JPEG Activity tracing in Perfetto (#108)

- Add JPEG activity track in perfetto trace
- Add JPEG decode tests to the examples
- Change existing videodecode test to include JPEG testing
- Rename videodecode test file to decode to include jpeg tests too
- Fix a bug in the test which checks for total activity of 0
- Disable rocDecode and rocJPEG samples from the github image files
This commit is contained in:
Sajina PK
2025-02-21 10:25:01 -05:00
committed by GitHub
parent c1d8799f92
commit 59d3399901
21 changed files with 1637 additions and 72 deletions
+2
View File
@@ -106,6 +106,7 @@ ROCPROFSYS_DEFINE_CATEGORY(category, rocm_smi_temp, ROCPROFSYS_CATEGORY_ROCM_SMI
ROCPROFSYS_DEFINE_CATEGORY(category, rocm_smi_power, ROCPROFSYS_CATEGORY_ROCM_SMI_POWER, "device_power", "Power consumption of a GPU device")
ROCPROFSYS_DEFINE_CATEGORY(category, rocm_smi_memory_usage, ROCPROFSYS_CATEGORY_ROCM_SMI_MEMORY_USAGE, "device_memory_usage", "Memory usage of a GPU device")
ROCPROFSYS_DEFINE_CATEGORY(category, rocm_smi_vcn_activity, ROCPROFSYS_CATEGORY_ROCM_SMI_VCN_ACTIVITY, "device_vcn_activity", "VCN Activity of a GPU device")
ROCPROFSYS_DEFINE_CATEGORY(category, rocm_smi_jpeg_activity, ROCPROFSYS_CATEGORY_ROCM_SMI_JPEG_ACTIVITY, "device_jpeg_activity", "JPEG Activity of a GPU device")
ROCPROFSYS_DEFINE_CATEGORY(category, rocm_rocdecode_api, ROCPROFSYS_CATEGORY_ROCM_ROCDECODE_API, "rocm_rocdecode_api", "ROCm ROCDecode API")
ROCPROFSYS_DEFINE_CATEGORY(category, rocm_rccl, ROCPROFSYS_CATEGORY_ROCM_RCCL, "rccl", "ROCm Communication Collectives Library (RCCL) regions")
ROCPROFSYS_DEFINE_CATEGORY(category, pthread, ROCPROFSYS_CATEGORY_PTHREAD, "pthread", "POSIX threading functions")
@@ -171,6 +172,7 @@ using name = perfetto_category<Tp...>;
ROCPROFSYS_PERFETTO_CATEGORY(category::rocm_smi_power), \
ROCPROFSYS_PERFETTO_CATEGORY(category::rocm_smi_memory_usage), \
ROCPROFSYS_PERFETTO_CATEGORY(category::rocm_smi_vcn_activity), \
ROCPROFSYS_PERFETTO_CATEGORY(category::rocm_smi_jpeg_activity), \
ROCPROFSYS_PERFETTO_CATEGORY(category::rocm_rocdecode_api), \
ROCPROFSYS_PERFETTO_CATEGORY(category::rocm_rccl), \
ROCPROFSYS_PERFETTO_CATEGORY(category::pthread), \
+13
View File
@@ -82,6 +82,8 @@ struct backtrace_gpu_memory
{};
struct backtrace_gpu_vcn
{};
struct backtrace_gpu_jpeg
{};
using sampling_wall_clock = data_tracker<double, backtrace_wall_clock>;
using sampling_cpu_clock = data_tracker<double, backtrace_cpu_clock>;
using sampling_percent = data_tracker<double, backtrace_fraction>;
@@ -90,6 +92,7 @@ using sampling_gpu_temp = data_tracker<double, backtrace_gpu_temp>;
using sampling_gpu_power = data_tracker<double, backtrace_gpu_power>;
using sampling_gpu_memory = data_tracker<double, backtrace_gpu_memory>;
using sampling_gpu_vcn = data_tracker<double, backtrace_gpu_vcn>;
using sampling_gpu_jpeg = data_tracker<double, backtrace_gpu_jpeg>;
