experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Additional memory profile annotations in BFCAllocator.

Browse Source 
PiperOrigin-RevId: 295243569
Change-Id: Iac2a4beb8654fa74ecaabbc9ce82fa2dd75b049e
This commit is contained in:
A. Unique TensorFlower 2020-02-14 15:44:36 -08:00 committed by TensorFlower Gardener
parent 1e74a754af
commit d6741e0994
4 changed files with 16 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
tensorflow/core/common_runtime/bfc_allocator.cc
View File

@ -382,7 +382,7 @@ void* BFCAllocator::AllocateRawInternal(size_t unused_alignment,
} }
void* ptr = FindChunkPtr(bin_num, rounded_bytes, num_bytes, freed_before); void* ptr = FindChunkPtr(bin_num, rounded_bytes, num_bytes, freed_before);
if (ptr != nullptr) { if (ptr != nullptr) {
AddTraceMe("MemoryAllocation"); AddTraceMe("MemoryAllocation", num_bytes);
return ptr; return ptr;
} }
@ -390,7 +390,7 @@ void* BFCAllocator::AllocateRawInternal(size_t unused_alignment,
if (Extend(unused_alignment, rounded_bytes)) { if (Extend(unused_alignment, rounded_bytes)) {
ptr = FindChunkPtr(bin_num, rounded_bytes, num_bytes, freed_before); ptr = FindChunkPtr(bin_num, rounded_bytes, num_bytes, freed_before);
if (ptr != nullptr) { if (ptr != nullptr) {
AddTraceMe("MemoryAllocation"); AddTraceMe("MemoryAllocation", num_bytes);
return ptr; return ptr;
} }
} }
@ -403,7 +403,7 @@ void* BFCAllocator::AllocateRawInternal(size_t unused_alignment,
if (MergeTimestampedChunks(rounded_bytes)) { if (MergeTimestampedChunks(rounded_bytes)) {
ptr = FindChunkPtr(bin_num, rounded_bytes, num_bytes, freed_before); ptr = FindChunkPtr(bin_num, rounded_bytes, num_bytes, freed_before);
if (ptr != nullptr) { if (ptr != nullptr) {
AddTraceMe("MemoryAllocation"); AddTraceMe("MemoryAllocation", num_bytes);
return ptr; return ptr;
} }
} }
@ -417,7 +417,7 @@ void* BFCAllocator::AllocateRawInternal(size_t unused_alignment,
Extend(unused_alignment, rounded_bytes)) { Extend(unused_alignment, rounded_bytes)) {
ptr = FindChunkPtr(bin_num, rounded_bytes, num_bytes, freed_before); ptr = FindChunkPtr(bin_num, rounded_bytes, num_bytes, freed_before);
if (ptr != nullptr) { if (ptr != nullptr) {
AddTraceMe("MemoryAllocation"); AddTraceMe("MemoryAllocation", num_bytes);
return ptr; return ptr;
} }
} }
@ -441,7 +441,8 @@ void* BFCAllocator::AllocateRawInternal(size_t unused_alignment,
return nullptr; return nullptr;
} }
void BFCAllocator::AddTraceMe(absl::string_view traceme_name) { void BFCAllocator::AddTraceMe(absl::string_view traceme_name,
int64 requested_bytes) {
tensorflow::profiler::TraceMe trace_me( tensorflow::profiler::TraceMe trace_me(
[&]() EXCLUSIVE_LOCKS_REQUIRED(lock_) { [&]() EXCLUSIVE_LOCKS_REQUIRED(lock_) {
AllocatorStats stats = stats_; AllocatorStats stats = stats_;
@ -452,6 +453,10 @@ void BFCAllocator::AddTraceMe(absl::string_view traceme_name) {
",bytes_allocated=", stats.bytes_in_use, ",bytes_allocated=", stats.bytes_in_use,
",bytes_available=", bytes_available, ",bytes_available=", bytes_available,
",peak_bytes_in_use=", stats.peak_bytes_in_use, ",peak_bytes_in_use=", stats.peak_bytes_in_use,
",requested_bytes=", requested_bytes,
#ifdef TENSORFLOW_MEM_DEBUG
",tf_op=", pending_op_name, ",id=", pending_step_id,
#endif
"#"); "#");
}, },
/*level=*/profiler::TraceMeLevel::kInfo); /*level=*/profiler::TraceMeLevel::kInfo);
@ -589,6 +594,7 @@ void BFCAllocator::DeallocateRawInternal(void* ptr) {
BFCAllocator::ChunkHandle h = region_manager_.get_handle(ptr); BFCAllocator::ChunkHandle h = region_manager_.get_handle(ptr);
CHECK(h != kInvalidChunkHandle); CHECK(h != kInvalidChunkHandle);
int64 requested_bytes = ChunkFromHandle(h)->requested_size;
MarkFree(h); MarkFree(h);
// Consider coalescing it. // Consider coalescing it.
@ -603,7 +609,7 @@ void BFCAllocator::DeallocateRawInternal(void* ptr) {
LOG(INFO) << "F: " << RenderOccupancy(); LOG(INFO) << "F: " << RenderOccupancy();
} }
AddTraceMe("MemoryDeallocation"); AddTraceMe("MemoryDeallocation", -requested_bytes);
} }
// Merges h1 and h2 when Chunk(h1)->next is h2 and Chunk(h2)->prev is c1. // Merges h1 and h2 when Chunk(h1)->next is h2 and Chunk(h2)->prev is c1.

4
tensorflow/core/common_runtime/bfc_allocator.h
View File

@ -116,8 +116,8 @@ class BFCAllocator : public Allocator {
EXCLUSIVE_LOCKS_REQUIRED(lock_); EXCLUSIVE_LOCKS_REQUIRED(lock_);
// Add TraceMe (in memory allocation and deallocation) for memory stats // Add TraceMe (in memory allocation and deallocation) for memory stats
// profiling. // profiling. The requested_bytes can be negative if it's a deallocation.
void AddTraceMe(absl::string_view traceme_name) void AddTraceMe(absl::string_view traceme_name, int64 requested_bytes)
EXCLUSIVE_LOCKS_REQUIRED(lock_); EXCLUSIVE_LOCKS_REQUIRED(lock_);
// A ChunkHandle is an index into the chunks_ vector in BFCAllocator // A ChunkHandle is an index into the chunks_ vector in BFCAllocator

1
tensorflow/core/profiler/utils/xplane_schema.cc
View File

@ -114,6 +114,7 @@ const StatTypeMap& GetStatTypeMap() {
{"bytes_available", kBytesAvailable}, {"bytes_available", kBytesAvailable},
{"fragmentation", kFragmentation}, {"fragmentation", kFragmentation},
{"peak_bytes_in_use", kPeakBytesInUse}, {"peak_bytes_in_use", kPeakBytesInUse},
{"requested_bytes", kRequestedBytes},
{"shape", kTensorShapes}, {"shape", kTensorShapes},
// Device trace arguments. // Device trace arguments.
{"device_id", kDeviceId}, {"device_id", kDeviceId},

1
tensorflow/core/profiler/utils/xplane_schema.h
View File

@ -106,6 +106,7 @@ enum StatType {
kBytesAvailable, kBytesAvailable,
kFragmentation, kFragmentation,
kPeakBytesInUse, kPeakBytesInUse,
kRequestedBytes,
kTensorShapes, kTensorShapes,
// Device trace arguments. // Device trace arguments.
kDeviceId, kDeviceId,