kernel iteration filtering for counter collection (#911)

* kernel filtering for counter collection

* fixing trace tests

* removing print statements

* fix CI fail

* handling preload and updating docs

* minor fix

* misc fix

* misc fix

* Typo fix

* Update rocprofv3 + input schema

- "application_passes" -> "jobs"
- removed nesting in YAML/JSON inputs
- improved customAction (now booleanArgAction)
  - supports --<name> (defaults to true)
  - supports --<name>=<truth-value>
  - supports --<name> <truth-value>
- added --kernel-iteration-range to command-line
- automatically support new command-line options in YAML/JSON input
- standardized PMC return from text input to match PMC from YAML/JSON input
- added support for --log-level env
- updated various input*.(yml|json) to modified schema

* Update config.cpp

- added recommended code to get_kernel_filter_range

* Fixing iteration

* misc fix

* support only [-] for iteration

* bug fix

* Fix using-rocprofv3.rst

* Update config.cpp

- patch get_kernel_filter_range

---------

Co-authored-by: Jonathan R. Madsen <jonathanrmadsen@gmail.com>
This commit is contained in:
SrirakshaNag
2024-07-26 21:46:53 -05:00
committato da GitHub
parent dc671497da
commit ace34abd11
23 ha cambiato i file con 1303 aggiunte e 275 eliminazioni
+31 -28
Vedi File
@@ -27,8 +27,11 @@
#include "lib/common/demangle.hpp"
#include "lib/common/environment.hpp"
#include "lib/common/filesystem.hpp"
#include "lib/common/logging.hpp"
#include "lib/common/utility.hpp"
#include <rocprofiler-sdk/cxx/details/delimit.hpp>
#include <fmt/core.h>
#include <unistd.h>
@@ -132,14 +135,6 @@ handle_special_chars(std::string& str)
str.at(pos) = ' ';
}
bool
has_kernel_name_format(std::string const& str)
{
return std::find_if(str.begin(), str.end(), [](unsigned char ch) {
return (isalnum(ch) != 0 || ch == '_');
}) != str.end();
}
bool
has_counter_format(std::string const& str)
{
@@ -149,34 +144,40 @@ has_counter_format(std::string const& str)
}
// validate kernel names
auto
parse_kernel_names(const std::string& line)
std::unordered_set<uint32_t>
get_kernel_filter_range(const std::string& kernel_filter)
{
auto kernel_names_v = std::vector<std::string>{};
if(line.empty()) return kernel_names_v;
if(kernel_filter.empty()) return {};
auto kernel_names = std::set<std::string>{};
trim(line);
auto input_line = std::stringstream{line};
auto kernel_name = std::string{};
while(getline(input_line, kernel_name, ','))
auto delim = rocprofiler::sdk::parse::tokenize(kernel_filter, ",");
auto range_set = std::unordered_set<uint32_t>{};
for(const auto& itr : delim)
{
if(has_kernel_name_format(kernel_name))
if(itr.find('-') != std::string::npos && itr.find('[') != std::string::npos &&
itr.find(']') != std::string::npos)
{
ROCP_INFO << "kernel name " << kernel_names.size() << ": " << kernel_name;
kernel_names.emplace(kernel_name);
auto drange = rocprofiler::sdk::parse::tokenize(itr, "[-] ");
ROCP_FATAL_IF(drange.size() != 2)
<< "bad range format for '" << itr << "'. Expected [A-B] where A and B are numbers";
uint32_t start_range = std::stoul(drange.front());
uint32_t end_range = std::stoul(drange.back());
for(auto i = start_range; i <= end_range; i++)
range_set.emplace(i);
}
else
{
ROCP_ERROR << "invalid kernel name: " << kernel_name;
auto dval = rocprofiler::sdk::parse::tokenize(itr, " ");
ROCP_ERROR_IF(dval.empty()) << "kernel range value '" << itr << "' produced no numbers";
