/external/tensorflow/tensorflow/compiler/xla/service/ |
D | ar_crs_combiner_test.cc | 454 auto replica_groups_before = crs_before->replica_groups(); in TEST_F() 464 auto replica_groups_after = crs_after->replica_groups(); in TEST_F() 504 auto replica_groups_before = crs_before->replica_groups(); in TEST_F() 513 auto replica_groups_after = crs_after->replica_groups(); in TEST_F() 572 auto replica_groups_before = crs_before->replica_groups(); in TEST_F() 582 auto replica_groups_after = crs_after->replica_groups(); in TEST_F() 640 auto replica_groups_before = crs_before->replica_groups(); in TEST_F() 651 auto replica_groups_after = crs_after->replica_groups(); in TEST_F() 682 auto replica_groups_before = crs_before->replica_groups(); in TEST_F() 692 auto replica_groups_after = crs_after->replica_groups(); in TEST_F() [all …]
|
D | collective_ops_utils.cc | 47 int64 device_ordinal, absl::Span<const ReplicaGroup> replica_groups, in GetParticipatingReplicas() argument 53 if (replica_groups.empty()) { in GetParticipatingReplicas() 65 for (const ReplicaGroup& g : replica_groups) { in GetParticipatingReplicas()
|
D | bfloat16_normalization_test.cc | 286 std::vector<ReplicaGroup> replica_groups(1); in TEST_F() local 287 replica_groups[0].add_replica_ids(0); in TEST_F() 288 replica_groups[0].add_replica_ids(1); in TEST_F() 291 replica_groups, absl::nullopt)); in TEST_F() 315 std::vector<ReplicaGroup> replica_groups(1); in TEST_F() local 316 replica_groups[0].add_replica_ids(0); in TEST_F() 317 replica_groups[0].add_replica_ids(1); in TEST_F() 320 replica_groups, absl::nullopt)); in TEST_F()
|
D | all_reduce_simplifier.cc | 60 if (all_reduce->replica_groups().empty()) { in Run() 64 for (const auto& group : all_reduce->replica_groups()) { in Run()
|
D | hlo_replication_analysis.cc | 66 return hlo->replica_groups().empty() || hlo->replica_groups().size() == 1; in DetermineHloInstructionIsReplicated() 72 return hlo->replica_groups().empty() || hlo->replica_groups().size() == 1; in DetermineHloInstructionIsReplicated()
|
D | hlo_verifier_test.cc | 853 string ReplicaGroupsStr(std::vector<std::vector<int64>> replica_groups) { in ReplicaGroupsStr() argument 855 for (const auto& g : replica_groups) { in ReplicaGroupsStr() 863 std::vector<std::vector<int64>> replica_groups) { in MakeAllReduceComputation() argument 876 kTemplate, {{"REPLICA_GROUPS", ReplicaGroupsStr(replica_groups)}})); in MakeAllReduceComputation() 911 std::vector<std::vector<int64>> replica_groups) { in MakeAllToAllComputation() argument 925 kTemplate, {{"REPLICA_GROUPS", ReplicaGroupsStr(replica_groups)}})); in MakeAllToAllComputation()
|
D | hlo_instructions.cc | 515 const std::vector<ReplicaGroup>& replica_groups, in HloCollectiveInstruction() argument 518 replica_groups_(replica_groups) { in HloCollectiveInstruction() 536 StrCat("replica_groups=", ReplicaGroupsToString(replica_groups()))); in ExtraAttributesToStringImpl() 547 absl::c_equal(replica_groups(), casted_other.replica_groups(), in IdenticalSlowPath() 556 const std::vector<ReplicaGroup>& replica_groups, bool constrain_layout, in HloAllReduceInstruction() argument 559 replica_groups, channel_id), in HloAllReduceInstruction() 565 for (auto replica_group : replica_groups()) { in IsNoop() 604 shape, new_operands, to_apply(), replica_groups(), constrain_layout(), in CloneWithNewOperandsImpl() 610 const std::vector<ReplicaGroup>& replica_groups, in HloAllToAllInstruction() argument 614 replica_groups, channel_id), in HloAllToAllInstruction() [all …]
|
