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

Fine-grained memory profiling

Browse Source 
Add residual_bytes, peak_bytes and output_bytes.
Allow to order/select/filter by
accelerator_micros/cpu_micros/peak_bytes/residual_bytes/output_bytes

Also updated the testdata.

PiperOrigin-RevId: 164079214
This commit is contained in:
A. Unique TensorFlower 2017-08-02 21:29:03 -07:00 committed by Benoit Steiner
parent d8689f3241
commit 1848070f47
40 changed files with 1949 additions and 1745 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

6
tensorflow/core/profiler/README.md
View File

@ -106,7 +106,7 @@ _TFProfRoot (--/930.58k params)
### Show the most expensive operation types.
```
tfprof> op -select micros,bytes,occurrence -order_by micros
node name | output bytes | total execution time | accelerator execution time | cpu execution time | op occurrence (run|defined)
node name | requested bytes | total execution time | accelerator execution time | cpu execution time | op occurrence (run|defined)
SoftmaxCrossEntropyWithLogits 36.58MB (100.00%, 0.05%), 1.37sec (100.00%, 26.68%), 0us (100.00%, 0.00%), 1.37sec (100.00%, 30.75%), 30|30
MatMul 2720.57MB (99.95%, 3.66%), 708.14ms (73.32%, 13.83%), 280.76ms (100.00%, 41.42%), 427.39ms (69.25%, 9.62%), 2694|3450
ConcatV2 741.37MB (96.29%, 1.00%), 389.63ms (59.49%, 7.61%), 31.80ms (58.58%, 4.69%), 357.83ms (59.63%, 8.05%), 4801|6098
@ -192,7 +192,7 @@ Open a Chrome browser, enter URL chrome://tracing and load the timeline file.
******************************************************
```
<left>
[Timeline](g3doc/graph_timeline.png)
![Timeline](g3doc/graph_timeline.png)
</left>
```
@ -213,7 +213,7 @@ pprof -png --nodecount=20 --sample_index=1 <filename>
```
<left>
[PprofGraph](g3doc/pprof.jpg)
![PprofGraph](g3doc/pprof.jpg)
</left>
### Feature Request and Bug Report

17
tensorflow/core/profiler/g3doc/options.md
View File

@ -48,7 +48,18 @@ In graph view, in means the number of hops in the <b>graph</b>.
`-min_bytes`: Show nodes that request at least this number of bytes.
`-min_micros`: Show nodes that spend at least this number of microseconds to run.
`-min_peak_bytes`: Show nodes that using at least this number of bytes during peak memory usage.
`-min_residual_bytes`: Show nodes that have at least this number of bytes not being de-allocated after Compute.
`-min_output_bytes`: Show nodes that have at least this number of bytes output (no necessarily allocated by the nodes).
`-min_micros`: Show nodes that spend at least this number of microseconds to run. It sums
accelerator_micros and cpu_micros. Note: cpu and accelerator can run in parallel.
`-min_accelerator_micros`: Show nodes that spend at least this number of microseconds to run on accelerator (e.g. GPU).
`-min_cpu_micros`: Show nodes that spend at least this number of microseconds to run on CPU.
`-min_params`: Show nodes that contains at least this number of parameters.
@ -58,7 +69,7 @@ In graph view, in means the number of hops in the <b>graph</b>.
`-step`: Show the stats of the this step when multiple steps of RunMetadata were added. By default, show the average of all steps."
`-order_by`: Order the results by [name|depth|bytes|micros|accelerator_micros|cpu_micros|params|float_ops|occurrence]
`-order_by`: Order the results by [name|depth|bytes|peak_bytes|residual_bytes|output_bytes|micros|accelerator_micros|cpu_micros|params|float_ops|occurrence]
`-account_type_regexes`: Account and display the nodes whose types match one of the type regexes specified. tfprof allow user to define extra operation types for graph nodes through tensorflow.tfprof.OpLogProto proto. regexes are comma-sperated.
@ -76,7 +87,7 @@ In graph view, in means the number of hops in the <b>graph</b>.
Notes: See <b>overview</b> sesion on how does above options play with each other to decide the output and counting.
`-select`: Comma-separated list of attributes to show. Supported attributes:
[bytes|micros|accelerator_micros|cpu_micros|params|float_ops|occurrence|tensor_value|device|op_types|input_shapes].
[bytes|peak_bytes|residual_bytes|output_bytes|micros|accelerator_micros|cpu_micros|params|float_ops|occurrence|tensor_value|device|op_types|input_shapes].
`-output`: Output results as stdout, file or timeline.
The format is ```output_type:key=value,key=value```.

7
tensorflow/core/profiler/g3doc/profile_memory.md
View File

@ -15,7 +15,6 @@ Open a Chrome browser, enter URL chrome://tracing and load the timeline file.
```
<left>
TODO(xpan): Show the image correctly in github.
![Timeline](graph_timeline.png)
</left>
@ -26,7 +25,7 @@ TODO(xpan): Show the image correctly in github.
# With op view, it shows you the aggregated output tensor bytes of each
# operation type.
tfprof> op -select bytes -order_by bytes
node name | output bytes
node name | requested bytes
Identity 32515.37MB (100.00%, 27.02%)
FusedBatchNormGrad 10802.14MB (72.98%, 8.98%)
FusedBatchNorm 10517.52MB (64.01%, 8.74%)
@ -41,7 +40,7 @@ AddN 2741.49MB (8.56%, 2.28%)
# With scope view, you can see the operations that outputs largest tensors.
tfprof> scope -order_by bytes -select bytes -min_bytes 100000000
node name | output bytes
node name | requested bytes
_TFProfRoot (--/120356.38MB)
tower_3/SepConv2d_2b_3x3/separable_conv2d (346.85MB/854.00MB)
tower_3/SepConv2d_2b_3x3/separable_conv2d/depthwise (507.15MB/507.15MB)
@ -61,7 +60,7 @@ _TFProfRoot (--/120356.38MB)
# code view.
tfprof> code -max_depth 10 -select bytes -order_by bytes -start_name_regexes .*seq2seq.* -min_bytes 1
node name | output bytes
node name | requested bytes
_TFProfRoot (--/74148.60MB)
seq2seq_attention.py'>:168:run_filename_from...:none (0B/74148.60MB)
seq2seq_attention.py'>:33:_run_code_in_main:none (0B/74148.60MB)

12
tensorflow/core/profiler/internal/advisor/expensive_operation_checker.h
View File

@ -47,8 +47,8 @@ class ExpensiveOperationChecker : public Checker {
fprintf(stderr, "Missing run_meta for %s\n", name().c_str());
return;
}
Options opts(3, 0, 1, 0, 0, 0, -1, "micros", {".*"}, {".*"}, {}, {".*"}, {},
false, {"micros", "occurrence"}, "none", {});
Options opts(3, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, -1, "micros", {".*"}, {".*"},
{}, {".*"}, {}, false, {"micros", "occurrence"}, "none", {});
const MultiGraphNodeProto root = stats->ShowMultiGraphNode("op", opts);
if (root.children_size() == 0) {
return;
@ -74,8 +74,8 @@ class ExpensiveOperationChecker : public Checker {
fprintf(stderr, "Missing op_log (code traces) for %s\n", name().c_str());
return;
}
Options opts(100, 0, 1, 0, 0, 0, -1, "micros", {".*"}, {".*"}, {}, {".*"},
{}, false, {"micros"}, "none", {});
Options opts(100, 0, 0, 0, 0, 1, 0, 0, 0, 0, 0, -1, "micros", {".*"},
{".*"}, {}, {".*"}, {}, false, {"micros"}, "none", {});
const MultiGraphNodeProto root = stats->ShowMultiGraphNode("code", opts);
const MultiGraphNodeProto* node = &root;
// A trick here is: Usually, codes in library file are usually referenced
@ -93,8 +93,8 @@ class ExpensiveOperationChecker : public Checker {
}
void CheckScopeView(const TFStats* stats) {
Options opts(100, 0, 100, 0, 0, 0, -1, "micros", {".*"}, {".*"}, {}, {".*"},
{}, false, {"micros"}, "none", {});
Options opts(100, 0, 0, 0, 0, 100, 0, 0, 0, 0, 0, -1, "micros", {".*"},
{".*"}, {}, {".*"}, {}, false, {"micros"}, "none", {});
const GraphNodeProto root = stats->ShowGraphNode("scope", opts);
if (root.children_size() == 0) {
return;

BIN
tensorflow/core/profiler/internal/testdata/ckpt.data-00000-of-00001 vendored
View File

Binary file not shown.

BIN
tensorflow/core/profiler/internal/testdata/ckpt.index vendored
View File

Binary file not shown.

BIN
tensorflow/core/profiler/internal/testdata/ckpt.meta vendored
View File

Binary file not shown.

1903
tensorflow/core/profiler/internal/testdata/graph.pbtxt vendored
View File

File diff suppressed because it is too large Load Diff

BIN
tensorflow/core/profiler/internal/testdata/run_meta vendored
View File

Binary file not shown.

26
tensorflow/core/profiler/internal/testdata/tfprof_log vendored
View File

@ -1,17 +1,11 @@

conv2d_2/BiasAdd 

conv2d/BiasAddè
%
conv2d_1/bias_trainable_variables

conv2d_2/convolutionÀp

conv2d/convolutionð—
#
conv2d/bias_trainable_variables
'
conv2d_1/kernel_trainable_variables
%
conv2d/kernel_trainable_variables

DW2_trainable_variables

ScalarW_trainable_variables

DW_trainable_variables
Conv2DÈ-
Conv2D_1€$

62
tensorflow/core/profiler/internal/tfprof_code.cc
View File

@ -191,6 +191,13 @@ class Samples {
} else if (type == kShown[0]) {
sample_pb->mutable_value()->Add(
gn->requested_bytes(node->node->step()));
} else if (type == kShown[11]) {
sample_pb->mutable_value()->Add(gn->peak_bytes(node->node->step()));
} else if (type == kShown[12]) {
sample_pb->mutable_value()->Add(
gn->residual_bytes(node->node->step()));
} else if (type == kShown[13]) {
sample_pb->mutable_value()->Add(gn->output_bytes(node->node->step()));
} else if (type == kShown[2]) {
sample_pb->mutable_value()->Add(gn->parameters());
} else if (type == kShown[3]) {
@ -296,9 +303,21 @@ class PprofProfileImpl : public PprofProfile {
string_table_.GetIndex("CPU execution time."));
}
} else if (type == kShown[0]) {
sample_type->set_unit(string_table_.GetIndex("bytes"));
sample_type->set_unit(string_table_.GetIndex("requested bytes"));
profile_pb->mutable_comment()->Add(
string_table_.GetIndex("Sum of operation output memory."));
string_table_.GetIndex("Sum of operation total requested memory."));
} else if (type == kShown[11]) {
sample_type->set_unit(string_table_.GetIndex("peak bytes"));
profile_pb->mutable_comment()->Add(
string_table_.GetIndex("Sum of operation peak memory usage."));
} else if (type == kShown[12]) {
sample_type->set_unit(string_table_.GetIndex("residual bytes"));
profile_pb->mutable_comment()->Add(string_table_.GetIndex(
"Sum of operation allocated memory after finish."));
} else if (type == kShown[13]) {
sample_type->set_unit(string_table_.GetIndex("output bytes"));
profile_pb->mutable_comment()->Add(
string_table_.GetIndex("Sum of operation output size."));
} else if (type == kShown[2]) {
sample_type->set_unit(string_table_.GetIndex("count"));
profile_pb->mutable_comment()->Add(
@ -370,7 +389,8 @@ const ShowMultiNode* TFCode::ShowInternal(const Options& opts,
}
string select = *opts.select.begin();
if (select != kShown[0] && select != kShown[1] && select != kShown[2] &&
select != kShown[3] && select != kShown[9] && select != kShown[10]) {
select != kShown[3] && select != kShown[9] && select != kShown[10] &&
select != kShown[11] && select != kShown[12] && select != kShown[13]) {
fprintf(stderr, "pprof doesn't support -select=%s\n", select.c_str());
return root_.get();
}
@ -522,17 +542,37 @@ std::vector<CodeNode*> TFCode::Account(const std::vector<CodeNode*>& roots,
return act_nodes;
}
string TFCode::FormatNode(CodeNode* node, const Options& opts, int64 indent) {
string TFCode::FormatNodeMemory(CodeNode* node, int64 bytes,
int64 total_bytes) const {
string memory = FormatMemory(total_bytes);
if (node->account) {
memory = FormatMemory(bytes) + "/" + memory;
} else {
memory = "--/" + memory;
}
return memory;
}
string TFCode::FormatNode(CodeNode* node, const Options& opts,
int64 indent) const {
std::vector<string> attrs;
if (opts.select.find(kShown[0]) != opts.select.end()) {
string memory = FormatMemory(node->proto().total_requested_bytes());
if (node->account) {
memory = FormatMemory(node->proto().requested_bytes()) + "/" + memory;
} else {
memory = "--/" + memory;
}
attrs.push_back(memory);
attrs.push_back(FormatNodeMemory(node, node->proto().requested_bytes(),
node->proto().total_requested_bytes()));
}
if (opts.select.find(kShown[11]) != opts.select.end()) {
attrs.push_back(FormatNodeMemory(node, node->proto().peak_bytes(),
node->proto().total_peak_bytes()));
}
if (opts.select.find(kShown[12]) != opts.select.end()) {
attrs.push_back(FormatNodeMemory(node, node->proto().residual_bytes(),
node->proto().total_residual_bytes()));
}
if (opts.select.find(kShown[13]) != opts.select.end()) {
attrs.push_back(FormatNodeMemory(node, node->proto().output_bytes(),
node->proto().total_output_bytes()));
}
std::vector<string> time_attrs = FormatTimes(node, opts);
attrs.insert(attrs.end(), time_attrs.begin(), time_attrs.end());

3
tensorflow/core/profiler/internal/tfprof_code.h
View File

@ -79,7 +79,8 @@ class TFCode : public TFMultiShow {
const Options& opts, string* display_str,
MultiGraphNodeProto* proto, std::vector<uint64>* call_ids);
string FormatNode(CodeNode* node, const Options& opts, int64 indent);
string FormatNode(CodeNode* node, const Options& opts, int64 indent) const;
string FormatNodeMemory(CodeNode* node, int64 bytes, int64 total_bytes) const;
std::unique_ptr<CodeNode> root_;
std::unique_ptr<TFMultiGraphNode> graph_root_;

16
tensorflow/core/profiler/internal/tfprof_node.cc
View File

@ -110,9 +110,11 @@ void ExecStep::AddMemoryStats(const string& dev,
uint64 output_ptr =
output.tensor_description().allocation_description().ptr();
total_output_bytes += output_bytes;
output_bytes_[output.slot()] = std::make_pair(output_bytes, output_ptr);
output_memory_[output.slot()] = std::make_pair(output_bytes, output_ptr);
}
}
output_bytes_ = total_output_bytes;
if (step_stat.has_memory_stats()) {
host_temp_bytes_ += step_stat.memory_stats().host_temp_memory_size();
host_persistent_bytes_ +=
@ -122,7 +124,17 @@ void ExecStep::AddMemoryStats(const string& dev,
accelerator_persistent_bytes_ +=
step_stat.memory_stats().device_persistent_memory_size();
}
requested_bytes_ = total_output_bytes;
int64 residual_bytes = 0;
int64 requested_bytes = 0;
int64 peak_bytes = 0;
for (const auto& mem : step_stat.memory()) {
residual_bytes += mem.live_bytes();
requested_bytes += mem.total_bytes();
peak_bytes += mem.peak_bytes();
}
requested_bytes_ = requested_bytes;
residual_bytes_ = residual_bytes;
peak_bytes_ = peak_bytes;
}
void TFGraphNode::AddStepStat(int64 step, const string& device,

79
tensorflow/core/profiler/internal/tfprof_node.h
View File

@ -51,6 +51,9 @@ class ExecStep {
latest_end_micros_(0),
mem_initiated_(false),
requested_bytes_(0),
peak_bytes_(0),
residual_bytes_(0),
output_bytes_(0),
host_temp_bytes_(0),
host_persistent_bytes_(0),
accelerator_temp_bytes_(0),
@ -78,14 +81,17 @@ class ExecStep {
int64 latest_end_micros() const { return latest_end_micros_; }
int64 requested_bytes() const { return requested_bytes_; }
int64 peak_bytes() const { return peak_bytes_; }
int64 residual_bytes() const { return residual_bytes_; }
int64 output_bytes() const { return output_bytes_; }
int64 accelerator_temp_bytes() const { return accelerator_temp_bytes_; }
int64 host_temp_bytes() const { return host_temp_bytes_; }
int64 accelerator_persistent_bytes() const {
return accelerator_persistent_bytes_;
}
int64 host_persistent_bytes() const { return host_persistent_bytes_; }
const std::map<int64, std::pair<int64, uint64>>& output_bytes() const {
return output_bytes_;
const std::map<int64, std::pair<int64, uint64>>& output_memory() const {
return output_memory_;
}
int64 allocator_bytes_in_use() const { return allocator_bytes_in_use_; }
@ -111,8 +117,14 @@ class ExecStep {
std::set<string> devices_;
bool mem_initiated_;
// Total output bytes requested by the op.
// Total bytes requested by the op.
int64 requested_bytes_;
// Total bytes requested by the op and released before op end.
int64 peak_bytes_;
// Total bytes requested by the op and not released after op end.
int64 residual_bytes_;
// Total bytes output by the op (not necessarily requested by the op).
int64 output_bytes_;
// Total temporary bytes allocated and released by the op.
int64 host_temp_bytes_;
// Total persistent bytes (e.g. variable) allocated by the op.