template <typename ApiT, typename StartFuncT = default_functor_t,
typename StopFuncT = default_functor_t>
@@ -123,6 +126,7 @@ ROCPROFSYS_DEFINE_CONCRETE_TRAIT(is_available, component::sampling_gpu_temp, fal
ROCPROFSYS_DEFINE_CONCRETE_TRAIT(is_available, component::sampling_gpu_power, false_type)
ROCPROFSYS_DEFINE_CONCRETE_TRAIT(is_available, component::sampling_gpu_memory, false_type)
ROCPROFSYS_DEFINE_CONCRETE_TRAIT(is_available, component::sampling_gpu_vcn, false_type)
ROCPROFSYS_DEFINE_CONCRETE_TRAIT(is_available, component::sampling_gpu_jpeg, false_type)
#endif
TIMEMORY_SET_COMPONENT_API(rocprofsys::component::sampling_wall_clock,
@@ -151,6 +155,9 @@ TIMEMORY_SET_COMPONENT_API(rocprofsys::component::sampling_gpu_temp, project::ro
TIMEMORY_SET_COMPONENT_API(rocprofsys::component::sampling_gpu_vcn, project::rocprofsys,
tpls::rocm, device::gpu, os::supports_linux,
category::sampling, category::process_sampling)
TIMEMORY_SET_COMPONENT_API(rocprofsys::component::sampling_gpu_jpeg, project::rocprofsys,
tpls::rocm, device::gpu, os::supports_linux,
category::sampling, category::process_sampling)
TIMEMORY_METADATA_SPECIALIZATION(rocprofsys::component::sampling_wall_clock,
"sampling_wall_clock", "Wall-clock timing",
@@ -179,6 +186,10 @@ TIMEMORY_METADATA_SPECIALIZATION(rocprofsys::component::sampling_gpu_vcn,
"sampling_gpu_vcn",
"GPU VCN Utilization (% activity) via ROCm-SMI",
"Derived from sampling")
TIMEMORY_METADATA_SPECIALIZATION(rocprofsys::component::sampling_gpu_jpeg,
"sampling_gpu_jpeg",
"GPU JPEG Utilization (% activity) via ROCm-SMI",
"Derived from sampling")
// statistics type
TIMEMORY_STATISTICS_TYPE(rocprofsys::component::sampling_wall_clock, double)
@@ -188,6 +199,7 @@ TIMEMORY_STATISTICS_TYPE(rocprofsys::component::sampling_gpu_temp, double)
TIMEMORY_STATISTICS_TYPE(rocprofsys::component::sampling_gpu_power, double)
TIMEMORY_STATISTICS_TYPE(rocprofsys::component::sampling_gpu_memory, double)
TIMEMORY_STATISTICS_TYPE(rocprofsys::component::sampling_gpu_vcn, double)
TIMEMORY_STATISTICS_TYPE(rocprofsys::component::sampling_gpu_jpeg, double)
TIMEMORY_STATISTICS_TYPE(rocprofsys::component::comm_data_tracker_t, float)
// enable timing units
@@ -220,6 +232,7 @@ ROCPROFSYS_DEFINE_CONCRETE_TRAIT(report_sum, component::sampling_gpu_temp, false
ROCPROFSYS_DEFINE_CONCRETE_TRAIT(report_sum, component::sampling_gpu_power, false_type)
ROCPROFSYS_DEFINE_CONCRETE_TRAIT(report_sum, component::sampling_gpu_memory, false_type)
ROCPROFSYS_DEFINE_CONCRETE_TRAIT(report_sum, component::sampling_gpu_vcn, false_type)
ROCPROFSYS_DEFINE_CONCRETE_TRAIT(report_sum, component::sampling_gpu_jpeg, false_type)
// reporting categories (mean)
ROCPROFSYS_DEFINE_CONCRETE_TRAIT(report_mean, component::sampling_percent, false_type)
+9 -9
View File
@@ -315,10 +315,10 @@ configure_settings(bool _init)
