Revert D24010742: [pytorch][PR] Add callgrind collection to Timer

Test Plan: revert-hammer

Differential Revision:
D24010742 (pytorch@9b27e09)

Original commit changeset: df6bc765f8ef

fbshipit-source-id: 4c1edd57ea932896f7052716427059c924222501

.circleci/docker/common/install_base.sh

@@ -47,7 +47,6 @@ install_ubuntu() {
     software-properties-common \
     sudo \
     wget \
-    valgrind \
     vim
 
   # TODO: THIS IS A HACK!!!
@@ -93,7 +92,6 @@ install_centos() {
     opencv-devel \
     sudo \
     wget \
-    valgrind \
     vim
 
   # Cleanup
@@ -133,3 +131,4 @@ sudo make install
 cd ../../
 rm -rf valgrind_build
 alias valgrind="/usr/local/bin/valgrind"
+

.circleci/scripts/binary_ios_build.sh

@@ -16,7 +16,6 @@ source ~/anaconda/bin/activate
 
 # Install dependencies
 conda install numpy ninja pyyaml mkl mkl-include setuptools cmake cffi typing requests --yes
-conda install -c conda-forge valgrind
 export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
 
 # sync submodules

.gitmodules

@@ -130,7 +130,3 @@
     ignore = dirty
     path = third_party/tensorpipe
     url = https://github.com/pytorch/tensorpipe.git
-[submodule "third_party/valgrind"]
-    ignore = dirty
-	path = third_party/valgrind
-	url = https://sourceware.org/git/valgrind.git

test/test_utils.py

@@ -15,7 +15,7 @@
 import torch.hub as hub
 from torch.autograd._functions.utils import check_onnx_broadcast
 from torch.onnx.symbolic_opset9 import _prepare_onnx_paddings
-from torch.testing._internal.common_utils import load_tests, retry, IS_SANDCASTLE, IS_WINDOWS, slowTest
+from torch.testing._internal.common_utils import load_tests, retry, IS_SANDCASTLE, IS_WINDOWS
 from urllib.error import URLError
 import numpy as np
 
@@ -768,16 +768,6 @@ class MockCudaTimer(benchmark_utils.Timer):
             self.assertEqual(len(measurement.times), repeats)
             self.assertEqual(measurement.number_per_run, number_per_run)
 
-    @slowTest
-    @unittest.skipIf(IS_WINDOWS, "Valgrind is not supported on Windows.")
-    def test_collect_callgrind(self):
-        timer = benchmark_utils.Timer("y = torch.ones((1,)) + 1")
-
-        # Don't collect baseline to speed up unit test by ~30 seconds.
-        stats = timer.collect_callgrind(number=1000, collect_baseline=False)
-
-        self.assertEqual(stats.counts(include_lookdict_unicode=False), 38803198, atol=0, rtol=0.0001)
-
     def test_compare(self):
         # Simulate several approaches.
         costs = (

torch/CMakeLists.txt

@@ -66,9 +66,6 @@ set(TORCH_PYTHON_INCLUDE_DIRECTORIES
     ${CMAKE_BINARY_DIR}/third_party
     ${CMAKE_BINARY_DIR}/third_party/onnx
 
-    ${TORCH_ROOT}/third_party/valgrind/callgrind
-    ${TORCH_ROOT}/third_party/valgrind/include
-
     ${TORCH_ROOT}/third_party/gloo
     ${TORCH_ROOT}/third_party/onnx
     ${pybind11_INCLUDE_DIRS}

torch/_C/__init__.pyi.in

@@ -388,10 +388,6 @@ def _vmapmode_increment_nesting() -> _int: ...  # THPModule_vmapmode_increment_n
 def _vmapmode_decrement_nesting() -> _int: ...  # THPModule_vmapmode_decrement_nesting
 def _log_api_usage_once(str) -> None: ...  # LogAPIUsageOnceFromPython
 
