tt-perf-report 1.1.1__tar.gz → 1.1.3__tar.gz

{tt_perf_report-1.1.1/src/tt_perf_report.egg-info → tt_perf_report-1.1.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tt-perf-report
-Version: 1.1.1
+Version: 1.1.3
 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-1.1.1 → tt_perf_report-1.1.3}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "tt-perf-report"
-version = "1.1.1"
+version = "1.1.3"
 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-1.1.1 → tt_perf_report-1.1.3}/src/tt_perf_report/perf_report.py

@@ -4,10 +4,10 @@
 # SPDX-FileCopyrightText: © 2025 Tenstorrent AI ULC
 import argparse
 import csv
-from collections import defaultdict
 import os
 import re
 import sys
+from collections import defaultdict
 from typing import Any, Optional, Union
 
 import matplotlib.pyplot as plt
@@ -168,11 +168,32 @@ def pad_string(string, length, align="left"):
     return padding + string if align == "right" else string + padding
 
 
-def evaluate_fidelity(input_0_datatype, input_1_datatype, output_datatype, math_fidelity):
-    mantissa_bits = {"BFLOAT16": 8, "BFLOAT8_B": 7, "BFLOAT4_B": 3}
-    in0_bits = mantissa_bits[input_0_datatype]  # activations -> srcB (7 bits)
-    in1_bits = mantissa_bits[input_1_datatype]  # weights -> srcA (5 bits)
-    out_bits = mantissa_bits[output_datatype]
+def evaluate_fidelity(
+    input_0_datatype, input_1_datatype, output_datatype, math_fidelity
+):
+    integer_types = {"UINT8", "UINT16", "INT32", "UINT32"}
+
+    if (
+        input_0_datatype in integer_types
+        or input_1_datatype in integer_types
+        or output_datatype in integer_types
+    ):
+        return (
+            "not_applicable",
+            "Fidelity evaluation is not applicable for integer datatypes (UINT8, UINT16, INT32, UINT32).",
+        )
+
+    mantissa_bits = {"FLOAT32": 23, "BFLOAT16": 8, "BFLOAT8_B": 7, "BFLOAT4_B": 3}
+    try:
+        in0_bits = mantissa_bits[input_0_datatype]  # activations -> srcB (7 bits)
+        in1_bits = mantissa_bits[input_1_datatype]  # weights -> srcA (5 bits)
+        out_bits = mantissa_bits[output_datatype]
+    except KeyError as e:
+        return (
+            "unknown",
+            f"Datatype {e.args[0]} is not supported for fidelity evaluation.",
+        )
+
     if in0_bits == 8 and out_bits >= 7:
         if math_fidelity == "HiFi4":
             return (
@@ -435,7 +456,12 @@ def analyze_op(row, prev_row, csv_format="v2"):
     output_datatype_cell = Cell(output_datatype)
     input_0_datatype_cell = Cell(input_0_datatype)
     input_1_datatype_cell = Cell(input_1_datatype)
-    short_name = lambda n: {"BFLOAT16": "BF16", "BFLOAT8_B": "BFP8", "BFLOAT4_B": "BFP4"}.get(n, n)
+    short_name = lambda n: {
+        "FLOAT32": "FP32",
+        "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)
@@ -1173,6 +1199,10 @@ def generate_perf_report(csv_file, signpost, ignore_signposts, min_percentage,
                     row["Advice"] = " • ".join(advice)
                 csv_writer.writerow(row)
     else:
+        if not rows:
+            print(colored("No operations to display after applying filters.", "yellow"))
+            return
+
         col_widths = [
             max(max(visible_length(str(row[header])) for row in rows), visible_length(header))
             for header in visible_headers
@@ -1182,7 +1212,7 @@ def generate_perf_report(csv_file, signpost, ignore_signposts, min_percentage,
             print_advice_section(rows, visible_headers, col_widths)
 
     # handle stacked report generation
-    if not(no_stacked_report):
+    if not(no_stacked_report) and rows:
         stacked_report = generate_stacked_report(rows, visible_headers, not(no_stack_by_in0))
 
         if not(csv_output_file):

{tt_perf_report-1.1.1 → tt_perf_report-1.1.3/src/tt_perf_report.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: tt-perf-report
-Version: 1.1.1
+Version: 1.1.3
 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