0fd465c057
This new function should deduplicate some common code, be slightly faster in the most common case, and remove a potential source of error / make it more future proof (supporting multiple computations). In the simple case (no replica groups and only one computation), we can avoid looking up the replica ID or materializing the intermediate replica IDs list, and just return the devices directly. Otherwise, if we're looking up the replica ID, we get the computation ID for free, so we might as well use it, avoiding a potential source of bugs where multiple computations are used. PiperOrigin-RevId: 345642270 Change-Id: If4da9fe00b104dd754dc1f6d18ba3c56d397f4b5
109 lines
4.2 KiB
C++
109 lines
4.2 KiB
C++
/* Copyright 2019 The TensorFlow Authors. All Rights Reserved.
|
|
|
|
Licensed under the Apache License, Version 2.0 (the "License");
|
|
you may not use this file except in compliance with the License.
|
|
You may obtain a copy of the License at
|
|
|
|
http://www.apache.org/licenses/LICENSE-2.0
|
|
|
|
Unless required by applicable law or agreed to in writing, software
|
|
distributed under the License is distributed on an "AS IS" BASIS,
|
|
WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
|
|
See the License for the specific language governing permissions and
|
|
limitations under the License.
|
|
==============================================================================*/
|
|
|
|
#include "tensorflow/compiler/xla/service/collective_ops_utils.h"
|
|
|
|
#include "tensorflow/compiler/xla/service/global_device_id.h"
|
|
|
|
namespace xla {
|
|
|
|
absl::optional<ReductionKind> MatchReductionComputation(
|
|
const HloComputation* computation) {
|
|
namespace m = match;
|
|
const HloInstruction* root = computation->root_instruction();
|
|
|
|
auto match_opcode = [&](HloOpcode opcode) {
|
|
return Match(
|
|
root, m::Op()
|
|
.WithOpcode(opcode)
|
|
.WithBinaryOperandsAnyOrder(m::Parameter(0), m::Parameter(1))
|
|
.WithShape(m::Shape().IsEffectiveScalar()));
|
|
};
|
|
|
|
// Match the operation to a reduction kind. We can represent and/or of pred as
|
|
// min/max. This works because pred is stored as an 8-bit int of value 0 or 1.
|
|
PrimitiveType type = computation->root_instruction()->shape().element_type();
|
|
if (match_opcode(HloOpcode::kAdd)) {
|
|
return ReductionKind::SUM;
|
|
} else if (match_opcode(HloOpcode::kMultiply)) {
|
|
return ReductionKind::PRODUCT;
|
|
} else if (match_opcode(HloOpcode::kMinimum) ||
|
|
(type == PRED && match_opcode(HloOpcode::kAnd))) {
|
|
return ReductionKind::MIN;
|
|
} else if (match_opcode(HloOpcode::kMaximum) ||
|
|
(type == PRED && match_opcode(HloOpcode::kOr))) {
|
|
return ReductionKind::MAX;
|
|
} else {
|
|
return absl::nullopt;
|
|
}
|
|
}
|
|
|
|
StatusOr<std::vector<int>> GetParticipatingReplicas(
|
|
int replica_id, int total_replica_count,
|
|
absl::Span<const ReplicaGroup> replica_groups) {
|
|
// Empty replica_groups() means that all replicas participate.
|
|
if (replica_groups.empty()) {
|
|
std::vector<int> all_replicas(total_replica_count);
|
|
absl::c_iota(all_replicas, 0);
|
|
return all_replicas;
|
|
}
|
|
|
|
// Figure out the other replicas that go together with this one.
|
|
absl::optional<ReplicaGroup> replica_group;
|
|
for (const ReplicaGroup& g : replica_groups) {
|
|
if (absl::c_linear_search(g.replica_ids(), replica_id)) {
|
|
TF_RET_CHECK(!replica_group.has_value())
|
|
<< "Replica " << replica_id << " appears twice in replica groups";
|
|
replica_group = g;
|
|
}
|
|
}
|
|
TF_RET_CHECK(replica_group.has_value())
|
|
<< "Replica " << replica_id << " doesn't appear in replica groups";
|
|
return std::vector<int>(replica_group->replica_ids().begin(),
|
|
replica_group->replica_ids().end());
|
|
}
|
|
|
|
StatusOr<std::vector<GlobalDeviceId>> GetParticipatingDevices(
|
|
GlobalDeviceId device_id, const DeviceAssignment& device_assignment,
|
|
int total_replica_count, absl::Span<const ReplicaGroup> replica_groups) {
|
|
std::vector<GlobalDeviceId> participants;
|
|
// Fast path for common case, avoiding logical IDs lookup.
|
|
if (replica_groups.empty() && device_assignment.computation_count() == 1) {
|
|
participants.reserve(total_replica_count);
|
|
for (int replica_id = 0; replica_id < total_replica_count; ++replica_id) {
|
|
participants.emplace_back(
|
|
device_assignment(replica_id, /*computation_id=*/0));
|
|
}
|
|
return participants;
|
|
}
|
|
|
|
std::pair<int, int> logical_ids;
|
|
TF_ASSIGN_OR_RETURN(logical_ids,
|
|
device_assignment.LogicalIdsForDevice(device_id));
|
|
int replica_id = logical_ids.first;
|
|
int computation_id = logical_ids.second;
|
|
TF_ASSIGN_OR_RETURN(std::vector<int> participating_replicas,
|
|
GetParticipatingReplicas(replica_id, total_replica_count,
|
|
replica_groups));
|
|
|
|
participants.reserve(participating_replicas.size());
|
|
for (int replica_id : participating_replicas) {
|
|
participants.emplace_back(device_assignment(replica_id, computation_id));
|
|
}
|
|
return participants;
|
|
}
|
|
|
|
} // end namespace xla
|