@ -122,9 +134,27 @@ class ExecStep {
// The total number of bytes currently allocated by the allocator if >0.
int64 allocator_bytes_in_use_;
// output_idx -> {output_bytes, memory_ptr}
std::map<int64, std::pair<int64, uint64>> output_bytes_;
std::map<int64, std::pair<int64, uint64>> output_memory_;
};
#define GRAPH_NODE_BYTES(type) \
do { \
if (execs_.empty()) { \
return 0; \
} \
if (step >= 0) { \
auto exec = execs_.find(step); \
CHECK(exec != execs_.end()) << "unknown step " << step; \
return exec->second.type##_bytes(); \
} \
\
int64 bytes = 0; \
for (const auto& exec : execs_) { \
bytes += exec.second.type##_bytes(); \
} \
return bytes / execs_.size(); \
} while (0)
class TFGraphNode {
public:
TFGraphNode(const NodeDef* node)
@ -270,22 +300,10 @@ class TFGraphNode {
return total_micros / execs_.size();
}
int64 requested_bytes(int64 step) const {
if (execs_.empty()) {
return 0;
}
if (step >= 0) {
auto exec = execs_.find(step);
CHECK(exec != execs_.end()) << "unknown step " << step;
return exec->second.requested_bytes();
}
int64 requested_bytes = 0;
for (const auto& exec : execs_) {
requested_bytes += exec.second.requested_bytes();
}
return requested_bytes / execs_.size();
}
int64 requested_bytes(int64 step) const { GRAPH_NODE_BYTES(requested); }
int64 peak_bytes(int64 step) const { GRAPH_NODE_BYTES(peak); }
int64 residual_bytes(int64 step) const { GRAPH_NODE_BYTES(residual); }
int64 output_bytes(int64 step) const { GRAPH_NODE_BYTES(output); }
int64 all_start_micros(int64 step) const {
auto exec = execs_.find(step);
@ -328,11 +346,11 @@ class TFGraphNode {
CHECK(exec != execs_.end()) << "unknown step " << step;
return exec->second.host_persistent_bytes();
}
const std::map<int64, std::pair<int64, uint64>>& output_bytes(
const std::map<int64, std::pair<int64, uint64>>& output_memory(
int64 step) const {
auto exec = execs_.find(step);
CHECK(exec != execs_.end()) << "unknown step " << step;
return exec->second.output_bytes();
return exec->second.output_memory();
}
int64 allocator_bytes_in_use(int64 step) const {
auto exec = execs_.find(step);
@ -427,6 +445,9 @@ class TFMultiGraphNode {
accelerator_exec_micros_(0),
cpu_exec_micros_(0),
requested_bytes_(0),
peak_bytes_(0),
residual_bytes_(0),
output_bytes_(0),
float_ops_(0),
parameters_(0) {}
@ -437,6 +458,10 @@ class TFMultiGraphNode {
cpu_exec_micros_ = 0;
requested_bytes_ = 0;
peak_bytes_ = 0;
residual_bytes_ = 0;
output_bytes_ = 0;
float_ops_ = 0;
parameters_ = 0;
op_types_.clear();
@ -460,6 +485,10 @@ class TFMultiGraphNode {
cpu_exec_micros_ += node->cpu_exec_micros(step);
requested_bytes_ += node->requested_bytes(step);
peak_bytes_ += node->peak_bytes(step);
residual_bytes_ += node->residual_bytes(step);
output_bytes_ += node->output_bytes(step);
float_ops_ += node->float_ops(step);
parameters_ += node->parameters();
if (node->shape().size() > 0) {
@ -492,6 +521,9 @@ class TFMultiGraphNode {
int64 cpu_exec_micros() const { return cpu_exec_micros_; }
int64 requested_bytes() const { return requested_bytes_; }
int64 peak_bytes() const { return peak_bytes_; }
int64 residual_bytes() const { return residual_bytes_; }
int64 output_bytes() const { return output_bytes_; }
int64 float_ops() const { return float_ops_; }
@ -540,6 +572,9 @@ class TFMultiGraphNode {
int64 cpu_exec_micros_;
int64 requested_bytes_;
int64 peak_bytes_;
int64 residual_bytes_;
int64 output_bytes_;
int64 float_ops_;
int64 parameters_;
std::set<string> devices_;

43
tensorflow/core/profiler/internal/tfprof_node_show.cc
View File

@ -38,6 +38,10 @@ void ShowNode::ReInit(int64 step) {
mutable_proto()->set_cpu_exec_micros(node->cpu_exec_micros(step));
mutable_proto()->set_requested_bytes(node->requested_bytes(step));
mutable_proto()->set_peak_bytes(node->peak_bytes(step));
mutable_proto()->set_residual_bytes(node->residual_bytes(step));
mutable_proto()->set_output_bytes(node->output_bytes(step));
mutable_proto()->set_float_ops(node->float_ops(step));
mutable_proto()->clear_input_shapes();
@ -68,6 +72,12 @@ void ShowNode::AggregateTotalStats(ShowNode* node) {
mutable_proto()->set_total_requested_bytes(proto().total_requested_bytes() +
node_pb->total_requested_bytes());
mutable_proto()->set_total_peak_bytes(proto().total_peak_bytes() +
node_pb->total_peak_bytes());
mutable_proto()->set_total_residual_bytes(proto().total_residual_bytes() +
node_pb->total_residual_bytes());
mutable_proto()->set_total_output_bytes(proto().total_output_bytes() +
node_pb->total_output_bytes());
mutable_proto()->set_total_parameters(proto().total_parameters() +
node_pb->total_parameters());
mutable_proto()->set_total_float_ops(proto().total_float_ops() +
@ -89,6 +99,13 @@ void ShowNode::AddSelfToTotalStats() {
mutable_proto()->set_total_requested_bytes(proto().total_requested_bytes() +
proto().requested_bytes());
mutable_proto()->set_total_peak_bytes(proto().total_peak_bytes() +
proto().peak_bytes());
mutable_proto()->set_total_residual_bytes(proto().total_residual_bytes() +
proto().residual_bytes());
mutable_proto()->set_total_output_bytes(proto().total_output_bytes() +
proto().output_bytes());
mutable_proto()->set_total_parameters(proto().total_parameters() +
proto().parameters());
mutable_proto()->set_total_float_ops(proto().total_float_ops() +
@ -105,6 +122,10 @@ void ShowNode::ResetTotalStats() {
mutable_proto()->set_total_cpu_exec_micros(0);
mutable_proto()->set_total_requested_bytes(0);
mutable_proto()->set_total_peak_bytes(0);
mutable_proto()->set_total_residual_bytes(0);
mutable_proto()->set_total_output_bytes(0);
mutable_proto()->set_total_parameters(0);
mutable_proto()->set_total_float_ops(0);
mutable_proto()->mutable_children()->Clear();
@ -135,6 +156,10 @@ bool ShowMultiNode::ReInit(int64 step,
mutable_proto()->set_cpu_exec_micros(node->cpu_exec_micros());
mutable_proto()->set_requested_bytes(node->requested_bytes());
mutable_proto()->set_peak_bytes(node->peak_bytes());
mutable_proto()->set_residual_bytes(node->residual_bytes());
mutable_proto()->set_output_bytes(node->output_bytes());
mutable_proto()->set_float_ops(node->float_ops());
mutable_proto()->set_parameters(node->parameters());
@ -157,6 +182,13 @@ void ShowMultiNode::AggregateTotalStats(ShowMultiNode* node) {
mutable_proto()->set_total_requested_bytes(proto().total_requested_bytes() +
node_pb->total_requested_bytes());
mutable_proto()->set_total_peak_bytes(proto().total_peak_bytes() +
node_pb->total_peak_bytes());
mutable_proto()->set_total_residual_bytes(proto().total_residual_bytes() +
node_pb->total_residual_bytes());
mutable_proto()->set_total_output_bytes(proto().total_output_bytes() +
node_pb->total_output_bytes());
mutable_proto()->set_total_parameters(proto().total_parameters() +
node_pb->total_parameters());
mutable_proto()->set_total_float_ops(proto().total_float_ops() +
@ -174,6 +206,13 @@ void ShowMultiNode::AddSelfToTotalStats() {
mutable_proto()->set_total_requested_bytes(proto().total_requested_bytes() +
proto().requested_bytes());
mutable_proto()->set_total_peak_bytes(proto().total_peak_bytes() +
proto().peak_bytes());
mutable_proto()->set_total_residual_bytes(proto().total_residual_bytes() +
proto().residual_bytes());
mutable_proto()->set_total_output_bytes(proto().total_output_bytes() +
proto().output_bytes());
mutable_proto()->set_total_parameters(proto().total_parameters() +
proto().parameters());
mutable_proto()->set_total_float_ops(proto().total_float_ops() +
@ -187,6 +226,10 @@ void ShowMultiNode::ResetTotalStats() {
mutable_proto()->set_total_cpu_exec_micros(0);
mutable_proto()->set_total_requested_bytes(0);
mutable_proto()->set_total_peak_bytes(0);
mutable_proto()->set_total_residual_bytes(0);
mutable_proto()->set_total_output_bytes(0);
mutable_proto()->set_total_parameters(0);
mutable_proto()->set_total_float_ops(0);
mutable_proto()->mutable_children()->Clear();

48
tensorflow/core/profiler/internal/tfprof_op.cc
View File

@ -211,24 +211,44 @@ int64 TFOp::SearchRoot(const std::vector<OpNode*> nodes,
return i;
}
string TFOp::FormatMemoryNode(int64 node_total_bytes, int64 root_total_bytes,
int64 node_bytes) const {
double accu_pct = 0.0;
double pct = 0.0;
if (node_bytes > 0) {
accu_pct = 100.0 * node_total_bytes / root_total_bytes;
pct = 100.0 * node_bytes / root_total_bytes;
}
return strings::Printf(
"%30s", strings::Printf("%s (%.2f%%, %.2f%%)",
FormatMemory(node_bytes).c_str(), accu_pct, pct)
.c_str());
}
string TFOp::FormatNode(OpNode* node, OpNode* root, const Options& opts) const {
std::vector<string> attrs;
if (opts.select.find(kShown[0]) != opts.select.end()) {
double accu_pct = 0.0;
double pct = 0.0;
if (node->proto().requested_bytes() > 0) {
accu_pct = 100.0 * node->proto().total_requested_bytes() /
root->proto().total_requested_bytes();
pct = 100.0 * node->proto().requested_bytes() /
root->proto().total_requested_bytes();
}
attrs.push_back(strings::Printf(
"%30s",
strings::Printf("%s (%.2f%%, %.2f%%)",
FormatMemory(node->proto().requested_bytes()).c_str(),
accu_pct, pct)
.c_str()));
attrs.push_back(FormatMemoryNode(node->proto().total_requested_bytes(),
root->proto().total_requested_bytes(),
node->proto().requested_bytes()));
}
if (opts.select.find(kShown[11]) != opts.select.end()) {
attrs.push_back(FormatMemoryNode(node->proto().total_peak_bytes(),
root->proto().total_peak_bytes(),
node->proto().peak_bytes()));
}
if (opts.select.find(kShown[12]) != opts.select.end()) {
attrs.push_back(FormatMemoryNode(node->proto().total_residual_bytes(),
root->proto().total_residual_bytes(),
node->proto().residual_bytes()));
}
if (opts.select.find(kShown[13]) != opts.select.end()) {
attrs.push_back(FormatMemoryNode(node->proto().total_output_bytes(),
root->proto().total_output_bytes(),
node->proto().output_bytes()));
}
if (opts.select.find(kShown[1]) != opts.select.end()) {

2
tensorflow/core/profiler/internal/tfprof_op.h
View File

@ -65,6 +65,8 @@ class TFOp : public TFMultiShow {
}
string FormatNode(OpNode* node, OpNode* root, const Options& opts) const;
string FormatMemoryNode(int64 node_total_bytes, int64 root_total_bytes,
int64 node_bytes) const;
std::unique_ptr<OpNode> root_;
std::map<string, std::unique_ptr<OpNode>> cnodes_map_;

38
tensorflow/core/profiler/internal/tfprof_options.cc
View File

@ -151,9 +151,11 @@ tensorflow::Status Options::FromProtoStr(const string& opts_proto_str,
}
*opts = Options(
opts_pb.max_depth(), opts_pb.min_bytes(), opts_pb.min_micros(),
opts_pb.min_params(), opts_pb.min_float_ops(), opts_pb.min_occurrence(),
opts_pb.step(), opts_pb.order_by(),
opts_pb.max_depth(), opts_pb.min_bytes(), opts_pb.min_peak_bytes(),
opts_pb.min_residual_bytes(), opts_pb.min_output_bytes(),
opts_pb.min_micros(), opts_pb.min_accelerator_micros(),
opts_pb.min_cpu_micros(), opts_pb.min_params(), opts_pb.min_float_ops(),
opts_pb.min_occurrence(), opts_pb.step(), opts_pb.order_by(),
std::vector<string>(opts_pb.account_type_regexes().begin(),
opts_pb.account_type_regexes().end()),
std::vector<string>(opts_pb.start_name_regexes().begin(),
@ -179,6 +181,11 @@ string Options::ToString() const {
"%-28s%lld\n"
"%-28s%lld\n"
"%-28s%lld\n"
"%-28s%lld\n"
"%-28s%lld\n"
"%-28s%lld\n"
"%-28s%lld\n"
"%-28s%lld\n"
"%-28s%s\n"
"%-28s%s\n"
"%-28s%s\n"
@ -188,17 +195,20 @@ string Options::ToString() const {
"%-28s%s\n"
"%-28s%s\n"
"%-28s%s:%s\n",
kOptions[0], max_depth, kOptions[1], min_bytes, kOptions[2], min_micros,
kOptions[3], min_params, kOptions[4], min_float_ops, kOptions[5],
min_occurrence, kOptions[6], step, kOptions[7], order_by.c_str(),
kOptions[8], str_util::Join(account_type_regexes, ",").c_str(),
kOptions[9], str_util::Join(start_name_regexes, ",").c_str(),
kOptions[10], str_util::Join(trim_name_regexes, ",").c_str(),
kOptions[11], str_util::Join(show_name_regexes, ",").c_str(),
kOptions[12], str_util::Join(hide_name_regexes, ",").c_str(),
kOptions[13], (account_displayed_op_only ? "true" : "false"),
kOptions[14], str_util::Join(select, ",").c_str(), kOptions[15],
output_type.c_str(), KeyValueToStr(output_options).c_str());
kOptions[0], max_depth, kOptions[1], min_bytes, kOptions[2],
min_peak_bytes, kOptions[3], min_residual_bytes, kOptions[4],
min_output_bytes, kOptions[5], min_micros, kOptions[6],
min_accelerator_micros, kOptions[7], min_cpu_micros, kOptions[8],
min_params, kOptions[9], min_float_ops, kOptions[10], min_occurrence,
kOptions[11], step, kOptions[12], order_by.c_str(), kOptions[13],
str_util::Join(account_type_regexes, ",").c_str(), kOptions[14],
str_util::Join(start_name_regexes, ",").c_str(), kOptions[15],
str_util::Join(trim_name_regexes, ",").c_str(), kOptions[16],
str_util::Join(show_name_regexes, ",").c_str(), kOptions[17],
str_util::Join(hide_name_regexes, ",").c_str(), kOptions[18],
(account_displayed_op_only ? "true" : "false"), kOptions[19],
str_util::Join(select, ",").c_str(), kOptions[20], output_type.c_str(),
KeyValueToStr(output_options).c_str());
return s;
}

41
tensorflow/core/profiler/internal/tfprof_options.h
View File

@ -29,7 +29,12 @@ namespace tfprof {
static const char* const kOptions[] = {
"-max_depth",
"-min_bytes",
"-min_peak_bytes",
"-min_residual_bytes",
"-min_output_bytes",
"-min_micros",
"-min_accelerator_micros",
"-min_cpu_micros",
"-min_params",
"-min_float_ops",
"-min_occurrence",
@ -46,17 +51,21 @@ static const char* const kOptions[] = {
};
static const char* const kOrderBy[] = {
"name", "bytes", "micros", "accelerator_micros",
"cpu_micros", "params", "float_ops", "occurrence",
"name", "bytes", "peak_bytes", "residual_bytes",
"output_bytes", "micros", "accelerator_micros", "cpu_micros",
"params", "float_ops", "occurrence",
};
// Append Only.
