saymrwulf/onnxruntime
1
0
Fork 0
mirror of https://github.com/saymrwulf/onnxruntime.git synced 2026-05-29 23:06:41 +00:00
Code Issues Projects Releases Packages Wiki Activity Actions

Turn off peak memory logging and fix memory pattern generation bug. (#5676)

Browse source 
* Turn off peak memory log lines and fix memory pattern generation bug.

* Turn off peak memory log lines and fix memory pattern generation bug.

Co-authored-by: Ubuntu <OrtTrainingDev3@OrtTrainingDev3.af05slrtruoetgaxwwjv5nsq5e.px.internal.cloudapp.net>
This commit is contained in:
M. Zeeshan Siddiqui 2020-11-03 08:44:15 -08:00 committed by GitHub
parent 5d66cf017c
commit 9b010963b7
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
3 changed files with 24 additions and 13 deletions
Show stats Download patch file Download diff file Expand all files Collapse all files

2
onnxruntime/core/framework/allocation_planner.cc
View file

@ -787,7 +787,7 @@ class PlannerImpl {
for (size_t index = 0; index < current_plan.program_counter_start.size(); index += 1) {
ORT_ENFORCE((current_plan.program_counter_start[index] > start) || (start == 0));
ORT_ENFORCE(current_plan.program_counter_start[index] <= current_plan.program_counter_end[index]);
ORT_ENFORCE((current_plan.program_counter_start[index] < SIZE_MAX) || (index == 0));
ORT_ENFORCE(current_plan.program_counter_start[index] < SIZE_MAX);
ORT_ENFORCE((current_plan.program_counter_end[index] > 0) || (index == 0));
start = current_plan.program_counter_start[index];

2
onnxruntime/core/framework/execution_frame.cc
View file

@ -292,7 +292,7 @@ ExecutionFrame::ExecutionFrame(const std::vector<int>& feed_mlvalue_idxs, const
}
// log size of activation. Keep it commented out for now to avoid log flooding.
VLOGS(session_state_.Logger(), 1) << "**** Allocated memory for activations, size: " <<mem_patterns_->patterns[i].PeakSize();
// VLOGS(session_state_.Logger(), 1) << "**** Allocated memory for activations, size: " <<mem_patterns_->patterns[i].PeakSize();
}
}
}

33
onnxruntime/core/framework/mem_pattern_planner.h
View file

@ -70,11 +70,7 @@ class MemPatternPlanner {
size_t current = 0;
size_t waste_bytes = std::numeric_limits<size_t>::max();
size_t best_offset = 0;
if (!blocks_.empty()) {
auto last_block = allocs_[*blocks_.rbegin()];
best_offset = last_block.block_.offset_ + last_block.block_.size_;
}
bool best_offset_found = false;
for (auto it = blocks_.begin(); it != blocks_.end(); it++) {
// Memory block can be re-used as long as there is no overlap between their time schedules.
if (allocs_[*it].reuse_ && !OverlappingTimeSchedules(program_counter_start, program_counter_end,
@ -87,16 +83,25 @@ class MemPatternPlanner {
if (gap >= size && (gap - size) < waste_bytes) {
waste_bytes = gap - size;
best_offset = current;
best_offset_found = true;
}
}
current = std::max(current, allocs_[*it].block_.offset_ + allocs_[*it].block_.size_);
}
ORT_ENFORCE(current <= buffer_size_);
if (current < buffer_size_) {
auto gap = buffer_size_ - current;
if ((gap >= size) && ((gap - size) < waste_bytes))
if ((gap >= size) && ((gap - size) < waste_bytes)) {
best_offset = current;
best_offset_found = true;
}
}
if (!best_offset_found) {
best_offset = current;
}
// we only need to bounds check the addition of size to best_offset as that is the only time we extend
@ -128,10 +133,7 @@ class MemPatternPlanner {
size_t current = 0;
size_t waste_bytes = std::numeric_limits<size_t>::max();
size_t best_offset = 0;
if (!blocks_.empty()) {
auto last_block = allocs_[*blocks_.rbegin()];
best_offset = last_block.block_.offset_ + last_block.block_.size_;
}
bool best_offset_found = false;
for (auto it = blocks_.begin(); it != blocks_.end(); it++) {
if (allocs_[*it].block_.offset_ >= current) {
@ -139,15 +141,24 @@ class MemPatternPlanner {
if (gap >= size && (gap - size) < waste_bytes) {
waste_bytes = gap - size;
best_offset = current;
best_offset_found = true;
}
}
current = std::max(current, allocs_[*it].block_.offset_ + allocs_[*it].block_.size_);
}
ORT_ENFORCE(current <= buffer_size_);
if (current < buffer_size_) {
auto gap = buffer_size_ - current;
if ((gap >= size) && ((gap - size) < waste_bytes))
if ((gap >= size) && ((gap - size) < waste_bytes)) {
best_offset = current;
best_offset_found = true;
}
}
if (!best_offset_found) {
best_offset = current;
}
// we only need to bounds check the addition of size to best_offset as that is the only time we extend