Add data formats to perf report (#15170)

Add data formats to perf report; added to math_fidelity column which was already there and previously showing data formats for only matmuls.
tenstorrent · Nov 20, 2024 · 145050a · 145050a
1 parent be20580
commit 145050a
Showing 1 changed file with 15 additions and 19 deletions.
diff --git a/models/perf/perf_report.py b/models/perf/perf_report.py
@@ -263,10 +263,6 @@ def analyze_matmul(row):
     dram_percentage = (dram_speed_gb_s / 288) * 100 if dram_speed_gb_s is not None else None
     flops_percentage = (flops / peak_flops_value) * 100
 
-    input_0_datatype = row["INPUT_0_DATATYPE"]
-    input_1_datatype = row["INPUT_1_DATATYPE"]
-    output_datatype = row["OUTPUT_0_DATATYPE"]
-
     return (
         dram_speed_gb_s,
         dram_percentage,
@@ -275,10 +271,7 @@ def analyze_matmul(row):
         size,
         memory_info,
         math_fidelity,
-        output_datatype,
         is_dram_sharded,
-        input_0_datatype,
-        input_1_datatype,
         core_count,  # Return the potentially adjusted core count
     )
 
@@ -301,6 +294,14 @@ def analyze_op(row, prev_row):
     else:
         dispatch_time = Cell(None, unit="us", decimals=0)
 
+    output_datatype = row["OUTPUT_0_DATATYPE"]
+    input_0_datatype = row["INPUT_0_DATATYPE"]
+    input_1_datatype = row["INPUT_1_DATATYPE"]
+    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)
+
     if "Matmul" in op_code.raw_value:
         (
             dram_speed,
@@ -310,10 +311,7 @@ def analyze_op(row, prev_row):
             size,
             memory_info,
             math_fidelity,
-            output_datatype,
             is_dram_sharded,
-            input_0_datatype,
-            input_1_datatype,
             adjusted_core_count,  # Get the potentially adjusted core count
         ) = analyze_matmul(row)
         op_code = Cell(f"{op_code.raw_value} {size}")
@@ -322,26 +320,24 @@ def analyze_op(row, prev_row):
         flops = Cell(flops / 1e12 if pd.notna(flops) else None, unit="TFLOPs", decimals=1)
         flops_percentage = Cell(flops_percentage, unit="%", decimals=1)
 
-        short_name = lambda n: {"BFLOAT16": "BF16", "BFLOAT8_B": "BFP8", "BFLOAT4_B": "BFP4"}.get(n, n)
         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_datatype_cell = Cell(output_datatype)
-        input_0_datatype_cell = Cell(input_0_datatype)
-        input_1_datatype_cell = Cell(input_1_datatype)
     else:
         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, output_datatype = "", None
-        math_fidelity_cell = Cell(None)
-        output_datatype_cell = Cell(None)
+
+        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
-        input_0_datatype_cell = Cell(None)
-        input_1_datatype_cell = Cell(None)
 
     output = {
         "ID": None,