adding benchmarking

.gitignore

@@ -2,3 +2,5 @@
 __pycache__/
 *.py[cod]
 .pytest_cache
+**/.cache
+**/meta-llama/**/*

benchmarking/__init__.py

@@ -0,0 +1,2 @@
+from .profiler import Profiler
+from .benchmark_utils import compare_benchmarks

benchmarking/benchmark_utils.py

@@ -0,0 +1,24 @@
+from typing import Any, Dict
+import pandas as pd
+
+
+def compare_benchmarks(benchmarks: Dict[str, Dict[str, Any]]) -> Dict[str, Any]:
+    series_dict = {k: pd.Series(v.values()) for k, v in benchmarks.items()}
+    series_dict["kernel_path"] = pd.Series(benchmarks[list(benchmarks.keys())[0]].keys())
+    series_dict["kernel"] = pd.Series([k.split(".")[-1] for k in series_dict["kernel_path"]])
+    df = pd.DataFrame()
+
+    for k, v in series_dict.items():
+        df[k] = v
+    columns = [c for c in df.columns if not "kernel" in c]
+    for i in range(len(columns)):
+        for j in range(i+1, len(columns)):
+            # calculate the difference between the two columns
+            diff_col_name = f"{columns[i]}-{columns[j]}"
+            df[diff_col_name] = df[columns[i]] - df[columns[j]]
+    df.sort_values(by = 'kernel_path', inplace = True)
+    columns = [c for c in df.columns if not "kernel" in c]
+    columns = ["kernel", "kernel_path"] + columns
+    df = df[columns]
+    df.set_index("kernel", inplace=True)
+    return df

benchmarking/profiler.py

@@ -0,0 +1,65 @@
+import torch
+import contextlib
+import time
+from collections import defaultdict
+
+
+class Profiler:
+    _instance = None
+
+    def __new__(cls, should_profile: bool = False, benchmark: bool = False):
+        if cls._instance is None:
+            cls._instance = super().__new__(cls)
+            cls._instance.profiler = torch.profiler.profile(record_shapes=True, with_flops=True, profile_memory=True, with_stack=True, with_modules=True) if should_profile else None
+            cls._instance.benchmark = benchmark
+            cls._instance.benchmark_vals = defaultdict(list)
+            cls._instance.function_stack = []
+
+        return cls._instance
+
+    @classmethod
+    def reset(cls):
+        cls._instance = None
+
+    @classmethod
+    def profiling_decorator(cls, record_name: str = None, skip_profiling: bool = False, skip_benchmark: bool = False):
+        def inner(func):
+            def wrapper(*args, **kwargs):
+                if not cls._instance or (skip_profiling and skip_benchmark):
+                    return func(*args, **kwargs)
+                cls._instance.function_stack.append(record_name or func.__name__)
+                name = ".".join(cls._instance.function_stack)
+                if cls._instance.profiler and not skip_profiling:
+                    cls._instance.profiler.start()
+                start_time = time.perf_counter()
+
+                with torch.profiler.record_function(name):
+                    result = func(*args, **kwargs)
+
+                end_time = time.perf_counter()
+                if cls._instance.benchmark and not skip_benchmark:
+                    cls._instance.benchmark_vals[name].append(end_time - start_time)
+                if cls._instance.profiler and not skip_profiling:
+                    cls._instance.profiler.stop()
+                cls._instance.function_stack.pop()
+                return result
+            return wrapper
+        return inner
+
+    @classmethod
+    def step(cls):
+        if cls._instance and cls._instance.profiler:
+            cls._instance.profiler.step()
+
+    @classmethod
+    def get_benchmark_vals(cls):
+        if cls._instance and cls._instance.benchmark:
+            return {k: sum(v) / len(v) for k, v in cls._instance.benchmark_vals.items()}
+        return None
+
+    @classmethod
+    def get_profiling_data(cls):
+        if cls._instance and  cls._instance.profiler:
+            return self.profiler.key_averages()
+        return None
+

kernels/__init__.py

@@ -4,4 +4,12 @@
 from .flash_attention import attention
 from .matmul import _matmul, get_higher_dtype, matmul
 
-__all__ = ["blocksparse", "_cross_entropy", "cross_entropy", "_matmul", "matmul", "attention", "get_higher_dtype"]
+__all__ = [
+    "blocksparse",
+    "_cross_entropy",
+    "cross_entropy",
+    "_matmul",
+    "matmul",
+    "attention",
+    "get_higher_dtype",
+]