tt-perf-report 1.0.5__tar.gz → 1.0.7__tar.gz

{tt_perf_report-1.0.5/src/tt_perf_report.egg-info → tt_perf_report-1.0.7}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.2
+Metadata-Version: 2.4
 Name: tt-perf-report
-Version: 1.0.5
+Version: 1.0.7
 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
@@ -218,6 +218,7 @@ Description-Content-Type: text/markdown
 License-File: LICENSE
 License-File: LICENSE_understanding.txt
 Requires-Dist: pandas
+Dynamic: license-file
 
 # Performance Report Analysis Tool
 

{tt_perf_report-1.0.5 → tt_perf_report-1.0.7}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "tt-perf-report"
-version = "1.0.5"
+version = "1.0.7"
 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-1.0.5 → tt_perf_report-1.0.7}/src/tt_perf_report/perf_report.py

@@ -71,7 +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 or "OptimizedConvNew" in self.raw_value):
+        if isinstance(self.raw_value, str) and ("Matmul" in self.raw_value or "OptimizedConvNew" in self.raw_value or "HaloDeviceOperation" in self.raw_value):
             parts = self.raw_value.split(maxsplit=1)
             op_name = parts[0]
             size = parts[1] if len(parts) > 1 else ""
@@ -276,6 +276,42 @@ def analyze_matmul(row):
         core_count,  # Return the potentially adjusted core count
     )
 
+def analyze_halo(row):
+    attributes = row["ATTRIBUTES"] if pd.notna(row["ATTRIBUTES"]) else ""
+
+    try:
+        window_hw = attributes.split("window_hw=")[1].split(";")[0:2]
+        window_hw = ",".join(window_hw[0:2])
+    except (IndexError, AttributeError):
+        window_hw = "x"
+
+    try:
+        stride_hw = attributes.split("stride_hw=")[1].split(";")[0:2]
+        stride_hw = ",".join(stride_hw[0:2])
+    except (IndexError, AttributeError):
+        stride_hw = "x"
+
+    try:
+        pad_hw = attributes.split("padding=")[1].split(";")[0:4]
+        pad_hw = ",".join(pad_hw[0:4])
+    except (IndexError, AttributeError):
+        pad_hw = "x"
+
+    try:
+        dilation_hw = attributes.split("dilation_hw=")[1].split(";")[0:2]
+        dilation_hw = ",".join(dilation_hw[0:2])
+    except (IndexError, AttributeError):
+        dilation_hw = "x"
+
+    try:
+        memory_layout = attributes.split("memory_layout=")[1].split(";")[0].split("::")[1]
+    except (IndexError, AttributeError):
+        memory_layout = "x"
+
+    config = f"w={window_hw} s={stride_hw} p={pad_hw} d={dilation_hw} | {memory_layout}"
+
+    return config
+
 def analyze_conv(row):
     duration_s = row["DEVICE KERNEL DURATION [ns]"] * 1e-9
 
@@ -300,12 +336,40 @@ def analyze_conv(row):
 
     flops_percentage = (flops / peak_flops_value) * 100
 
+    try:
+        act_block_h_ntiles = int(attributes.split("act_block_h_ntiles")[1][1:].split(";")[0])
+    except (IndexError, ValueError):
+        act_block_h_ntiles = "x"
+
+    try:
+        enable_act_double_buffer = "true" == attributes.split("enable_act_double_buffer': '")[1].split("'")[0]
+    except (IndexError, ValueError):
+        enable_act_double_buffer = "x"
+
+    try:
+        enable_split_reader = "true" == attributes.split("enable_split_reader': '")[1].split("'")[0]
+    except (IndexError, ValueError):
+        enable_split_reader = "x"
+
+    try:
+        per_core_out_matrix_height_ntile = int(attributes.split("per_core_out_matrix_height_ntile")[1][1:].split(";")[0])
+    except (IndexError, ValueError):
+        per_core_out_matrix_height_ntile = "x"
+
+    config = f"[ABH={per_core_out_matrix_height_ntile}|{act_block_h_ntiles}"
+    if (enable_act_double_buffer):
+        config += " ADB"
+    if (enable_split_reader):
+        config += " SR"
+    config += "]"
+
     return (
         flops,
         flops_percentage,
         size,
         memory_info,
-        math_fidelity
+        math_fidelity,
+        config,
     )
 