for(const auto& ditr : dval)
{
ROCP_FATAL_IF(ditr.find_first_not_of("0123456789") != std::string::npos)
<< "expected integer for " << itr << ". Non-integer value detected";
range_set.emplace(std::stoul(ditr));
}
}
}
kernel_names_v.reserve(kernel_names.size());
for(const auto& itr : kernel_names)
kernel_names_v.emplace_back(itr);
return kernel_names_v;
return range_set;
}
std::set<std::string>
@@ -240,7 +241,8 @@ get_mpi_rank()
}
config::config()
: kernel_names{parse_kernel_names(get_env("ROCPROF_KERNEL_NAMES", std::string{}))}
: kernel_filter_range{get_kernel_filter_range(
get_env("ROCPROF_KERNEL_FILTER_RANGE", std::string{}))}
, counters{parse_counters(get_env("ROCPROF_COUNTERS", std::string{}))}
{
auto output_format = get_env("ROCPROF_OUTPUT_FORMAT", "CSV");
@@ -281,6 +283,7 @@ config::config()
LOG_IF(FATAL, supported_formats.count(itr) == 0)
<< "Unsupported output format type: " << itr;
}
if(kernel_filter_include.empty()) kernel_filter_include = std::string(".*");
}
std::vector<output_key>
@@ -28,6 +28,7 @@
#include <set>
#include <string>
#include <unordered_set>
#include <vector>
namespace rocprofiler
@@ -81,8 +82,13 @@ struct config
std::string output_path = get_env("ROCPROF_OUTPUT_PATH", fs::current_path().string());
std::string output_file = get_env("ROCPROF_OUTPUT_FILE_NAME", std::to_string(getpid()));
std::string tmp_directory = get_env("ROCPROF_TMPDIR", output_path);
std::vector<std::string> kernel_names = {};
std::set<std::string> counters = {};
std::string kernel_filter_include =
get_env("ROCPROF_KERNEL_FILTER_INCLUDE_REGEX", std::string{".*"});
std::string kernel_filter_exclude =
get_env("ROCPROF_KERNEL_FILTER_EXCLUDE_REGEX", std::string{});
std::unordered_set<uint32_t> kernel_filter_range = {};
std::set<std::string> counters = {};
};
template <config_context ContextT = config_context::global>
+73 -37
Vedi File
@@ -177,40 +177,79 @@ as_pointer()
}
using code_object_data_map_t = std::unordered_map<uint64_t, rocprofiler_code_object_data_t>;
using targeted_kernels_set_t = std::unordered_set<rocprofiler_kernel_id_t>;
using targeted_kernels_map_t =
std::unordered_map<rocprofiler_kernel_id_t, std::unordered_set<uint32_t>>;
using counter_dimension_info_map_t =
std::unordered_map<uint64_t, std::vector<rocprofiler_record_dimension_info_t>>;
using agent_info_map_t = std::unordered_map<rocprofiler_agent_id_t, rocprofiler_agent_t>;
using agent_info_map_t = std::unordered_map<rocprofiler_agent_id_t, rocprofiler_agent_t>;
using kernel_iteration_t = std::unordered_map<rocprofiler_kernel_id_t, uint32_t>;
auto code_obj_data = as_pointer<common::Synchronized<code_object_data_map_t, true>>();
auto* kernel_data = as_pointer<common::Synchronized<kernel_symbol_data_map_t, true>>();
auto* marker_msg_data = as_pointer<common::Synchronized<marker_message_map_t, true>>();
auto counter_dimension_data = common::Synchronized<counter_dimension_info_map_t, true>{};
auto target_kernels = common::Synchronized<targeted_kernels_set_t>{};
auto target_kernels = common::Synchronized<targeted_kernels_map_t>{};
auto* buffered_name_info = as_pointer(get_buffer_id_names());
auto* callback_name_info = as_pointer(get_callback_id_names());
auto* agent_info = as_pointer(agent_info_map_t{});
auto* tool_functions = as_pointer(tool_table{});
auto* stats_timestamp = as_pointer(timestamps_t{});
auto kernel_iteration = common::Synchronized<kernel_iteration_t, true>{};
