[Pytorch] Add python binding to use mobile cpu allocator. (#52323)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/52323

Using default cpu allocator for ops executed on qnnpack backend will result in
asan failures with heap overflow since qnnpack (and xnnpack) can access input
beyond their and/beginning.

Here we are enabling this feature specifically to enable dynamic sparse linear op test
using qnnpack engine. In dynamic linear op, the fp32 bias is not packed and
hence can result in out-of-bound access.

Test Plan: test_set_default_mobile_cpu_allocator.py

Reviewed By: z-a-f

Differential Revision: D26263481

fbshipit-source-id: a49227cac7e6781b0db4a156ca734d7671972d9f

aten/src/ATen/Context.cpp

@@ -267,4 +267,21 @@ void Context::setDisplayVmapFallbackWarnings(bool enabled) {
   display_vmap_fallback_warnings_ = enabled;
 }
 
+void Context::setDefaultMobileCPUAllocator() {
+  TORCH_CHECK(prev_allocator_ptr_ == nullptr,
+      "Already within the scope of another non-default cpu allocator."
+      "Cannot set another allocator.");
+  // Setting the priority high to make sure no other allocator gets used instead of this.
+  prev_allocator_ptr_ = c10::GetCPUAllocator();
+  c10::SetCPUAllocator(c10::GetDefaultMobileCPUAllocator(), /*priority*/ 100);
+}
+
+void Context::unsetDefaultMobileCPUAllocator() {
+  TORCH_CHECK(prev_allocator_ptr_ != nullptr,
+      "setDefaultMobileCPUAllocator must have been called "
+      "before unsetDefaultMobileCPUAllocator.");
+  // Setting the priority high to make sure no other allocator gets used instead of this.
+  c10::SetCPUAllocator(prev_allocator_ptr_ , /*priority*/ 100);
+  prev_allocator_ptr_ = nullptr;
+}
 } // namespace at

aten/src/ATen/Context.h

@@ -199,6 +199,9 @@ class TORCH_API Context {
   void setDisplayVmapFallbackWarnings(bool enabled);
   bool areVmapFallbackWarningsEnabled() const;
 
+  void setDefaultMobileCPUAllocator();
+  void unsetDefaultMobileCPUAllocator();
+
  private:
   void initCUDAIfNeeded(DeviceType p) {
     if (p == DeviceType::CUDA) {
@@ -229,6 +232,8 @@ class TORCH_API Context {
   c10::optional<at::QEngine> quantized_engine = c10::nullopt;
   std::unique_ptr<THCState, void(*)(THCState*)> thc_state;
   std::unique_ptr<THHState, void(*)(THHState*)> thh_state;
+
+  Allocator* prev_allocator_ptr_{nullptr};
 };
 
 TORCH_API Context& globalContext();

test/run_test.py

@@ -60,6 +60,7 @@ TESTS = [
     'test_optim',
     'test_pytree',
     'test_mobile_optimizer',
+    'test_set_default_mobile_cpu_allocator',
     'test_xnnpack_integration',
     'test_vulkan',
     'test_sparse',

test/test_set_default_mobile_cpu_allocator.py

@@ -0,0 +1,27 @@
+import torch
+from torch.testing._internal.common_utils import TestCase, run_tests
+
+class TestSetDefaultMobileCPUAllocator(TestCase):
+    def test_no_exception(self):
+        torch._C._set_default_mobile_cpu_allocator()
+        torch._C._unset_default_mobile_cpu_allocator()
+
+    def test_exception(self):
+        with self.assertRaises(Exception):
+            torch._C._unset_default_mobile_cpu_allocator()
+
+        with self.assertRaises(Exception):
+            torch._C._set_default_mobile_cpu_allocator()
+            torch._C._set_default_mobile_cpu_allocator()
+
+        # Must reset to good state
+        # For next test.
+        torch._C._unset_default_mobile_cpu_allocator()
+
+        with self.assertRaises(Exception):
+            torch._C._set_default_mobile_cpu_allocator()
+            torch._C._unset_default_mobile_cpu_allocator()
+            torch._C._unset_default_mobile_cpu_allocator()
+
+if __name__ == '__main__':
+    run_tests()

torch/csrc/Module.cpp

@@ -589,6 +589,26 @@ PyObject *THPModule_isEnabledXNNPACK(PyObject *_unused, PyObject *noargs)
   else Py_RETURN_FALSE;
 }
 
+PyObject *THPModule_setDefaultMobileCPUAllocator(PyObject *_unused, PyObject *noargs)
+{
+  try {
+    at::globalContext().setDefaultMobileCPUAllocator();
+  } catch (c10::Error& e) {
+    THPUtils_setError(e.what());
+  }
+  Py_RETURN_NONE;
+}
+
+PyObject *THPModule_unsetDefaultMobileCPUAllocator(PyObject *_unused, PyObject *noargs)
+{
+  try {
+    at::globalContext().unsetDefaultMobileCPUAllocator();
+  } catch (c10::Error& e) {
+    THPUtils_setError(e.what());
+  }
+  Py_RETURN_NONE;
+}
+
 static PyObject * THPModule_vmapmode_increment_nesting(PyObject* _unused, PyObject *arg) {
   HANDLE_TH_ERRORS
   return THPUtils_packInt64(at::impl::VmapMode::increment_nesting());
@@ -673,6 +693,8 @@ static PyMethodDef TorchMethods[] = {
   {"_set_qengine", THPModule_setQEngine, METH_O, nullptr},
   {"_supported_qengines", THPModule_supportedQEngines, METH_NOARGS, nullptr},
   {"_is_xnnpack_enabled", THPModule_isEnabledXNNPACK, METH_NOARGS, nullptr},
+  {"_set_default_mobile_cpu_allocator", THPModule_setDefaultMobileCPUAllocator, METH_NOARGS, nullptr},
+  {"_unset_default_mobile_cpu_allocator", THPModule_unsetDefaultMobileCPUAllocator, METH_NOARGS, nullptr},
   {"_is_torch_function_enabled", THPModule_isEnabledTorchFunction, METH_NOARGS, nullptr},
   {"_disabled_torch_function_impl", THPModule_disable_torch_function, METH_VARARGS, nullptr},
   {"_has_torch_function", THPModule_has_torch_function, METH_O, nullptr},