// TODO(xpan): As we are adding more fields to be selected, we
// need to have a way to tell users what fields are available in which view.
static const char* const kShown[] = {
"bytes", "micros", "params", "float_ops", "tensor_value",
"device", "op_types", "occurrence", "input_shapes", "accelerator_micros",
"cpu_micros"};
static const char* const kShown[] = {"bytes", "micros",
"params", "float_ops",
"tensor_value", "device",
"op_types", "occurrence",
"input_shapes", "accelerator_micros",
"cpu_micros", "peak_bytes",
"residual_bytes", "output_bytes"};
static const char* const kCmds[] = {
"scope", "graph", "code", "op", "advise", "set", "help",
@ -94,11 +103,15 @@ struct Options {
virtual ~Options() {}
Options()
: Options(0, 0, 0, 0, 0, 0, 0, "", {}, {}, {}, {}, {}, false, {}, "",
{}) {}
: Options(0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, "", {}, {}, {}, {}, {},
false, {}, "", {}) {}
Options(int max_depth, tensorflow::int64 min_bytes,
tensorflow::int64 min_micros, tensorflow::int64 min_params,
tensorflow::int64 min_peak_bytes,
tensorflow::int64 min_residual_bytes,
tensorflow::int64 min_output_bytes, tensorflow::int64 min_micros,
tensorflow::int64 min_accelerator_micros,
tensorflow::int64 min_cpu_micros, tensorflow::int64 min_params,
tensorflow::int64 min_float_ops, tensorflow::int64 min_occurrence,
tensorflow::int64 step, const string& order_by,
const std::vector<string>& account_type_regexes,
@ -111,7 +124,12 @@ struct Options {
const std::map<string, string>& output_options)
: max_depth(max_depth),
min_bytes(min_bytes),
min_peak_bytes(min_peak_bytes),
min_residual_bytes(min_residual_bytes),
min_output_bytes(min_output_bytes),
min_micros(min_micros),
min_accelerator_micros(min_accelerator_micros),
min_cpu_micros(min_cpu_micros),
min_params(min_params),
min_float_ops(min_float_ops),
min_occurrence(min_occurrence),
@ -131,7 +149,12 @@ struct Options {
int max_depth;
tensorflow::int64 min_bytes;
tensorflow::int64 min_peak_bytes;
tensorflow::int64 min_residual_bytes;
tensorflow::int64 min_output_bytes;
tensorflow::int64 min_micros;
tensorflow::int64 min_accelerator_micros;
tensorflow::int64 min_cpu_micros;
tensorflow::int64 min_params;
tensorflow::int64 min_float_ops;
tensorflow::int64 min_occurrence;

53
tensorflow/core/profiler/internal/tfprof_show.cc
View File

@ -73,8 +73,14 @@ bool TFShow::ShouldShow(const ShowNode* node, const Options& opts,
// Always show kTFProfRoot.
if (node->name() == kTFProfRoot) return true;
if (node->proto().requested_bytes() < opts.min_bytes ||
node->proto().exec_micros() < opts.min_micros ||
if (node->proto().total_requested_bytes() < opts.min_bytes ||
node->proto().total_peak_bytes() < opts.min_peak_bytes ||
node->proto().total_residual_bytes() < opts.min_residual_bytes ||
node->proto().total_output_bytes() < opts.min_output_bytes ||
node->proto().total_exec_micros() < opts.min_micros ||
node->proto().total_accelerator_exec_micros() <
opts.min_accelerator_micros ||
node->proto().total_cpu_exec_micros() < opts.min_cpu_micros ||
node->proto().parameters() < opts.min_params ||
node->proto().float_ops() < opts.min_float_ops ||
node->proto().run_count() < opts.min_occurrence ||
@ -128,6 +134,17 @@ bool TFShow::ReAccount(ShowNode* node, const Options& opts) {
return false;
}
string TFShow::FormatNodeMemory(ShowNode* node, int64 bytes,
int64 total_bytes) const {
string memory = FormatMemory(total_bytes);
if (node->account) {
memory = FormatMemory(bytes) + "/" + memory;
} else {
memory = "--/" + memory;
}
return memory;
}
string TFShow::FormatNode(ShowNode* node, const Options& opts) const {
std::vector<string> info;
if (opts.select.find(kShown[2]) != opts.select.end()) {
@ -152,15 +169,22 @@ string TFShow::FormatNode(ShowNode* node, const Options& opts) const {
}
info.push_back(fops);
}
std::vector<string> attrs;
if (opts.select.find(kShown[0]) != opts.select.end()) {
string memory = FormatMemory(node->proto().total_requested_bytes());
if (node->account) {
memory = FormatMemory(node->proto().requested_bytes()) + "/" + memory;
} else {
memory = "--/" + memory;
}
info.push_back(memory);
info.push_back(FormatNodeMemory(node, node->proto().requested_bytes(),
node->proto().total_requested_bytes()));
}
if (opts.select.find(kShown[11]) != opts.select.end()) {
info.push_back(FormatNodeMemory(node, node->proto().peak_bytes(),
node->proto().total_peak_bytes()));
}
if (opts.select.find(kShown[12]) != opts.select.end()) {
info.push_back(FormatNodeMemory(node, node->proto().residual_bytes(),
node->proto().total_residual_bytes()));
}
if (opts.select.find(kShown[13]) != opts.select.end()) {
info.push_back(FormatNodeMemory(node, node->proto().output_bytes(),
node->proto().total_output_bytes()));
}
if (opts.select.find(kShown[1]) != opts.select.end()) {
info.push_back(FormatTotalExecTime(node, opts));
@ -225,6 +249,15 @@ string TFShow::FormatLegend(const Options& opts) const {
legends.push_back("# float_ops");
}
if (opts.select.find(kShown[0]) != opts.select.end()) {
legends.push_back("requested bytes");
}
if (opts.select.find(kShown[11]) != opts.select.end()) {
legends.push_back("peak bytes");
}
if (opts.select.find(kShown[12]) != opts.select.end()) {
legends.push_back("residual bytes");
}
if (opts.select.find(kShown[13]) != opts.select.end()) {
legends.push_back("output bytes");
}
if (opts.select.find(kShown[1]) != opts.select.end()) {

17
tensorflow/core/profiler/internal/tfprof_show.h
View File

@ -67,6 +67,7 @@ class TFShow {
bool ReAccount(ShowNode* node, const Options& opts);
string FormatNode(ShowNode* node, const Options& opts) const;
string FormatNodeMemory(ShowNode* node, int64 bytes, int64 total_bytes) const;
string FormatLegend(const Options& opts) const;
@ -87,17 +88,25 @@ class TFShow {
return n1->proto().total_requested_bytes() >
n2->proto().total_requested_bytes();
} else if (opts.order_by == kOrderBy[2]) {
return n1->proto().total_peak_bytes() > n2->proto().total_peak_bytes();
} else if (opts.order_by == kOrderBy[3]) {
return n1->proto().total_residual_bytes() >
n2->proto().total_residual_bytes();
} else if (opts.order_by == kOrderBy[4]) {
return n1->proto().total_output_bytes() >
n2->proto().total_output_bytes();
} else if (opts.order_by == kOrderBy[5]) {
return n1->proto().total_exec_micros() >
n2->proto().total_exec_micros();
} else if (opts.order_by == kOrderBy[3]) {
} else if (opts.order_by == kOrderBy[6]) {
return n1->proto().total_accelerator_exec_micros() >
n2->proto().total_accelerator_exec_micros();
} else if (opts.order_by == kOrderBy[4]) {
} else if (opts.order_by == kOrderBy[7]) {
return n1->proto().total_cpu_exec_micros() >
n2->proto().total_cpu_exec_micros();
} else if (opts.order_by == kOrderBy[5]) {
} else if (opts.order_by == kOrderBy[8]) {
return n1->proto().total_parameters() > n2->proto().total_parameters();
} else if (opts.order_by == kOrderBy[6]) {
} else if (opts.order_by == kOrderBy[9]) {
return n1->proto().total_float_ops() > n2->proto().total_float_ops();
}
return name_cmp;

15
tensorflow/core/profiler/internal/tfprof_show_multi.cc
View File

@ -65,7 +65,13 @@ bool TFMultiShow::ShouldShow(const ShowMultiNode* node, const Options& opts,
// want to see the middle code traces (i.e. their own codes.), instead
// of the TensorFlow internal codes traces.
if (node->proto().total_requested_bytes() < opts.min_bytes ||
node->proto().total_peak_bytes() < opts.min_peak_bytes ||
node->proto().total_residual_bytes() < opts.min_residual_bytes ||
node->proto().total_output_bytes() < opts.min_output_bytes ||
node->proto().total_exec_micros() < opts.min_micros ||
node->proto().total_accelerator_exec_micros() <
opts.min_accelerator_micros ||
node->proto().total_cpu_exec_micros() < opts.min_cpu_micros ||
node->proto().total_parameters() < opts.min_params ||
node->proto().total_float_ops() < opts.min_float_ops ||
depth > opts.max_depth || !ShouldShowIfExtra(node, opts, depth)) {
@ -109,6 +115,15 @@ bool TFMultiShow::ReAccount(ShowMultiNode* node, const Options& opts) {
string TFMultiShow::FormatLegend(const Options& opts) const {
std::vector<string> legends;
if (opts.select.find(kShown[0]) != opts.select.end()) {
legends.push_back("requested bytes");
}
if (opts.select.find(kShown[11]) != opts.select.end()) {
legends.push_back("peak bytes");
}
if (opts.select.find(kShown[12]) != opts.select.end()) {
legends.push_back("residual bytes");
}
if (opts.select.find(kShown[13]) != opts.select.end()) {
legends.push_back("output bytes");
}
if (opts.select.find(kShown[1]) != opts.select.end()) {

19
tensorflow/core/profiler/internal/tfprof_show_multi.h
View File

@ -90,21 +90,30 @@ class TFMultiShow {
return n1->proto().total_requested_bytes() >
n2->proto().total_requested_bytes();
} else if (opts.order_by == kOrderBy[2]) {
return n1->proto().total_peak_bytes() >
n2->proto().total_peak_bytes();
} else if (opts.order_by == kOrderBy[3]) {
return n1->proto().total_residual_bytes() >
n2->proto().total_residual_bytes();
} else if (opts.order_by == kOrderBy[4]) {
return n1->proto().total_output_bytes() >
n2->proto().total_output_bytes();
} else if (opts.order_by == kOrderBy[5]) {
return n1->proto().total_exec_micros() >
n2->proto().total_exec_micros();
} else if (opts.order_by == kOrderBy[3]) {
} else if (opts.order_by == kOrderBy[6]) {
return n1->proto().total_accelerator_exec_micros() >
n2->proto().total_accelerator_exec_micros();
} else if (opts.order_by == kOrderBy[4]) {
} else if (opts.order_by == kOrderBy[7]) {
return n1->proto().total_cpu_exec_micros() >
n2->proto().total_cpu_exec_micros();
} else if (opts.order_by == kOrderBy[5]) {
} else if (opts.order_by == kOrderBy[8]) {
return n1->proto().total_parameters() >
n2->proto().total_parameters();
} else if (opts.order_by == kOrderBy[6]) {
} else if (opts.order_by == kOrderBy[9]) {
return n1->proto().total_float_ops() >
n2->proto().total_float_ops();
} else if (opts.order_by == kOrderBy[7]) {
} else if (opts.order_by == kOrderBy[10]) {
return n1->node->graph_nodes().size() >
n2->node->graph_nodes().size();
}

166
tensorflow/core/profiler/internal/tfprof_show_test.cc
View File

@ -22,12 +22,12 @@ limitations under the License.
#include "tensorflow/core/lib/io/path.h"
#include "tensorflow/core/platform/env.h"
#include "tensorflow/core/platform/test.h"
#include "tensorflow/core/protobuf/config.pb.h"
#include "tensorflow/core/profiler/internal/tfprof_constants.h"
#include "tensorflow/core/profiler/internal/tfprof_options.h"
#include "tensorflow/core/profiler/internal/tfprof_utils.h"
#include "tensorflow/core/profiler/tfprof_log.pb.h"
#include "tensorflow/core/profiler/tfprof_output.pb.h"
#include "tensorflow/core/protobuf/config.pb.h"
namespace tensorflow {
namespace tfprof {
@ -72,91 +72,80 @@ class TFProfShowTest : public ::testing::Test {
};
TEST_F(TFProfShowTest, DumpScopeMode) {
string dump_file = io::JoinPath(testing::TmpDir(), "dump");
Options opts(5, 0, 0, 0, 0, 0, -1, "name",
{"VariableV2"}, // accout_type_regexes
{".*"}, {""}, {".*"}, {""}, false,
{"params", "bytes", "micros", "float_ops"}, "file",
{{"outfile", dump_file}});
tf_stats_->ShowGraphNode("scope", opts);
string dump_str;
TF_CHECK_OK(ReadFileToString(Env::Default(), dump_file, &dump_str));
EXPECT_EQ(
"node name | # parameters | # float_ops | output bytes | total execution "
"time | accelerator execution time | cpu execution time\n_TFProfRoot "
"(--/370 params, --/0 flops, --/1.48KB, --/5us, --/0us, --/5us)\n "
"conv2d (--/140 params, --/0 flops, --/560B, --/2us, --/0us, --/2us)\n "
" conv2d/bias (5, 5/5 params, 0/0 flops, 20B/20B, 1us/1us, 0us/0us, "
"1us/1us)\n conv2d/kernel (3x3x3x5, 135/135 params, 0/0 flops, "
"540B/540B, 1us/1us, 0us/0us, 1us/1us)\n conv2d_1 (--/230 params, --/0 "
"flops, --/920B, --/3us, --/0us, --/3us)\n conv2d_1/bias (5, 5/5 "
"params, 0/0 flops, 20B/20B, 1us/1us, 0us/0us, 1us/1us)\n "
"conv2d_1/kernel (3x3x5x5, 225/225 params, 0/0 flops, 900B/900B, "
"2us/2us, 0us/0us, 2us/2us)\n",
dump_str);
}
TEST_F(TFProfShowTest, DumpAcceleratorAndCPUMicros) {
string dump_file = io::JoinPath(testing::TmpDir(), "dump");
Options opts(
5, 0, 0, 0, 0, 0, -1, "cpu_micros", {".*"}, // accout_type_regexes
{".*"}, {""}, {".*"}, {""}, false, {"accelerator_micros", "cpu_micros"},
5, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, "name",
{"VariableV2"}, // accout_type_regexes
{".*"}, {""}, {".*"}, {""}, false,
{"params", "bytes", "peak_bytes", "residual_bytes", "output_bytes",
"micros", "accelerator_micros", "cpu_micros", "float_ops"},
"file", {{"outfile", dump_file}});
tf_stats_->ShowGraphNode("scope", opts);
string dump_str;
TF_CHECK_OK(ReadFileToString(Env::Default(), dump_file, &dump_str));
EXPECT_EQ(
"node name | # parameters | # float_ops | requested bytes | peak bytes | "
"residual bytes | output bytes | total execution time | accelerator "
"execution time | cpu execution time\n_TFProfRoot (--/451 params, --/0 "
"flops, --/0B, --/0B, --/0B, --/2.56KB, --/13us, --/0us, --/13us)\n DW "
"(3x3x3x6, 162/162 params, 0/0 flops, 0B/0B, 0B/0B, 0B/0B, "
"1.28KB/1.28KB, 2us/2us, 0us/0us, 2us/2us)\n DW2 (2x2x6x12, 288/288 "
"params, 0/0 flops, 0B/0B, 0B/0B, 0B/0B, 1.28KB/1.28KB, 11us/11us, "
"0us/0us, 11us/11us)\n ScalarW (1, 1/1 params, 0/0 flops, 0B/0B, 0B/0B, "
"0B/0B, 0B/0B, 0us/0us, 0us/0us, 0us/0us)\n",
dump_str);
}
TEST_F(TFProfShowTest, DumpAcceleratorAndCPUMicros) {
string dump_file = io::JoinPath(testing::TmpDir(), "dump");
Options opts(5, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, "cpu_micros",
{".*"}, // accout_type_regexes
{".*"}, {""}, {".*"}, {""}, false,
{"accelerator_micros", "cpu_micros"}, "file",
{{"outfile", dump_file}});
tf_stats_->ShowGraphNode("scope", opts);
string dump_str;
TF_CHECK_OK(ReadFileToString(Env::Default(), dump_file, &dump_str));
EXPECT_EQ(
"node name | accelerator execution time | cpu execution "
"time\n_TFProfRoot (--/0us, --/97us)\n conv2d (0us/0us, 0us/76us)\n "
"conv2d/convolution (0us/0us, 60us/60us)\n conv2d/convolution/Shape "
"(0us/0us, 0us/0us)\n conv2d/convolution/dilation_rate (0us/0us, "
"0us/0us)\n conv2d/BiasAdd (0us/0us, 12us/12us)\n conv2d/bias "
"(0us/0us, 1us/2us)\n conv2d/bias/read (0us/0us, 1us/1us)\n "
"conv2d/bias/Assign (0us/0us, 0us/0us)\n conv2d/bias/Initializer "
"(0us/0us, 0us/0us)\n conv2d/bias/Initializer/Const (0us/0us, "
"0us/0us)\n conv2d/kernel (0us/0us, 1us/2us)\n "
"conv2d/kernel/read (0us/0us, 1us/1us)\n conv2d/kernel/Assign "
"(0us/0us, 0us/0us)\n conv2d/kernel/Initializer (0us/0us, "
"0us/0us)\n conv2d/kernel/Initializer/random_uniform (0us/0us, "
"0us/0us)\n conv2d_2 (0us/0us, 0us/15us)\n conv2d_2/convolution "
"(0us/0us, 13us/13us)\n conv2d_2/convolution/Shape (0us/0us, "
"0us/0us)\n conv2d_2/convolution/dilation_rate (0us/0us, 0us/0us)\n "
" conv2d_2/BiasAdd (0us/0us, 2us/2us)\n conv2d_1 (0us/0us, 0us/5us)\n "
" conv2d_1/kernel (0us/0us, 2us/3us)\n conv2d_1/kernel/read "
"(0us/0us, 1us/1us)\n conv2d_1/kernel/Assign (0us/0us, 0us/0us)\n "
" conv2d_1/kernel/Initializer (0us/0us, 0us/0us)\n "
"conv2d_1/kernel/Initializer/random_uniform (0us/0us, 0us/0us)\n "
"conv2d_1/bias (0us/0us, 1us/2us)\n conv2d_1/bias/read (0us/0us, "
"1us/1us)\n conv2d_1/bias/Assign (0us/0us, 0us/0us)\n "
"conv2d_1/bias/Initializer (0us/0us, 0us/0us)\n "
"conv2d_1/bias/Initializer/Const (0us/0us, 0us/0us)\n zeros (0us/0us, "
"1us/1us)\n init (0us/0us, 0us/0us)\n save (0us/0us, 0us/0us)\n "
"save/Assign (0us/0us, 0us/0us)\n save/Assign_1 (0us/0us, 0us/0us)\n "
" save/Assign_2 (0us/0us, 0us/0us)\n save/Assign_3 (0us/0us, "
"0us/0us)\n save/Const (0us/0us, 0us/0us)\n save/RestoreV2 "
"(0us/0us, 0us/0us)\n save/RestoreV2/shape_and_slices (0us/0us, "
"0us/0us)\n save/RestoreV2/tensor_names (0us/0us, 0us/0us)\n "
"save/RestoreV2_1 (0us/0us, 0us/0us)\n "
"save/RestoreV2_1/shape_and_slices (0us/0us, 0us/0us)\n "
"save/RestoreV2_1/tensor_names (0us/0us, 0us/0us)\n save/RestoreV2_2 "
"(0us/0us, 0us/0us)\n save/RestoreV2_2/shape_and_slices (0us/0us, "
"0us/0us)\n save/RestoreV2_2/tensor_names (0us/0us, 0us/0us)\n "
"save/RestoreV2_3 (0us/0us, 0us/0us)\n "
"save/RestoreV2_3/shape_and_slices (0us/0us, 0us/0us)\n "
"save/RestoreV2_3/tensor_names (0us/0us, 0us/0us)\n save/SaveV2 "
"(0us/0us, 0us/0us)\n save/SaveV2/shape_and_slices (0us/0us, "
"0us/0us)\n save/SaveV2/tensor_names (0us/0us, 0us/0us)\n "
"save/control_dependency (0us/0us, 0us/0us)\n save/restore_all "
"(0us/0us, 0us/0us)\n",
"time\n_TFProfRoot (--/404us, --/4.50ms)\n Conv2D (226us/226us, "
"4.07ms/4.07ms)\n Conv2D_1 (178us/178us, 419us/419us)\n DW2 (0us/0us, "
"11us/11us)\n DW2/Assign (0us/0us, 0us/0us)\n DW2/Initializer "
"(0us/0us, 0us/0us)\n DW2/Initializer/random_normal (0us/0us, "
"0us/0us)\n DW2/Initializer/random_normal/RandomStandardNormal "
"(0us/0us, 0us/0us)\n DW2/Initializer/random_normal/mean "
"(0us/0us, 0us/0us)\n DW2/Initializer/random_normal/mul (0us/0us, "
"0us/0us)\n DW2/Initializer/random_normal/shape (0us/0us, "
"0us/0us)\n DW2/Initializer/random_normal/stddev (0us/0us, "
"0us/0us)\n DW2/read (0us/0us, 0us/0us)\n DW (0us/0us, 2us/2us)\n "
"DW/Assign (0us/0us, 0us/0us)\n DW/Initializer (0us/0us, 0us/0us)\n "
" DW/Initializer/random_normal (0us/0us, 0us/0us)\n "
"DW/Initializer/random_normal/RandomStandardNormal (0us/0us, 0us/0us)\n "
" DW/Initializer/random_normal/mean (0us/0us, 0us/0us)\n "
"DW/Initializer/random_normal/mul (0us/0us, 0us/0us)\n "
"DW/Initializer/random_normal/shape (0us/0us, 0us/0us)\n "
"DW/Initializer/random_normal/stddev (0us/0us, 0us/0us)\n DW/read "
"(0us/0us, 0us/0us)\n zeros (0us/0us, 2us/2us)\n ScalarW (0us/0us, "
"0us/0us)\n ScalarW/Assign (0us/0us, 0us/0us)\n "
"ScalarW/Initializer (0us/0us, 0us/0us)\n "
"ScalarW/Initializer/random_normal (0us/0us, 0us/0us)\n "
"ScalarW/Initializer/random_normal/RandomStandardNormal (0us/0us, "
"0us/0us)\n ScalarW/Initializer/random_normal/mean (0us/0us, "
"0us/0us)\n ScalarW/Initializer/random_normal/mul (0us/0us, "
"0us/0us)\n ScalarW/Initializer/random_normal/shape (0us/0us, "
"0us/0us)\n ScalarW/Initializer/random_normal/stddev (0us/0us, "
"0us/0us)\n ScalarW/read (0us/0us, 0us/0us)\n init (0us/0us, "
"0us/0us)\n",
dump_str);
}
TEST_F(TFProfShowTest, DumpOpMode) {
string dump_file = io::JoinPath(testing::TmpDir(), "dump");
Options opts(
5, 0, 0, 0, 0, 4, -1, "params", {".*"}, // accout_type_regexes
5, 0, 0, 0, 0, 0, 0, 0, 0, 0, 1, -1, "params",
{".*"}, // accout_type_regexes
{".*"}, {""}, {".*"}, {""}, false,
{"params", "bytes", "micros", "float_ops", "occurrence", "input_shapes"},
"file", {{"outfile", dump_file}});
@ -165,17 +154,32 @@ TEST_F(TFProfShowTest, DumpOpMode) {
string dump_str;
TF_CHECK_OK(ReadFileToString(Env::Default(), dump_file, &dump_str));
EXPECT_EQ(
"nodename|outputbytes|totalexecutiontime|acceleratorexecutiontime|"
"nodename|requestedbytes|totalexecutiontime|acceleratorexecutiontime|"
"cpuexecutiontime|#parameters|#float_ops|opoccurrence(run|defined)|"
"inputshapes\nVariableV21.48KB(100.00%,17.10%),5us(100.00%,5.15%),0us(0."
