Google Git
Sign in
android / platform / external / tensorflow / ea9d358b5aea41f40e47cf85900259689f90ae07 / . / tensorflow / compiler / xla / service / buffer_assignment.cc
blob: 3ae7235d8871ecef23a5f29598fff2d824d5d64e [file] [log] [blame]
/* Copyright 2017 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.
==============================================================================*/
// Defines the data returned by the XLA buffer assignment packages.
#include "tensorflow/compiler/xla/service/buffer_assignment.h"
#include <algorithm>
#include <deque>
#include <ostream>
#include <utility>
#include "absl/container/flat_hash_map.h"
#include "absl/container/flat_hash_set.h"
#include "absl/memory/memory.h"
#include "absl/strings/str_cat.h"
#include "absl/strings/str_format.h"
#include "tensorflow/compiler/xla/map_util.h"
#include "tensorflow/compiler/xla/service/buffer_value_containers.h"
#include "tensorflow/compiler/xla/service/heap_simulator.h"
#include "tensorflow/compiler/xla/service/hlo.pb.h"
#include "tensorflow/compiler/xla/service/hlo_alias_analysis.h"
#include "tensorflow/compiler/xla/service/hlo_buffer.h"
#include "tensorflow/compiler/xla/service/hlo_opcode.h"
#include "tensorflow/compiler/xla/shape_util.h"
#include "tensorflow/compiler/xla/status_macros.h"
#include "tensorflow/compiler/xla/types.h"
#include "tensorflow/compiler/xla/util.h"
#include "tensorflow/core/lib/core/errors.h"
#include "tensorflow/core/lib/hash/hash.h"
#include "tensorflow/core/lib/strings/numbers.h"
namespace xla {
namespace {
using absl::flat_hash_map;
using absl::flat_hash_set;
using absl::StrAppend;
using absl::StrAppendFormat;
using ::tensorflow::strings::HumanReadableNumBytes;
// Given the interference map of a graph (the list of interfering node indices
// for each node), perform graph coloring such that interfering nodes are
// assigned to different colors. Returns the assigned color of the nodes, where
// the colors are represented as integer values [0, color_count).
std::vector<int64> ColorInterferenceGraph(
const std::vector<std::vector<int64>>& interference_map) {
const int64 node_count = interference_map.size();
// Sort the nodes such that we assign nodes with more interference first. This
// relies on the common heuristic of assigning the most constrained node
// first, but it would be good to investigate other ordering heuristics too.
std::vector<int64> nodes(node_count);
std::iota(nodes.begin(), nodes.end(), 0);
absl::c_sort(nodes, [&interference_map](const int64 i, const int64 j) {
return interference_map[i].size() > interference_map[j].size();
});
const int64 kColorUnassigned = -1;
std::vector<int64> assigned_colors(node_count, kColorUnassigned);
for (int64 node : nodes) {
// Mark the colors that are already assigned to the neighbors.
std::vector<bool> available_colors(node_count, true);
for (int64 neighbor : interference_map[node]) {
int64 color = assigned_colors[neighbor];
if (color != kColorUnassigned) {
available_colors[color] = false;
}
}
// Find the color that is not yet assigned to the neighbors.
int64 color = kColorUnassigned;
for (color = 0; color < available_colors.size(); ++color) {
if (available_colors[color]) {
break;
}
}
CHECK_NE(color, kColorUnassigned);
assigned_colors[node] = color;
}
return assigned_colors;
}
// If an hlo buffer contains an entry parameter, the buffer is read-only unless
// it is aliased with an output.
bool HloBufferIsReadOnly(const HloBuffer& buffer) {
for (const HloValue* value : buffer.values()) {
const HloInstruction* instruction = value->instruction();
const HloModule* module = instruction->parent()->parent();
const bool is_entry_parameter =
instruction->opcode() == HloOpcode::kParameter &&
instruction->parent() == module->entry_computation();
if (is_entry_parameter) {
bool parameter_has_alias =
module->input_output_alias_config().ParameterHasAlias(
instruction->parameter_number(), value->index());
// The parameter doesn't have an alias, it must be read-only.
if (!parameter_has_alias) {
return true;
}
}
}
return false;
}
} // namespace
Status GatherComputationsByAllocationType(
const HloModule* module,
std::vector<const HloComputation*>* thread_local_computations,
std::vector<const HloComputation*>* global_computations) {
// Create a worklist of computations paired with whether the allocation must
// be thread-local.
std::deque<std::pair<const HloComputation*, bool>> worklist;
worklist.push_back(std::make_pair(module->entry_computation(),
/*is_thread_local*/ false));
// Sets for quickly checking membership. Computations are returned in vectors
// for stable iteration.
flat_hash_set<const HloComputation*> thread_local_set;
flat_hash_set<const HloComputation*> global_set;
while (!worklist.empty()) {
auto worklist_front = worklist.front();
worklist.pop_front();
const HloComputation* computation = worklist_front.first;
bool is_thread_local = worklist_front.second;
bool in_thread_local_set = thread_local_set.contains(computation);
bool in_global_set = global_set.contains(computation);
// If the computation has already been added to the respective set, then
// nothing to do.
if ((is_thread_local && in_thread_local_set) ||
(!is_thread_local && in_global_set)) {
continue;
}
// If the computation has already been added to the other set this is an
// error condition because the global call to the computation (eg,
// while/call) may return a reference to one of the thread-local buffers to
// the calling computation which will become a dangling reference when the
// thread-local is deallocated with the call return.
if ((is_thread_local && in_global_set) ||
(!is_thread_local && in_thread_local_set)) {
return InvalidArgument(
"computation %s has conflicting allocation requirements (global "
"and thread-local)",
computation->name());
}
if (is_thread_local) {
thread_local_set.insert(computation);
} else {
global_set.insert(computation);
}
for (auto* instruction : computation->instructions()) {
for (HloComputation* subcomputation :
instruction->called_computations()) {
switch (instruction->opcode()) {
case HloOpcode::kCall:
case HloOpcode::kConditional:
case HloOpcode::kWhile:
// Call and while must be called from a computation with global
// allocations as they may return references to buffers inside the
// called computation which cannot be thread-local.
if (is_thread_local) {
return InvalidArgument(
"computation %s cannot contain call/while op because it "
"requires thread-local buffer allocations",
computation->name());
}
worklist.push_back(std::make_pair(subcomputation,
false)); // Not thread local.
break;
case HloOpcode::kAllReduce:
case HloOpcode::kMap:
case HloOpcode::kReduce:
case HloOpcode::kReduceWindow:
case HloOpcode::kScatter:
case HloOpcode::kSelectAndScatter:
case HloOpcode::kSort:
case HloOpcode::kFusion:
// Map/reduce etc computations are always thread-local.
worklist.push_back(std::make_pair(subcomputation,
true)); // Thread local.
break;
default:
return InternalError("Unexpected calling opcode: %s",
HloOpcodeString(instruction->opcode()));
}
}
}
}
// Add the computations to the vectors in post order.
for (auto* computation : module->MakeComputationPostOrder()) {
if (thread_local_set.contains(computation)) {
thread_local_computations->push_back(computation);
} else if (global_set.contains(computation)) {
global_computations->push_back(computation);
}
// If the computation is not reachable from the entry computation, then it
// will not appear in either thread_local_set or global_set. We don't bother
// assigning buffers for these.
