tt-perf-report 1.0.3__tar.gz → 1.0.5__tar.gz

tt_perf_report-1.0.5/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-1.0.3 → tt_perf_report-1.0.5}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: tt-perf-report
-Version: 1.0.3
+Version: 1.0.5
 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-1.0.3 → tt_perf_report-1.0.5}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "tt-perf-report"
-version = "1.0.3"
+version = "1.0.5"
 description = "This tool analyzes performance traces from TT-Metal operations, providing insights into throughput, bottlenecks, and optimization opportunities."
 license = {file = "LICENSE"}
 readme = "README.md"

{tt_perf_report-1.0.3 → tt_perf_report-1.0.5}/src/tt_perf_report/perf_report.py

@@ -38,6 +38,7 @@ def colored(text, color):
             "yellow": "\033[38;5;11m",
             "blue": "\033[38;5;12m",
             "magenta": "\033[38;5;13m",
+            "orange": "\033[38;5;208m",
             "cyan": "\033[38;5;14m",
             "white": "\033[38;5;15m",
             "end": "\033[0m",
@@ -70,7 +71,7 @@ class Cell:
         if self.raw_value is None or pd.isna(self.raw_value):
             return ""
 
-        if isinstance(self.raw_value, str) and "Matmul" in self.raw_value:
+        if isinstance(self.raw_value, str) and ("Matmul" in self.raw_value or "OptimizedConvNew" in self.raw_value):
             parts = self.raw_value.split(maxsplit=1)
             op_name = parts[0]
             size = parts[1] if len(parts) > 1 else ""
@@ -275,6 +276,37 @@ def analyze_matmul(row):
         core_count,  # Return the potentially adjusted core count
     )
 
+def analyze_conv(row):
+    duration_s = row["DEVICE KERNEL DURATION [ns]"] * 1e-9
+
+    core_count = 64 # we decided to normalize to the max core count
+    math_fidelity = row["MATH FIDELITY"]
+
+    # Check for DRAM-sharded program config
+    attributes = row["ATTRIBUTES"] if pd.notna(row["ATTRIBUTES"]) else ""
+
+    peak_flops_value = tflops_per_core(math_fidelity) * 1e12 * core_count
+
+    NHW = int(row["OUTPUT_0_Y"])
+    CH_IN = int(row["INPUT_0_X"])
+    W = [int(x) for x in (attributes.split("window_hw")[1].split("; ")[0][2:-1].split(";"))]
+    CH_OUT = int(row["INPUT_1_X"])
+
+    M, K, N = NHW, CH_IN * W[0] * W[1], CH_OUT
+    flops = (M * K * N * 2) / duration_s
+
+    size = f"{M} x {K} x {N}"
+    memory_info = f"({row['INPUT_0_DATATYPE']} {row['INPUT_0_MEMORY'].replace('DEV_0_', '')} @ {row['INPUT_1_DATATYPE']} {row['INPUT_1_MEMORY'].replace('DEV_0_', '')} => {row['OUTPUT_0_DATATYPE']} {row['OUTPUT_0_MEMORY'].replace('DEV_0_', '')})"
+
+    flops_percentage = (flops / peak_flops_value) * 100
+
+    return (
+        flops,
+        flops_percentage,
+        size,
+        memory_info,
+        math_fidelity
+    )
 
 def analyze_op(row, prev_row):
     op_code = Cell(row["OP CODE"])
@@ -305,6 +337,19 @@ def analyze_op(row, prev_row):
     input_1_datatype_cell = Cell(input_1_datatype)
     short_name = lambda n: {"BFLOAT16": "BF16", "BFLOAT8_B": "BFP8", "BFLOAT4_B": "BFP4"}.get(n, n)
 