D | hlo_parser.cc | 283 bool ParseReplicaGroupsOnly(std::vector<ReplicaGroup>* replica_groups); 413 std::vector<ReplicaGroup> replica_groups; in CreateReplicaGroups() local 414 absl::c_transform(groups, std::back_inserter(replica_groups), in CreateReplicaGroups() 420 return replica_groups; in CreateReplicaGroups() 865 std::vector<ReplicaGroup> replica_groups; in ParseInstructionRhs() local 867 replica_groups = CreateReplicaGroups(*tmp_groups); in ParseInstructionRhs() 870 shape, operands, *to_apply, replica_groups, in ParseInstructionRhs() 887 std::vector<ReplicaGroup> replica_groups; in ParseInstructionRhs() local 889 replica_groups = CreateReplicaGroups(*tmp_groups); in ParseInstructionRhs() 896 shape, operands, replica_groups, channel_id, split_dimension)); in ParseInstructionRhs() [all …]
|
D | hlo_parser_test.cc | 1410 ROOT crs = f32[8]{0} all-reduce(input), replica_groups={}, to_apply=add 1428 ROOT all-reduce = f32[128,32]{0,1} all-reduce(input), replica_groups={{0,1},{2,3}}, to_apply=add 1446 ROOT crs = f32[8]{0} all-reduce(input), replica_groups={}, constrain_layout=true, to_apply=add 1464 crs.1 = f32[8]{0} all-reduce(input), channel_id=1, replica_groups={{0}}, to_apply=add 1465 ROOT crs.0 = f32[8]{0} all-reduce(input), channel_id=1, replica_groups={{0}}, to_apply=add 1477 ROOT a2a = (f32[128,32]{0,1}) all-to-all(input), replica_groups={} 1490 ROOT a2a = (f32[128,32]{0,1}, f32[128,32]{0,1}) all-to-all(p0, p1), replica_groups={{1,2},{3,0}}
|
D | hlo_instruction.cc | 401 std::vector<ReplicaGroup>(proto.replica_groups().begin(), in CreateFromProto() 402 proto.replica_groups().end()), in CreateFromProto() 424 std::vector<ReplicaGroup>(proto.replica_groups().begin(), in CreateFromProto() 425 proto.replica_groups().end()), in CreateFromProto() 913 const std::vector<ReplicaGroup>& replica_groups, bool constrain_layout, in CreateAllReduce() argument 916 shape, operands, reduce_computation, replica_groups, constrain_layout, in CreateAllReduce() 922 const std::vector<ReplicaGroup>& replica_groups, in CreateAllToAll() argument 926 shape, operands, replica_groups, channel_id, split_dimension); in CreateAllToAll() 3394 string ReplicaGroupsToString(const std::vector<ReplicaGroup>& replica_groups) { in ReplicaGroupsToString() argument 3396 replica_group_str.reserve(replica_groups.size()); in ReplicaGroupsToString() [all …]
|
D | hlo_instruction.h | 610 const std::vector<ReplicaGroup>& replica_groups, bool constrain_layout, 643 const std::vector<ReplicaGroup>& replica_groups, 1705 const std::vector<ReplicaGroup>& replica_groups() const; 2032 string ReplicaGroupsToString(const std::vector<ReplicaGroup>& replica_groups);
|
D | hlo_verifier.cc | 192 for (const ReplicaGroup& g : hlo->replica_groups()) { in CheckReplicaGroups() 231 if (hlo->replica_groups().empty()) { in HandleAllToAll() 241 const int64 split_count = hlo->replica_groups().empty() in HandleAllToAll() 243 : hlo->replica_groups()[0].replica_ids_size(); in HandleAllToAll() 244 for (const ReplicaGroup& g : hlo->replica_groups()) { in HandleAllToAll()
|
D | collective_ops_utils.h | 42 int64 device_ordinal, absl::Span<const ReplicaGroup> replica_groups,
|
D | hlo_instructions.h | 311 const std::vector<ReplicaGroup>& replica_groups() const { in replica_groups() function 319 const std::vector<ReplicaGroup>& replica_groups, 339 const std::vector<ReplicaGroup>& replica_groups, bool constrain_layout, 384 const std::vector<ReplicaGroup>& replica_groups,
|
D | ar_crs_combiner.cc | 76 instruction->replica_groups().size() == num_replicas_ && in MatchesArCrsPattern()
|
/external/tensorflow/tensorflow/compiler/xla/python/ |
