diff options
Diffstat (limited to 'tensorflow/compiler/xla/service/gpu/ir_emission_utils.cc')
| -rw-r--r-- | tensorflow/compiler/xla/service/gpu/ir_emission_utils.cc | 19 |
1 files changed, 11 insertions, 8 deletions
diff --git a/tensorflow/compiler/xla/service/gpu/ir_emission_utils.cc b/tensorflow/compiler/xla/service/gpu/ir_emission_utils.cc index 388aa35d7d..2799baab41 100644 --- a/tensorflow/compiler/xla/service/gpu/ir_emission_utils.cc +++ b/tensorflow/compiler/xla/service/gpu/ir_emission_utils.cc @@ -242,15 +242,17 @@ llvm::Value* EmitPrintf(tensorflow::StringPiece fmt, arguments_ptr}); } -llvm::Value* EmitShuffleDown(llvm::Value* value, llvm::Value* offset, - llvm::IRBuilder<>* builder) { +llvm::Value* EmitFullWarpShuffleDown(llvm::Value* value, llvm::Value* offset, + llvm::IRBuilder<>* builder) { int bit_width = value->getType()->getPrimitiveSizeInBits(); + llvm::Value* all_warps_mask = builder->getInt32(-1); // Special case for efficiency if (value->getType()->isFloatTy() && bit_width == 32) { return llvm_ir::EmitCallToIntrinsic( - llvm::Intrinsic::nvvm_shfl_down_f32, - {value, offset, builder->getInt32(kWarpSize - 1)}, {}, builder); + llvm::Intrinsic::nvvm_shfl_sync_down_f32, + {all_warps_mask, value, offset, builder->getInt32(kWarpSize - 1)}, {}, + builder); } // We must split values wider than 32 bits as the "shfl" instruction operates @@ -264,10 +266,11 @@ llvm::Value* EmitShuffleDown(llvm::Value* value, llvm::Value* offset, for (int i = 0; i < num_segments; ++i) { x = builder->CreateInsertElement( x, - llvm_ir::EmitCallToIntrinsic(llvm::Intrinsic::nvvm_shfl_down_i32, - {builder->CreateExtractElement(x, i), - offset, builder->getInt32(kWarpSize - 1)}, - {}, builder), + llvm_ir::EmitCallToIntrinsic( + llvm::Intrinsic::nvvm_shfl_sync_down_i32, + {all_warps_mask, builder->CreateExtractElement(x, i), offset, + builder->getInt32(kWarpSize - 1)}, + {}, builder), i); } return builder->CreateBitCast( |