"Enable ROCm API and kernel tracing", true, "backend",
"rocm");
ROCPROFSYS_CONFIG_SETTING(
bool, "ROCPROFSYS_USE_ROCM_SMI",
"Enable sampling GPU power, temp, utilization, vcn_activity and memory usage",
true, "backend", "rocm_smi", "rocm", "process_sampling");
ROCPROFSYS_CONFIG_SETTING(bool, "ROCPROFSYS_USE_ROCM_SMI",
"Enable sampling GPU power, temp, utilization, "
"vcn_activity, jpeg_activity and memory usage",
true, "backend", "rocm_smi", "rocm", "process_sampling");
ROCPROFSYS_CONFIG_SETTING(bool, "ROCPROFSYS_USE_SAMPLING",
"Enable statistical sampling of call-stack", false,
@@ -626,11 +626,11 @@ configure_settings(bool _init)
rocprofiler_sdk::config_settings(_config);
ROCPROFSYS_CONFIG_SETTING(
std::string, "ROCPROFSYS_ROCM_SMI_METRICS",
"rocm-smi metrics to collect: busy, temp, power, vcn_activity, mem_usage",
"busy,temp,power,vcn_activity,mem_usage", "backend", "rocm_smi", "rocm",
"process_sampling", "advanced");
ROCPROFSYS_CONFIG_SETTING(std::string, "ROCPROFSYS_ROCM_SMI_METRICS",
"rocm-smi metrics to collect: busy, temp, power, "
"vcn_activity, jpeg_activity, mem_usage",
"busy,temp,power,mem_usage", "backend", "rocm_smi", "rocm",
"process_sampling", "advanced");
ROCPROFSYS_CONFIG_SETTING(size_t, "ROCPROFSYS_PERFETTO_SHMEM_SIZE_HINT_KB",
"Hint for shared-memory buffer size in perfetto (in KB)",
@@ -57,8 +57,9 @@ extern "C"
ROCPROFSYS_CATEGORY_ROCM_SMI_TEMP,
ROCPROFSYS_CATEGORY_ROCM_SMI_POWER,
ROCPROFSYS_CATEGORY_ROCM_SMI_MEMORY_USAGE,
ROCPROFSYS_CATEGORY_ROCM_SMI_VCN_ACTIVITY,
ROCPROFSYS_CATEGORY_ROCM_ROCDECODE_API,
ROCPROFSYS_CATEGORY_ROCM_SMI_VCN_ACTIVITY,
ROCPROFSYS_CATEGORY_ROCM_SMI_JPEG_ACTIVITY,
ROCPROFSYS_CATEGORY_ROCM_RCCL,
ROCPROFSYS_CATEGORY_SAMPLING,
ROCPROFSYS_CATEGORY_PTHREAD,
+68 -16
View File
@@ -160,12 +160,15 @@ data::sample(uint32_t _dev_id)
&m_power, &power_type)
ROCPROFSYS_RSMI_GET(get_settings(m_dev_id).mem_usage, rsmi_dev_memory_usage_get,
_dev_id, RSMI_MEM_TYPE_VRAM, &m_mem_usage);
ROCPROFSYS_RSMI_GET(get_settings(m_dev_id).vcn_activity,
rsmi_dev_gpu_metrics_info_get, _dev_id, &_gpu_metrics);
ROCPROFSYS_ROCM_SMI_CALL(rsmi_dev_gpu_metrics_info_get(_dev_id, &_gpu_metrics));
for(const auto& activity : _gpu_metrics.vcn_activity)
for(const auto& v_activity : _gpu_metrics.vcn_activity)
{
if(activity != UINT16_MAX) m_vcn_metrics.push_back(activity);
if(v_activity != UINT16_MAX) m_vcn_metrics[_dev_id].push_back(v_activity);
}
for(const auto& j_activity : _gpu_metrics.jpeg_activity)
{
if(j_activity != UINT16_MAX) m_jpeg_metrics[_dev_id].push_back(j_activity);
}
#undef ROCPROFSYS_RSMI_GET
@@ -262,6 +265,7 @@ void
data::post_process(uint32_t _dev_id)
{
using component::sampling_gpu_busy;
using component::sampling_gpu_jpeg;
using component::sampling_gpu_memory;
using component::sampling_gpu_power;
using component::sampling_gpu_temp;
@@ -282,7 +286,7 @@ data::post_process(uint32_t _dev_id)
auto _settings = get_settings(_dev_id);