"00%,0.00%),5us(100.00%,5.15%),370params(100.00%,100.00%),0float_ops(100."
"00%,0.00%),4|4\n\ninput_type:\t(run*4|defined*4)\texec_time:"
"5us\n\nAssign0B(0.00%,0.00%),0us(94.85%,0.00%),0us(0.00%,0.00%),0us(94."
"85%,0.00%),0params(0.00%,0.00%),0float_ops(100.00%,0.00%),0|8\n\ninput_"
"type:0:unknown,\t1:unknown\t(run*0|defined*8)\texec_time:0us\n\nConst1."
"54KB(58.87%,17.74%),1us(80.41%,1.03%),0us(0.00%,0.00%),1us(80.41%,1.03%)"
",0params(0.00%,0.00%),0float_ops(98.49%,0.00%),1|24\n\ninput_type:\t("
"run*1|defined*24)\texec_time:1us\n\n",
"inputshapes\nVariableV20B(0.00%,0.00%),13us(100.00%,0.27%),0us(100.00%,"
"0.00%),13us(100.00%,0.29%),451params(100.00%,100.00%),0float_ops(100.00%"
",0.00%),2|3\n\ninput_type:\t(run*2|defined*3)\texec_time:13us\n\nAdd0B("
"0.00%,0.00%),0us(99.73%,0.00%),0us(100.00%,0.00%),0us(99.71%,0.00%),"
"0params(0.00%,0.00%),0float_ops(100.00%,0.00%),0|3\n\ninput_type:0:1,"
"\t1:1\t(run*0|defined*1)\texec_time:0us\ninput_type:0:2x2x6x12,\t1:1\t("
"run*0|defined*1)\texec_time:0us\ninput_type:0:3x3x3x6,\t1:1\t(run*0|"
"defined*1)\texec_time:0us\n\nAssign0B(0.00%,0.00%),0us(99.73%,0.00%),"
"0us(100.00%,0.00%),0us(99.71%,0.00%),0params(0.00%,0.00%),0float_ops("
"100.00%,0.00%),0|3\n\ninput_type:0:1,\t1:1\t(run*0|defined*1)\texec_"
"time:0us\ninput_type:0:2x2x6x12,\t1:2x2x6x12\t(run*0|defined*1)\texec_"
"time:0us\ninput_type:0:3x3x3x6,\t1:3x3x3x6\t(run*0|defined*1)\texec_"
"time:0us\n\nConst0B(0.00%,0.00%),2us(99.73%,0.04%),0us(100.00%,0.00%),"
"2us(99.71%,0.04%),0params(0.00%,0.00%),0float_ops(100.00%,0.00%),1|"
"10\n\ninput_type:\t(run*1|defined*10)\texec_time:2us\n\nConv2D14.59KB("
"100.00%,100.00%),4.89ms(99.69%,99.69%),404us(100.00%,100.00%),4.49ms(99."
"67%,99.67%),0params(0.00%,0.00%),10.44kfloat_ops(100.00%,100.00%),2|"
"2\n\ninput_type:0:2x3x3x6,\t1:2x2x6x12\t(run*1|defined*1)\texec_time:"
"597us\ninput_type:0:2x6x6x3,\t1:3x3x3x6\t(run*1|defined*1)\texec_time:4."
"29ms\n\nIdentity0B(0.00%,0.00%),0us(0.00%,0.00%),0us(0.00%,0.00%),0us(0."
"00%,0.00%),0params(0.00%,0.00%),0float_ops(0.00%,0.00%),0|3\n\ninput_"
"type:0:1\t(run*0|defined*1)\texec_time:0us\ninput_type:0:2x2x6x12\t(run*"
"0|defined*1)\texec_time:0us\ninput_type:0:3x3x3x6\t(run*0|defined*1)"
"\texec_time:0us\n\n",
StringReplace(dump_str, " ", ""));
}
} // namespace tfprof

383
tensorflow/core/profiler/internal/tfprof_stats_test.cc
View File

@ -23,12 +23,12 @@ limitations under the License.
#include "tensorflow/core/platform/env.h"
#include "tensorflow/core/platform/protobuf.h"
#include "tensorflow/core/platform/test.h"
#include "tensorflow/core/protobuf/config.pb.h"
#include "tensorflow/core/profiler/internal/tfprof_constants.h"
#include "tensorflow/core/profiler/internal/tfprof_options.h"
#include "tensorflow/core/profiler/internal/tfprof_utils.h"
#include "tensorflow/core/profiler/tfprof_log.pb.h"
#include "tensorflow/core/profiler/tfprof_output.pb.h"
#include "tensorflow/core/protobuf/config.pb.h"
namespace tensorflow {
namespace tfprof {
@ -73,7 +73,7 @@ class TFProfStatsTest : public ::testing::Test {
};
TEST_F(TFProfStatsTest, CustomOpType) {
Options opts(3, 0, 0, 0, 0, 0, -1, "name",
Options opts(3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, "name",
{kTrainableVarType}, // accout_type_regexes
{".*"}, {""}, {".*"}, {""}, false,
{"params", "bytes", "micros", "float_ops"}, "", {});
@ -81,62 +81,27 @@ TEST_F(TFProfStatsTest, CustomOpType) {
GraphNodeProto expected;
CHECK(protobuf::TextFormat::ParseFromString(
"name: \"_TFProfRoot\"\nexec_micros: 0\nrequested_bytes: "
"0\ntotal_exec_micros: 5\ntotal_requested_bytes: 1480\ntotal_parameters: "
"370\nchildren {\n name: \"conv2d\"\n exec_micros: 0\n "
"requested_bytes: 0\n total_exec_micros: 2\n total_requested_bytes: "
"560\n total_parameters: 140\n children {\n name: \"conv2d/bias\"\n "
" exec_micros: 1\n requested_bytes: 20\n parameters: 5\n "
"total_exec_micros: 1\n total_requested_bytes: 20\n "
"total_parameters: 5\n devices: "
"\"/job:localhost/replica:0/task:0/cpu:0\"\n float_ops: 0\n "
"total_float_ops: 0\n accelerator_exec_micros: 0\n "
"cpu_exec_micros: 1\n total_accelerator_exec_micros: 0\n "
"total_cpu_exec_micros: 1\n run_count: 1\n total_run_count: 1\n "
"total_definition_count: 1\n }\n children {\n name: "
"\"conv2d/kernel\"\n exec_micros: 1\n requested_bytes: 540\n "
"parameters: 135\n total_exec_micros: 1\n total_requested_bytes: "
"540\n total_parameters: 135\n devices: "
"\"/job:localhost/replica:0/task:0/cpu:0\"\n float_ops: 0\n "
"total_float_ops: 0\n accelerator_exec_micros: 0\n "
"cpu_exec_micros: 1\n total_accelerator_exec_micros: 0\n "
"total_cpu_exec_micros: 1\n run_count: 1\n total_run_count: 1\n "
"total_definition_count: 1\n }\n float_ops: 0\n total_float_ops: 0\n "
"accelerator_exec_micros: 0\n cpu_exec_micros: 0\n "
"total_accelerator_exec_micros: 0\n total_cpu_exec_micros: 2\n "
"run_count: 0\n total_run_count: 2\n total_definition_count: "
"3\n}\nchildren {\n name: \"conv2d_1\"\n exec_micros: 0\n "
"requested_bytes: 0\n total_exec_micros: 3\n total_requested_bytes: "
"920\n total_parameters: 230\n children {\n name: "
"\"conv2d_1/bias\"\n exec_micros: 1\n requested_bytes: 20\n "
"parameters: 5\n total_exec_micros: 1\n total_requested_bytes: "
"20\n total_parameters: 5\n devices: "
"\"/job:localhost/replica:0/task:0/cpu:0\"\n float_ops: 0\n "
"total_float_ops: 0\n accelerator_exec_micros: 0\n "
"cpu_exec_micros: 1\n total_accelerator_exec_micros: 0\n "
"total_cpu_exec_micros: 1\n run_count: 1\n total_run_count: 1\n "
"total_definition_count: 1\n }\n children {\n name: "
"\"conv2d_1/kernel\"\n exec_micros: 2\n requested_bytes: 900\n "
"parameters: 225\n total_exec_micros: 2\n total_requested_bytes: "
"900\n total_parameters: 225\n devices: "
"\"/job:localhost/replica:0/task:0/cpu:0\"\n float_ops: 0\n "
"total_float_ops: 0\n accelerator_exec_micros: 0\n "
"cpu_exec_micros: 2\n total_accelerator_exec_micros: 0\n "
"total_cpu_exec_micros: 2\n run_count: 1\n total_run_count: 1\n "
"total_definition_count: 1\n }\n float_ops: 0\n total_float_ops: 0\n "
"accelerator_exec_micros: 0\n cpu_exec_micros: 0\n "
"total_accelerator_exec_micros: 0\n total_cpu_exec_micros: 3\n "
"run_count: 0\n total_run_count: 2\n total_definition_count: "
"3\n}\nfloat_ops: 0\ntotal_float_ops: 0\naccelerator_exec_micros: "
"0\ncpu_exec_micros: 0\ntotal_accelerator_exec_micros: "
"0\ntotal_cpu_exec_micros: 5\nrun_count: 0\ntotal_run_count: "
"4\ntotal_definition_count: 6\n",
"name: \"_TFProfRoot\"\ntotal_exec_micros: 13\ntotal_parameters: "
"451\nchildren {\n name: \"DW\"\n exec_micros: 2\n parameters: 162\n "
"total_exec_micros: 2\n total_parameters: 162\n devices: "
"\"/job:localhost/replica:0/task:0/gpu:0\"\n cpu_exec_micros: 2\n "
"total_cpu_exec_micros: 2\n run_count: 1\n total_run_count: 1\n "
"total_definition_count: 1\n output_bytes: 1280\n total_output_bytes: "
"1280\n}\nchildren {\n name: \"DW2\"\n exec_micros: 11\n parameters: "
"288\n total_exec_micros: 11\n total_parameters: 288\n devices: "
"\"/job:localhost/replica:0/task:0/gpu:0\"\n cpu_exec_micros: 11\n "
"total_cpu_exec_micros: 11\n run_count: 1\n total_run_count: 1\n "
"total_definition_count: 1\n output_bytes: 1280\n total_output_bytes: "
"1280\n}\nchildren {\n name: \"ScalarW\"\n parameters: 1\n "
"total_parameters: 1\n total_definition_count: "
"1\n}\ntotal_cpu_exec_micros: 13\ntotal_run_count: "
"2\ntotal_definition_count: 3\ntotal_output_bytes: 2560\n",
&expected));
EXPECT_EQ(expected.DebugString(), root.DebugString());
}
TEST_F(TFProfStatsTest, CheckPointOpType) {
Options opts(3, 0, 0, 0, 0, 0, -1, "name",
Options opts(3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, "name",
{kCkptVarType}, // accout_type_regexes
{".*"}, {""}, {".*"}, {""}, false,
{"params", "bytes", "micros", "float_ops"}, "", {});
@ -144,169 +109,235 @@ TEST_F(TFProfStatsTest, CheckPointOpType) {
GraphNodeProto expected;
CHECK(protobuf::TextFormat::ParseFromString(
"name: \"_TFProfRoot\"\nexec_micros: 0\nrequested_bytes: "
"0\ntotal_exec_micros: 5\ntotal_requested_bytes: 1480\ntotal_parameters: "
"370\nchildren {\n name: \"conv2d\"\n exec_micros: 0\n "
"requested_bytes: 0\n total_exec_micros: 2\n total_requested_bytes: "
"560\n total_parameters: 140\n children {\n name: \"conv2d/bias\"\n "
" exec_micros: 1\n requested_bytes: 20\n parameters: 5\n "
"total_exec_micros: 1\n total_requested_bytes: 20\n "
"total_parameters: 5\n devices: "
"\"/job:localhost/replica:0/task:0/cpu:0\"\n float_ops: 0\n "
"total_float_ops: 0\n accelerator_exec_micros: 0\n "
"cpu_exec_micros: 1\n total_accelerator_exec_micros: 0\n "
"total_cpu_exec_micros: 1\n run_count: 1\n total_run_count: 1\n "
"total_definition_count: 1\n }\n children {\n name: "
"\"conv2d/kernel\"\n exec_micros: 1\n requested_bytes: 540\n "
"parameters: 135\n total_exec_micros: 1\n total_requested_bytes: "
"540\n total_parameters: 135\n devices: "
"\"/job:localhost/replica:0/task:0/cpu:0\"\n float_ops: 0\n "
"total_float_ops: 0\n accelerator_exec_micros: 0\n "
"cpu_exec_micros: 1\n total_accelerator_exec_micros: 0\n "
"total_cpu_exec_micros: 1\n run_count: 1\n total_run_count: 1\n "
"total_definition_count: 1\n }\n float_ops: 0\n total_float_ops: 0\n "
"accelerator_exec_micros: 0\n cpu_exec_micros: 0\n "
"total_accelerator_exec_micros: 0\n total_cpu_exec_micros: 2\n "
"run_count: 0\n total_run_count: 2\n total_definition_count: "
"3\n}\nchildren {\n name: \"conv2d_1\"\n exec_micros: 0\n "
"requested_bytes: 0\n total_exec_micros: 3\n total_requested_bytes: "
"920\n total_parameters: 230\n children {\n name: "
"\"conv2d_1/bias\"\n exec_micros: 1\n requested_bytes: 20\n "
"parameters: 5\n total_exec_micros: 1\n total_requested_bytes: "
"20\n total_parameters: 5\n devices: "
"\"/job:localhost/replica:0/task:0/cpu:0\"\n float_ops: 0\n "
"total_float_ops: 0\n accelerator_exec_micros: 0\n "
"cpu_exec_micros: 1\n total_accelerator_exec_micros: 0\n "
"total_cpu_exec_micros: 1\n run_count: 1\n total_run_count: 1\n "
"total_definition_count: 1\n }\n children {\n name: "
"\"conv2d_1/kernel\"\n exec_micros: 2\n requested_bytes: 900\n "
"parameters: 225\n total_exec_micros: 2\n total_requested_bytes: "
"900\n total_parameters: 225\n devices: "
"\"/job:localhost/replica:0/task:0/cpu:0\"\n float_ops: 0\n "
"total_float_ops: 0\n accelerator_exec_micros: 0\n "
"cpu_exec_micros: 2\n total_accelerator_exec_micros: 0\n "
"total_cpu_exec_micros: 2\n run_count: 1\n total_run_count: 1\n "
"total_definition_count: 1\n }\n float_ops: 0\n total_float_ops: 0\n "
"accelerator_exec_micros: 0\n cpu_exec_micros: 0\n "
"total_accelerator_exec_micros: 0\n total_cpu_exec_micros: 3\n "
"run_count: 0\n total_run_count: 2\n total_definition_count: "
"3\n}\nfloat_ops: 0\ntotal_float_ops: 0\naccelerator_exec_micros: "
"0\ncpu_exec_micros: 0\ntotal_accelerator_exec_micros: "
"0\ntotal_cpu_exec_micros: 5\nrun_count: 0\ntotal_run_count: "