}
return Status::OK();
}
string BufferAllocation::Slice::ToString() const {
return absl::StrCat("{index:", index(), ", offset:", offset_,
", size:", size_, "}");
}
BufferAllocation::Slice BufferAllocation::GetSlice(
const HloValue& buffer) const {
const OffsetSize os = FindOrDie(assigned_buffers_, &buffer);
return Slice(this, os.offset, os.size);
}
void BufferAllocation::AddAssignment(const HloValue& buffer, int64 offset,
int64 size) {
VLOG(4) << "Adding the following buffer to allocation #" << index() << ": "
<< buffer;
CHECK(!assigned_buffers_.contains(&buffer))
<< "LogicalBuffer " << buffer << " already assigned to allocation "
<< index_;
CHECK_LE(offset, size_) << "LogicalBuffer " << buffer
<< " offset out of range";
CHECK_LE(offset + size, size_)
<< "LogicalBuffer " << buffer
<< " size out of range at offset: " << offset << " with size: " << size;
CHECK_EQ(buffer.color(), color())
<< "Buffer color " << buffer.color() << " for buffer " << buffer
<< " does not match allocation color " << color() << ".";
OffsetSize offset_size;
offset_size.offset = offset;
offset_size.size = size;
assigned_buffers_.emplace(&buffer, offset_size);
}
BufferAllocationProto BufferAllocation::ToProto() const {
BufferAllocationProto proto;
proto.set_index(index_);
proto.set_size(size_);
proto.set_is_thread_local(is_thread_local_);
proto.set_is_tuple(is_tuple_);
proto.set_color(color_.value());
if (is_entry_computation_parameter_) {
proto.set_is_entry_computation_parameter(true);
for (int64 idx : param_shape_index()) {
proto.add_parameter_shape_index(idx);
}
proto.set_parameter_number(parameter_number_);
}
proto.set_is_constant(is_constant_);
proto.set_maybe_live_out(maybe_live_out_);
for (const auto& buffer_offset_size : assigned_buffers_) {
BufferAllocationProto::Assigned* proto_assigned = proto.add_assigned();
proto_assigned->set_logical_buffer_id(buffer_offset_size.first->id());
proto_assigned->set_offset(buffer_offset_size.second.offset);
proto_assigned->set_size(buffer_offset_size.second.size);
}
absl::c_sort(*proto.mutable_assigned(),
[](const BufferAllocationProto::Assigned& assign1,
const BufferAllocationProto::Assigned& assign2) {
return assign1.logical_buffer_id() <
assign2.logical_buffer_id();
});
return proto;
}
string BufferAllocation::ToString() const {
string output;
StrAppendFormat(&output, "allocation %d: %p, size %d", index_, this, size());
if (color().value() != 0) {
StrAppend(&output, ", color ", color().value());
}
if (is_entry_computation_parameter()) {
StrAppend(&output, ", parameter ", parameter_number(), " at ShapeIndex ",
param_shape_index().ToString());
}
if (is_constant()) {
StrAppend(&output, ", constant");
}
if (is_thread_local()) {
StrAppend(&output, ", thread-local");
}
if (maybe_live_out()) {
StrAppend(&output, ", maybe-live-out");
}
if (IsPreallocatedTempBuffer()) {
StrAppend(&output, ", preallocated-temp");
}
StrAppend(&output, ":\n");
// Dump the assigned buffers ordered by id.
std::vector<const HloValue*> sorted_buffers;
for (const auto& buffer_offset_size : assigned_buffers_) {
sorted_buffers.push_back(buffer_offset_size.first);
}
absl::c_sort(sorted_buffers, [](const HloValue* a, const HloValue* b) {
return a->id() < b->id();
});
for (const HloValue* buffer : sorted_buffers) {
const OffsetSize& offset_size = FindOrDie(assigned_buffers_, buffer);
StrAppend(&output, absl::StrFormat(
" %s [%d,%d]: %s\n", buffer->ToString(),
offset_size.offset, offset_size.size,
ShapeUtil::HumanStringWithLayout(buffer->shape())));
}
return output;
}
std::ostream& operator<<(std::ostream& out, const BufferAllocation& buffer) {
out << buffer.ToString();
return out;
}
std::ostream& operator<<(std::ostream& out, const BufferAllocation::Slice& s) {
out << s.ToString();
return out;
}
bool BufferAssignment::HasAllocation(const HloValue& value) const {
return allocation_index_for_value_.contains(&value);
}
bool BufferAssignment::HasAllocation(const HloBuffer& buffer) const {
return allocation_index_for_value_.contains(buffer.values()[0]);
}
const BufferAllocation& BufferAssignment::GetAssignedAllocation(
const HloValue& value) const {
CHECK(HasAllocation(value));
return GetAllocation(allocation_index_for_value_.at(&value));
}
const BufferAllocation& BufferAssignment::GetAssignedAllocation(
const HloBuffer& hlo_buffer) const {
return GetAssignedAllocation(*hlo_buffer.values()[0]);
}
BufferAllocation* BufferAssignment::GetMutableAssignedAllocation(
const HloBuffer& buffer) {
return const_cast<BufferAllocation*>(&GetAssignedAllocation(buffer));
}
std::set<BufferAllocation::Slice> BufferAssignment::GetAllSlices(
const HloInstruction* instruction, const ShapeIndex& index) const {
std::set<BufferAllocation::Slice> result;
for (const HloValue* value :
dataflow_analysis().GetValueSet(instruction, index).values()) {
if (HasAllocation(*value)) {
result.insert(GetAssignedAllocation(*value).GetSlice(*value));
}
}
return result;
}
const BufferAllocation& BufferAssignment::GetAllocation(
BufferAllocation::Index index) const {
CHECK_GE(index, 0);
CHECK_LT(index, allocations_.size());
return allocations_[index];
}
const BufferAllocation* BufferAssignment::GetInstructionAllocation(
const HloInstruction* hlo, const ShapeIndex& shape_index) const {
const HloValue* value =
dataflow_analysis().GetValueSet(hlo, shape_index).values()[0];
if (!HasAllocation(*value)) {
return nullptr;
}
const BufferAllocation& instruction_allocation =
GetAssignedAllocation(*value);
return &instruction_allocation;
}
BufferAllocation* BufferAssignment::GetMutableAllocation(
BufferAllocation::Index index) {
return const_cast<BufferAllocation*>(&GetAllocation(index));
}
bool BufferAssignment::HasAllocationAt(const HloInstruction* instruction,
const ShapeIndex& index) const {
for (const HloValue* value :
dataflow_analysis().GetValueSet(instruction, index).values()) {
if (allocation_index_for_value_.contains(value)) {
return true;
}
}
return false;
}
bool BufferAssignment::HasTopLevelAllocation(
const HloInstruction* instruction) const {
return HasAllocationAt(instruction, /*index=*/{});
}
StatusOr<BufferAllocation::Slice> BufferAssignment::GetUniqueSlice(
const HloInstruction* instruction, const ShapeIndex& index) const {
VLOG(3) << "Trying to find unique slice for " << instruction->name() << " ["
<< index << "]";
BufferAllocation::Slice result;
for (const HloValue* value :
dataflow_analysis().GetValueSet(instruction, index).values()) {
VLOG(3) << "Examining value " << *value;
if (HasAllocation(*value)) {
VLOG(3) << "Has allocation";
const BufferAllocation::Slice slice =
GetAssignedAllocation(*value).GetSlice(*value);
if (result.allocation() == nullptr) {
result = slice;
} else if (result != slice) {
return FailedPrecondition(
"BufferAllocation::Slice for instruction %s at index %s cannot "
"be determined at compile-time.",
instruction->name(), index.ToString());
}
} else {
VLOG(3) << "No allocation";
}
}
if (result.allocation() == nullptr) {
return FailedPrecondition(
"BufferAllocation::Slice not assigned for instruction %s at index %s",
instruction->name(), index.ToString());
}
return result;
}
StatusOr<BufferAllocation::Slice> BufferAssignment::GetUniqueTopLevelSlice(
const HloInstruction* instruction) const {
return GetUniqueSlice(instruction, /*index=*/{});
}
bool BufferAssignment::SharesSliceAtIndex(
const HloInstruction* hlo_a, const ShapeIndex& shape_index_a,
const HloInstruction* hlo_b, const ShapeIndex& shape_index_b) const {
return GetUniqueSlice(hlo_a, shape_index_a).ConsumeValueOrDie() ==
GetUniqueSlice(hlo_b, shape_index_b).ConsumeValueOrDie();
}
bool BufferAssignment::HaveDisjointSlices(const HloInstruction* hlo_a,
const HloInstruction* hlo_b) const {
using SliceSet = flat_hash_set<BufferAllocation::Slice>;
// Gets the slices all of instr's subshapes. If any subshape doesn't have an
// assigned slice, returns the empty set.
