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

Populate the memory access breakdown information in OpMetrics.

Browse Source 
PiperOrigin-RevId: 313469967
Change-Id: Ib8783f7a4a81cbda79a32410440866946972a1f2
This commit is contained in:
Jiho Choi 2020-05-27 14:57:12 -07:00 committed by TensorFlower Gardener
parent b3925cc9b6
commit 3bb7d689d3
2 changed files with 13 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

15
tensorflow/core/profiler/utils/op_utils.cc
View File

@ -63,13 +63,11 @@ void HostOpMetricsDbBuilder::UpdateHostInfeedEnqInfo(
start_timestamp_ps_diff); start_timestamp_ps_diff);
} }
void DeviceOpMetricsDbBuilder::EnterOp(uint64 program_id, void DeviceOpMetricsDbBuilder::EnterOp(
absl::string_view name, uint64 program_id, absl::string_view name, absl::string_view category,
absl::string_view category, absl::string_view provenance, bool is_eager, uint64 occurrences,
absl::string_view provenance, uint64 time_ps, uint64 children_time_ps, int64 flops, int64 bytes_accessed,
bool is_eager, uint64 occurrences, const std::vector<OpMetrics::MemoryAccessed>& memory_accessed_breakdown) {
uint64 time_ps, uint64 children_time_ps,
int64 flops, int64 bytes_accessed) {
uint64 self_time_ps = time_ps - children_time_ps; uint64 self_time_ps = time_ps - children_time_ps;
DCHECK_GE(time_ps, self_time_ps); DCHECK_GE(time_ps, self_time_ps);
OpMetrics* op_metrics = LookupOrInsertNewOpMetrics(program_id, name); OpMetrics* op_metrics = LookupOrInsertNewOpMetrics(program_id, name);
@ -89,6 +87,9 @@ void DeviceOpMetricsDbBuilder::EnterOp(uint64 program_id,
op_metrics->bytes_accessed() + op_metrics->bytes_accessed() +
GetCappedPerf(bytes_accessed * occurrences, self_time_ps, GetCappedPerf(bytes_accessed * occurrences, self_time_ps,
peak_hbm_bw_giga_bytes_per_second_ / 1000)); peak_hbm_bw_giga_bytes_per_second_ / 1000));
for (const auto& memory_accessed : memory_accessed_breakdown) {
*op_metrics->add_memory_accessed_breakdown() = memory_accessed;
}
db()->set_total_op_time_ps(db()->total_op_time_ps() + self_time_ps); db()->set_total_op_time_ps(db()->total_op_time_ps() + self_time_ps);
} }

6
tensorflow/core/profiler/utils/op_utils.h
View File

@ -69,10 +69,14 @@ class DeviceOpMetricsDbBuilder : public OpMetricsDbBuilder {
// picoseconds. // picoseconds.
// flops = the number of floating-point operations computed. // flops = the number of floating-point operations computed.
// bytes_accessed = the sum of bytes read and bytes written by this OP. // bytes_accessed = the sum of bytes read and bytes written by this OP.
// memory_accessed_breakdown = the breakdown of memory accessed by operation
// type and memory space.
void EnterOp(uint64 program_id, absl::string_view name, void EnterOp(uint64 program_id, absl::string_view name,
absl::string_view category, absl::string_view provenance, absl::string_view category, absl::string_view provenance,
bool is_eager, uint64 occurrences, uint64 time_ps, bool is_eager, uint64 occurrences, uint64 time_ps,
uint64 children_time_ps, int64 flops, int64 bytes_accessed); uint64 children_time_ps, int64 flops, int64 bytes_accessed,
const std::vector<OpMetrics::MemoryAccessed>&
memory_accessed_breakdown = {});
protected: protected:
// Peak performance of a TensorCore or a GPU in TFLOP/s. // Peak performance of a TensorCore or a GPU in TFLOP/s.