"4\ntotal_definition_count: 6\n",
"name: \"_TFProfRoot\"\ntotal_exec_micros: 13\ntotal_parameters: "
"451\nchildren {\n name: \"DW\"\n exec_micros: 2\n parameters: 162\n "
"total_exec_micros: 2\n total_parameters: 162\n devices: "
"\"/job:localhost/replica:0/task:0/gpu:0\"\n cpu_exec_micros: 2\n "
"total_cpu_exec_micros: 2\n run_count: 1\n total_run_count: 1\n "
"total_definition_count: 1\n output_bytes: 1280\n total_output_bytes: "
"1280\n}\nchildren {\n name: \"DW2\"\n exec_micros: 11\n parameters: "
"288\n total_exec_micros: 11\n total_parameters: 288\n devices: "
"\"/job:localhost/replica:0/task:0/gpu:0\"\n cpu_exec_micros: 11\n "
"total_cpu_exec_micros: 11\n run_count: 1\n total_run_count: 1\n "
"total_definition_count: 1\n output_bytes: 1280\n total_output_bytes: "
"1280\n}\nchildren {\n name: \"ScalarW\"\n parameters: 1\n "
"total_parameters: 1\n total_definition_count: "
"1\n}\ntotal_cpu_exec_micros: 13\ntotal_run_count: "
"2\ntotal_definition_count: 3\ntotal_output_bytes: 2560\n",
&expected));
EXPECT_EQ(expected.DebugString(), root.DebugString());
}
TEST_F(TFProfStatsTest, TestGraph) {
Options opts(100, 0, 10000, 0, 0, 0, -1, "name", {".*"},
{"cost.*"}, // start_name_regexes
Options opts(100, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, "name", {".*"},
{"DW/Initializer/random_normal/mul"}, // start_name_regexes
{""}, {".*"}, {""}, false,
{"params", "bytes", "micros", "float_ops"}, "", {});
const GraphNodeProto& root = tf_stats_->ShowGraphNode("graph", opts);
GraphNodeProto expected;
CHECK(protobuf::TextFormat::ParseFromString(
"name: \"_TFProfRoot\"\nexec_micros: 0\nrequested_bytes: "
"0\ntotal_exec_micros: 97\ntotal_requested_bytes: "
"8656\ntotal_parameters: 370\nfloat_ops: 0\ntotal_float_ops: "
"34360\naccelerator_exec_micros: 0\ncpu_exec_micros: "
"0\ntotal_accelerator_exec_micros: 0\ntotal_cpu_exec_micros: "
"97\nrun_count: 0\ntotal_run_count: 13\ntotal_definition_count: 60\n",
"name: \"_TFProfRoot\"\ntotal_exec_micros: 4904\ntotal_requested_bytes: "
"14592\ntotal_parameters: 451\nchildren {\n name: "
"\"DW/Initializer/random_normal/mul\"\n children {\n name: "
"\"DW/Initializer/random_normal/RandomStandardNormal\"\n children {\n "
" name: \"DW/Initializer/random_normal/shape\"\n "
"total_definition_count: 1\n }\n input_shapes {\n key: 0\n "
" value {\n dim {\n size: 4\n }\n }\n "
"}\n total_definition_count: 2\n }\n children {\n name: "
"\"DW/Initializer/random_normal/stddev\"\n total_definition_count: "
"1\n }\n input_shapes {\n key: 0\n value {\n dim {\n "
"size: 3\n }\n dim {\n size: 3\n }\n dim {\n "
" size: 3\n }\n dim {\n size: 6\n }\n }\n "
"}\n input_shapes {\n key: 1\n value {\n dim {\n "
"size: 1\n }\n }\n }\n total_definition_count: "
"4\n}\ntotal_float_ops: 10440\ntotal_accelerator_exec_micros: "
"404\ntotal_cpu_exec_micros: 4500\ntotal_run_count: "
"5\ntotal_definition_count: 31\ntotal_peak_bytes: "
"9984\ntotal_residual_bytes: 1280\ntotal_output_bytes: 4864\n",
&expected));
EXPECT_EQ(expected.DebugString(), root.DebugString());
}
TEST_F(TFProfStatsTest, TestFloatOps) {
Options opts(10, 0, 0, 0, 1, 0, -1, "name", {".*"}, {".*"}, {""}, {".*"},
{""}, false, {"float_ops"}, "", {});
Options opts(10, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, -1, "name", {".*"}, {".*"},
{""}, {".*"}, {""}, false, {"float_ops"}, "", {});
const GraphNodeProto& root = tf_stats_->ShowGraphNode("scope", opts);
GraphNodeProto expected;
CHECK(protobuf::TextFormat::ParseFromString(
"name: \"_TFProfRoot\"\nexec_micros: 0\nrequested_bytes: "
"0\ntotal_exec_micros: 97\ntotal_requested_bytes: "
"8656\ntotal_parameters: 370\nchildren {\n name: \"conv2d/BiasAdd\"\n "
"exec_micros: 12\n requested_bytes: 1440\n total_exec_micros: 12\n "
"total_requested_bytes: 1440\n total_parameters: 0\n devices: "
"\"/job:localhost/replica:0/task:0/cpu:0\"\n float_ops: 360\n "
"total_float_ops: 360\n input_shapes {\n key: 0\n value {\n "
"unknown_rank: true\n }\n }\n input_shapes {\n key: 1\n value "
"{\n unknown_rank: true\n }\n }\n accelerator_exec_micros: 0\n "
" cpu_exec_micros: 12\n total_accelerator_exec_micros: 0\n "
"total_cpu_exec_micros: 12\n run_count: 1\n total_run_count: 1\n "
"total_definition_count: 1\n}\nchildren {\n name: "
"\"conv2d/convolution\"\n exec_micros: 60\n requested_bytes: 1440\n "
"total_exec_micros: 60\n total_requested_bytes: 1440\n "
"total_parameters: 0\n devices: "
"\"/job:localhost/replica:0/task:0/cpu:0\"\n float_ops: 19440\n "
"total_float_ops: 19440\n input_shapes {\n key: 0\n value {\n "
" unknown_rank: true\n }\n }\n input_shapes {\n key: 1\n "
"value {\n unknown_rank: true\n }\n }\n "
"accelerator_exec_micros: 0\n cpu_exec_micros: 60\n "
"total_accelerator_exec_micros: 0\n total_cpu_exec_micros: 60\n "
"run_count: 1\n total_run_count: 1\n total_definition_count: "
"3\n}\nchildren {\n name: \"conv2d_2/BiasAdd\"\n exec_micros: 2\n "
"requested_bytes: 640\n total_exec_micros: 2\n total_requested_bytes: "
"640\n total_parameters: 0\n devices: "
"\"/job:localhost/replica:0/task:0/cpu:0\"\n float_ops: 160\n "
"total_float_ops: 160\n input_shapes {\n key: 0\n value {\n "
"unknown_rank: true\n }\n }\n input_shapes {\n key: 1\n value "
"{\n unknown_rank: true\n }\n }\n accelerator_exec_micros: 0\n "
" cpu_exec_micros: 2\n total_accelerator_exec_micros: 0\n "
"total_cpu_exec_micros: 2\n run_count: 1\n total_run_count: 1\n "
"total_definition_count: 1\n}\nchildren {\n name: "
"\"conv2d_2/convolution\"\n exec_micros: 13\n requested_bytes: 640\n "
"total_exec_micros: 13\n total_requested_bytes: 640\n "
"total_parameters: 0\n devices: "
"\"/job:localhost/replica:0/task:0/cpu:0\"\n float_ops: 14400\n "
"total_float_ops: 14400\n input_shapes {\n key: 0\n value {\n "
" unknown_rank: true\n }\n }\n input_shapes {\n key: 1\n "
"value {\n unknown_rank: true\n }\n }\n "
"accelerator_exec_micros: 0\n cpu_exec_micros: 13\n "
"total_accelerator_exec_micros: 0\n total_cpu_exec_micros: 13\n "
"run_count: 1\n total_run_count: 1\n total_definition_count: "
"3\n}\nfloat_ops: 0\ntotal_float_ops: 34360\naccelerator_exec_micros: "
"0\ncpu_exec_micros: 0\ntotal_accelerator_exec_micros: "
"0\ntotal_cpu_exec_micros: 97\nrun_count: 0\ntotal_run_count: "
"13\ntotal_definition_count: 68\n",
"name: \"_TFProfRoot\"\ntotal_exec_micros: 4904\ntotal_requested_bytes: "
"14592\ntotal_parameters: 451\nchildren {\n name: \"Conv2D\"\n "
"exec_micros: 4292\n requested_bytes: 9472\n total_exec_micros: 4292\n "
" total_requested_bytes: 9472\n devices: "
"\"/job:localhost/replica:0/task:0/gpu:0\"\n float_ops: 5832\n "
"total_float_ops: 5832\n input_shapes {\n key: 0\n value {\n "
"dim {\n size: 2\n }\n dim {\n size: 6\n "
"}\n dim {\n size: 6\n }\n dim {\n size: "
"3\n }\n }\n }\n input_shapes {\n key: 1\n value {\n "
" dim {\n size: 3\n }\n dim {\n size: 3\n "
"}\n dim {\n size: 3\n }\n dim {\n size: "
"6\n }\n }\n }\n accelerator_exec_micros: 226\n "
"cpu_exec_micros: 4066\n total_accelerator_exec_micros: 226\n "
"total_cpu_exec_micros: 4066\n run_count: 1\n total_run_count: 1\n "
"total_definition_count: 1\n peak_bytes: 5888\n residual_bytes: 768\n "
"output_bytes: 768\n total_peak_bytes: 5888\n total_residual_bytes: "
"768\n total_output_bytes: 768\n}\nchildren {\n name: \"Conv2D_1\"\n "
"exec_micros: 597\n requested_bytes: 5120\n total_exec_micros: 597\n "
"total_requested_bytes: 5120\n devices: "
"\"/job:localhost/replica:0/task:0/gpu:0\"\n float_ops: 4608\n "
"total_float_ops: 4608\n input_shapes {\n key: 0\n value {\n "
"dim {\n size: 2\n }\n dim {\n size: 3\n "
"}\n dim {\n size: 3\n }\n dim {\n size: "
"6\n }\n }\n }\n input_shapes {\n key: 1\n value {\n "
" dim {\n size: 2\n }\n dim {\n size: 2\n "
"}\n dim {\n size: 6\n }\n dim {\n size: "
"12\n }\n }\n }\n accelerator_exec_micros: 178\n "
"cpu_exec_micros: 419\n total_accelerator_exec_micros: 178\n "
"total_cpu_exec_micros: 419\n run_count: 1\n total_run_count: 1\n "
"total_definition_count: 1\n peak_bytes: 4096\n residual_bytes: 512\n "
"output_bytes: 512\n total_peak_bytes: 4096\n total_residual_bytes: "
"512\n total_output_bytes: 512\n}\ntotal_float_ops: "
"10440\ntotal_accelerator_exec_micros: 404\ntotal_cpu_exec_micros: "
"4500\ntotal_run_count: 5\ntotal_definition_count: 34\ntotal_peak_bytes: "
"9984\ntotal_residual_bytes: 1280\ntotal_output_bytes: 4864\n",
&expected));
EXPECT_EQ(expected.DebugString(), root.DebugString());
}
TEST_F(TFProfStatsTest, TestAccountShownNameOnly) {
Options opts(100, 0, 0, 0, 0, 0, -1, "name", {".*"}, {".*"}, {""},
{"unit_2_1.*DW"}, // show_name_regexes.
{""}, true, // account_displayed_op_only.
Options opts(100, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, "name", {".*"}, {".*"},
{""}, {"Conv2D_1"}, // show_name_regexes.
{""}, true, // account_displayed_op_only.
{"params"}, "", {});
const GraphNodeProto& root = tf_stats_->ShowGraphNode("scope", opts);
GraphNodeProto expected;
CHECK(protobuf::TextFormat::ParseFromString(
"name: \"_TFProfRoot\"\nexec_micros: 0\nrequested_bytes: "
"0\ntotal_exec_micros: 0\ntotal_requested_bytes: 0\ntotal_parameters: "
"0\nfloat_ops: 0\ntotal_float_ops: 0\naccelerator_exec_micros: "
"0\ncpu_exec_micros: 0\ntotal_accelerator_exec_micros: "
"0\ntotal_cpu_exec_micros: 0\nrun_count: 0\ntotal_run_count: "
"0\ntotal_definition_count: 1\n",
"name: \"_TFProfRoot\"\ntotal_exec_micros: 597\ntotal_requested_bytes: "
"5120\nchildren {\n name: \"Conv2D_1\"\n exec_micros: 597\n "
"requested_bytes: 5120\n total_exec_micros: 597\n "
"total_requested_bytes: 5120\n devices: "
"\"/job:localhost/replica:0/task:0/gpu:0\"\n float_ops: 4608\n "
"total_float_ops: 4608\n input_shapes {\n key: 0\n value {\n "
"dim {\n size: 2\n }\n dim {\n size: 3\n "
"}\n dim {\n size: 3\n }\n dim {\n size: "
"6\n }\n }\n }\n input_shapes {\n key: 1\n value {\n "
" dim {\n size: 2\n }\n dim {\n size: 2\n "
"}\n dim {\n size: 6\n }\n dim {\n size: "
"12\n }\n }\n }\n accelerator_exec_micros: 178\n "
"cpu_exec_micros: 419\n total_accelerator_exec_micros: 178\n "
"total_cpu_exec_micros: 419\n run_count: 1\n total_run_count: 1\n "
"total_definition_count: 1\n peak_bytes: 4096\n residual_bytes: 512\n "
"output_bytes: 512\n total_peak_bytes: 4096\n total_residual_bytes: "
"512\n total_output_bytes: 512\n}\ntotal_float_ops: "
"4608\ntotal_accelerator_exec_micros: 178\ntotal_cpu_exec_micros: "
"419\ntotal_run_count: 1\ntotal_definition_count: 2\ntotal_peak_bytes: "
"4096\ntotal_residual_bytes: 512\ntotal_output_bytes: 512\n",
&expected));
EXPECT_EQ(expected.DebugString(), root.DebugString());
}
TEST_F(TFProfStatsTest, TestShowTensorValue) {
Options opts(10, 0, 0, 0, 0, 0, -1, "name", {".*"}, {".*"}, {""},
{"unit_1_0.*gamma"}, {""}, false,
Options opts(10, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, "name", {".*"}, {".*"},
{""}, {"DW"}, {""}, false,
{"tensor_value"}, // Show tensor value from checkpoint.
