AI/rocm-systems
2
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

OMPT Target Offload Support (#17)

Browse Source 
- Porting from https://github.com/ROCm/omnitrace/pull/411
- Improve OMPT support
- Add OpenMP target example to testing
- Update Timemory submodule to use ROCm/Timemory rather than NERSC/Timemory
- Update `actions/upload-artifacts` to v4
- Standardize the `cmake_minimum_required` to 3.18.4 across workflows, project, and examples
- Updated Ubuntu 20.04 workflows
This commit is contained in:
David Galiffi
2024-11-07 16:49:32 -05:00
committed by GitHub
parent 2cb8a82101
commit 7dce5926a7
34 changed files with 800 additions and 124 deletions
Download Patch File Download Diff File Expand all files Collapse all files
source/lib/rocprof-sys/library/roctracer.cpp
+5 -42
View File
@@ -163,55 +163,18 @@ get_clock_skew()
static auto _use = tim::get_env("ROCPROFSYS_USE_ROCTRACER_CLOCK_SKEW", true);
if(!_use) return 0;
static auto _v = []() {
namespace cpu = tim::cpu;
// synchronize timestamps
// We'll take a CPU timestamp before and after taking a GPU timestmp, then
// take the average of those two, hoping that it's roughly at the same time
// as the GPU timestamp.
static auto _cpu_now = []() {
cpu::fence();
return comp::wall_clock::record();
};
static auto _gpu_now = []() {
cpu::fence();
auto _gpu_now = []() {
uint64_t _ts = 0;
ROCPROFSYS_ROCTRACER_CALL(roctracer_get_timestamp(&_ts));
roctracer_get_timestamp(&_ts);
return _ts;
};
do
{
// warm up cache and allow for any static initialization
(void) _cpu_now();
(void) _gpu_now();
} while(false);
// discard (warm-up)
(void) tracing::get_clock_skew(_gpu_now, 1);
auto _compute = [](volatile uint64_t& _cpu_ts, volatile uint64_t& _gpu_ts) {
_cpu_ts = 0;
_gpu_ts = 0;
_cpu_ts += _cpu_now() / 2;
_gpu_ts += _gpu_now() / 1;
_cpu_ts += _cpu_now() / 2;
return static_cast<int64_t>(_cpu_ts) - static_cast<int64_t>(_gpu_ts);
};
constexpr int64_t _n = 10;
int64_t _cpu_ave = 0;
int64_t _gpu_ave = 0;
int64_t _diff = 0;
for(int64_t i = 0; i < _n; ++i)
{
volatile uint64_t _cpu_ts = 0;
volatile uint64_t _gpu_ts = 0;
_diff += _compute(_cpu_ts, _gpu_ts);
_cpu_ave += _cpu_ts / _n;
_gpu_ave += _gpu_ts / _n;
}
ROCPROFSYS_BASIC_VERBOSE(2, "CPU timestamp: %li\n", _cpu_ave);
ROCPROFSYS_BASIC_VERBOSE(2, "HIP timestamp: %li\n", _gpu_ave);
auto _diff = tracing::get_clock_skew(_gpu_now, 10);
ROCPROFSYS_BASIC_VERBOSE(1, "CPU/HIP timestamp skew: %li (used: %s)\n", _diff,
_use ? "yes" : "no");
_diff /= _n;
return _diff;
}();
return _v;