add cpu objdump to LLVM/CLANG (#4537)

test/external/external_benchmark_schedule.py

@@ -19,5 +19,15 @@ if __name__ == "__main__":
 
   with Profiling(PROFILE):
     with Timing("***** model lower in "):
-      ei = list(lower_schedule(sched))
+      eis = list(lower_schedule(sched))
+
+  # random makes this slow
+  #with Profiling(PROFILE):
+  #  with Timing("***** model run in "):
+  #    for ei in eis: ei.run()
+
+  # this is all wait
+  #with Profiling(PROFILE):
+  #  with Timing("***** model finish in "):
+  #    out.data()
 

tinygrad/helpers.py

@@ -1,6 +1,6 @@
 from __future__ import annotations
 import os, functools, platform, time, re, contextlib, operator, hashlib, pickle, sqlite3, cProfile, pstats, tempfile, pathlib, string, ctypes
-import itertools, urllib.request
+import itertools, urllib.request, subprocess
 from tqdm import tqdm
 from typing import Dict, Tuple, Union, List, ClassVar, Optional, Iterable, Any, TypeVar, TYPE_CHECKING, Callable, Sequence
 if TYPE_CHECKING:  # TODO: remove this and import TypeGuard from typing once minimum python supported version is 3.10
@@ -220,6 +220,11 @@ def cpu_time_execution(cb, enable):
   cb()
   if enable: return time.perf_counter()-st
 
+def cpu_objdump(lib):
+  with tempfile.NamedTemporaryFile(delete=True) as f:
+    pathlib.Path(f.name).write_bytes(lib)
+    print(subprocess.check_output(['objdump', '-d', f.name]).decode('utf-8'))
+
 # *** ctypes helpers
 
 # TODO: make this work with read only memoryviews (if possible)

tinygrad/runtime/ops_clang.py

@@ -1,6 +1,6 @@
 import ctypes, subprocess, pathlib, tempfile
 from tinygrad.device import Compiled, Compiler, MallocAllocator
-from tinygrad.helpers import cpu_time_execution
+from tinygrad.helpers import cpu_time_execution, DEBUG, cpu_objdump
 from tinygrad.renderer.cstyle import ClangRenderer
 
 class ClangCompiler(Compiler):
@@ -13,6 +13,7 @@ class ClangCompiler(Compiler):
 
 class ClangProgram:
   def __init__(self, name:str, lib:bytes):
+    if DEBUG >= 6: cpu_objdump(lib)
     self.name, self.lib = name, lib
     # write to disk so we can load it
     with tempfile.NamedTemporaryFile(delete=True) as cached_file_path:

tinygrad/runtime/ops_llvm.py

@@ -2,7 +2,7 @@ from __future__ import annotations
 import ctypes, functools
 from typing import Tuple
 from tinygrad.device import Compiled, Compiler, MallocAllocator
-from tinygrad.helpers import DEBUG, cpu_time_execution
+from tinygrad.helpers import DEBUG, cpu_time_execution, cpu_objdump
 from tinygrad.renderer.llvmir import LLVMRenderer
 import llvmlite.binding as llvm
 
@@ -19,6 +19,7 @@ class LLVMCompiler(Compiler):
 
 class LLVMProgram:
   def __init__(self, device:LLVMDevice, name:str, lib:bytes):
+    if DEBUG >= 6: cpu_objdump(lib)
     self.name, self.lib = name, lib
     device.engine.add_object_file(llvm.object_file.ObjectFileRef.from_data(lib))
     self.fxn = device.engine.get_function_address(name)