"", {});
const GraphNodeProto& root = tf_stats_->ShowGraphNode("scope", opts);
GraphNodeProto expected;
CHECK(protobuf::TextFormat::ParseFromString(
"name: \"_TFProfRoot\"\nexec_micros: 0\nrequested_bytes: "
"0\ntotal_exec_micros: 97\ntotal_requested_bytes: "
"8656\ntotal_parameters: 370\nfloat_ops: 0\ntotal_float_ops: "
"34360\naccelerator_exec_micros: 0\ncpu_exec_micros: "
"0\ntotal_accelerator_exec_micros: 0\ntotal_cpu_exec_micros: "
"97\nrun_count: 0\ntotal_run_count: 13\ntotal_definition_count: 68\n",
"name: \"_TFProfRoot\"\ntotal_exec_micros: 4904\ntotal_requested_bytes: "
"14592\ntotal_parameters: 451\nchildren {\n name: \"DW\"\n "
"exec_micros: 2\n parameters: 162\n total_exec_micros: 2\n "
"total_parameters: 162\n devices: "
"\"/job:localhost/replica:0/task:0/gpu:0\"\n tensor_value {\n dtype: "
"DT_FLOAT\n value_double: -0.000534315\n value_double: "
"-0.00089602\n value_double: -0.000417239\n value_double: "
"0.00041444\n value_double: 0.000780691\n value_double: "
"-0.000559057\n value_double: -0.000234623\n value_double: "
"0.00013393\n value_double: -0.00187574\n value_double: "
"0.000785666\n value_double: 0.000673294\n value_double: "
"0.000653368\n value_double: 0.000924489\n value_double: "
"-0.000318373\n value_double: -0.000385202\n value_double: "
"-7.92661e-05\n value_double: 2.70287e-05\n value_double: "
"0.00152302\n value_double: 8.04435e-05\n value_double: "
"-0.00058102\n value_double: 0.000244291\n value_double: "
"-0.000438045\n value_double: -0.000110199\n value_double: "
"0.000731663\n value_double: -0.0012326\n value_double: "
"0.00064065\n value_double: -0.00135203\n value_double: "
"-6.42784e-05\n value_double: -0.0011857\n value_double: "
"-0.000487383\n value_double: 3.41493e-05\n value_double: "
"-0.00158447\n value_double: 0.00168448\n value_double: "
"0.00160946\n value_double: -0.000600483\n value_double: "
"0.000650259\n value_double: -0.00109938\n value_double: "
"-0.000842166\n value_double: -0.0022673\n value_double: "
"-0.00101941\n value_double: -0.0011169\n value_double: "
"-0.0013557\n value_double: -1.46354e-05\n value_double: "
"-1.05487e-05\n value_double: -0.00092014\n value_double: "
"0.00272874\n value_double: 5.13942e-05\n value_double: "
"-0.00223472\n value_double: -0.000250875\n value_double: "
"-0.00180747\n value_double: -0.00234714\n value_double: "
"-0.00113523\n value_double: -0.00112635\n value_double: "
"-0.000843118\n value_double: -6.84256e-05\n value_double: "
"0.000243336\n value_double: 0.00119151\n value_double: "
"0.00131022\n value_double: 0.000768038\n value_double: "
"-8.90095e-05\n value_double: -0.000626427\n value_double: "
"-7.0617e-05\n value_double: -0.0021988\n value_double: "
"-0.00221544\n value_double: -0.000393118\n value_double: "
"0.000159464\n value_double: -0.000874746\n value_double: "
"-0.00131239\n value_double: -0.00135747\n value_double: "
"-0.00179753\n value_double: -0.00101005\n value_double: "
"-0.000107518\n value_double: -0.000616882\n value_double: "
"-0.000360923\n value_double: -0.00026896\n value_double: "
"-0.000142548\n value_double: 0.000577227\n value_double: "
"0.000536027\n value_double: 0.00126907\n value_double: "
"-0.00122712\n value_double: -3.60499e-05\n value_double: "
"0.000151026\n value_double: 0.00107658\n value_double: "
"0.00116475\n value_double: -0.00145312\n value_double: "
"0.000233326\n value_double: -0.00020198\n value_double: "
"0.00179029\n value_double: 0.00150048\n value_double: "
"-0.000884775\n value_double: 0.000409188\n value_double: "
"2.97176e-05\n value_double: -0.000506118\n value_double: "
"-2.33992e-05\n value_double: -0.00037212\n value_double: "
"0.000862773\n value_double: 0.00174046\n value_double: "
"-0.000240207\n value_double: 0.000663976\n value_double: "
"-0.00134747\n value_double: 0.00115585\n value_double: "
"0.000555869\n value_double: 0.00176722\n value_double: "
"-0.000518409\n value_double: 0.00101051\n value_double: "
"0.000129399\n value_double: -0.000916389\n value_double: "
"-0.00137693\n value_double: -0.00152412\n value_double: "
"7.32515e-05\n value_double: -0.000190811\n value_double: "
"-0.000158692\n value_double: -5.7791e-05\n value_double: "
"0.000671785\n value_double: -0.00152924\n value_double: "
"0.00117314\n value_double: -0.000384202\n value_double: "
"0.00176709\n value_double: -0.000181703\n value_double: "
"-0.000460994\n value_double: 0.000643716\n value_double: "
"4.76719e-05\n value_double: -0.00101037\n value_double: "
"0.00159621\n value_double: 0.00186758\n value_double: "
"0.00100001\n value_double: -0.00121831\n value_double: "
"0.00132231\n value_double: 0.0013511\n value_double: 0.00106659\n "
" value_double: 0.00018091\n value_double: 0.00155925\n "
"value_double: 4.26087e-05\n value_double: 0.000243264\n "
"value_double: -0.0017202\n value_double: -0.000218897\n "
"value_double: 0.00118693\n value_double: 0.00258909\n "
"value_double: 0.000641913\n value_double: -0.0013211\n "
"value_double: -0.00171943\n value_double: 0.00089151\n "
"value_double: -0.00114969\n value_double: -0.000196331\n "
"value_double: 0.00109994\n value_double: 0.000302616\n "
"value_double: 0.000675812\n value_double: 0.00112222\n "
"value_double: 0.000516456\n value_double: 0.00133357\n "
"value_double: 0.000298491\n value_double: 0.00145934\n "
"value_double: -0.00159102\n value_double: -0.000819061\n "
"value_double: 0.000120583\n value_double: 0.0006108\n "
"value_double: 0.00124132\n value_double: 0.000764859\n "
"value_double: 0.000374641\n value_double: -0.00149603\n "
"value_double: -0.000317367\n value_double: -0.000417829\n }\n "
"cpu_exec_micros: 2\n total_cpu_exec_micros: 2\n run_count: 1\n "
"total_run_count: 1\n total_definition_count: 10\n output_bytes: "
"1280\n total_output_bytes: 1280\n}\ntotal_float_ops: "
"10440\ntotal_accelerator_exec_micros: 404\ntotal_cpu_exec_micros: "
"4500\ntotal_run_count: 5\ntotal_definition_count: 34\ntotal_peak_bytes: "
"9984\ntotal_residual_bytes: 1280\ntotal_output_bytes: 4864\n",
&expected));
EXPECT_EQ(expected.DebugString(), root.DebugString());
}

113
tensorflow/core/profiler/internal/tfprof_tensor.h
View File

@ -51,6 +51,33 @@ class TFProfTensor {
void Build();
template <typename T>
bool AddValue(const T& value, TFProfTensorProto* dim) {
std::ostringstream sstream;
sstream << value;
if (typeid(value) == typeid(double)) {
double double_val;
CHECK(strings::safe_strtod(sstream.str().c_str(), &double_val));
dim->add_value_double(double_val);
formatted_str_ += strings::Printf(
"%.2f ", dim->value_double(dim->value_double_size() - 1));
} else if (typeid(value) == typeid(int64)) {
int64 int64_val;
CHECK(strings::safe_strto64(sstream.str().c_str(), &int64_val));
dim->add_value_int64(int64_val);
formatted_str_ += strings::Printf(
"%lld ",
static_cast<int64>(dim->value_int64(dim->value_int64_size() - 1)));
} else if (typeid(value) == typeid(string)) {
dim->add_value_str(sstream.str());
formatted_str_ =
strings::StrCat(formatted_str_, "'",
dim->value_str(dim->value_str_size() - 1) + "' ");
} else {
CHECK(false) << "Unsupported type: " << typeid(value).name();
}
}
// It assumes the flatten values are stored in row-major, which is mentioned
// indirectly at various places:
// TODO(xpan): Further verifying it.
@ -59,37 +86,65 @@ class TFProfTensor {
TFProfTensorProto* dim) {
formatted_str_ += "[";
int64 nstart = start;
for (int i = 0; i < tensor_->dim_size(depth); i++) {
// Last dimension, pull the values.
if (depth == tensor_->dims() - 1) {
std::ostringstream sstream;
sstream << values[nstart];
if (tensor_->dims() == 0 && values.size() == 1) {
std::ostringstream sstream;
sstream << values[nstart];
if (typeid(values[nstart]) == typeid(double)) {
double double_val;
CHECK(strings::safe_strtod(sstream.str().c_str(), &double_val));
dim->add_value_double(double_val);
formatted_str_ += strings::Printf(
"%.2f ", dim->value_double(dim->value_double_size() - 1));
} else if (typeid(values[nstart]) == typeid(int64)) {
int64 int64_val;
CHECK(strings::safe_strto64(sstream.str().c_str(), &int64_val));
dim->add_value_int64(int64_val);
formatted_str_ += strings::Printf(
"%lld ", static_cast<int64>(
dim->value_int64(dim->value_int64_size() - 1)));
} else if (typeid(values[nstart]) == typeid(string)) {
dim->add_value_str(sstream.str());
formatted_str_ =
strings::StrCat(formatted_str_, "'",
dim->value_str(dim->value_str_size() - 1) + "' ");
} else {
CHECK(false) << "Unsupported type: " << typeid(values[nstart]).name();
}
++nstart;
if (typeid(values[nstart]) == typeid(double)) {
double double_val;
CHECK(strings::safe_strtod(sstream.str().c_str(), &double_val));
dim->add_value_double(double_val);
formatted_str_ += strings::Printf(
"%.2f ", dim->value_double(dim->value_double_size() - 1));
} else if (typeid(values[nstart]) == typeid(int64)) {
int64 int64_val;
CHECK(strings::safe_strto64(sstream.str().c_str(), &int64_val));
dim->add_value_int64(int64_val);
formatted_str_ += strings::Printf(
"%lld ",
static_cast<int64>(dim->value_int64(dim->value_int64_size() - 1)));
} else if (typeid(values[nstart]) == typeid(string)) {
dim->add_value_str(sstream.str());
formatted_str_ =
strings::StrCat(formatted_str_, "'",
dim->value_str(dim->value_str_size() - 1) + "' ");
} else {
// Not-last dimension. Drill deeper.
nstart = BuildOutput<T>(nstart, depth + 1, values, dim);
CHECK(false) << "Unsupported type: " << typeid(values[nstart]).name();
}
} else {
for (int i = 0; i < tensor_->dim_size(depth); i++) {
// Last dimension, pull the values.
if (depth == tensor_->dims() - 1) {
std::ostringstream sstream;
sstream << values[nstart];
if (typeid(values[nstart]) == typeid(double)) {
double double_val;
CHECK(strings::safe_strtod(sstream.str().c_str(), &double_val));
dim->add_value_double(double_val);
formatted_str_ += strings::Printf(
"%.2f ", dim->value_double(dim->value_double_size() - 1));
} else if (typeid(values[nstart]) == typeid(int64)) {
int64 int64_val;
CHECK(strings::safe_strto64(sstream.str().c_str(), &int64_val));
dim->add_value_int64(int64_val);
formatted_str_ += strings::Printf(
"%lld ", static_cast<int64>(
dim->value_int64(dim->value_int64_size() - 1)));
} else if (typeid(values[nstart]) == typeid(string)) {
dim->add_value_str(sstream.str());
formatted_str_ = strings::StrCat(
formatted_str_, "'",
dim->value_str(dim->value_str_size() - 1) + "' ");
} else {
CHECK(false) << "Unsupported type: "
<< typeid(values[nstart]).name();
}
++nstart;
} else {
// Not-last dimension. Drill deeper.
nstart = BuildOutput<T>(nstart, depth + 1, values, dim);
}
}
}
if (formatted_str_.length() > kTFProfTenosrMaxDisplayLen) {

245
tensorflow/core/profiler/internal/tfprof_tensor_test.cc
View File

@ -18,12 +18,12 @@ limitations under the License.
#include "tensorflow/core/lib/io/path.h"
#include "tensorflow/core/platform/protobuf.h"
#include "tensorflow/core/platform/test.h"
#include "tensorflow/core/protobuf/config.pb.h"
#include "tensorflow/core/profiler/internal/tfprof_options.h"
#include "tensorflow/core/profiler/internal/tfprof_stats.h"
#include "tensorflow/core/profiler/internal/tfprof_utils.h"
#include "tensorflow/core/profiler/tfprof_log.pb.h"
#include "tensorflow/core/profiler/tfprof_output.pb.h"
#include "tensorflow/core/protobuf/config.pb.h"
namespace tensorflow {
namespace tfprof {
@ -57,244 +57,19 @@ class TFProfTensorTest : public ::testing::Test {
};
TEST_F(TFProfTensorTest, Basics) {
Options opts(3, 0, 0, 0, 0, 0, -1, "name", {"VariableV2"}, {".*"}, {""},
{".*"}, {""}, false, {"tensor_value"}, // show the tensor value.
Options opts(3, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, -1, "name", {"VariableV2"},
{".*"}, {""}, {".*"}, {""}, false,
{"tensor_value"}, // show the tensor value.
"", {});
const GraphNodeProto& root = tf_stats_->ShowGraphNode("scope", opts);
GraphNodeProto expected;
CHECK(protobuf::TextFormat::ParseFromString(
"name: \"_TFProfRoot\"\nexec_micros: 0\nrequested_bytes: "
"0\ntotal_exec_micros: 0\ntotal_requested_bytes: 0\ntotal_parameters: "
"370\nchildren {\n name: \"conv2d\"\n exec_micros: 0\n "
"requested_bytes: 0\n total_exec_micros: 0\n total_requested_bytes: "
"0\n total_parameters: 140\n children {\n name: \"conv2d/bias\"\n "
" exec_micros: 0\n requested_bytes: 0\n parameters: 5\n "
"total_exec_micros: 0\n total_requested_bytes: 0\n "
"total_parameters: 5\n float_ops: 0\n total_float_ops: 0\n "
"tensor_value {\n dtype: DT_FLOAT\n value_double: 0\n "
"value_double: 0\n value_double: 0\n value_double: 0\n "
"value_double: 0\n }\n accelerator_exec_micros: 0\n "
"cpu_exec_micros: 0\n total_accelerator_exec_micros: 0\n "
"total_cpu_exec_micros: 0\n run_count: 0\n total_run_count: 0\n "
"total_definition_count: 1\n }\n children {\n name: "
"\"conv2d/kernel\"\n exec_micros: 0\n requested_bytes: 0\n "
"parameters: 135\n total_exec_micros: 0\n total_requested_bytes: "
"0\n total_parameters: 135\n float_ops: 0\n total_float_ops: "
"0\n tensor_value {\n dtype: DT_FLOAT\n value_double: "
"-0.113138\n value_double: 0.261431\n value_double: 0.215777\n "
" value_double: 0.24135\n value_double: -0.113195\n "
"value_double: -0.212639\n value_double: -0.0907301\n "
"value_double: 0.0221634\n value_double: 0.21821\n "
"value_double: 0.22715\n value_double: -0.108698\n "
"value_double: 0.240911\n value_double: -0.138626\n "
"value_double: -0.144752\n value_double: -0.00962037\n "
"value_double: 0.0971008\n value_double: 0.00264764\n "
"value_double: -0.272929\n value_double: 0.0129845\n "
"value_double: 0.0466554\n value_double: -0.229184\n "
"value_double: 0.153576\n value_double: -0.169218\n "
"value_double: -0.112991\n value_double: 0.205739\n "
"value_double: 0.257844\n value_double: 0.107455\n "
"value_double: -0.207914\n value_double: 0.15211\n "
"value_double: 0.277932\n value_double: 0.145986\n "
"value_double: -0.0883989\n value_double: 0.167506\n "
"value_double: 0.10237\n value_double: 0.0542143\n "
"value_double: 0.0334378\n value_double: 0.159489\n "
"value_double: 0.246583\n value_double: 0.0154283\n "
"value_double: 0.0872411\n value_double: -0.25732\n "
"value_double: 0.0499355\n value_double: 0.0266221\n "
"value_double: 0.088801\n value_double: -0.0794552\n "
"value_double: -0.00383255\n value_double: -0.165267\n "
"value_double: 0.0271328\n value_double: 0.0729822\n "
"value_double: 0.200795\n value_double: 0.100276\n "
"value_double: 0.285254\n value_double: -0.171945\n "
"value_double: -0.0187411\n value_double: -0.218729\n "
"value_double: 0.233753\n value_double: 0.109184\n "
"value_double: 0.247875\n value_double: -0.224632\n "
"value_double: 0.0940739\n value_double: 0.00663087\n "
"value_double: -0.075786\n value_double: -0.179992\n "
"value_double: -0.276016\n value_double: 0.261207\n "
"value_double: -0.0658191\n value_double: -0.0747132\n "
"value_double: -0.0839638\n value_double: -0.0825393\n "
"value_double: 0.0915958\n value_double: -0.195425\n "
"value_double: -0.255836\n value_double: -0.08745\n "
"value_double: -0.181623\n value_double: -0.235936\n "
"value_double: 0.0205423\n value_double: 0.185447\n "
"value_double: -0.0691599\n value_double: -0.0451089\n "
"value_double: -0.153922\n value_double: -0.0279411\n "
"value_double: 0.148915\n value_double: -0.018026\n "
"value_double: -0.144903\n value_double: 0.0370046\n "
"value_double: 0.0764987\n value_double: 0.0586488\n "
"value_double: -0.222919\n value_double: 0.0238447\n "
"value_double: -0.106012\n value_double: -0.102202\n "
"value_double: -0.159347\n value_double: -0.0232876\n "
"value_double: 0.109855\n value_double: -0.141833\n "
"value_double: 0.1376\n value_double: -0.12413\n value_double: "
"-0.208968\n value_double: 0.0758635\n value_double: "
"-0.217672\n value_double: -0.20153\n value_double: "
"-0.195414\n value_double: -0.18549\n value_double: "
"0.00298014\n value_double: -0.279283\n value_double: "
"0.200084\n value_double: -0.0968328\n value_double: -0.243\n "
" value_double: 0.239319\n value_double: -0.236288\n "
"value_double: 0.169477\n value_double: 0.126673\n "
"value_double: 0.182215\n value_double: -0.028243\n "
"value_double: 0.282762\n value_double: -0.165548\n "
"value_double: -0.0641245\n value_double: -0.186382\n "
"value_double: 0.0329038\n value_double: 0.271848\n "
"value_double: 0.084653\n value_double: -0.108163\n "
"value_double: 0.247094\n value_double: 0.192687\n "
"value_double: 0.171922\n value_double: -0.187649\n "
"value_double: 0.251253\n value_double: 0.272077\n "
"value_double: 0.19068\n value_double: 0.220352\n "
"value_double: -0.255741\n value_double: 0.110853\n "
"value_double: 0.146625\n value_double: 0.167754\n "
"value_double: 0.249554\n }\n accelerator_exec_micros: 0\n "
"cpu_exec_micros: 0\n total_accelerator_exec_micros: 0\n "
"total_cpu_exec_micros: 0\n run_count: 0\n total_run_count: 0\n "
"total_definition_count: 1\n }\n float_ops: 0\n total_float_ops: 0\n "
"accelerator_exec_micros: 0\n cpu_exec_micros: 0\n "
"total_accelerator_exec_micros: 0\n total_cpu_exec_micros: 0\n "
"run_count: 0\n total_run_count: 0\n total_definition_count: "
"3\n}\nchildren {\n name: \"conv2d_1\"\n exec_micros: 0\n "
"requested_bytes: 0\n total_exec_micros: 0\n total_requested_bytes: "
"0\n total_parameters: 230\n children {\n name: \"conv2d_1/bias\"\n "
" exec_micros: 0\n requested_bytes: 0\n parameters: 5\n "
"total_exec_micros: 0\n total_requested_bytes: 0\n "
"total_parameters: 5\n float_ops: 0\n total_float_ops: 0\n "
"tensor_value {\n dtype: DT_FLOAT\n value_double: 0\n "
"value_double: 0\n value_double: 0\n value_double: 0\n "
