tt-perf-report 1.0.1__tar.gz → 1.0.4__tar.gz

{tt_perf_report-1.0.1 → tt_perf_report-1.0.4}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: tt-perf-report
-Version: 1.0.1
+Version: 1.0.4
 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

{tt_perf_report-1.0.1 → tt_perf_report-1.0.4}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "tt-perf-report"
-version = "1.0.1"
+version = "1.0.4"
 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-1.0.1 → tt_perf_report-1.0.4}/src/tt_perf_report/perf_report.py

@@ -2,7 +2,7 @@
 # SPDX-License-Identifier: Apache-2.0
 
 # SPDX-FileCopyrightText: © 2025 Tenstorrent AI ULC
-
+import csv
 import sys
 import argparse
 import re
@@ -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
@@ -582,88 +634,98 @@ def print_op_to_op_gap_advice(rows, headers, col_widths):
         )
 
 
+def is_matmul_op(op_data):
+    return "Matmul" in op_data["OP Code"].raw_value
+
+
 def print_matmul_advice(rows, headers, col_widths):
-    matmul_ops = [op_data for op_data in rows if "Matmul" in op_data["OP Code"].raw_value]
+    matmul_ops = [op_data for op_data in rows if is_matmul_op(op_data)]
 
     if matmul_ops:
         print("Matmul Optimization\n-------------------")
         for op_data in matmul_ops:
             print_row(op_data, col_widths, headers)
-            advice = []
+            advice = generate_matmul_advice(op_data)
             color = "grey" if op_data["OP Code"].color == "grey" else "white"
 
-            math_fidelity = (
-                op_data["Math Fidelity"].raw_value.split()[0] if op_data["Math Fidelity"].raw_value else None
+            if advice:
+                for item in advice:
+                    print(colored(f"- {item}", color))
+            else:
+                print(colored("✅ Optimized", color))
+            print()  # Add a blank line between matmuls
+
+
+def generate_matmul_advice(op_data):
+    advice = []
+
+    math_fidelity = (
+        op_data["Math Fidelity"].raw_value.split()[0] if op_data["Math Fidelity"].raw_value else None
+    )
+    output_datatype = op_data["Output Datatype"].raw_value
+    input_0_datatype = op_data["Input 0 Datatype"].raw_value
+    input_1_datatype = op_data["Input 1 Datatype"].raw_value
+    cores = op_data["Cores"].raw_value
+    fidelity_evaluation, fidelity_advice = evaluate_fidelity(
+        input_0_datatype, input_1_datatype, output_datatype, math_fidelity
+    )
+
+    if op_data["Bound"].raw_value in ["DRAM", "BOTH"]:
+        if not op_data["DRAM Sharded"].raw_value:
+            advice.append(
+                "Try a DRAM-sharded program config (MatmulMultiCoreReuseMultiCastDRAMShardedProgramConfig) to improve throughput further"
             )
-            output_datatype = op_data["Output Datatype"].raw_value
-            input_0_datatype = op_data["Input 0 Datatype"].raw_value
-            input_1_datatype = op_data["Input 1 Datatype"].raw_value
-            cores = op_data["Cores"].raw_value
-            fidelity_evaluation, fidelity_advice = evaluate_fidelity(
-                input_0_datatype, input_1_datatype, output_datatype, math_fidelity
+        if fidelity_evaluation == "too_low" and op_data["FLOPs %"].raw_value < 40:
+            advice.append(fidelity_advice)
+        if fidelity_evaluation == "too_high":
+            advice.append(fidelity_advice)
+    elif op_data["Bound"].raw_value in ["FLOP", "BOTH"]:
+        if cores < 64:
+            advice.append(f"Increase grid size (currently using {cores})")
+        if fidelity_evaluation == "too_high":
+            advice.append(fidelity_advice)
+    elif op_data["Bound"].raw_value == "SLOW":
+        input_0_memory = op_data["Input 0 Memory"].raw_value
+        if input_0_memory and "L1" not in input_0_memory:
+            advice.append(f"If possible place input 0 in L1 (currently in {input_0_memory})")
+
+        inner_dim_block = op_data["Inner Dim Block Size"].raw_value
+        out_h = op_data["Output Subblock H"].raw_value
+        out_w = op_data["Output Subblock W"].raw_value
+
+        if inner_dim_block is None and out_h is None and out_w is None:
+            advice.append(
+                "No program_config specified, try using one to override in0_block_w and out_subblock_h/w"
             )
+        else:
+            all_good = True
+            if inner_dim_block is not None:
+                if inner_dim_block < 2:
+                    advice.append(f"in0_block_w={inner_dim_block} is small, try in0_block_w=2 or above")
+                    all_good = False
+            else:
+                advice.append("No inner dim block size found")
+                all_good = False
 