auto collect_slices = [&](const HloInstruction* instr) -> SliceSet {
SliceSet slices;
Status status = ShapeUtil::ForEachSubshapeWithStatus(
instr->shape(),
[&](const Shape& /*subshape*/, const ShapeIndex& index) {
auto shape_slices = GetAllSlices(instr, index);
if (shape_slices.empty()) {
return InvalidArgument("No slices assigned to part of instr.");
}
slices.insert(shape_slices.begin(), shape_slices.end());
return Status::OK();
});
if (!status.ok()) {
return {};
}
return slices;
};
SliceSet slices_a = collect_slices(hlo_a);
SliceSet slices_b = collect_slices(hlo_b);
// hlo_a and hlo_b have disjoint slices if collect_slices succeeded (i.e.
// didn't return the empty set) for both HLOs, and the two resulting sets of
// slices are disjoint.
return !slices_a.empty() && !slices_b.empty() &&
absl::c_none_of(slices_a, [&](const BufferAllocation::Slice& slice) {
return slices_b.contains(slice);
});
}
StatusOr<BufferAllocation::Slice>
BufferAssignment::GetUniqueTopLevelOutputSlice() const {
return GetUniqueTopLevelSlice(
module_->entry_computation()->root_instruction());
}
BufferAllocation* BufferAssignment::NewEmptyAllocation(
int64 size, LogicalBuffer::Color color) {
BufferAllocation::Index index = allocations_.size();
allocations_.emplace_back(index, size, color);
BufferAllocation* allocation = &allocations_.back();
return allocation;
}
BufferAllocation* BufferAssignment::NewAllocation(const HloBuffer& buffer,
int64 size) {
BufferAllocation* allocation = NewEmptyAllocation(size, buffer.color());
AddAssignment(allocation, buffer, /*offset=*/0, size);
allocation->peak_buffers_.push_back(buffer.values()[0]);
return allocation;
}
void BufferAssignment::AddAssignment(BufferAllocation* allocation,
const HloBuffer& buffer, int64 offset,
int64 size) {
CHECK(allocation->is_reusable() || allocation->assigned_buffers().empty())
<< "Non-reusable allocation already assigned a buffer: "
<< allocation->ToString();
for (const HloValue* buffer_value : buffer.values()) {
CHECK(!allocation_index_for_value_.contains(buffer_value))
<< "BufferValue " << buffer_value << " already has an allocation.";
allocation->AddAssignment(*buffer_value, offset, size);
allocation_index_for_value_[buffer_value] = allocation->index();
}
if (alias_analysis().BufferLivesOut(buffer)) {
VLOG(3) << "HloBuffer lives out" << buffer.ToString();
VLOG(3) << "Set maybe live out: " << allocation->ToString();
allocation->set_maybe_live_out(true);
}
}
void BufferAssignment::AddAssignment(BufferAllocation* allocation,
const HloValue& value, int64 offset,
int64 size) {
allocation->AddAssignment(value, offset, size);
allocation_index_for_value_[&value] = allocation->index();
const HloValue& hlo_value =
*CHECK_NOTNULL(dynamic_cast<const HloValue*>(&value));
if (alias_analysis().ValueLivesOut(hlo_value)) {
VLOG(3) << "HloValue lives out: " << hlo_value.ToString();
VLOG(3) << "Set maybe live out: " << allocation->ToString();
allocation->set_maybe_live_out(true);
}
}
// Combines allocations of temporary buffers of the same color into one big
// BufferAllocation.
void BufferAssignment::CombineTempAllocations() {
VLOG(1) << "CombineTempAllocations()";
flat_hash_map<BufferValue::Color, BufferAllocation,
BufferValue::Color::Hasher>
combined_allocation_map;
// Move all temp allocations into a single run at the end of the allocations
// vector.
const auto first_temp_it =
std::partition(allocations_.begin(), allocations_.end(),
[](const BufferAllocation& allocation) {
return !allocation.IsPreallocatedTempBuffer();
});
// Walk over the run of temp allocations, collecting the allocations belonging
// to the same color.
if (first_temp_it != allocations_.end()) {
for (auto it = first_temp_it; it != allocations_.end(); ++it) {
const BufferAllocation& temp_allocation = *it;
BufferValue::Color color = temp_allocation.color();
auto combined_it = combined_allocation_map.find(color);
if (combined_it == combined_allocation_map.end()) {
// We have found the first temp allocation of this color. Collect
// the other temp allocations of the same color into it.
VLOG(1) << "Combined temp allocation for color " << color
<< " is: " << temp_allocation;
combined_allocation_map.emplace(color, temp_allocation);
continue;
}
auto* combined_allocation = &combined_it->second;
VLOG(1) << "Combined allocation absorbing temp allocation: "
<< temp_allocation;
// Each temp allocation is placed end-to-end, accounting for alignment.
