fix crash for running rocprofv3 on mi100 (#557)

* initial hack to fix for v3 stucking on mi100 becasue of -m parameter and missing counter csv file

* proper formating

* refactored profiler option function to take soc arch

* resolve missing step that casued error for profiler option

* fix typo of arch name

* change method of putting soc info into profiler option

* isort and black format

* add comment for the part that handles missing counter csv file

* remove unncecessary import

---------

Co-authored-by: YANG WANG <[email protected]>

src/rocprof_compute_profile/profiler_base.py

@@ -63,7 +63,7 @@ def __init__(self, args, profiler_mode, soc):
     def get_args(self):
         return self.__args
 
-    def get_profiler_options(self, fname):
+    def get_profiler_options(self, fname, soc):
         """Fetch any version specific arguments required by profiler"""
         # assume no SoC specific options and return empty list by default
         return []
@@ -361,9 +361,7 @@ def run_profiling(self, version: str, prog: str):
                     console_debug(output)
             console_log("profiling", "Current input file: %s" % fname)
 
-            # Fetch any SoC/profiler specific profiling options
-            options = self._soc.get_profiler_options()
-            options += self.get_profiler_options(fname)
+            options = self.get_profiler_options(fname, self._soc)
             if (
                 self.__profiler == "rocprofv1"
                 or self.__profiler == "rocprofv2"

src/rocprof_compute_profile/profiler_rocprof_v1.py

@@ -38,10 +38,16 @@ def __init__(self, profiling_args, profiler_mode, soc):
             or not self.get_args().roof_only
         )
 
-    def get_profiler_options(self, fname):
+    def get_profiler_options(self, fname, soc):
         fbase = Path(fname).stem
         app_cmd = self.get_args().remaining
-        args = [
+
+        args = []
+        # can be removed in the future. It supports gfx908 + v1
+        if soc.get_arch() == "gfx908":
+            args += ["-m", soc.get_workload_perfmon_dir() + "/" + "metrics.xml"]
+
+        args += [
             # v1 requires request for timestamps
             "--timestamp",
             "on",

src/rocprof_compute_profile/profiler_rocprof_v2.py

@@ -39,10 +39,16 @@ def __init__(self, profiling_args, profiler_mode, soc):
             or not self.get_args().roof_only
         )
 
-    def get_profiler_options(self, fname):
+    def get_profiler_options(self, fname, soc):
         fbase = Path(fname).stem
         app_cmd = shlex.split(self.get_args().remaining)
-        args = [
+
+        args = []
+        # can be removed in the future. It supports gfx908 + v2
+        if soc.get_arch() == "gfx908":
+            args += ["-m", soc.get_workload_perfmon_dir() + "/" + "metrics.xml"]
+
+        args += [
             # v2 requires output directory argument
             "-d",
             self.get_args().path + "/" + "out",

src/rocprof_compute_profile/profiler_rocprof_v3.py

@@ -40,7 +40,7 @@ def __init__(self, profiling_args, profiler_mode, soc):
             or not self.get_args().roof_only
         )
 
-    def get_profiler_options(self, fname):
+    def get_profiler_options(self, fname, soc):
         app_cmd = shlex.split(self.get_args().remaining)
         trace_option = "--kernel-trace"
         rocprof_out_format = "json"

src/rocprof_compute_soc/soc_base.py

@@ -97,12 +97,6 @@ def set_compatible_profilers(self, profiler_names: list):
     def get_compatible_profilers(self):
         return self.__compatible_profilers
 
-    @demarcate
-    def get_profiler_options(self):
-        """Fetch any SoC specific arguments required by the profiler"""
-        # assume no SoC specific options and return empty list by default
-        return []
-
     @demarcate
     def populate_mspec(self):
         from utils.specs import run, search, total_sqc, total_xcds

src/rocprof_compute_soc/soc_gfx908.py

@@ -69,11 +69,6 @@ def __init__(self, args, mspec):
             self._mspec.max_mclk = 1200
             self._mspec.cur_mclk = 1200
 
-    @demarcate
-    def get_profiler_options(self):
-        # Mi100 requires a custom xml config
-        return ["-m", self.get_workload_perfmon_dir() + "/" + "metrics.xml"]
-
     # -----------------------
     # Required child methods
     # -----------------------

src/utils/utils.py

@@ -596,6 +596,9 @@ def run_prof(
         new_env = os.environ.copy()
         new_env["ROCPROFILER_INDIVIDUAL_XCC_MODE"] = "1"
 
+    is_timestamps = False
+    if path(fname).name == "timestamps.txt":
+        is_timestamps = True
     time_1 = time.time()
 
     # profile the app
@@ -686,10 +689,14 @@ def run_prof(
                 results_files_csv = glob.glob(
                     workload_dir + "/out/pmc_1/*/*_converted.csv"
                 )
-            else:
+            elif is_timestamps:
+                # when the input is timestamps, we know counter csv file is not generated and will instead parse kernel trace file
                 results_files_csv = glob.glob(
                     workload_dir + "/out/pmc_1/*/*_kernel_trace.csv"
                 )
+            else:
+                # when the input is not for timestamps, and counter csv file is not generated, we assume failed rocprof run and will completely bypass the file generation and merging for current pmc
+                return
 
         else:
             console_error("The output file of rocprofv3 can only support json or csv!!!")