D | xla_client.py | 1008 def AllReduce(self, operand, computation, replica_groups=None): argument 1022 replica_groups_protos = _get_replica_groups_protos(replica_groups) 1030 replica_groups=None): argument 1046 replica_groups_protos = _get_replica_groups_protos(replica_groups) 1047 if not replica_groups: 1050 split_count = len(replica_groups[0]) 1051 if not all(split_count == len(g) for g in replica_groups): 1056 def CrossReplicaSum(self, operand, replica_groups=None): argument 1069 replica_groups_protos = _get_replica_groups_protos(replica_groups) 1901 def _get_replica_groups_protos(replica_groups): argument [all …]
|
/external/tensorflow/tensorflow/compiler/xla/tests/ |
D | collective_ops_test.cc | 44 const Shape& shape, std::vector<std::vector<int64>> replica_groups, in MakeCrsModule() argument 65 for (const auto& g : replica_groups) { in MakeCrsModule()
|
/external/tensorflow/tensorflow/compiler/xla/client/ |
D | xla_builder.h | 512 absl::Span<const ReplicaGroup> replica_groups = {}); 516 absl::Span<const ReplicaGroup> replica_groups = {}, 522 const std::vector<ReplicaGroup>& replica_groups); 923 absl::Span<const ReplicaGroup> replica_groups); 925 absl::Span<const ReplicaGroup> replica_groups, 930 const std::vector<ReplicaGroup>& replica_groups); 1654 absl::Span<const ReplicaGroup> replica_groups = {}); 1676 absl::Span<const ReplicaGroup> replica_groups = {}, 1683 const std::vector<ReplicaGroup>& replica_groups = {});
|
D | xla_builder.cc | 2084 XlaOp operand, absl::Span<const ReplicaGroup> replica_groups) { in CrossReplicaSum() argument 2108 return AllReduce(operand, computation, replica_groups, in CrossReplicaSum() 2114 absl::Span<const ReplicaGroup> replica_groups, in AllReduce() argument 2166 for (const ReplicaGroup& group : replica_groups) { in AllReduce() 2191 const std::vector<ReplicaGroup>& replica_groups) { in AllToAll() argument 2227 for (const ReplicaGroup& group : replica_groups) { in AllToAll() 3302 absl::Span<const ReplicaGroup> replica_groups) { in CrossReplicaSum() argument 3303 return operand.builder()->CrossReplicaSum(operand, replica_groups); in CrossReplicaSum() 3307 absl::Span<const ReplicaGroup> replica_groups, in AllReduce() argument 3310 return operand.builder()->AllReduce(operand, computation, replica_groups, in AllReduce() [all …]
|
/external/tensorflow/tensorflow/compiler/mlir/xla/ |
D | mlir_hlo_to_hlo.cc | 487 auto replica_groups = Convert_replica_groups(op.replica_groups()); in ExportXlaOp() local 490 xla::AllReduce(value_map[op.operand()], computation, replica_groups, in ExportXlaOp() 496 replica_groups, channel_id); in ExportXlaOp()
|
/external/tensorflow/tensorflow/compiler/mlir/xla/tests/translate/ |
D | export.mlir | 25 replica_groups = dense<[[0, 2, 4, 6], [1, 3, 5, 7]]> : tensor<2x4xi64>, 39 // CHECK-SAME: replica_groups={{[{][{]}}0,2,4,6},{1,3,5,7{{[}][}]}} 346 …%1 = "xla_hlo.cross-replica-sum"(%arg0) {replica_groups = dense<[[0, 2, 4, 6], [1, 3, 5, 7]]> : te… 356 // CHECK-SAME: replica_groups={{[{][{]}}0,2,4,6},{1,3,5,7{{[}][}]}}
|
/external/tensorflow/tensorflow/compiler/xla/service/gpu/ |
D | nccl_all_reduce_thunk.cc | 497 GetParticipatingReplicas(device_ordinal, instr->replica_groups(), in ExecuteOnStream()
|
/external/tensorflow/tensorflow/compiler/mlir/xla/ir/ |
D | hlo_ops.td | 527 I64ElementsAttr:$replica_groups, 544 I64ElementsAttr:$replica_groups 861 I64ElementsAttr:$replica_groups
|
D | hlo_ops_base.td | 685 - replica_groups: each ReplicaGroup contains a list of replica id 689 `replica_groups` = {{1,2,3}, {4,5,0}} means that an AllToAll will be applied 693 also 4, 5, 0. If `replica_groups` is empty, all replicas belong to one
|
/external/tensorflow/tensorflow/compiler/mlir/xla/tests/ |
D | ops.mlir | 22 replica_groups = dense<[[0, 1, 2, 3]]> : tensor<1x4xi64> 35 replica_groups = dense<[[0, 1, 2, 3, 4]]> : tensor<1x5xi64> 48 replica_groups = dense<[[0, 1, 2, 3, 4]]> : tensor<1x5xi64>
|