// The offset of each buffer in the combined allocation is computed from
// the base offset of the allocation.
int64 alignment = color_alignment_(color);
const int64 base =
RoundUpToNearest(combined_allocation->size(), alignment);
combined_allocation->set_size(base + temp_allocation.size());
for (const auto& buffer_offset_size : temp_allocation.assigned_buffers_) {
const HloValue* value = buffer_offset_size.first;
const int64 offset = buffer_offset_size.second.offset;
const int64 size = buffer_offset_size.second.size;
combined_allocation->AddAssignment(*value, base + offset, size);
}
if (!temp_allocation.HeapTraces().empty()) {
CHECK_EQ(temp_allocation.HeapTraces().size(), 1);
combined_allocation->AddHeapTrace(temp_allocation.HeapTraces().front());
}
combined_allocation->peak_buffers_.insert(
combined_allocation->peak_buffers_.end(),
temp_allocation.peak_buffers_.begin(),
temp_allocation.peak_buffers_.end());
}
// Replace all existing temporary allocations with the new combined
// allocations.
allocations_.erase(first_temp_it, allocations_.end());
for (auto& combined : combined_allocation_map) {
allocations_.push_back(combined.second);
temp_allocation_total_size_ += combined.second.size();
}
}
// Update allocation indices to their new positions.
allocation_index_for_value_.erase(allocation_index_for_value_.begin(),
allocation_index_for_value_.end());
for (size_t index = 0; index < allocations_.size(); ++index) {
BufferAllocation* allocation = &allocations_[index];
allocation->set_index(index);
for (const auto& buffer_offset_size : allocation->assigned_buffers_) {
const HloValue* value = buffer_offset_size.first;
allocation_index_for_value_[value] = index;
}
}
}
Status BufferAssignment::ComputeSummaryStats() {
for (auto& allocation : Allocations()) {
if (allocation.is_entry_computation_parameter()) {
stats_.parameter_allocation_count++;
stats_.parameter_allocation_bytes += allocation.size();
}
if (allocation.is_constant()) {
stats_.constant_allocation_count++;
stats_.constant_allocation_bytes += allocation.size();
}
if (allocation.maybe_live_out()) {
stats_.maybe_live_out_allocation_count++;
stats_.maybe_live_out_allocation_bytes += allocation.size();
}
if (allocation.IsPreallocatedTempBuffer()) {
stats_.preallocated_temp_allocation_count++;
stats_.preallocated_temp_allocation_bytes += allocation.size();
}
stats_.total_allocation_count++;
stats_.total_allocation_bytes += allocation.size();
}
// Only compute total fragmentation if all computations have schedules.
HloSchedule schedule(module_);
bool schedule_complete = true;
for (const auto& computation : module_->computations()) {
if (!computation->IsFusionComputation()) {
const HloInstructionSequence* sequence =
hlo_ordering().SequentialOrder(*computation);
if (sequence == nullptr) {
schedule_complete = false;
} else {
schedule.set_sequence(computation, *sequence);
}
}
}
if (schedule_complete) {
TF_RETURN_IF_ERROR(schedule.Verify());
TF_ASSIGN_OR_RETURN(
const int64 min_size,
HeapSimulator::MinimumMemoryForModule(schedule, buffer_size_));
stats_.total_fragmentation_bytes = stats_.total_allocation_bytes - min_size;
}
return Status::OK();
}
string BufferAssignment::Stats::ToString() const {
string s;
StrAppendFormat(&s, "BufferAssignment stats:\n");
StrAppendFormat(&s, " parameter allocation: %10s\n",
HumanReadableNumBytes(parameter_allocation_bytes));
StrAppendFormat(&s, " constant allocation: %10s\n",
HumanReadableNumBytes(constant_allocation_bytes));
StrAppendFormat(&s, " maybe_live_out allocation: %10s\n",
HumanReadableNumBytes(maybe_live_out_allocation_bytes));
StrAppendFormat(&s, " preallocated temp allocation: %10s\n",
HumanReadableNumBytes(preallocated_temp_allocation_bytes));
if (preallocated_temp_fragmentation_bytes >= 0) {
const double percent = 100. * preallocated_temp_fragmentation_bytes /
preallocated_temp_allocation_bytes;
StrAppendFormat(
&s, " preallocated temp fragmentation: %10s (%.2f%%)\n",
HumanReadableNumBytes(preallocated_temp_fragmentation_bytes), percent);
}
StrAppendFormat(&s, " total allocation: %10s\n",
HumanReadableNumBytes(total_allocation_bytes));
if (total_fragmentation_bytes >= 0) {
const double percent =
100. * total_fragmentation_bytes / total_allocation_bytes;
StrAppendFormat(&s, " total fragmentation: %10s (%.2f%%)\n",
HumanReadableNumBytes(total_fragmentation_bytes), percent);
}
return s;
}
string BufferAssignment::ToString() const {
string output;
absl::StrAppend(&output, "BufferAssignment:\n");
for (auto& allocation : allocations_) {
absl::StrAppend(&output, allocation.ToString());
}
return output;
}
BufferAssignmentProto BufferAssignment::ToProto() const {
BufferAssignmentProto proto;
// NOTE: DataflowAnalysis state is serialized here in BufferAssignment,
// because we need to do the HasAllocation check for each buffer. Otherwise
// the buffer_size_ call might fail for some backends.
const HloDataflowAnalysis& dataflow = this->dataflow_analysis();
for (BufferValue::Id id = 0; id < dataflow.values().size(); id++) {
auto& value = dataflow.values().at(id);
if (HasAllocation(*value)) {
LogicalBufferProto proto_buffer = value->ToProto(buffer_size_);
proto.add_logical_buffers()->Swap(&proto_buffer);
// Fill buffer aliases.
for (const HloValue* alias :
alias_analysis().GetBufferContainingValue(*value).values()) {
if (alias->instruction() == value->instruction() &&
alias->index() == value->index()) {
continue; // skip self-aliases
}
BufferAssignmentProto::BufferAlias* proto_alias =
proto.add_buffer_aliases();
LogicalBufferProto::Location proto_alias_location =
BufferValue::ToLocationProto(*alias->instruction(), alias->index());
proto_alias->set_source_buffer_id(value->id());
proto_alias->mutable_location()->Swap(&proto_alias_location);
}
}
}
for (const BufferAllocation& allocation : Allocations()) {
BufferAllocationProto proto_allocation = allocation.ToProto();
proto.add_buffer_allocations()->Swap(&proto_allocation);
for (const HeapSimulatorTrace& heap_trace : allocation.HeapTraces()) {
*proto.add_heap_simulator_traces() = heap_trace;
}
}
return proto;
}
/* static */
StatusOr<std::unique_ptr<BufferAssignment>> BufferAssigner::Run(
const HloModule* module, std::unique_ptr<HloOrdering> hlo_ordering,
BufferValue::SizeFunction buffer_size,
LogicalBuffer::AlignmentFunction color_alignment,
bool allocate_buffers_for_constants, BufferAssigner::Colorer colorer,
const absl::flat_hash_set<HloOpcode>& reuse_checker,
HloDataflowAnalysis::CanShareBuffer can_share_buffer) {
BufferAssigner assigner(allocate_buffers_for_constants, std::move(colorer),
reuse_checker);
return assigner.CreateAssignment(
module, std::move(hlo_ordering), std::move(buffer_size),
std::move(color_alignment), std::move(can_share_buffer));
}
bool BufferAssigner::MaybeAssignBuffer(BufferAllocation* allocation,
const HloBuffer& hlo_buffer,
BufferAssignment* assignment) {
CHECK(!assignment->HasAllocation(hlo_buffer))
<< "buffer " << hlo_buffer << " already has an allocation assigned.";
VLOG(4) << "Trying to assign " << hlo_buffer << " size "
<< assignment->HloBufferSize(hlo_buffer)
<< " to allocation: " << *allocation;
if (hlo_buffer.color() != allocation->color()) {
VLOG(4) << "Can't assign: buffer has color" << hlo_buffer.color()
<< " and allocation has color " << allocation->color() << ".";
return false;
}
if (assignment->HloBufferSize(hlo_buffer) > allocation->size()) {
VLOG(4) << "Can't assign: buffer is larger than allocation ("
<< assignment->HloBufferSize(hlo_buffer) << " > "
<< allocation->size() << ")";
return false;
}
if (allocation->is_readonly()) {
VLOG(4) << "Can't assign: allocation is readonly";
return false;
}
if (!must_not_live_out_.empty()) {
if (allocation->maybe_live_out()) {
// If a buffer maybe live out, the allocation cannot contain any node from
// the "must_not_live_out_" set.