-# Defined in `valgrind.h` and `callgrind.h` respecitively.
-def valgrind_supported_platform() -> _bool: ...  # NVALGRIND
-def valgrind_toggle() -> None: ...  # CALLGRIND_TOGGLE_COLLECT
-
 has_openmp: _bool
 has_mkl: _bool
 has_lapack: _bool

torch/csrc/Module.cpp

@@ -61,12 +61,6 @@
 #endif
 #endif
 
-#if (defined(_WIN32) || defined(_WIN64) || defined(FBCODE_CAFFE2) || defined(C10_MOBILE))
-#define NVALGRIND
-#else
-#include <callgrind.h>
-#endif
-
 #define WITH_NUMPY_IMPORT_ARRAY
 #include <torch/csrc/utils/numpy_stub.h>
 
@@ -827,26 +821,6 @@ Call this whenever a new thread is created in order to propagate values from
   ASSERT_TRUE(set_module_attr("has_mkl", at::hasMKL() ? Py_True : Py_False));
   ASSERT_TRUE(set_module_attr("has_lapack", at::hasLAPACK() ? Py_True : Py_False));
 
-  py_module.def(
-    "valgrind_supported_platform", [](){
-      #if defined(NVALGRIND)
-      return false;
-      #else
-      return true;
-      #endif
-    }
-  );
-
-  py_module.def(
-    "valgrind_toggle", [](){
-      #if defined(NVALGRIND)
-      TORCH_CHECK(false, "Valgrind is not supported.");
-      #else
-      CALLGRIND_TOGGLE_COLLECT;
-      #endif
-    }
-  );
-
 #ifdef USE_CUDA
   PyObject *has_cuda = Py_True;
 #else

torch/overrides.py

@@ -156,8 +156,6 @@ def get_ignored_functions() -> Set[Callable]:
         torch.is_deterministic,
         torch.set_deterministic,
         torch.unify_type_list,
-        torch.valgrind_supported_platform,
-        torch.valgrind_toggle,
         Tensor.__delitem__,
         Tensor.__dir__,
         Tensor.__getattribute__,

torch/utils/benchmark/utils/timer.py

@@ -6,7 +6,6 @@
 import numpy as np
 import torch
 from torch.utils.benchmark.utils import common
-from torch.utils.benchmark.utils.valgrind_wrapper import timer_interface as valgrind_timer_interface
 
 
 __all__ = ["Timer", "timer"]
@@ -43,7 +42,6 @@ def __init__(
         # specified as a convenience feature.
         globals = dict(globals or {})
         globals.setdefault("torch", torch)
-        self._globals = globals
 
         self._timer = self._timer_cls(stmt=stmt, setup=setup, timer=timer, globals=globals)
         self._task_spec = common.TaskSpec(
@@ -161,28 +159,3 @@ def stop_hook(times) -> bool:
             raw_times=times,
             task_spec=self._task_spec
         )
-
-    def collect_callgrind(self, number=100, collect_baseline=True):
-        if not isinstance(self._task_spec.stmt, str):
-            raise ValueError("`collect_callgrind` currently only supports string `stmt`")
-
-        # __init__ adds torch, and Timer adds __builtins__
-        allowed_keys = {"torch", "__builtins__"}
-        if any(k not in allowed_keys for k in self._globals.keys()):
-            raise ValueError(
-                "`collect_callgrind` does not currently support passing globals. "
-                "Please define a `setup` str instead.")
-
-        if self._globals.get("torch", torch) is not torch:
-            raise ValueError("`collect_callgrind` does not support mocking out `torch`.")
-
-        # Check that the statement is valid. It doesn't guarantee success, but it's much
-        # simpler and quicker to raise an exception for a faulty `stmt` or `setup` in
-        # the parent process rather than the valgrind subprocess.
-        self._timer.timeit(1)
-        return valgrind_timer_interface.wrapper_singleton().collect_callgrind(
-            stmt=self._task_spec.stmt,
-            setup=self._task_spec.setup,
-            number=number,
-            num_threads=self._task_spec.num_threads,
-            collect_baseline=collect_baseline)