Cherry-pick [Reduce number of calls to rocprof ROCm#384] into release…

…/rocm-rel-6.2 (ROCm#422)

* Improve perfmon coalescing

Signed-off-by: benrichard-amd <ben.richard@amd.com>

* Interleve TCC channel counters

Signed-off-by: benrichard-amd <ben.richard@amd.com>

* Remove duplicate normal counters
Interleve TCC channel counters in putput file  e.g.  TCC_HIT[0] TCC_ATOMIC[0] ... TCC_HIT[1] TCC_ATOMIC[1]

Signed-off-by: benrichard-amd <ben.richard@amd.com>

* Save accumulate counters to SQ_ files

Omniperf analyze expects the accumulate files to be in SQ_*.csv files.

Since these files also contain PMC counters (we are trying to
fit as many counters into each file as possible to minimize runs),
we need to include these SQ_*.csv files in pmc_perf.csv.

Signed-off-by: benrichard-amd <ben.richard@amd.com>

* Update to work with rocprof v1

Signed-off-by: benrichard-amd <ben.richard@amd.com>

* Remove unused method

Signed-off-by: benrichard-amd <ben.richard@amd.com>

* Set correct number of TCC channels for gfx942

Ran into rocprof error:
ROCProfiler: fatal error: input metric'TCC_EA0_RDREQ[16]' not supported on this hardware: gfx942

gfx942 has 16 channels, not 32.

Signed-off-by: benrichard-amd <ben.richard@amd.com>

* Fix code formatting

Signed-off-by: benrichard-amd <ben.richard@amd.com>

---------

Signed-off-by: benrichard-amd <ben.richard@amd.com>

src/omniperf_profile/profiler_base.py

@@ -62,43 +62,6 @@ def get_profiler_options(self, fname):
         # assume no SoC specific options and return empty list by default
         return []
 
-    @demarcate
-    def pmc_perf_split(self):
-        """Avoid default rocprof join utility by spliting each line into a separate input file"""
-        workload_perfmon_dir = os.path.join(self.__args.path, "perfmon")
-        lines = (
-            open(os.path.join(workload_perfmon_dir, "pmc_perf.txt"), "r")
-            .read()
-            .splitlines()
-        )
-
-        # Iterate over each line in pmc_perf.txt
-        mpattern = r"^pmc:(.*)"
-        i = 0
-        for line in lines:
-            # Verify no comments
-            stext = line.split("#")[0].strip()
-            if not stext:
-                continue
-
-            # all pmc counters start with  "pmc:"
-            m = re.match(mpattern, stext)
-            if m is None:
-                continue
-
-            # Create separate file for each line
-            fd = open(workload_perfmon_dir + "/pmc_perf_" + str(i) + ".txt", "w")
-            fd.write(stext + "\n\n")
-            fd.write("gpu:\n")
-            fd.write("range:\n")
-            fd.write("kernel:\n")
-            fd.close()
-
-            i += 1
-
-        # Remove old pmc_perf.txt input from perfmon dir
-        os.remove(workload_perfmon_dir + "/pmc_perf.txt")
-
     @demarcate
     def join_prof(self, out=None):
         """Manually join separated rocprof runs"""
@@ -107,6 +70,7 @@ def join_prof(self, out=None):
             if out is None:
                 out = self.__args.path + "/pmc_perf.csv"
             files = glob.glob(self.__args.path + "/" + "pmc_perf_*.csv")
+            files.extend(glob.glob(self.__args.path + "/" + "SQ_*.csv"))
         elif type(self.__args.path) == list:
             files = self.__args.path
         else:
@@ -261,7 +225,9 @@ def join_prof(self, out=None):
             df.to_csv(out, index=False)
             if not self.__args.verbose:
                 for file in files:
-                    os.remove(file)
+                    # Do not remove accumulate counter files
+                    if "SQ_" not in file:
+                        os.remove(file)
         else:
             return df
 

src/omniperf_profile/profiler_rocprof_v1.py

@@ -59,8 +59,6 @@ def get_profiler_options(self, fname):
     def pre_processing(self):
         """Perform any pre-processing steps prior to profiling."""
         super().pre_processing()
-        if self.ready_to_profile:
-            self.pmc_perf_split()
 
     @demarcate
     def run_profiling(self, version: str, prog: str):

src/omniperf_profile/profiler_rocprof_v2.py

@@ -63,8 +63,6 @@ def get_profiler_options(self, fname):
     def pre_processing(self):
         """Perform any pre-processing steps prior to profiling."""
         super().pre_processing()
-        if self.ready_to_profile:
-            self.pmc_perf_split()
 
     @demarcate
     def run_profiling(self, version, prog):