for (const HloValue* value : hlo_buffer.values()) {
if (must_not_live_out_.count(value->instruction()->opcode()) > 0) {
VLOG(4) << "Can't assign: " << value->instruction()->ToString()
<< " cannot live out of the module";
return false;
}
}
}
// The above check is not enough -- There could be the case where an
// allocation can be not live out and contains an instruction with opcode
// from the "must_not_live_out_" set, but assigning a live out buffer to
// that allocation makes the allocation live out and also contains
// instruction from the "must_not_live_out_" set.
if (assignment->alias_analysis().BufferLivesOut(hlo_buffer)) {
for (const auto& buffer_offset_size : allocation->assigned_buffers()) {
if (must_not_live_out_.count(
buffer_offset_size.first->instruction()->opcode()) > 0) {
VLOG(4) << "Can't assign: " << buffer_offset_size.first->instruction()
<< " cannot live out of the module";
return false;
}
}
}
}
if (!allocation->is_reusable()) {
VLOG(4) << "Can't assign: allocation is not reusable";
return false;
}
for (const auto& buffer_offset_size : allocation->assigned_buffers()) {
// Pairwise compare.
const HloValue& assigned_buffer =
*CHECK_NOTNULL(dynamic_cast<const HloValue*>(buffer_offset_size.first));
for (const HloValue* new_value : hlo_buffer.values()) {
if (assignment->hlo_ordering().MayInterfere(
assigned_buffer, *new_value, assignment->dataflow_analysis())) {
VLOG(4) << "Can't assign: assignee " << assigned_buffer
<< " may interfere with " << new_value;
return false;
}
for (const HloPosition& assigned_buffer_position :
assigned_buffer.positions()) {
// Copy instruction don't share a buffer with their input operand.
if (new_value->instruction()->IsUserOf(
assigned_buffer_position.instruction) &&
new_value->instruction()->opcode() == HloOpcode::kCopy) {
VLOG(4) << "Can't assign: assignee " << assigned_buffer
<< " is used at copy instruction " << new_value;
return false;
}
}
}
}
// If the buffer is live out of the computation then it should only be
// assigned a buffer which exactly fits the result to avoid wasting memory
// (result buffers can have arbitrary lifetimes).
if (assignment->alias_analysis().BufferLivesOut(hlo_buffer) &&
allocation->size() != assignment->HloBufferSize(hlo_buffer)) {
VLOG(4) << "Can't assign: buffer " << hlo_buffer
<< "is live out and size not the same as allocation";
return false;
}
assignment->AddAssignment(allocation, hlo_buffer, /*offset=*/0,
assignment->HloBufferSize(hlo_buffer));
return true;
} // namespace xla
Status BufferAssigner::MergeInplaceOpBuffers(BufferAssignment* assignment) {
// Try allocate same buffer for dynamic update slice's operand and output.
//
// TODO(yunxing): Moving this logic to alias analysis and add must-alias rule
// to operations that can be done in place.
for (HloComputation* computation : assignment->module().computations()) {
for (HloInstruction* instruction : computation->instructions()) {
if (!(instruction->opcode() == HloOpcode::kDynamicUpdateSlice ||
(instruction->opcode() == HloOpcode::kFusion &&
(instruction->fused_expression_root()->opcode() ==
HloOpcode::kDynamicUpdateSlice)))) {
continue;
}
if (instruction->parent()->IsFusionComputation()) {
continue;
}
if (instruction->operand_count() == 0) {
continue;
}
// The operand can't share the same buffer with the user based on dataflow
// analysis.
if (!assignment->dataflow_analysis().CanShareOperandBufferWithUser(
instruction->mutable_operand(0), {}, instruction, {})) {
continue;
}
HloBuffer& instruction_buffer =
assignment->alias_analysis().GetUniqueBufferAt(instruction, {});
HloBuffer& operand_buffer =
assignment->alias_analysis().GetUniqueBufferAt(
instruction->operand(0), {});
// Already have the same buffer. No need to merge those.
if (instruction_buffer.id() == operand_buffer.id()) {
continue;
}
// Do not perform in-place dynamic update slice if the operand buffer is
// read-only.
if (HloBufferIsReadOnly(operand_buffer)) {
continue;
}
bool interfere = false;
for (const HloValue* instruction_value : instruction_buffer.values()) {
for (const HloValue* operand_value : operand_buffer.values()) {
if (assignment->hlo_ordering().MayInterfere(
*instruction_value, *operand_value,
assignment->dataflow_analysis())) {
interfere = true;
break;
}
}
}
if (interfere) {
continue;
}
if (assignment->alias_analysis().BufferLivesOut(instruction_buffer)) {
continue;
}
if (instruction_buffer.color() != operand_buffer.color()) {
continue;
}
VLOG(3) << "Merging inplace " << instruction_buffer << " and "
<< operand_buffer;
assignment->alias_analysis().MergeBuffers(instruction_buffer,
operand_buffer);
}
}
return Status::OK();
}
Status BufferAssigner::AssignSingleHloBuffer(
const HloBuffer* hlo_buffer, bool is_thread_local,
absl::flat_hash_map<const HloComputation*,
absl::flat_hash_set<const HloValue*>>*
buffers_to_assign_sequentially,
std::vector<BufferAllocation::Index>* allocation_indices,
BufferAssignment* assignment) {
const int64 buffer_size = assignment->HloBufferSize(*hlo_buffer);
for (const HloValue* value : hlo_buffer->values()) {
if (value->instruction()->opcode() == HloOpcode::kConstant) {
if (allocate_buffers_for_constants_) {
BufferAllocation* allocation =
assignment->NewAllocation(*hlo_buffer, buffer_size);
allocation->set_constant(true);
VLOG(3) << "New allocation #" << allocation->index() << " for constant "
<< *hlo_buffer << " value ptr: " << value;
}
VLOG(3) << "Not allocating buffer for constant";
return Status::OK();
}
const HloInstruction* instruction = value->instruction();
const bool is_entry_parameter =
instruction->opcode() == HloOpcode::kParameter &&
instruction->parent() ==
instruction->parent()->parent()->entry_computation();
if (is_entry_parameter) {
bool parameter_has_alias =
assignment->module().input_output_alias_config().ParameterHasAlias(
instruction->parameter_number(), value->index());
// If the hlo buffer is part of an external parameter, creates a new
// allocation and sets its parameter number. Parameters of non-entry
// computations do not need special allocations because they live inside
// callers.
