[rocprofv3] Unconditionally collect stream and kernel rename data in rocprofv3 for rocpd (#171)

* Remove config checks for stream and kernel rename data collection

* Updated csv generation to check if kernel rename is on before calling get_kernel_name

* Update metadata to use kernel_rename bool argument

* Formatting + unconditionally store kernel name in rocpd

* Readded kernel rename parameter after rebase

* Fixed rebase conflicts

* Updated comment in line with github comments

* Added check in rocpd csv.cpp to output kernel name if region name is empty

* Add test for kernel rename

---------

Co-authored-by: Ian Trowbridge <Ian.Trowbridge@amd.com>
This commit is contained in:
systems-assistant[bot]ésIan Trowbridge authored and GitHub committed 2025-09-10 16:03:15 -05:00
1 szülő 156de36a92
commit b60c0ceddd
14 files changed
+441 -79

No files matched your search

@@ -288,11 +288,12 @@ generate_csv(const output_config&
{
for(auto record : data.get(ditr))
{
auto row_ss = std::stringstream{};
auto kernel_name = tool_metadata.get_kernel_name(record.dispatch_info.kernel_id,
record.correlation_id.external.value);
auto row_ss = std::stringstream{};
const auto* kernel_info =
tool_metadata.get_kernel_symbol(record.dispatch_info.kernel_id);
auto kernel_name = tool_metadata.get_kernel_name(record.dispatch_info.kernel_id,
cfg.kernel_rename,
record.correlation_id.external.value);
auto lds_block_size_v =
(kernel_info->group_segment_size + (lds_block_size - 1)) & ~(lds_block_size - 1);
@@ -637,7 +638,8 @@ generate_csv(const output_config& cfg,
record.thread_id,
magnitude(record.dispatch_data.dispatch_info.grid_size),
record.dispatch_data.dispatch_info.kernel_id,
tool_metadata.get_kernel_name(kernel_id, correlation_id.external.value),
tool_metadata.get_kernel_name(
kernel_id, cfg.kernel_rename, correlation_id.external.value),
magnitude(record.dispatch_data.dispatch_info.workgroup_size),
lds_block_size_v,
record.dispatch_data.dispatch_info.private_segment_size,
@@ -695,8 +695,8 @@ write_otf2(const output_config& cfg,
const auto* sym = _get_kernel_sym_data(info);
CHECK(sym != nullptr);
auto name =
tool_metadata.get_kernel_name(info.kernel_id, itr.correlation_id.external.value);
auto name = tool_metadata.get_kernel_name(
info.kernel_id, cfg.kernel_rename, itr.correlation_id.external.value);
_hash_data.emplace(
get_hash_id(name),
region_info{std::string{name}, OTF2_REGION_ROLE_FUNCTION, OTF2_PARADIGM_HIP});
@@ -647,6 +647,11 @@ write_rocpd(
itr.demangled_kernel_name,
itr.truncated_kernel_name);
for(const auto& itr : tool_metadata.kernel_rename_map.get())
{
add_string_entry(_metadata, itr.first);
}
for(const auto& itr : tool_metadata.get_code_objects())
if(itr.uri != nullptr) add_string_entry(_metadata, itr.uri);
@@ -1020,10 +1025,12 @@ write_rocpd(
}
dispatch_evt_ids.at(dispatch_id) = evt_id;
// Unconditionally collect kernel rename data if it is available. rocpd needs to be able
// to use kernel rename option after data has already been collected, so the kernel
// rename data needs to be stored in generated db.
auto region_name =
(corr_id.external.value > 0 && (enable_duplicate_check || kernel_id > 0))
? tool_metadata.get_kernel_name(kernel_id, corr_id.external.value)
? tool_metadata.get_kernel_name(kernel_id, true, corr_id.external.value)
: std::string_view{};
auto agent_node_id = tool_metadata.get_agent(info.agent_id)->node_id;
@@ -62,7 +62,7 @@ get_stats(const stats_map_t& data_v)
} // namespace
stats_entry_t
generate_stats(const output_config& /*cfg*/,
generate_stats(const output_config& cfg,
const metadata& tool_metadata,
const generator<tool_buffer_tracing_kernel_dispatch_ext_record_t>& data)
{
@@ -72,6 +72,7 @@ generate_stats(const output_config& /*cfg*/,
for(auto record : data.get(ditr))
{
auto kernel_name = tool_metadata.get_kernel_name(record.dispatch_info.kernel_id,
cfg.kernel_rename,
record.correlation_id.external.value);
kernel_stats[kernel_name] += (record.end_timestamp - record.start_timestamp);
@@ -696,19 +696,22 @@ metadata::get_marker_message(uint64_t corr_id) const
}
std::string_view
metadata::get_kernel_name(uint64_t kernel_id, uint64_t rename_id) const
metadata::get_kernel_name(uint64_t kernel_id, bool rename_kernel, uint64_t rename_id) const
{
auto string_entry = kernel_rename_map.rlock(
[](auto& _data, uint64_t _val) {
for(const auto& itr : _data)
if(itr.second == _val) return itr.first;
return std::string_view{};
},
rename_id);
if(!string_entry.empty())
if(rename_kernel)
{
if(const auto* _name = common::get_string_entry(string_entry))
return std::string_view{*_name};
auto string_entry = kernel_rename_map.rlock(
[](auto& _data, uint64_t _val) {
for(const auto& itr : _data)
if(itr.second == _val) return itr.first;
return std::string_view{};
},
rename_id);
if(!string_entry.empty())
{
if(const auto* _name = common::get_string_entry(string_entry))
return std::string_view{*_name};
}
}
const auto* _kernel_data = get_kernel_symbol(kernel_id);
@@ -231,7 +231,9 @@ struct metadata
const std::vector<std::string>& _command_line = {});
std::string_view get_marker_message(uint64_t corr_id) const;
std::string_view get_kernel_name(uint64_t kernel_id, uint64_t rename_id) const;
std::string_view get_kernel_name(uint64_t kernel_id,
bool rename_kernel,
uint64_t rename_id) const;
std::string_view get_kind_name(rocprofiler_callback_tracing_kind_t kind) const;
std::string_view get_kind_name(rocprofiler_buffer_tracing_kind_t kind) const;
std::string_view get_operation_name(rocprofiler_callback_tracing_kind_t kind,