tt-perf-report 1.0.4__py3-none-any.whl → 1.0.6__py3-none-any.whl

tt_perf_report/perf_report.py

@@ -742,8 +742,28 @@ def merge_device_rows(df):
     device_ids = sorted(block_by_device.keys())
     merged_blocks = []
 
-    for blocks in zip(*[block_by_device[device_id] for device_id in device_ids]):
-        op_name = blocks[0][0]
+    global_index = 0
+    while max(len(block_by_device[device_id]) for device_id in device_ids) > 0:
+        blocks = []
+        op_name = None
+        missing_devices = []
+        for device_id in device_ids:
+            if not len(block_by_device[device_id]):
+                print(colored(f"Warning: Device {device_id} is missing operation {op_name} at index {global_index}", "yellow"))
+                continue
+            if op_name is None:
+                op_name = block_by_device[device_id][0][0]
+            elif op_name != block_by_device[device_id][0][0]:
+                missing_devices.append(device_id)
+                continue
+
+            blocks.append(block_by_device[device_id].pop(0))
+
+        if missing_devices:
+            print(colored(f"Warning: {op_name} at index {global_index} not present in CSV for {len(missing_devices)} devices {missing_devices} - do not trust data for this op or directly subsequent ops with the same name", "yellow"))
+
+        if not blocks:
+            break
 
         if "AllGather" in op_name or "ReduceScatter" in op_name:
             # For collective ops, take the row with minimum duration
@@ -754,6 +774,8 @@ def merge_device_rows(df):
             max_duration_block = max(blocks, key=lambda x: x[1]["DEVICE KERNEL DURATION [ns]"])
             merged_blocks.append(max_duration_block[1])
 
+        global_index += 1
+
     return pd.DataFrame(merged_blocks)
 
 
@@ -792,11 +814,15 @@ def filter_by_id_range(rows, id_range):
     return rows
 
 
+def filter_host_ops(rows):
+    return [row for row in rows if not is_host_op(row)]
+
+
 def main():
     args, id_range = parse_args()
     generate_perf_report(
-        args.csv_file, args.signpost, args.ignore_signposts, args.min_percentage, id_range, args.csv, args.no_advice, args.tracing_mode, args.raw_op_codes,
-    )
+        args.csv_file, args.signpost, args.ignore_signposts, args.min_percentage, id_range, args.csv, args.no_advice,
+        args.tracing_mode, args.raw_op_codes, args.no_host_ops)
 
 
 def parse_args():
@@ -818,6 +844,8 @@ def parse_args():
     parser.add_argument("--no-advice", action="store_true", help="Only show the table section of the report")
     parser.add_argument("--tracing-mode", action="store_true", help="Do not sort when in tracing mode")
     parser.add_argument("--raw-op-codes", action="store_true", help="Include raw op codes in output")
+    parser.add_argument("--no-host-ops", action="store_true", help="Do not include host ops in output")
+
     args = parser.parse_args()
 
     # Set the global color_output variable
@@ -833,7 +861,9 @@ def parse_args():
     return args, id_range
 
 
-def generate_perf_report(csv_file, signpost, ignore_signposts, min_percentage, id_range, csv_output_file, no_advice, tracing_mode, raw_op_codes):
+def generate_perf_report(csv_file, signpost, ignore_signposts, min_percentage,
+                         id_range, csv_output_file, no_advice, tracing_mode,
+                         raw_op_codes, no_host_ops):
     df = pd.read_csv(csv_file, low_memory=False)
 
     # Add a column for original row numbers
@@ -867,7 +897,7 @@ def generate_perf_report(csv_file, signpost, ignore_signposts, min_percentage, i
         prev_row = row
 
         # Count device and host ops
-        if "(torch)" in op_data["OP Code"].raw_value:
+        if is_host_op(op_data):
             host_ops += 1
         else:
             device_ops += 1
@@ -878,6 +908,9 @@ def generate_perf_report(csv_file, signpost, ignore_signposts, min_percentage, i
     # Filter rows based on id_range
     rows = filter_by_id_range(rows, id_range)
 
+    if no_host_ops:
+        rows = filter_host_ops(rows)
+
     # Recalculate derived columns after filtering
     add_derived_columns(rows)
 
@@ -933,5 +966,9 @@ def generate_perf_report(csv_file, signpost, ignore_signposts, min_percentage, i
             print_advice_section(rows, visible_headers, col_widths)
 
 
+def is_host_op(op_data):
+    return "(torch)" in op_data["OP Code"].raw_value
+
+
 if __name__ == "__main__":
     main()

tt_perf_report-1.0.6.dist-info/LICENSE_understanding.txt

@@ -0,0 +1,3 @@
+For the avoidance of doubt, this software assists in programming Tenstorrent products.
+
+However, making, using, or selling hardware, models, or IP may require the license of rights (such as patent rights) from Tenstorrent or others.