"value_double: 0\n }\n accelerator_exec_micros: 0\n "
"cpu_exec_micros: 0\n total_accelerator_exec_micros: 0\n "
"total_cpu_exec_micros: 0\n run_count: 0\n total_run_count: 0\n "
"total_definition_count: 1\n }\n children {\n name: "
"\"conv2d_1/kernel\"\n exec_micros: 0\n requested_bytes: 0\n "
"parameters: 225\n total_exec_micros: 0\n total_requested_bytes: "
"0\n total_parameters: 225\n float_ops: 0\n total_float_ops: "
"0\n tensor_value {\n dtype: DT_FLOAT\n value_double: "
"-0.00170514\n value_double: 0.138601\n value_double: "
"-0.224822\n value_double: -0.0848449\n value_double: "
"0.170551\n value_double: 0.147666\n value_double: "
"-0.0570606\n value_double: -0.132805\n value_double: "
"-0.172013\n value_double: 0.249707\n value_double: 0.149734\n "
" value_double: 0.0365986\n value_double: -0.0923146\n "
"value_double: -0.17745\n value_double: -0.169978\n "
"value_double: -0.173298\n value_double: -0.110407\n "
"value_double: 0.1469\n value_double: 0.0419576\n "
"value_double: 0.0391093\n value_double: -0.137381\n "
"value_double: 0.212642\n value_double: -0.067034\n "
"value_double: -0.0727709\n value_double: -0.0276531\n "
"value_double: 0.218212\n value_double: 0.0596479\n "
"value_double: -0.0468102\n value_double: -0.0250467\n "
"value_double: -0.20391\n value_double: -0.233801\n "
"value_double: 0.135615\n value_double: -0.182124\n "
"value_double: 0.254205\n value_double: 0.0819146\n "
"value_double: -0.146696\n value_double: -0.20095\n "
"value_double: -0.250555\n value_double: -0.226406\n "
"value_double: 0.0421331\n value_double: 0.0361264\n "
"value_double: -0.188558\n value_double: -0.0222711\n "
"value_double: -0.128226\n value_double: -0.148305\n "
"value_double: -0.137598\n value_double: -0.041647\n "
"value_double: -0.0574933\n value_double: 0.122506\n "
"value_double: 0.0415936\n value_double: 0.244957\n "
"value_double: 0.00372121\n value_double: -0.139939\n "
"value_double: 0.250411\n value_double: -0.23848\n "
"value_double: -0.0717569\n value_double: -0.00884159\n "
"value_double: 0.135616\n value_double: -0.0493895\n "
"value_double: 0.254308\n value_double: -0.181419\n "
"value_double: -0.114829\n value_double: -0.172638\n "
"value_double: 0.06984\n value_double: -0.086704\n "
"value_double: 0.168515\n value_double: -0.152275\n "
"value_double: -0.230775\n value_double: -0.254366\n "
"value_double: -0.115397\n value_double: 0.0418207\n "
"value_double: -0.199607\n value_double: -0.167001\n "
"value_double: -0.187238\n value_double: 0.0196097\n "
"value_double: 0.201653\n value_double: -0.143758\n "
"value_double: 0.167187\n value_double: -0.129141\n "
"value_double: 0.230154\n value_double: -0.119968\n "
"value_double: -0.121843\n value_double: -0.0118565\n "
"value_double: 0.0285747\n value_double: -0.0593699\n "
"value_double: -0.175214\n value_double: -0.211524\n "
"value_double: 0.167042\n value_double: -0.216357\n "
"value_double: -0.0218886\n value_double: -0.244211\n "
"value_double: 0.175301\n value_double: 0.0654932\n "
"value_double: -0.0419763\n value_double: -0.103275\n "
"value_double: -0.0848433\n value_double: -0.0845421\n "
"value_double: -0.00269318\n value_double: -0.145978\n "
"value_double: -0.217061\n value_double: -0.0937043\n "
"value_double: 0.235796\n value_double: -0.0893372\n "
"value_double: 0.000827968\n value_double: 0.0172743\n "
"value_double: -0.234205\n value_double: -0.0867703\n "
"value_double: 0.131704\n value_double: 0.134143\n "
"value_double: -0.162257\n value_double: -0.129706\n "
"value_double: 0.0763288\n value_double: 0.156988\n "
"value_double: 0.220033\n value_double: -0.179884\n "
"value_double: 0.066697\n value_double: 0.212322\n "
"value_double: -0.0961226\n value_double: -0.11223\n "
"value_double: 0.249944\n value_double: 0.115673\n "
"value_double: -0.100203\n value_double: 0.125645\n "
"value_double: -0.256104\n value_double: 0.0996534\n "
"value_double: 0.167306\n value_double: -0.00700775\n "
"value_double: 0.242145\n value_double: 0.088406\n "
"value_double: 0.0975334\n value_double: -0.0309525\n "
"value_double: -0.0422794\n value_double: 0.20739\n "
"value_double: 0.113992\n value_double: 0.253818\n "
"value_double: -0.0857835\n value_double: 0.223902\n "
"value_double: 0.10291\n value_double: 0.103091\n "
"value_double: -0.177502\n value_double: -0.0258242\n "
"value_double: -0.130567\n value_double: -0.15999\n "
"value_double: -0.101484\n value_double: 0.0188813\n "
"value_double: 0.160626\n value_double: 0.0467491\n "
"value_double: 0.193634\n value_double: -0.0910993\n "
"value_double: 0.0440249\n value_double: -0.255389\n "
"value_double: -0.240244\n value_double: -0.213171\n "
"value_double: 0.175978\n value_double: -0.0251202\n "
"value_double: 0.0943941\n value_double: -0.196194\n "
"value_double: 0.163395\n value_double: -0.010777\n "
"value_double: -0.0626751\n value_double: -0.246234\n "
"value_double: 0.0662063\n value_double: 0.120589\n "
"value_double: 0.237322\n value_double: 0.0849243\n "
"value_double: -0.066591\n value_double: 0.0512236\n "
"value_double: -0.144309\n value_double: -0.235415\n "
"value_double: -0.0565311\n value_double: 0.0882529\n "
"value_double: -0.215923\n value_double: -0.0873292\n "
"value_double: -0.0691103\n value_double: -0.00238678\n "
"value_double: 0.147789\n value_double: -0.124451\n "
"value_double: 0.205044\n value_double: -0.0596834\n "
"value_double: 0.0268479\n value_double: 0.0857448\n "
"value_double: -0.0923855\n value_double: -0.0960547\n "
"value_double: 0.169869\n value_double: 0.16988\n "
"value_double: -0.032271\n value_double: -0.120731\n "
"value_double: -0.199086\n value_double: 0.181199\n "
"value_double: 0.00897732\n value_double: -0.257469\n "
"value_double: -0.135556\n value_double: -0.149663\n "
"value_double: -0.00990398\n value_double: 0.221165\n "
"value_double: 0.0327134\n value_double: -0.0392821\n "
"value_double: -0.0614503\n value_double: 0.246602\n "
"value_double: -0.171692\n value_double: -0.150835\n "
"value_double: -0.13854\n value_double: -0.244668\n "
"value_double: 0.0790781\n value_double: 0.212678\n "
"value_double: 0.0782059\n value_double: -0.177888\n "
"value_double: -0.165914\n value_double: -0.164251\n "
"value_double: 0.165007\n value_double: 0.239615\n "
"value_double: -0.217642\n value_double: -0.219843\n "
"value_double: 0.0828398\n value_double: 0.00272235\n "
"value_double: -0.0323662\n value_double: -0.255953\n "
"value_double: 0.237298\n value_double: -0.0896481\n "
"value_double: -0.0605349\n value_double: 0.231679\n "
"value_double: -0.123842\n value_double: 0.0858642\n "
"value_double: 0.23111\n value_double: 0.0491742\n }\n "
"accelerator_exec_micros: 0\n cpu_exec_micros: 0\n "
"total_accelerator_exec_micros: 0\n total_cpu_exec_micros: 0\n "
"run_count: 0\n total_run_count: 0\n total_definition_count: 1\n "
"}\n float_ops: 0\n total_float_ops: 0\n accelerator_exec_micros: 0\n "
" cpu_exec_micros: 0\n total_accelerator_exec_micros: 0\n "
"total_cpu_exec_micros: 0\n run_count: 0\n total_run_count: 0\n "
"total_definition_count: 3\n}\nfloat_ops: 0\ntotal_float_ops: "
"0\naccelerator_exec_micros: 0\ncpu_exec_micros: "
"0\ntotal_accelerator_exec_micros: 0\ntotal_cpu_exec_micros: "
"0\nrun_count: 0\ntotal_run_count: 0\ntotal_definition_count: 6\n",
&expected));
EXPECT_EQ(expected.DebugString(), root.DebugString());
EXPECT_EQ(root.children(0).name(), "DW");
EXPECT_GT(root.children(0).tensor_value().value_double_size(), 10);
EXPECT_EQ(root.children(1).name(), "DW2");
EXPECT_GT(root.children(1).tensor_value().value_double_size(), 10);
EXPECT_EQ(root.children(2).name(), "ScalarW");
EXPECT_EQ(root.children(2).tensor_value().value_double_size(), 1);
}
} // namespace tfprof

4
tensorflow/core/profiler/internal/tfprof_timeline.cc
View File

@ -147,8 +147,8 @@ void MemoryTracker::TrackNodeConnection(int64 step, const GraphNode* node,
if (output_idx == node->node->src_output_idx().end()) {
return;
}
const auto& output = src->node->output_bytes(step).find(output_idx->second);
if (output == src->node->output_bytes(step).end()) {
const auto& output = src->node->output_memory(step).find(output_idx->second);
if (output == src->node->output_memory(step).end()) {
return;
}
int64 output_bytes = output->second.first;

10
tensorflow/core/profiler/internal/tfprof_timeline_test.cc
View File

@ -62,7 +62,8 @@ class TFProfTimelineTest : public ::testing::Test {
// manually check it's correct
TEST_F(TFProfTimelineTest, GraphView) {
string dump_file = io::JoinPath(testing::TmpDir(), "dump");
Options opts(10000, 0, 0, 0, 0, 0, 0, "name", {".*"}, // accout_type_regexes
Options opts(10000, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, "name",
{".*"}, // accout_type_regexes
{".*"}, {""}, {".*"}, {""}, false,
{"params", "bytes", "micros", "float_ops"}, "timeline",
{{"outfile", dump_file}});
@ -70,12 +71,13 @@ TEST_F(TFProfTimelineTest, GraphView) {
string dump_str;
TF_CHECK_OK(ReadFileToString(Env::Default(), dump_file, &dump_str));
EXPECT_EQ(5576767607271035974ull, Hash64(dump_str));
EXPECT_EQ(16947107375505024864ull, Hash64(dump_str));
}
TEST_F(TFProfTimelineTest, ScopeView) {
string dump_file = io::JoinPath(testing::TmpDir(), "dump");
Options opts(5, 0, 0, 0, 0, 0, 0, "name", {".*"}, // accout_type_regexes
Options opts(5, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, "name",
{".*"}, // accout_type_regexes
{".*"}, {""}, {".*"}, {""}, false,
{"params", "bytes", "micros", "float_ops"}, "timeline",
{{"outfile", dump_file}});
@ -83,7 +85,7 @@ TEST_F(TFProfTimelineTest, ScopeView) {
string dump_str;
TF_CHECK_OK(ReadFileToString(Env::Default(), dump_file, &dump_str));
EXPECT_EQ(10135186027625211652ull, Hash64(dump_str));
EXPECT_EQ(2710044785377031280ull, Hash64(dump_str));
}
// TODO(xpan): tfprof_log is too large to include in testdata when adding

57
tensorflow/core/profiler/internal/tfprof_utils.cc
View File

@ -140,35 +140,66 @@ tensorflow::Status ParseCmdLine(const string& line, string* cmd,
++i;
} else if (pieces[i] == tensorflow::tfprof::kOptions[2]) {
if (pieces.size() <= i + 1 ||
!strings::safe_strto64(pieces[i + 1], &opts->min_micros)) {
!strings::safe_strto64(pieces[i + 1], &opts->min_peak_bytes)) {
return ReturnError(pieces, i);
}
++i;
} else if (pieces[i] == tensorflow::tfprof::kOptions[3]) {
if (pieces.size() <= i + 1 ||
!strings::safe_strto64(pieces[i + 1], &opts->min_params)) {
!strings::safe_strto64(pieces[i + 1], &opts->min_residual_bytes)) {
return ReturnError(pieces, i);
}
++i;
} else if (pieces[i] == tensorflow::tfprof::kOptions[4]) {
if (pieces.size() <= i + 1 ||
!strings::safe_strto64(pieces[i + 1], &opts->min_float_ops)) {
!strings::safe_strto64(pieces[i + 1], &opts->min_output_bytes)) {
return ReturnError(pieces, i);
}
++i;
} else if (pieces[i] == tensorflow::tfprof::kOptions[5]) {
if (pieces.size() <= i + 1 ||
!strings::safe_strto64(pieces[i + 1], &opts->min_occurrence)) {
!strings::safe_strto64(pieces[i + 1], &opts->min_micros)) {
return ReturnError(pieces, i);
}
++i;
} else if (pieces[i] == tensorflow::tfprof::kOptions[6]) {
if (pieces.size() <= i + 1 ||
!strings::safe_strto64(pieces[i + 1], &opts->step)) {
!strings::safe_strto64(pieces[i + 1],
&opts->min_accelerator_micros)) {
return ReturnError(pieces, i);
}
++i;
} else if (pieces[i] == tensorflow::tfprof::kOptions[7]) {
if (pieces.size() <= i + 1 ||
!strings::safe_strto64(pieces[i + 1], &opts->min_cpu_micros)) {
return ReturnError(pieces, i);
}
++i;
} else if (pieces[i] == tensorflow::tfprof::kOptions[8]) {
if (pieces.size() <= i + 1 ||
!strings::safe_strto64(pieces[i + 1], &opts->min_params)) {
return ReturnError(pieces, i);
}
++i;
} else if (pieces[i] == tensorflow::tfprof::kOptions[9]) {
if (pieces.size() <= i + 1 ||
!strings::safe_strto64(pieces[i + 1], &opts->min_float_ops)) {
return ReturnError(pieces, i);
}
++i;
} else if (pieces[i] == tensorflow::tfprof::kOptions[10]) {
if (pieces.size() <= i + 1 ||
!strings::safe_strto64(pieces[i + 1], &opts->min_occurrence)) {
return ReturnError(pieces, i);
}
++i;
} else if (pieces[i] == tensorflow::tfprof::kOptions[11]) {
if (pieces.size() <= i + 1 ||
!strings::safe_strto64(pieces[i + 1], &opts->step)) {
return ReturnError(pieces, i);
}
++i;
} else if (pieces[i] == tensorflow::tfprof::kOptions[12]) {
if (pieces.size() <= i + 1) {
return ReturnError(pieces, i);
}
@ -180,42 +211,42 @@ tensorflow::Status ParseCmdLine(const string& line, string* cmd,
}
opts->order_by = *order_by;
++i;
} else if (pieces[i] == tensorflow::tfprof::kOptions[8]) {
} else if (pieces[i] == tensorflow::tfprof::kOptions[13]) {
if (pieces.size() <= i + 1) {
return ReturnError(pieces, i);
}
opts->account_type_regexes = str_util::Split(StripQuote(pieces[i + 1]),
',', str_util::SkipEmpty());
++i;
} else if (pieces[i] == tensorflow::tfprof::kOptions[9]) {
} else if (pieces[i] == tensorflow::tfprof::kOptions[14]) {
if (pieces.size() <= i + 1) {
return ReturnError(pieces, i);
}
opts->start_name_regexes = str_util::Split(StripQuote(pieces[i + 1]), ',',
str_util::SkipEmpty());
++i;
} else if (pieces[i] == tensorflow::tfprof::kOptions[10]) {
} else if (pieces[i] == tensorflow::tfprof::kOptions[15]) {
if (pieces.size() <= i + 1) {
return ReturnError(pieces, i);
}
opts->trim_name_regexes = str_util::Split(StripQuote(pieces[i + 1]), ',',
str_util::SkipEmpty());
++i;
} else if (pieces[i] == tensorflow::tfprof::kOptions[11]) {
} else if (pieces[i] == tensorflow::tfprof::kOptions[16]) {
if (pieces.size() <= i + 1) {
return ReturnError(pieces, i);
}
opts->show_name_regexes = str_util::Split(StripQuote(pieces[i + 1]), ',',
str_util::SkipEmpty());
++i;
} else if (pieces[i] == tensorflow::tfprof::kOptions[12]) {
} else if (pieces[i] == tensorflow::tfprof::kOptions[17]) {
if (pieces.size() <= i + 1) {
return ReturnError(pieces, i);
}
opts->hide_name_regexes = str_util::Split(StripQuote(pieces[i + 1]), ',',
str_util::SkipEmpty());
++i;
} else if (pieces[i] == tensorflow::tfprof::kOptions[13]) {
} else if (pieces[i] == tensorflow::tfprof::kOptions[18]) {
if ((pieces.size() > i + 1 && pieces[i + 1].find("-") == 0) ||
pieces.size() == i + 1) {
opts->account_displayed_op_only = true;
@ -225,7 +256,7 @@ tensorflow::Status ParseCmdLine(const string& line, string* cmd,
} else {
++i;
}
} else if (pieces[i] == tensorflow::tfprof::kOptions[14]) {
} else if (pieces[i] == tensorflow::tfprof::kOptions[19]) {
if (pieces.size() <= i + 1) {
return ReturnError(pieces, i);
}
@ -242,7 +273,7 @@ tensorflow::Status ParseCmdLine(const string& line, string* cmd,
}
opts->select = requested_set;
++i;
} else if (pieces[i] == tensorflow::tfprof::kOptions[15]) {
} else if (pieces[i] == tensorflow::tfprof::kOptions[20]) {
if (pieces.size() <= i + 1) {
return ReturnError(pieces, i);
}

26
tensorflow/core/profiler/profiler.cc
View File

@ -72,7 +72,12 @@ int Run(int argc, char** argv) {
string FLAGS_checkpoint_path = "";
int32 FLAGS_max_depth = 10;
int64 FLAGS_min_bytes = 0;
int64 FLAGS_min_peak_bytes = 0;
int64 FLAGS_min_residual_bytes = 0;
int64 FLAGS_min_output_bytes = 0;
int64 FLAGS_min_micros = 0;
int64 FLAGS_min_accelerator_micros = 0;
int64 FLAGS_min_cpu_micros = 0;
int64 FLAGS_min_params = 0;
int64 FLAGS_min_float_ops = 0;
int64 FLAGS_min_occurrence = 0;
@ -101,7 +106,14 @@ int Run(int argc, char** argv) {
"TensorFlow Checkpoint file name"),
Flag("max_depth", &FLAGS_max_depth, "max depth"),
Flag("min_bytes", &FLAGS_min_bytes, "min_bytes"),
Flag("min_peak_bytes", &FLAGS_min_peak_bytes, "min_peak_bytes"),
Flag("min_residual_bytes", &FLAGS_min_residual_bytes,
"min_residual_bytes"),
Flag("min_output_bytes", &FLAGS_min_output_bytes, "min_output_bytes"),
Flag("min_micros", &FLAGS_min_micros, "min micros"),
Flag("min_accelerator_micros", &FLAGS_min_accelerator_micros,
"min acclerator_micros"),
Flag("min_cpu_micros", &FLAGS_min_cpu_micros, "min_cpu_micros"),
Flag("min_params", &FLAGS_min_params, "min params"),
Flag("min_float_ops", &FLAGS_min_float_ops, "min float ops"),
Flag("min_occurrence", &FLAGS_min_occurrence, "min occurrence"),
@ -214,12 +226,14 @@ int Run(int argc, char** argv) {
return 0;
}
Options opts(FLAGS_max_depth, FLAGS_min_bytes, FLAGS_min_micros,
FLAGS_min_params, FLAGS_min_float_ops, FLAGS_min_occurrence,
FLAGS_step, FLAGS_order_by, account_type_regexes,
start_name_regexes, trim_name_regexes, show_name_regexes,
hide_name_regexes, FLAGS_account_displayed_op_only, select,
output_type, output_options);
Options opts(
FLAGS_max_depth, FLAGS_min_bytes, FLAGS_min_peak_bytes,
FLAGS_min_residual_bytes, FLAGS_min_output_bytes, FLAGS_min_micros,
FLAGS_min_accelerator_micros, FLAGS_min_cpu_micros, FLAGS_min_params,
FLAGS_min_float_ops, FLAGS_min_occurrence, FLAGS_step, FLAGS_order_by,
account_type_regexes, start_name_regexes, trim_name_regexes,
show_name_regexes, hide_name_regexes, FLAGS_account_displayed_op_only,
select, output_type, output_options);
if (cmd == kCmds[2] || cmd == kCmds[3]) {
tf_stat.BuildView(cmd);

5
tensorflow/core/profiler/tfprof_options.proto
View File

@ -7,7 +7,12 @@ package tensorflow.tfprof;
message OptionsProto {
int64 max_depth = 1;
int64 min_bytes = 2;
int64 min_peak_bytes = 19;
int64 min_residual_bytes = 20;
int64 min_output_bytes = 21;
int64 min_micros = 3;
int64 min_accelerator_micros = 22;
int64 min_cpu_micros = 23;
int64 min_params = 4;
int64 min_float_ops = 5;
int64 min_occurrence = 17;

24
tensorflow/core/profiler/tfprof_output.proto
View File

@ -28,8 +28,15 @@ message GraphNodeProto {
int64 accelerator_exec_micros = 17;
int64 cpu_exec_micros = 18;
// Total requested bytes by the op.