+    dram_speed = Cell(None, unit="GB/s", decimals=0)
+    dram_percentage = Cell(None, unit="%", decimals=1)
+    flops = Cell(None, unit="TFLOPs", decimals=1)
+    flops_percentage = Cell(None, unit="%", decimals=1)
+
+    math_fidelity = ""
+    math_fidelity += f"{short_name(input_0_datatype)}" if pd.notna(input_0_datatype) else ""
+    math_fidelity += f", {short_name(input_1_datatype)}" if pd.notna(input_1_datatype) else ""
+    math_fidelity += f" => {short_name(output_datatype)}" if pd.notna(output_datatype) else ""
+    math_fidelity_cell = Cell(math_fidelity.strip())
+
+    is_dram_sharded = False
+
     if "Matmul" in op_code.raw_value:
         (
             dram_speed,
@@ -329,19 +374,24 @@ def analyze_op(row, prev_row):
             if math_fidelity
             else None
         )
-    else:
+    elif "OptimizedConvNew" in op_code.raw_value:
+        (
+            flops,
+            flops_percentage,
+            size,
+            memory_info,
+            math_fidelity,
+        ) = analyze_conv(row)
+        op_code = Cell(f"{op_code.raw_value} {size}")
         dram_speed = Cell(None, unit="GB/s", decimals=0)
         dram_percentage = Cell(None, unit="%", decimals=1)
-        flops = Cell(None, unit="TFLOPs", decimals=1)
-        flops_percentage = Cell(None, unit="%", decimals=1)
-
-        math_fidelity = ""
-        math_fidelity += f"{short_name(input_0_datatype)}" if pd.notna(input_0_datatype) else ""
-        math_fidelity += f", {short_name(input_1_datatype)}" if pd.notna(input_1_datatype) else ""
-        math_fidelity += f" => {short_name(output_datatype)}" if pd.notna(output_datatype) else ""
-        math_fidelity_cell = Cell(math_fidelity.strip())
-
-        is_dram_sharded = False
+        flops = Cell(flops / 1e12 if pd.notna(flops) else None, unit="TFLOPs", decimals=1)
+        flops_percentage = Cell(flops_percentage, unit="%", decimals=1)
+        math_fidelity_cell = Cell(
+            f"{math_fidelity} {short_name(input_0_datatype)} x {short_name(input_1_datatype)} => {short_name(output_datatype)}".strip()
+            if math_fidelity
+            else None
+        )
 
     output = {
         "ID": None,
@@ -434,6 +484,7 @@ def color_row(op_data, percentage, min_percentage):
         op_colors = {
             "(torch)": "red",
             "Matmul": "magenta",
+            "OptimizedConvNew" : "orange",
             "LayerNorm": "cyan",
             "AllGather": "cyan",
             "AllReduce": "cyan",
@@ -484,7 +535,8 @@ def color_row(op_data, percentage, min_percentage):
         if op_data["Op-to-Op Gap"].raw_value is not None and op_data["Op-to-Op Gap"].raw_value > 6.5:
             op_data["Op-to-Op Gap"].color = "red"
 
-        if "Matmul" in op_data["OP Code"].raw_value and op_data["Math Fidelity"].raw_value:
+        if ("Matmul" in op_data["OP Code"].raw_value 
+            or "OptimizedConvNew" in op_data["OP Code"].raw_value) and op_data["Math Fidelity"].raw_value:
             math_fidelity = op_data["Math Fidelity"].raw_value.split()[0]
             input_0_datatype = op_data["Input 0 Datatype"].raw_value
             input_1_datatype = op_data["Input 1 Datatype"].raw_value
@@ -690,8 +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
@@ -702,6 +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)
 
 
@@ -743,7 +817,7 @@ def filter_by_id_range(rows, id_range):
 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.csv_file, args.signpost, args.ignore_signposts, args.min_percentage, id_range, args.csv, args.no_advice, args.tracing_mode, args.raw_op_codes,
     )
 
 
@@ -765,6 +839,7 @@ def parse_args():
     parser.add_argument("--csv", type=str, help="Output filename for CSV format", metavar="OUTPUT_FILE")
     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")
     args = parser.parse_args()
 
     # Set the global color_output variable
@@ -780,7 +855,7 @@ 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):
+def generate_perf_report(csv_file, signpost, ignore_signposts, min_percentage, id_range, csv_output_file, no_advice, tracing_mode, raw_op_codes):
     df = pd.read_csv(csv_file, low_memory=False)
 
     # Add a column for original row numbers
@@ -808,6 +883,8 @@ def generate_perf_report(csv_file, signpost, ignore_signposts, min_percentage, i
     for _, row in df.iterrows():
         op_data, current_gap = analyze_op(row, prev_row)
         op_data["ID"] = Cell(row["ORIGINAL_ROW"])  # Use the original row number
+        if raw_op_codes:
+            op_data["Raw OP Code"] = Cell(row["OP CODE"])
         rows.append(op_data)
         prev_row = row
 
@@ -856,6 +933,8 @@ def generate_perf_report(csv_file, signpost, ignore_signposts, min_percentage, i
         ]
         if not no_advice:
             all_headers.append("Advice")
+        if raw_op_codes:
+            all_headers.append("Raw OP Code")
         print(colored(f"Writing CSV output to {csv_output_file}", "cyan"))
         with open(csv_output_file, "w") as f:
             csv_writer = csv.DictWriter(f, fieldnames=all_headers)

{tt_perf_report-1.0.3 → tt_perf_report-1.0.5}/src/tt_perf_report.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: tt-perf-report
-Version: 1.0.3
+Version: 1.0.5
 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-1.0.3 → tt_perf_report-1.0.5}/src/tt_perf_report.egg-info/SOURCES.txt

@@ -1,4 +1,5 @@
 LICENSE
+LICENSE_understanding.txt
 README.md
 pyproject.toml
 src/tt_perf_report/__init__.py