bool
add_kernel_target(uint64_t _kern_id)
add_kernel_target(uint64_t _kern_id, const std::unordered_set<uint32_t>& range)
{
return target_kernels
.wlock([](targeted_kernels_set_t& _targets_v,
uint64_t _kern_id_v) { return _targets_v.emplace(_kern_id_v); },
_kern_id)
.wlock(
[](targeted_kernels_map_t& _targets_v,
uint64_t _kern_id_v,
std::unordered_set<uint32_t> _range) {
return _targets_v.emplace(_kern_id_v, _range);
},
_kern_id,
range)
.second;
}
bool
is_targeted_kernel(uint64_t _kern_id)
{
return target_kernels.rlock(
[](const targeted_kernels_set_t& _targets_v, uint64_t _kern_id_v) {
return (_targets_v.count(_kern_id_v) > 0);
bool is_target_kernel = false;
std::unordered_set<uint32_t> range = {};
is_target_kernel = target_kernels.rlock(
[&range](const auto& _targets_v, uint64_t _kern_id_v) {
if(_targets_v.find(_kern_id_v) != _targets_v.end())
{
range = _targets_v.at(_kern_id_v);
return true;
}
return false;
},
_kern_id);
if(is_target_kernel)
{
kernel_iteration.rlock(
[&](const auto& _kernel_iter,
uint64_t _kernel_id,
std::unordered_set<uint32_t> _range) {
auto itr = _kernel_iter.at(_kernel_id);
// If the iteration range is not given then all iterations of the kernel is profiled
if(_range.empty())
is_target_kernel = true;
else if(_range.find(itr) != _range.end())
{
is_target_kernel = true;
}
else
is_target_kernel = false;
},
_kern_id,
range);
}
return is_target_kernel;
}
auto&
@@ -496,34 +535,19 @@ code_object_tracing_callback(rocprofiler_callback_tracing_record_t record,
{
// if kernel name is provided by user then by default all kernels in the application
// are targeted
if(tool::get_config().kernel_names.empty())
const auto& kernel_info = itr.first->second;
auto kernel_filter_include = tool::get_config().kernel_filter_include;
auto kernel_filter_exclude = tool::get_config().kernel_filter_exclude;
auto kernel_filter_range = tool::get_config().kernel_filter_range;
std::regex include_regex(kernel_filter_include);
std::regex exclude_regex(kernel_filter_exclude);
if(std::regex_search(kernel_info.formatted_kernel_name, include_regex))
{
add_kernel_target(sym_data->kernel_id);
}
else
{
const auto& kernel_info = itr.first->second;
for(const auto& name : tool::get_config().kernel_names)
{
if(name == kernel_info.truncated_kernel_name)
{
add_kernel_target(itr.first->first);
break;
}
else
{
auto dkernel_name = std::string_view{kernel_info.demangled_kernel_name};
auto pos = dkernel_name.find(name);
// if the demangled kernel name contains name and the next character is
// '(' then mark as found
if(pos != std::string::npos && (pos + 1) < dkernel_name.size() &&
dkernel_name.at(pos + 1) == '(')
{
add_kernel_target(itr.first->first);
break;
}
}
}
if(kernel_filter_exclude.empty())
add_kernel_target(sym_data->kernel_id, kernel_filter_range);
else if(!std::regex_search(kernel_info.formatted_kernel_name, exclude_regex))
add_kernel_target(sym_data->kernel_id, kernel_filter_range);
}
}
}
@@ -882,6 +906,18 @@ dispatch_callback(rocprofiler_profile_counting_dispatch_data_t dispatch_data,
auto kernel_id = dispatch_data.dispatch_info.kernel_id;
auto agent_id = dispatch_data.dispatch_info.agent_id;
kernel_iteration.wlock(
[](auto& _kernel_iter, rocprofiler_kernel_id_t _kernel_id) {
auto itr = _kernel_iter.find(_kernel_id);
if(itr == _kernel_iter.end())
_kernel_iter.emplace(_kernel_id, 1);
else
{
itr->second++;
}
},
kernel_id);
if(!is_targeted_kernel(kernel_id))
{
return;