// Total bytes requested by the op.
int64 requested_bytes = 3;
// Max bytes allocated and being used by the op at a point.
int64 peak_bytes = 24;
// Total bytes requested by the op and not released before end.
int64 residual_bytes = 25;
// Total bytes output by the op (not necessarily allocated by the op).
int64 output_bytes = 26;
// Number of parameters if available.
int64 parameters = 4;
// Number of float operations.
@ -49,6 +56,10 @@ message GraphNodeProto {
int64 total_cpu_exec_micros = 20;
int64 total_requested_bytes = 7;
int64 total_peak_bytes = 27;
int64 total_residual_bytes = 28;
int64 total_output_bytes = 29;
int64 total_parameters = 8;
int64 total_float_ops = 14;
@ -81,6 +92,13 @@ message MultiGraphNodeProto {
// Total requested bytes by the code.
int64 requested_bytes = 3;
// Max bytes allocated and being used by the op at a point.
int64 peak_bytes = 16;
// Total bytes requested by the op and not released before end.
int64 residual_bytes = 17;
// Total bytes output by the op (not necessarily allocated by the op).
int64 output_bytes = 18;
// Number of parameters if available.
int64 parameters = 4;
// Number of float operations.
@ -93,6 +111,10 @@ message MultiGraphNodeProto {
int64 total_cpu_exec_micros = 15;
int64 total_requested_bytes = 7;
int64 total_peak_bytes = 19;
int64 total_residual_bytes = 20;
int64 total_output_bytes = 21;
int64 total_parameters = 8;
int64 total_float_ops = 9;

5
tensorflow/python/profiler/model_analyzer.py
View File

@ -53,7 +53,12 @@ def _build_options(options):
opts = tfprof_options_pb2.OptionsProto()
opts.max_depth = options.get('max_depth', 10)
opts.min_bytes = options.get('min_bytes', 0)
opts.min_peak_bytes = options.get('min_peak_bytes', 0)
opts.min_residual_bytes = options.get('min_residual_bytes', 0)
opts.min_output_bytes = options.get('min_output_bytes', 0)
opts.min_micros = options.get('min_micros', 0)
opts.min_accelerator_micros = options.get('min_accelerator_micros', 0)
opts.min_cpu_micros = options.get('min_cpu_micros', 0)
opts.min_params = options.get('min_params', 0)
opts.min_float_ops = options.get('min_float_ops', 0)
opts.min_occurrence = options.get('min_occurrence', 0)

128
tensorflow/python/profiler/model_analyzer_test.py
View File

@ -20,6 +20,7 @@ from __future__ import print_function
import gzip
import io
import os
import random
from tensorflow.core.profiler import profile_pb2
from tensorflow.core.protobuf import config_pb2
@ -118,7 +119,7 @@ class PrintModelAnalysisTest(test.TestCase):
with gfile.Open(outfile, 'r') as f:
# pylint: disable=line-too-long
self.assertEqual(
'node name | output bytes | # parameters | # float_ops | assigned devices | input',
'node name | requested bytes | # parameters | # float_ops | assigned devices | in',
f.read()[0:80])
# pylint: enable=line-too-long
@ -243,7 +244,9 @@ class PrintModelAnalysisTest(test.TestCase):
.with_accounted_types(['.*'])
.with_min_occurrence(10)
.order_by('occurrence')
.select(['params', 'micros', 'occurrence', 'input_shapes']).build())
.select(['params', 'micros', 'bytes',
'peak_bytes', 'residual_bytes',
'output_bytes', 'occurrence', 'input_shapes']).build())
with session.Session() as sess:
x = lib.BuildFullModel()
@ -261,8 +264,8 @@ class PrintModelAnalysisTest(test.TestCase):
with gfile.Open(outfile, 'r') as f:
# pylint: disable=line-too-long
self.assertEqual(
'nodename|totalexecutiontime|acceleratorexecutiontime|cpuexecutiontime|#parameters|opoccurrence(run|defined)|inputshapes\n',
f.read().replace('\t', '').replace(' ', '')[0:120])
'nodename|requestedbytes|peakbytes|residualbytes|outputbytes|totalexecutiontime|acceleratorexecutiontime|cpuexecutiontime|#parameters|opoccurrence(run|defined)|inputshapes\nConst0B(0',
f.read().replace('\t', '').replace(' ', '')[0:180])
# pylint: enable=line-too-long
total_children = 0
@ -370,6 +373,123 @@ class PrintModelAnalysisTest(test.TestCase):
for attr in ['op_types', 'device', 'input_shapes']:
self.pprof_test_helper(attr, True)
def testMinOption(self):
ops.reset_default_graph()
def check_min(nodes, mm=0, mam=0, mcm=0, mb=0, mpb=0, mrb=0, mob=0):
for n in nodes:
if mm > 0:
self.assertGreaterEqual(n.exec_micros, mm)
if mam > 0:
self.assertGreaterEqual(n.accelerator_exec_micros, mam)
if mcm > 0:
self.assertGreaterEqual(n.cpu_exec_micros, mcm)
if mb > 0:
self.assertGreaterEqual(n.requested_bytes, mb)
if mpb > 0:
self.assertGreaterEqual(n.peak_bytes, mpb)
if mrb > 0:
self.assertGreaterEqual(n.residual_bytes, mrb)
if mob > 0:
self.assertGreaterEqual(n.output_bytes, mob)
check_min(n.children, mm, mam, mcm, mb, mpb, mrb, mob)
with session.Session() as sess:
x = lib.BuildSmallModel()
sess.run(variables.global_variables_initializer())
run_meta = config_pb2.RunMetadata()
_ = sess.run(x,
options=config_pb2.RunOptions(
trace_level=config_pb2.RunOptions.FULL_TRACE),
run_metadata=run_meta)
min_val = random.randint(0, 10000)
opts = builder(builder.time_and_memory(min_micros=min_val)
).with_empty_output().build()
tfprof_node = model_analyzer.profile(
sess.graph, run_meta=run_meta, options=opts)
check_min(tfprof_node.children, mm=min_val)
opts = builder(builder.time_and_memory(min_accelerator_micros=min_val)
).with_empty_output().build()
tfprof_node = model_analyzer.profile(
sess.graph, run_meta=run_meta, options=opts)
check_min(tfprof_node.children, mam=min_val)
opts = builder(builder.time_and_memory(min_cpu_micros=min_val)
).with_empty_output().build()
tfprof_node = model_analyzer.profile(
sess.graph, run_meta=run_meta, options=opts)
check_min(tfprof_node.children, mcm=min_val)
opts = builder(builder.time_and_memory(min_bytes=min_val)
).with_empty_output().build()
tfprof_node = model_analyzer.profile(
sess.graph, run_meta=run_meta, options=opts)
check_min(tfprof_node.children, mb=min_val)
opts = builder(builder.time_and_memory(min_peak_bytes=min_val)
).with_empty_output().build()
tfprof_node = model_analyzer.profile(
sess.graph, run_meta=run_meta, options=opts)
check_min(tfprof_node.children, mpb=min_val)
opts = builder(builder.time_and_memory(min_residual_bytes=min_val)
).with_empty_output().build()
tfprof_node = model_analyzer.profile(
sess.graph, run_meta=run_meta, options=opts)
check_min(tfprof_node.children, mrb=min_val)
opts = builder(builder.time_and_memory(min_output_bytes=min_val)
).with_empty_output().build()
tfprof_node = model_analyzer.profile(
sess.graph, run_meta=run_meta, options=opts)
check_min(tfprof_node.children, mob=min_val)
def testSelectOption(self):
ops.reset_default_graph()
outfile = os.path.join(test.get_temp_dir(), 'dump')
def check_selection(selected, not_selected):
with gfile.Open(outfile, 'r') as f:
s = f.read()
for attr in selected:
self.assertTrue(s.find(attr) > 0, s)
for attr in not_selected:
self.assertFalse(s.find(attr) > 0, s)
with session.Session() as sess:
x = lib.BuildSmallModel()
sess.run(variables.global_variables_initializer())
run_meta = config_pb2.RunMetadata()
_ = sess.run(x,
options=config_pb2.RunOptions(
trace_level=config_pb2.RunOptions.FULL_TRACE),
run_metadata=run_meta)
opts = builder(builder.time_and_memory()
).with_file_output(outfile).select(['micros']).build()
_ = model_analyzer.profile(
sess.graph, run_meta=run_meta, options=opts)
check_selection(['total execution time', 'accelerator execution time'],
['bytes'])
opts = builder(builder.time_and_memory()
).with_file_output(outfile).select(['bytes']).build()
_ = model_analyzer.profile(
sess.graph, run_meta=run_meta, options=opts)
check_selection(['requested bytes'],
['peak bytes', 'residual bytes', 'output bytes'])
opts = builder(builder.time_and_memory()).with_file_output(
outfile).select(
['peak_bytes', 'residual_bytes', 'output_bytes']).build()
_ = model_analyzer.profile(
sess.graph, run_meta=run_meta, options=opts)
check_selection(['peak bytes', 'residual bytes', 'output bytes'],
['requested_bytes'])
if __name__ == '__main__':
test.main()

65
tensorflow/python/profiler/option_builder.py
View File

@ -139,18 +139,41 @@ class ProfileOptionBuilder(object):
'output': 'stdout'}
@staticmethod
def time_and_memory(min_micros=1, min_bytes=1):
def time_and_memory(min_micros=1, min_bytes=1, min_accelerator_micros=0,
min_cpu_micros=0, min_peak_bytes=0, min_residual_bytes=0,
min_output_bytes=0):
"""Show operation time and memory consumptions.
Args:
min_micros: Only show profiler nodes with more execution time than this.
min_bytes: Only show profiler nodes consuming more memory than this.
min_micros: Only show profiler nodes with execution time
no less than this. It sums accelerator and cpu times.
min_bytes: Only show profiler nodes requested to allocate no less bytes
than this.
min_accelerator_micros: Only show profiler nodes spend no less than
this time on accelerator (e.g. GPU).
min_cpu_micros: Only show profiler nodes spend no less than
this time on cpu.
min_peak_bytes: Only show profiler nodes using no less than this bytes
at peak (high watermark). For profiler nodes consist of multiple
graph nodes, it sums the graph nodes' peak_bytes.
min_residual_bytes: Only show profiler nodes have no less than
this bytes not being de-allocated after Compute() ends. For
profiler nodes consist of multiple graph nodes, it sums the
graph nodes' residual_bytes.
min_output_bytes: Only show profiler nodes have no less than this bytes
output. The output are not necessarily allocated by this profiler
nodes.
Returns:
A dict of profiling options.
"""
return {'max_depth': 10000,
'min_bytes': min_bytes,
'min_peak_bytes': min_peak_bytes,
'min_residual_bytes': min_residual_bytes,
'min_output_bytes': min_output_bytes,
'min_micros': min_micros,
'min_accelerator_micros': min_accelerator_micros,
'min_cpu_micros': min_cpu_micros,
'min_params': 0,
'min_float_ops': 0,
'min_occurrence': 0,
@ -188,28 +211,54 @@ class ProfileOptionBuilder(object):
self._options['max_depth'] = max_depth
return self
def with_min_memory(self, min_bytes):
def with_min_memory(self,
min_bytes=0,
min_peak_bytes=0,
min_residual_bytes=0,
min_output_bytes=0):
"""Only show profiler nodes consuming no less than 'min_bytes'.
Args:
min_bytes: Only show profiler nodes with memory consumption
no less than this.
min_bytes: Only show profiler nodes requested to allocate no less bytes
than this.
min_peak_bytes: Only show profiler nodes using no less than this bytes
at peak (high watermark). For profiler nodes consist of multiple
graph nodes, it sums the graph nodes' peak_bytes.
min_residual_bytes: Only show profiler nodes have no less than
this bytes not being de-allocated after Compute() ends. For
profiler nodes consist of multiple graph nodes, it sums the
graph nodes' residual_bytes.
min_output_bytes: Only show profiler nodes have no less than this bytes
output. The output are not necessarily allocated by this profiler
nodes.
Returns:
self
"""
self._options['min_bytes'] = min_bytes
self._options['min_peak_bytes'] = min_peak_bytes
self._options['min_residual_bytes'] = min_residual_bytes
self._options['min_output_bytes'] = min_output_bytes
return self
def with_min_execution_time(self, min_micros):
def with_min_execution_time(self,
min_micros=0,
min_accelerator_micros=0,
min_cpu_micros=0):
"""Only show profiler nodes consuming no less than 'min_micros'.
Args:
min_micros: Only show profiler nodes with execution time
no less than this.
no less than this. It sums accelerator and cpu times.
min_accelerator_micros: Only show profiler nodes spend no less than
this time on accelerator (e.g. GPU).
min_cpu_micros: Only show profiler nodes spend no less than
this time on cpu.
Returns:
self
"""
self._options['min_micros'] = min_micros
self._options['min_accelerator_micros'] = min_accelerator_micros
self._options['min_cpu_micros'] = min_cpu_micros
return self
def with_min_parameters(self, min_params):

2
tensorflow/python/profiler/profiler_test.py
View File

@ -118,7 +118,7 @@ class ProfilerTest(test.TestCase):
def testMultiStepProfile(self):
ops.reset_default_graph()
opts = builder.time_and_memory()
opts = builder.time_and_memory(min_bytes=0)
with session.Session() as sess:
r1, r2, r3 = lib.BuildSplitableModel()

24
tensorflow/tools/api/golden/tensorflow.profiler.-graph-node-proto.pbtxt
View File

@ -46,14 +46,26 @@ tf_class {
name: "NAME_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "OUTPUT_BYTES_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "PARAMETERS_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "PEAK_BYTES_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "REQUESTED_BYTES_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "RESIDUAL_BYTES_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "RUN_COUNT_FIELD_NUMBER"
mtype: "<type \'int\'>"
@ -86,14 +98,26 @@ tf_class {
name: "TOTAL_FLOAT_OPS_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "TOTAL_OUTPUT_BYTES_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "TOTAL_PARAMETERS_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "TOTAL_PEAK_BYTES_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "TOTAL_REQUESTED_BYTES_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "TOTAL_RESIDUAL_BYTES_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "TOTAL_RUN_COUNT_FIELD_NUMBER"
mtype: "<type \'int\'>"

24
tensorflow/tools/api/golden/tensorflow.profiler.-multi-graph-node-proto.pbtxt
View File

@ -38,14 +38,26 @@ tf_class {
name: "NAME_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "OUTPUT_BYTES_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "PARAMETERS_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "PEAK_BYTES_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "REQUESTED_BYTES_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "RESIDUAL_BYTES_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "TOTAL_ACCELERATOR_EXEC_MICROS_FIELD_NUMBER"
mtype: "<type \'int\'>"
@ -62,14 +74,26 @@ tf_class {
name: "TOTAL_FLOAT_OPS_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "TOTAL_OUTPUT_BYTES_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "TOTAL_PARAMETERS_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "TOTAL_PEAK_BYTES_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "TOTAL_REQUESTED_BYTES_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member {
name: "TOTAL_RESIDUAL_BYTES_FIELD_NUMBER"
mtype: "<type \'int\'>"
}
member_method {
name: "ByteSize"
}

6
tensorflow/tools/api/golden/tensorflow.profiler.-profile-option-builder.pbtxt
View File

@ -28,7 +28,7 @@ tf_class {
}
member_method {
name: "time_and_memory"
argspec: "args=[\'min_micros\', \'min_bytes\'], varargs=None, keywords=None, defaults=[\'1\', \'1\'], "
argspec: "args=[\'min_micros\', \'min_bytes\', \'min_accelerator_micros\', \'min_cpu_micros\', \'min_peak_bytes\', \'min_residual_bytes\', \'min_output_bytes\'], varargs=None, keywords=None, defaults=[\'1\', \'1\', \'0\', \'0\', \'0\', \'0\', \'0\'], "
}
member_method {
name: "trainable_variables_parameter"
@ -52,7 +52,7 @@ tf_class {
}
member_method {
name: "with_min_execution_time"
argspec: "args=[\'self\', \'min_micros\'], varargs=None, keywords=None, defaults=None"
argspec: "args=[\'self\', \'min_micros\', \'min_accelerator_micros\', \'min_cpu_micros\'], varargs=None, keywords=None, defaults=[\'0\', \'0\', \'0\'], "
}
member_method {
name: "with_min_float_operations"
@ -60,7 +60,7 @@ tf_class {
}
member_method {
name: "with_min_memory"
argspec: "args=[\'self\', \'min_bytes\'], varargs=None, keywords=None, defaults=None"
argspec: "args=[\'self\', \'min_bytes\', \'min_peak_bytes\', \'min_residual_bytes\', \'min_output_bytes\'], varargs=None, keywords=None, defaults=[\'0\', \'0\', \'0\', \'0\'], "
}
member_method {
name: "with_min_occurrence"