auto _process_perfetto = [&]() {
auto _idx = std::array<uint64_t, 5>{};
auto _idx = std::array<uint64_t, 6>{};
{
_idx.fill(_idx.size());
uint64_t nidx = 0;
@@ -291,6 +295,7 @@ data::post_process(uint32_t _dev_id)
if(_settings.power) _idx.at(2) = nidx++;
if(_settings.mem_usage) _idx.at(3) = nidx++;
if(_settings.vcn_activity) _idx.at(4) = nidx++;
if(_settings.jpeg_activity) _idx.at(5) = nidx++;
}
for(auto& itr : _rocm_smi)
@@ -302,6 +307,18 @@ data::post_process(uint32_t _dev_id)
auto addendum = [&](const char* _v) {
return JOIN(" ", "GPU", _v, JOIN("", '[', _dev_id, ']'), "(S)");
};
auto addendum_blk = [&](std::size_t _i, const char* _metric) {
if(_i < 10)
{
return JOIN(" ", "GPU", JOIN("", '[', _dev_id, ']'), _metric,
JOIN("", "[0", _i, ']'), "(S)");
}
else
{
return JOIN(" ", "GPU", JOIN("", '[', _dev_id, ']'), _metric,
JOIN("", '[', _i, ']'), "(S)");
}
};
if(_settings.busy) counter_track::emplace(_dev_id, addendum("Busy"), "%");
if(_settings.temp)
@@ -313,11 +330,25 @@ data::post_process(uint32_t _dev_id)
"megabytes");
if(_settings.vcn_activity)
{
for(std::size_t i = 0; i < std::size(itr.m_vcn_metrics); ++i)
counter_track::emplace(
_dev_id,
addendum(("VCN Activity on " + std::to_string(i)).c_str()),
"%");
for(const auto& [dev_id, metrics] : itr.m_vcn_metrics)
{
for(std::size_t i = 0; i < std::size(metrics); ++i)
{
counter_track::emplace(
_dev_id, addendum_blk(i, " VCN Activity"), "%");
}
}
}
if(_settings.jpeg_activity)
{
for(const auto& [dev_id, metrics] : itr.m_jpeg_metrics)
{
for(std::size_t i = 0; i < std::size(metrics); ++i)
{
counter_track::emplace(_dev_id,
addendum_blk(i, "JPEG Activity"), "%");
}
}
}
}
uint64_t _ts = itr.m_ts;
@@ -342,12 +373,28 @@ data::post_process(uint32_t _dev_id)
counter_track::at(_dev_id, _idx.at(3)), _ts, _usage);
if(_settings.vcn_activity)
{
uint64_t idx = _idx.at(4);
for(const auto& temp : itr.m_vcn_metrics)
for(const auto& [dev_id, metrics] : itr.m_vcn_metrics)
{
TRACE_COUNTER("device_vcn_activity", counter_track::at(_dev_id, idx),
_ts, temp);
++idx;
for(std::size_t i = 0; i < std::size(metrics); ++i)
{
double _vcn_activity = metrics[i];
TRACE_COUNTER("device_vcn_activity",
counter_track::at(_dev_id, _idx.at(4) + i), _ts,
_vcn_activity);
}
}
}
if(_settings.jpeg_activity)
{
for(const auto& [dev_id, metrics] : itr.m_jpeg_metrics)
{
for(std::size_t i = 0; i < std::size(metrics); ++i)
{
double _jpeg_activity = metrics[i];
TRACE_COUNTER("device_jpeg_activity",
counter_track::at(_dev_id, _idx.at(5) + i), _ts,
_jpeg_activity);
}
}
}
}
@@ -440,9 +487,10 @@ setup()
key_pair_t{ "power", get_settings(dev_id).power },
key_pair_t{ "mem_usage", get_settings(dev_id).mem_usage },
key_pair_t{ "vcn_activity", get_settings(dev_id).vcn_activity },
key_pair_t{ "jpeg_activity", get_settings(dev_id).jpeg_activity },
};
get_settings(dev_id) = { false, false, false, false };
