[AIROCMLIR-1010] Separate --debug flag into --debug and --debug-quick-tune-data to fix ~50GB .debug file

mlir/test/perf-scripts/runtime/tuningRunner-gemm.py

@@ -0,0 +1,13 @@
+# Using a tiny GEMM with --debug-quick-tune-data. This emits a `.debug`
+# TSV of the per-config table entries (PerfConfig + TFlops) but, unlike the
+# full ``--debug`` flag, omits the heavy per-iteration ``MeasurementsMs``
+# arrays. Verify the debug file is produced, has the expected header and
+# per-config rows, and that the measurements column is absent.
+# RUN: rm -f %t2.tsv %t2.tsv.state %t2.tsv.debug
+# RUN: tuningRunner.py --op gemm --tuning-space=quick --debug-quick-tune-data \
+# RUN:     --config='-g 1 -m 64 -n 64 -k 64 -t f32 -out_datatype f32 -transA 0 -transB 0' \
+# RUN:     -q -o %t2.tsv
+# RUN: FileCheck %s --check-prefix=DEBUG --implicit-check-not=MeasurementsMs < %t2.tsv.debug
+#
+# DEBUG: PerfConfig{{.*}}TFlops
+# DEBUG: {{v[0-9]+:}}

mlir/utils/performance/tuningRunner.py

@@ -169,6 +169,7 @@ def get_gpu_logger(gpu_id: int) -> logging.LoggerAdapter:
 class Options:
     """Configuration options for the tuning process."""
     debug: bool
+    debug_quick_tune_data: bool
     tuning_space_kind: str
     quiet: bool
     verbose: bool
@@ -1640,8 +1641,9 @@ def tune_configs(ctx: TuningContext, status_only: bool) -> bool:
 
     has_errors = False
 
-    debug_enabled = ctx.options.debug and ctx.options.output != '-'
-    if ctx.options.debug and not debug_enabled:
+    debug_requested = ctx.options.debug or ctx.options.debug_quick_tune_data
+    debug_enabled = debug_requested and ctx.options.output != '-'
+    if debug_requested and not debug_enabled:
         logger.warning("Debug output disabled when writing to stdout")
 
     with (OutputFileWriter(ctx.options.output, ctx.options) as results_writer,
@@ -1939,7 +1941,13 @@ def parse_arguments(gpu_topology: GpuTopology,
                         "--debug",
                         action='store_true',
                         default=False,
-                        help="Enable debug output including detailed measurements")
+                        help="Enable debug output including detailed per-iteration measurements")
+
+    parser.add_argument("--debug-quick-tune-data",
+                        action='store_true',
+                        default=False,
+                        help="Enable debug output for quick tuning data generation without the "
+                        "detailed per-iteration measurement arrays")
 
     parser.add_argument("--tuning-space",
                         default="full",
@@ -2125,6 +2133,7 @@ def main(args=None):
                       num_cu=num_cu,
                       num_chiplets=num_chiplets,
                       debug=parsed_args.debug,
+                      debug_quick_tune_data=parsed_args.debug_quick_tune_data,
                       quiet=parsed_args.quiet,
                       verbose=parsed_args.verbose,
                       tuning_space_kind=parsed_args.tuning_space,