BufferAllocation* allocation =
assignment->NewAllocation(*hlo_buffer, buffer_size);
allocation->set_entry_computation_parameter(
instruction->parameter_number(), value->index(), parameter_has_alias);
if (parameter_has_alias) {
allocation_indices->push_back(allocation->index());
}
VLOG(3) << "New allocation #" << allocation->index()
<< " marked as entry computation parameter: " << *hlo_buffer;
return Status::OK();
}
}
if (is_thread_local) {
BufferAllocation* allocation =
assignment->NewAllocation(*hlo_buffer, buffer_size);
allocation->set_is_thread_local(true);
VLOG(3) << "New allocation #" << allocation->index()
<< " for thread-local: " << *hlo_buffer;
return Status::OK();
}
for (const HloValue* value : hlo_buffer->values()) {
if (value->shape().IsTuple()) {
BufferAllocation* allocation =
assignment->NewAllocation(*hlo_buffer, buffer_size);
allocation->set_is_tuple(true);
VLOG(3) << "New allocation #" << allocation->index()
<< " for tuple-shaped buffer: " << *hlo_buffer;
return Status::OK();
}
if (value->IsTopLevel() && !value->IsTuple()) {
const HloInstruction* instruction = value->instruction();
for (auto* operand : instruction->operands()) {
for (const auto& operand_slice :
assignment->GetAllSlices(operand, /*index=*/{})) {
BufferAllocation* allocation =
assignment->GetMutableAllocation(operand_slice.index());
if (MaybeAssignBuffer(allocation, *hlo_buffer, assignment)) {
VLOG(3) << "Reusing (operand) allocation #" << allocation->index()
<< " for: " << *hlo_buffer;
return Status::OK();
}
}
}
}
}
// Find the smallest buffer which can be reused iterating from end of
// allocation_indices (smallest) to beginning (largest).
for (int allocation_index = allocation_indices->size() - 1;
allocation_index >= 0; allocation_index--) {
BufferAllocation* allocation = assignment->GetMutableAllocation(
allocation_indices->at(allocation_index));
if (MaybeAssignBuffer(allocation, *hlo_buffer, assignment)) {
VLOG(3) << "Reusing allocation #" << allocation->index()
<< " for: " << *hlo_buffer;
return Status::OK();
}
}
if (!assignment->HasAllocation(*hlo_buffer) &&
!assignment->alias_analysis().BufferLivesOut(*hlo_buffer)) {
bool all_computations_have_sequential_order = true;
for (const HloValue* hlo_value : hlo_buffer->values()) {
HloComputation* computation = hlo_value->instruction()->parent();
const bool has_sequential_order =
assignment->hlo_ordering().SequentialOrder(*computation) != nullptr;
all_computations_have_sequential_order &= has_sequential_order;
}
if (all_computations_have_sequential_order) {
for (const HloValue* hlo_value : hlo_buffer->values()) {
HloComputation* computation = hlo_value->instruction()->parent();
// There is a sequential instruction ordering, so we delay assignment
// of temp buffers until after the loop. We do this right before we
// decide to create a new allocation, to ensure we've exhausted all
// the buffer re-use cases above.
//
// Entry parameters and thread local buffers were already handled
// earlier in this loop iteration. See
// BufferAllocation::IsPreallocatedTempBuffer for the definition of
// temp buffers.
(*buffers_to_assign_sequentially)[computation].insert(hlo_value);
VLOG(3) << "Delaying assignment of temp buffer: " << *hlo_value;
}
return Status::OK();
}
}
if (!assignment->HasAllocation(*hlo_buffer)) {
BufferAllocation* allocation =
assignment->NewAllocation(*hlo_buffer, buffer_size);
allocation_indices->push_back(allocation->index());
VLOG(3) << "New allocation #" << allocation->index()
<< " for: " << *hlo_buffer;
}
TF_RET_CHECK(assignment->HasAllocation(*hlo_buffer));
return Status::OK();
}
Status BufferAssigner::AssignBuffersForComputations(
const std::vector<const HloComputation*>& computations,
bool is_thread_local,
absl::flat_hash_map<const HloComputation*,
absl::flat_hash_set<const HloValue*>>*
buffers_to_assign_sequentially,
BufferAssignment* assignment) {
if (computations.empty()) {
return Status::OK();
}
std::vector<const HloBuffer*> sorted_buffers;
const HloAliasAnalysis& alias_analysis = assignment->alias_analysis();
for (const HloBuffer& buffer : alias_analysis.buffers()) {
TF_RET_CHECK(!buffer.values().empty());
const HloComputation* comp = buffer.values()[0]->instruction()->parent();
if (absl::c_linear_search(computations, comp)) {
sorted_buffers.push_back(&buffer);
}
}
// Generate a post order sort of instructions for sorting of the
// HloBuffers.
flat_hash_map<const HloInstruction*, int> post_order_position;
int position = 0;
std::vector<const HloComputation*> reverse_post_order_computations;
std::unique_ptr<CallGraph> call_graph =
CallGraph::Build(computations[0]->parent());
TF_RETURN_IF_ERROR(call_graph->VisitNodes([&](const CallGraphNode& node) {
if (absl::c_linear_search(computations, node.computation())) {
reverse_post_order_computations.push_back(node.computation());
}
return Status::OK();
}));
absl::c_reverse(reverse_post_order_computations);
for (auto* computation : reverse_post_order_computations) {
for (auto* instruction : computation->MakeInstructionPostOrder()) {
post_order_position.emplace(instruction, position);
position++;
}
}
for (const HloComputation* computation : computations) {
const bool has_sequential_order =
assignment->hlo_ordering().SequentialOrder(*computation) != nullptr;
if (has_sequential_order && buffers_to_assign_sequentially != nullptr) {
// Every sequential computation must get an entry in the
// buffers_to_assign_sequentially map, even if we end up with an empty
// set of buffers. This ensures we can correctly determine whether to
// run whole-module heap simulation.
buffers_to_assign_sequentially->emplace(computation,
flat_hash_set<const HloValue*>());
}
}
absl::c_sort(
sorted_buffers, [&post_order_position, &alias_analysis, assignment](
const HloBuffer* a, const HloBuffer* b) {
// Primary sort is by decreasing buffer size.
const int64 a_size = assignment->HloBufferSize(*a);
const int64 b_size = assignment->HloBufferSize(*b);
if (a_size != b_size) {
return a_size > b_size; // use ">" for decreasing size.