-            if op_data["Bound"].raw_value in ["DRAM", "BOTH"]:
-                if not op_data["DRAM Sharded"].raw_value:
-                    advice.append(
-                        "- Try a DRAM-sharded program config (MatmulMultiCoreReuseMultiCastDRAMShardedProgramConfig) to improve throughput further"
-                    )
-                if fidelity_evaluation == "too_low" and op_data["FLOPs %"].raw_value < 40:
-                    advice.append(f"- {fidelity_advice}")
-                if fidelity_evaluation == "too_high":
-                    advice.append(f"- {fidelity_advice}")
-            elif op_data["Bound"].raw_value in ["FLOP", "BOTH"]:
-                if cores < 64:
-                    advice.append(f"- Increase grid size (currently using {cores})")
-                if fidelity_evaluation == "too_high":
-                    advice.append(f"- {fidelity_advice}")
-            elif op_data["Bound"].raw_value == "SLOW":
-                input_0_memory = op_data["Input 0 Memory"].raw_value
-                if input_0_memory and "L1" not in input_0_memory:
-                    advice.append(f"- If possible place input 0 in L1 (currently in {input_0_memory})")
-
-                inner_dim_block = op_data["Inner Dim Block Size"].raw_value
-                out_h = op_data["Output Subblock H"].raw_value
-                out_w = op_data["Output Subblock W"].raw_value
-
-                if inner_dim_block is None and out_h is None and out_w is None:
+            if out_h is not None and out_w is not None:
+                out_area = out_h * out_w
+                if out_area < 2:
                     advice.append(
-                        "- No program_config specified, try using one to override in0_block_w and out_subblock_h/w"
+                        f"Output subblock {out_h}x{out_w} is small, try out_subblock_h * out_subblock_w >= 2 if possible"
                     )
-                else:
-                    all_good = True
-                    if inner_dim_block is not None:
-                        if inner_dim_block < 2:
-                            advice.append(f"- in0_block_w={inner_dim_block} is small, try in0_block_w=2 or above")
-                            all_good = False
-                    else:
-                        advice.append("- No inner dim block size found")
-                        all_good = False
-
-                    if out_h is not None and out_w is not None:
-                        out_area = out_h * out_w
-                        if out_area < 2:
-                            advice.append(
-                                f"- Output subblock {out_h}x{out_w} is small, try out_subblock_h * out_subblock_w >= 2 if possible"
-                            )
-                            all_good = False
-                    else:
-                        advice.append("- No output subblock size found")
-                        all_good = False
-
-                    if all_good:
-                        advice.append(
-                            f"- in0_block_w={inner_dim_block} and output subblock {out_h}x{out_w} look good 🤷"
-                        )
-                    if fidelity_advice:
-                        advice.append(f"- {fidelity_advice}")
-
-            if advice:
-                for item in advice:
-                    print(colored(item, color))
+                    all_good = False
             else:
-                print(colored("✅ Optimized", color))
-            print()  # Add a blank line between matmuls
+                advice.append("No output subblock size found")
+                all_good = False
+
+            if all_good:
+                advice.append(
+                    f"in0_block_w={inner_dim_block} and output subblock {out_h}x{out_w} look good 🤷"
+                )
+            if fidelity_advice:
+                advice.append(fidelity_advice)
+
+    return advice
 
 
 def merge_device_rows(df):
@@ -733,7 +795,7 @@ 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,
     )
 
 
@@ -755,6 +817,7 @@ 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
@@ -770,7 +833,7 @@ 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
@@ -798,6 +861,8 @@ 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
 
@@ -844,11 +909,20 @@ def generate_perf_report(csv_file, signpost, ignore_signposts, min_percentage, i
             "Output Subblock H",
             "Output Subblock W",
         ]
+        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:
-            f.write(",".join(all_headers) + "\n")
+            csv_writer = csv.DictWriter(f, fieldnames=all_headers)
+            csv_writer.writeheader()
             for op_data in rows:
-                f.write(",".join(str(op_data[header].raw_value) for header in all_headers) + "\n")
+                row = {header: op_data[header].raw_value for header in all_headers if header in op_data}
+                if not no_advice:
+                    advice = generate_matmul_advice(op_data) if is_matmul_op(op_data) else ""
+                    row["Advice"] = " • ".join(advice)
+                csv_writer.writerow(row)
     else:
         col_widths = [
             max(max(visible_length(str(row[header])) for row in rows), visible_length(header))

{tt_perf_report-1.0.1 → tt_perf_report-1.0.4}/src/tt_perf_report.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.2
 Name: tt-perf-report
-Version: 1.0.1
+Version: 1.0.4
 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