get_settings(dev_id) = { false, false, false, false, false, false };
for(const auto& metric : tim::delimit(*_metrics, ",;:\t\n "))
{
auto iitr = supported.find(metric);
@@ -524,3 +572,7 @@ ROCPROFSYS_INSTANTIATE_EXTERN_COMPONENT(
ROCPROFSYS_INSTANTIATE_EXTERN_COMPONENT(
TIMEMORY_ESC(data_tracker<double, rocprofsys::component::backtrace_gpu_vcn>), true,
double)
ROCPROFSYS_INSTANTIATE_EXTERN_COMPONENT(
TIMEMORY_ESC(data_tracker<double, rocprofsys::component::backtrace_gpu_jpeg>), true,
double)
+18 -12
View File
@@ -71,11 +71,12 @@ device_count();
struct settings
{
bool busy = true;
bool temp = true;
bool power = true;
bool mem_usage = true;
bool vcn_activity = true;
bool busy = true;
bool temp = true;
bool power = true;
bool mem_usage = true;
bool vcn_activity = true;
bool jpeg_activity = true;
};
struct data
@@ -100,13 +101,14 @@ struct data
static void post_process(uint32_t _dev_id);
uint32_t m_dev_id = std::numeric_limits<uint32_t>::max();
timestamp_t m_ts = 0;
busy_perc_t m_busy_perc = 0;
temp_t m_temp = 0;
power_t m_power = 0;
mem_usage_t m_mem_usage = 0;
std::vector<uint16_t> m_vcn_metrics = {};
uint32_t m_dev_id = std::numeric_limits<uint32_t>::max();
timestamp_t m_ts = 0;
busy_perc_t m_busy_perc = 0;
temp_t m_temp = 0;
power_t m_power = 0;
mem_usage_t m_mem_usage = 0;
std::unordered_map<uint32_t, std::vector<double>> m_vcn_metrics = {};
std::unordered_map<uint32_t, std::vector<double>> m_jpeg_metrics = {};
friend std::ostream& operator<<(std::ostream& _os, const data& _v)
{
@@ -185,5 +187,9 @@ ROCPROFSYS_DECLARE_EXTERN_COMPONENT(
TIMEMORY_ESC(data_tracker<double, rocprofsys::component::backtrace_gpu_vcn>), true,
double)
ROCPROFSYS_DECLARE_EXTERN_COMPONENT(
TIMEMORY_ESC(data_tracker<double, rocprofsys::component::backtrace_gpu_jpeg>), true,
double)
# endif
#endif
+8 -1
View File
@@ -126,6 +126,7 @@ using component::backtrace_wall_clock; // NOLINT
using component::callchain;
using component::sampling_cpu_clock;
using component::sampling_gpu_busy;
using component::sampling_gpu_jpeg;
using component::sampling_gpu_memory;
using component::sampling_gpu_power;
using component::sampling_gpu_temp;
@@ -1575,10 +1576,16 @@ struct sampling_initialization
sampling_gpu_temp::set_format_flags(sampling_gpu_temp::get_format_flags());
sampling_gpu_vcn::label() = "sampling_gpu_vcn_percent";
sampling_gpu_vcn::description() = "Utilization of VCN(s)";
sampling_gpu_vcn::description() = "VCN instance(s) activity";
sampling_gpu_vcn::set_precision(0);
sampling_gpu_vcn::set_format_flags(sampling_gpu_vcn::get_format_flags() &
std::ios_base::showpoint);
sampling_gpu_jpeg::label() = "sampling_gpu_jpeg_percent";
sampling_gpu_jpeg::description() = "JPEG instance(s) activity";
sampling_gpu_jpeg::set_precision(0);
sampling_gpu_jpeg::set_format_flags(sampling_gpu_jpeg::get_format_flags() &
std::ios_base::showpoint);
}
};
} // namespace