}
const bool a_live_out = alias_analysis.BufferLivesOut(*a);
const bool b_live_out = alias_analysis.BufferLivesOut(*b);
if (a_live_out != b_live_out) {
return a_live_out;
}
auto compare = [&post_order_position](const HloValue* value1,
const HloValue* value2) {
return post_order_position.at(value1->instruction()) <
post_order_position.at(value2->instruction());
};
const HloValue* a_min = *absl::c_min_element(a->values(), compare);
const HloValue* b_min = *absl::c_min_element(b->values(), compare);
return post_order_position.at(a_min->instruction()) <
post_order_position.at(b_min->instruction());
});
std::vector<BufferAllocation::Index> allocation_indices;
for (const HloBuffer* buffer : sorted_buffers) {
VLOG(3) << "=================================================";
VLOG(3) << "Assigning buffer for " << *buffer;
TF_RETURN_IF_ERROR(AssignSingleHloBuffer(buffer, is_thread_local,
buffers_to_assign_sequentially,
&allocation_indices, assignment));
}
return Status::OK();
}
flat_hash_map<LogicalBuffer::Color, flat_hash_set<const HloValue*>,
LogicalBuffer::Color::Hasher>
BufferAssigner::SplitBuffersByColor(
const flat_hash_set<const HloValue*>& buffers) {
flat_hash_map<LogicalBuffer::Color, flat_hash_set<const HloValue*>,
LogicalBuffer::Color::Hasher>
color_map;
for (auto buffer : buffers) {
color_map[buffer->color()].insert(buffer);
}
return color_map;
}
Status BufferAssigner::AssignBuffersWithSequentialOrdering(
const flat_hash_map<const HloComputation*, flat_hash_set<const HloValue*>>&
buffers_to_assign_sequentially,
bool run_whole_module_heap_simulation, BufferAssignment* assignment) {
// Run the sequence of instructions through the heap simulator. The
// heuristic that seems to give the best results is lazy-best-fit, with all
// runs of alloc / free calls sorted in decreasing size order.
const HloOrdering& hlo_ordering = assignment->hlo_ordering();
// Returns a heap algorithm that chooses the best result from several
// algorithms.
auto get_heap_algorithm = [&](int64 alignment) {
auto algorithms =
absl::make_unique<std::vector<std::unique_ptr<HeapAlgorithm>>>();
algorithms->push_back(absl::make_unique<GlobalDecreasingSizeBestFitHeap>(
alignment, GlobalDecreasingSizeBestFitHeap::kSpatial));
algorithms->push_back(absl::make_unique<GlobalDecreasingSizeBestFitHeap>(
alignment, GlobalDecreasingSizeBestFitHeap::kTemporal));
return absl::make_unique<ChooseBestHeapAlgorithm>(std::move(algorithms));
};
if (run_whole_module_heap_simulation) {
// Run the heap simulation over the whole module. This reduces memory
// usage, since buffers for kCall, kWhile, and kConditional
// sub-computations are only live for the duration of their calling
// instructions.
VLOG(1) << "Running whole-module heap simulation";
HloSchedule schedule(&assignment->module());
flat_hash_set<const HloValue*> all_buffers_to_assign;
for (const auto& pair : buffers_to_assign_sequentially) {
const HloComputation* computation = pair.first;
const flat_hash_set<const HloValue*>& buffers_to_assign = pair.second;
const HloInstructionSequence* instruction_sequence =
hlo_ordering.SequentialOrder(*computation);
CHECK(instruction_sequence != nullptr) << computation->name();
schedule.set_sequence(computation, *instruction_sequence);
all_buffers_to_assign.insert(buffers_to_assign.begin(),
buffers_to_assign.end());
}
auto color_map = SplitBuffersByColor(all_buffers_to_assign);
for (auto& single_colored_set : color_map) {
auto color = single_colored_set.first;
VLOG(2) << "Simulating heap for color " << color;
int64 alignment = assignment->color_alignment_(color);
HeapSimulator::Options options;
options.alloc_constants = allocate_buffers_for_constants_;
options.buffers_to_assign = &single_colored_set.second;
TF_ASSIGN_OR_RETURN(
HeapSimulator::Result result,
HeapSimulator::Run(
get_heap_algorithm(alignment), assignment->module(), schedule,
assignment->alias_analysis(), assignment->buffer_size_, options));
AssignBuffersFromHeapSimulator(result, assignment,
single_colored_set.first);
}
} else {
// Run the heap-simulation on a per-computation basis. Buffers for
// sub-computations are assigned disjoint BufferAllocations, assuming the
// worst-case that they may all be live concurrently.
VLOG(1) << "Running per-computation heap simulation";
for (const auto& pair : buffers_to_assign_sequentially) {
const HloComputation* computation = pair.first;
const flat_hash_set<const HloValue*>& buffers_to_assign = pair.second;
const HloInstructionSequence* instruction_sequence =
hlo_ordering.SequentialOrder(*computation);
CHECK(instruction_sequence != nullptr) << computation->name();
auto color_map = SplitBuffersByColor(buffers_to_assign);
for (auto& single_colored_set : color_map) {
auto color = single_colored_set.first;
VLOG(2) << "Simulating heap for color " << color;
int64 alignment = assignment->color_alignment_(color);
HeapSimulator::Options options;
options.buffers_to_assign = &single_colored_set.second;
TF_ASSIGN_OR_RETURN(
HeapSimulator::Result result,
HeapSimulator::Run(get_heap_algorithm(alignment), *computation,
*instruction_sequence,
assignment->alias_analysis(),
assignment->buffer_size_, options));
AssignBuffersFromHeapSimulator(result, assignment,
single_colored_set.first);
}
}
}
return Status::OK();
}
namespace {
// Computes and returns the set of logical buffers live at the point of
// maximal liveness in the given heap trace. LogicalBuffers are (stabily)
// sorted by id.
std::vector<const HloValue*> ComputePeakMemoryLogicalBuffers(
const BufferAllocation& allocation, const HeapSimulatorTrace& heap_trace) {
// Create a map from LogicalBuffer::Id to LogicalBuffer* for the logical
// buffers in this allocation.
absl::flat_hash_map<BufferValue::Id, const HloValue*> id_to_value;
absl::flat_hash_map<const HloValue*, int64> buffer_sizes;
for (const auto& pair : allocation.assigned_buffers()) {
const HloValue* value = pair.first;
const BufferAllocation::OffsetSize& offset_size = pair.second;
id_to_value[value->id()] = value;
buffer_sizes[value] = offset_size.size;
}
VLOG(1) << "Compute peak memory logical buffers";
// Returns how much the given event increases the total size of live
// buffers. Can be negative.
auto memory_delta = [&id_to_value, &buffer_sizes](
const HeapSimulatorTrace::Event& event) -> int64 {
const HloValue* buffer = id_to_value.at(event.buffer_id());
const int64 buffer_size = buffer_sizes.at(buffer);
if (event.kind() == HeapSimulatorTrace::Event::ALLOC ||
event.kind() == HeapSimulatorTrace::Event::SHARE_WITH) {
return buffer_size;
} else if (event.kind() == HeapSimulatorTrace::Event::FREE) {
return -1 * buffer_size;
}
LOG(FATAL) << "Unknown event kind: " << event.kind();
};
// First compute the size of the maximal live set.