{tt_perf_report-1.0.4.dist-info → tt_perf_report-1.0.6.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: tt-perf-report
-Version: 1.0.4
+Version: 1.0.6
 Summary: This tool analyzes performance traces from TT-Metal operations, providing insights into throughput, bottlenecks, and optimization opportunities.
 License:                                  Apache License
                                    Version 2.0, January 2004
@@ -216,6 +216,7 @@ Project-URL: Repository, https://github.com/tenstorrent/tt-perf-report
 Keywords: tenstorrent,tt-metal
 Description-Content-Type: text/markdown
 License-File: LICENSE
+License-File: LICENSE_understanding.txt
 Requires-Dist: pandas
 
 # Performance Report Analysis Tool

tt_perf_report-1.0.6.dist-info/RECORD

@@ -0,0 +1,9 @@
+tt_perf_report/__init__.py,sha256=-j4iFYebIwgdS8uphk8-M6zasRqGBL3CQGnJH9keRuI,92
+tt_perf_report/perf_report.py,sha256=OAHGjdxnq5BOYgodTR1v4Ec3GksoKuBXAGA6eeJi4zY,38265
+tt_perf_report-1.0.6.dist-info/LICENSE,sha256=6dZGjPECz_ULS-sf40FLlt6OmQFcrRvmzG5mJRZCQ5I,11825
+tt_perf_report-1.0.6.dist-info/LICENSE_understanding.txt,sha256=pymi-yb_RvYM9p2ZA4iSNsImcvhDBBxlGuJCY9dTq7M,233
+tt_perf_report-1.0.6.dist-info/METADATA,sha256=NJ8q_4zz2URY9Tey2zcpv_YccFRuCGtWM-0BTnXLxVw,18345
+tt_perf_report-1.0.6.dist-info/WHEEL,sha256=jB7zZ3N9hIM9adW7qlTAyycLYW9npaWKLRzaoVcLKcM,91
+tt_perf_report-1.0.6.dist-info/entry_points.txt,sha256=ReAziglcjbAkPbklqheUISkfoEVI5ptlFrBAJTIk5dI,67
+tt_perf_report-1.0.6.dist-info/top_level.txt,sha256=mEQ-BK3rRbmz9QyWitTCLy2xwmC5rmJno_TY_H9s9CE,15
+tt_perf_report-1.0.6.dist-info/RECORD,,

tt_perf_report-1.0.4.dist-info/RECORD

@@ -1,8 +0,0 @@
-tt_perf_report/__init__.py,sha256=-j4iFYebIwgdS8uphk8-M6zasRqGBL3CQGnJH9keRuI,92
-tt_perf_report/perf_report.py,sha256=MJHgosbw7kD7MrJ6-ZtLhdny_iX9bR_UB-ZaHoyjTsE,36952
-tt_perf_report-1.0.4.dist-info/LICENSE,sha256=6dZGjPECz_ULS-sf40FLlt6OmQFcrRvmzG5mJRZCQ5I,11825
-tt_perf_report-1.0.4.dist-info/METADATA,sha256=fZKE2o9edCU0N6v-_tkUqSZ39j8mimeGBU_SIuuFi84,18305
-tt_perf_report-1.0.4.dist-info/WHEEL,sha256=jB7zZ3N9hIM9adW7qlTAyycLYW9npaWKLRzaoVcLKcM,91
-tt_perf_report-1.0.4.dist-info/entry_points.txt,sha256=ReAziglcjbAkPbklqheUISkfoEVI5ptlFrBAJTIk5dI,67
-tt_perf_report-1.0.4.dist-info/top_level.txt,sha256=mEQ-BK3rRbmz9QyWitTCLy2xwmC5rmJno_TY_H9s9CE,15
-tt_perf_report-1.0.4.dist-info/RECORD,,