV50_IR_SUBOP_ATOM_CAS && -- 2.30.2