 def analyze_op(row, prev_row):
@@ -381,8 +445,9 @@ def analyze_op(row, prev_row):
             size,
             memory_info,
             math_fidelity,
+            config,
         ) = analyze_conv(row)
-        op_code = Cell(f"{op_code.raw_value} {size}")
+        op_code = Cell(f"{op_code.raw_value} {size} {config}")
         dram_speed = Cell(None, unit="GB/s", decimals=0)
         dram_percentage = Cell(None, unit="%", decimals=1)
         flops = Cell(flops / 1e12 if pd.notna(flops) else None, unit="TFLOPs", decimals=1)
@@ -392,6 +457,13 @@ def analyze_op(row, prev_row):
             if math_fidelity
             else None
         )
+    elif "HaloDeviceOperation" in op_code.raw_value:
+        config = analyze_halo(row)
+        op_code = Cell(f"{op_code.raw_value} {config}")
+        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)
 
     output = {
         "ID": None,
@@ -765,10 +837,17 @@ def merge_device_rows(df):
         if not blocks:
             break
 
-        if "AllGather" in op_name or "ReduceScatter" in op_name:
-            # For collective ops, take the row with minimum duration
-            min_duration_block = min(blocks, key=lambda x: x[1]["DEVICE KERNEL DURATION [ns]"])
-            merged_blocks.append(min_duration_block[1])
+        if "AllGather" in op_name or "ReduceScatter" in op_name or "AllReduce" in op_name:
+            # For collective ops, take the average duration over all rows within a block
+            device_kernel_durations = [d["DEVICE KERNEL DURATION [ns]"] 
+                             for _, d in blocks 
+                             if pd.notna(d["DEVICE KERNEL DURATION [ns]"])]
+            # Use the first block's data but update its duration with the average
+            base_block = blocks[0][1].copy()
+            base_block["DEVICE KERNEL DURATION [ns]"] = (
+                sum(device_kernel_durations) / len(device_kernel_durations) if device_kernel_durations else float("nan")
+            )
+            merged_blocks.append(base_block)
         else:
             # For non-collective ops, take the row with maximum duration
             max_duration_block = max(blocks, key=lambda x: x[1]["DEVICE KERNEL DURATION [ns]"])
@@ -814,11 +893,15 @@ 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():
@@ -840,6 +923,8 @@ 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
@@ -855,7 +940,9 @@ 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
@@ -883,13 +970,14 @@ 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
+        op_data["Global Call Count"] = Cell(row["GLOBAL CALL COUNT"])
         if raw_op_codes:
             op_data["Raw OP Code"] = Cell(row["OP CODE"])
         rows.append(op_data)
         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
@@ -900,6 +988,9 @@ 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)
 
@@ -930,6 +1021,7 @@ def generate_perf_report(csv_file, signpost, ignore_signposts, min_percentage, i
             "Inner Dim Block Size",
             "Output Subblock H",
             "Output Subblock W",
+            "Global Call Count",
         ]
         if not no_advice:
             all_headers.append("Advice")
@@ -955,5 +1047,9 @@ 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-1.0.5 → tt_perf_report-1.0.7/src/tt_perf_report.egg-info}/PKG-INFO

@@ -1,6 +1,6 @@
-Metadata-Version: 2.2
+Metadata-Version: 2.4
 Name: tt-perf-report
-Version: 1.0.5
+Version: 1.0.7
 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
@@ -218,6 +218,7 @@ Description-Content-Type: text/markdown
 License-File: LICENSE
 License-File: LICENSE_understanding.txt
 Requires-Dist: pandas
+Dynamic: license-file
 
 # Performance Report Analysis Tool