int64 max_live_size = 0;
int64 live_size = 0;
for (const auto& event : heap_trace.events()) {
live_size += memory_delta(event);
if (max_live_size < live_size) {
max_live_size = live_size;
}
}
// Next gather the set of logical buffers live at the earliest point of
// maximal live set size.
absl::flat_hash_set<const HloValue*> live_values;
live_size = 0;
for (const auto& event : heap_trace.events()) {
const HloValue* value = id_to_value.at(event.buffer_id());
if (event.kind() == HeapSimulatorTrace::Event::ALLOC ||
event.kind() == HeapSimulatorTrace::Event::SHARE_WITH) {
InsertOrDie(&live_values, value);
} else if (event.kind() == HeapSimulatorTrace::Event::FREE) {
CHECK(ContainsKey(live_values, value));
live_values.erase(value);
}
live_size += memory_delta(event);
if (live_size == max_live_size) {
break;
}
}
CHECK_EQ(live_size, max_live_size);
std::vector<const HloValue*> live_values_vector;
live_values_vector.insert(live_values_vector.end(), live_values.begin(),
live_values.end());
// Stabily sort the live buffers.
absl::c_sort(live_values_vector, [](const HloValue* a, const HloValue* b) {
return a->id() < b->id();
});
VLOG(4) << "Peak memory buffer:";
for (auto value : live_values_vector) {
VLOG(4) << " " << value->ToString();
}
return live_values_vector;
}
} // namespace
void BufferAssigner::AssignBuffersFromHeapSimulator(
const HeapSimulator::Result& result, BufferAssignment* assignment,
BufferValue::Color color) {
if (assignment->stats_.preallocated_temp_fragmentation_bytes == -1) {
assignment->stats_.preallocated_temp_fragmentation_bytes =
result.fragmentation_size;
} else {
assignment->stats_.preallocated_temp_fragmentation_bytes +=
result.fragmentation_size;
}
VLOG(1) << "Result size from heap simulator: " << result.heap_size;
BufferAllocation* allocation =
assignment->NewEmptyAllocation(result.heap_size, color);
for (const auto& buffer_chunk : result.chunk_map) {
const HloValue& value = *buffer_chunk.first;
const HeapSimulator::Chunk& chunk = buffer_chunk.second;
assignment->AddAssignment(allocation, value, chunk.offset, chunk.size);
}
allocation->peak_buffers_ =
ComputePeakMemoryLogicalBuffers(*allocation, result.debug_trace);
VLOG(1) << "Ran heap simulation for allocation: ";
XLA_VLOG_LINES(2, allocation->ToString());
allocation->AddHeapTrace(result.debug_trace);
}
StatusOr<std::unique_ptr<BufferAssignment>> BufferAssigner::CreateAssignment(
const HloModule* module, std::unique_ptr<HloOrdering> hlo_ordering,
BufferValue::SizeFunction buffer_size,
LogicalBuffer::AlignmentFunction color_alignment,
HloDataflowAnalysis::CanShareBuffer can_share_buffer) {
TF_ASSIGN_OR_RETURN(std::unique_ptr<HloAliasAnalysis> alias_analysis,
HloAliasAnalysis::Run(module, can_share_buffer));
VLOG(1) << "Assigning buffers to module " << module->name();
XLA_VLOG_LINES(3, module->ToString());
XLA_VLOG_LINES(3, alias_analysis->ToString());
XLA_VLOG_LINES(3, alias_analysis->dataflow_analysis().ToString());
VLOG(1) << "Number of buffers to assign: "
<< alias_analysis->buffers().size();
// Can't use absl::make_unique because BufferAssignment constructor is
// private.
std::unique_ptr<BufferAssignment> assignment(new BufferAssignment(
module, std::move(hlo_ordering), std::move(buffer_size),
std::move(color_alignment), std::move(alias_analysis)));
TF_RETURN_IF_ERROR(
colorer_(&assignment->alias_analysis(), assignment->hlo_ordering()));
VLOG(3) << "After coloring:";
XLA_VLOG_LINES(3,
assignment->alias_analysis().dataflow_analysis().ToString());
TF_RETURN_IF_ERROR(MergeInplaceOpBuffers(assignment.get()));
std::vector<const HloComputation*> thread_local_computations;
std::vector<const HloComputation*> global_computations;
TF_RETURN_IF_ERROR(GatherComputationsByAllocationType(
module, &thread_local_computations, &global_computations));
// First assign buffers for global computations. Temporary buffers for
// sequential computations are collected in
// 'buffers_to_assign_sequentially'.
flat_hash_map<const HloComputation*, flat_hash_set<const HloValue*>>
buffers_to_assign_sequentially;
TF_RETURN_IF_ERROR(AssignBuffersForComputations(
global_computations,
/*is_thread_local=*/false, &buffers_to_assign_sequentially,
assignment.get()));
// Assign buffers with sequential ordering, if any. If all global
// computations are sequential, we can run heap simuation on the whole
// module, which reduces memory usage.
const bool run_whole_module_heap_simulation =
buffers_to_assign_sequentially.size() == global_computations.size();
VLOG(2) << "Running whole module heap simulation"
<< run_whole_module_heap_simulation;
TF_RETURN_IF_ERROR(AssignBuffersWithSequentialOrdering(
buffers_to_assign_sequentially, run_whole_module_heap_simulation,
assignment.get()));
std::vector<const HloComputation*> thread_local_computations_no_fusion;
// Now assign buffers for thread-local computations. All LogicalBuffers get
// their own BufferAllocation.
for (auto* computation : thread_local_computations) {
TF_RET_CHECK(computation != module->entry_computation());
if (computation->IsFusionComputation()) {
continue;
}
thread_local_computations_no_fusion.push_back(computation);
}
TF_RETURN_IF_ERROR(AssignBuffersForComputations(
thread_local_computations_no_fusion,
/*is_thread_local=*/true,
/*buffers_to_assign_sequentially=*/nullptr, assignment.get()));
// Mark all buffers which may be live out of the entry computation as
// "liveout".
for (const HloBuffer* buffer :
assignment->alias_analysis().LiveOutBuffers()) {
VLOG(3) << "maybe_live_out LogicalBuffer: " << *buffer;
if (assignment->HasAllocation(*buffer)) {
BufferAllocation* alloc =
assignment->GetMutableAssignedAllocation(*buffer);
alloc->set_maybe_live_out(true);
VLOG(3) << "maybe_live_out BufferAllocation: " << *alloc;
}
}
// Combines allocations of temporary buffers into one big BufferAllocation.
// This can only be performed after all buffers have been assigned, and
// after maybe_live_out is marked, since it is used to determine whether an
// allocation contains temporary buffers or not.
assignment->CombineTempAllocations();
XLA_VLOG_LINES(2, assignment->ToString());
TF_RETURN_IF_ERROR(assignment->ComputeSummaryStats());
XLA_VLOG_LINES(1, assignment->GetStats().ToString());
VLOG(1) << "Buffer assignment done.";
return std::move(assignment);
}
} // namespace xla