Use TensorMethods.cpp (#37639)

Ilia Cherniavskii · facebook-github-bot · commit 68e62b9ab643 · 2020-05-31T17:11:12.000-07:00
Summary: Pull Request resolved: #37639 Changing TensorMethods.h to .cpp Necessary to avoid incomplete types in dispatcher Test Plan: CI Imported from OSS checked mobile size, no change, small reduction in size in fbios fbios: Succeeded Change in Download Size for arm64 + 3x assets variation: -18.2 KiB Change in Uncompressed Size for arm64 + 3x assets variation: -8.8 KiB reran benchmark, no stat. significant difference buck run mode/opt caffe2/caffe2/fb/high_perf_models/pytorch/benchmark_framework_overheads:benchmark_torchscript_model -- --model_file caffe2/caffe2/fb/high_perf_models/pytorch/benchmark_framework_overheads/addmodule.pt --num_runs 3 ╷ @ 68592d0d 41 minutes ago iliacher D21374247 ╭─╯ Use TensorMethods.cpp Created 3 benchmark runs on aibench for caffe2/caffe2/fb/high_perf_models/pytorch/benchmark_framework_overheads/addmodule.pt. Links to the results: * Adhoc run: https://our.intern.facebook.com/intern/aibench/details/1729113760 * Adhoc run: https://our.intern.facebook.com/intern/aibench/details/3867976782 * Adhoc run: https://our.intern.facebook.com/intern/aibench/details/2782186766 hg prev @ 7f501b42 Thursday at 14:26 bvaughan D21764704 ╷ short-circuit pow for complex 1 and 0 exponents Created 3 benchmark runs on aibench for caffe2/caffe2/fb/high_perf_models/pytorch/benchmark_framework_overheads/addmodule.pt. Links to the results: * Adhoc run: https://our.intern.facebook.com/intern/aibench/details/2155256332 * Adhoc run: https://our.intern.facebook.com/intern/aibench/details/1802057074 * Adhoc run: https://our.intern.facebook.com/intern/aibench/details/4119590830 Differential Revision: D21374247 fbshipit-source-id: 076964415079cf84fb57f1f7b43d087afed86e1d
diff --git a/BUILD.bazel b/BUILD.bazel
@@ -140,7 +140,7 @@ genrule(
         "aten/src/ATen/TypeDefault.h",
         "aten/src/ATen/TypeDefault.cpp",
         "aten/src/ATen/core/TensorBody.h",
-        "aten/src/ATen/core/TensorMethods.h",
+        "aten/src/ATen/core/TensorMethods.cpp",
         "aten/src/ATen/core/ATenOpList.cpp",
     ],
     cmd = "$(location :gen) --source-path aten/src/ATen --install_dir `dirname $(location aten/src/ATen/Declarations.yaml)` aten/src/ATen/Declarations.cwrap aten/src/THCUNN/generic/THCUNN.h aten/src/ATen/nn.yaml aten/src/ATen/native/native_functions.yaml",
diff --git a/aten/src/ATen/Tensor.h b/aten/src/ATen/Tensor.h
@@ -1,12 +1,3 @@
 #pragma once
 
-/*
- * We split Tensor.h into TensorBody.h and TensorMethods.h because we want
- * all TensorMethods to be inlined, but they depend on the Dispatcher,
- * which in turn depends on many other things, which then depend back on Tensor.
- *
- * We can break this dependency chain by having the dispatcher only depend on
- * TensorBody.h and not TensorMethods.h.
- */
 #include <ATen/core/TensorBody.h>
-#include <ATen/core/TensorMethods.h>
diff --git a/aten/src/ATen/core/Tensor.h b/aten/src/ATen/core/Tensor.h
@@ -1,12 +1,3 @@
 #pragma once
 
-/*
- * We split Tensor.h into TensorBody.h and TensorMethods.h because we want
- * all TensorMethods to be inlined, but they depend on the Dispatcher,
- * which in turn depends on many other things, which then depend back on Tensor.
- *
- * We can break this dependency chain by having the dispatcher only depend on
- * TensorBody.h and not TensorMethods.h.
- */
 #include <ATen/core/TensorBody.h>
-#include <ATen/core/TensorMethods.h>
diff --git a/aten/src/ATen/core/boxing/impl/kernel_function_legacy_test.cpp b/aten/src/ATen/core/boxing/impl/kernel_function_legacy_test.cpp
@@ -9,6 +9,8 @@
 #include <ATen/core/Tensor.h>
 #include <torch/csrc/jit/frontend/function_schema_parser.h>
 
+#include <ATen/core/LegacyTypeDispatch.h>
+
 /**
  * This file tests the legacy function-based API for registering kernels.
  *
diff --git a/aten/src/ATen/core/boxing/impl/kernel_function_test.cpp b/aten/src/ATen/core/boxing/impl/kernel_function_test.cpp
@@ -5,6 +5,8 @@
 #include <ATen/core/Tensor.h>
 #include <torch/csrc/jit/frontend/function_schema_parser.h>
 
+#include <ATen/core/LegacyTypeDispatch.h>
+
 using c10::RegisterOperators;
 using c10::DispatchKey;
 using c10::Stack;
diff --git a/aten/src/ATen/core/boxing/impl/kernel_lambda_legacy_test.cpp b/aten/src/ATen/core/boxing/impl/kernel_lambda_legacy_test.cpp
@@ -9,6 +9,8 @@
 #include <ATen/core/Tensor.h>
 #include <torch/csrc/jit/frontend/function_schema_parser.h>
 
+#include <ATen/core/LegacyTypeDispatch.h>
+
 /**
  * This file tests the legacy lambda-based API for registering kernels:
  *
diff --git a/aten/src/ATen/core/boxing/impl/kernel_lambda_test.cpp b/aten/src/ATen/core/boxing/impl/kernel_lambda_test.cpp
@@ -5,6 +5,8 @@
 #include <ATen/core/Tensor.h>
 #include <torch/csrc/jit/frontend/function_schema_parser.h>
 
+#include <ATen/core/LegacyTypeDispatch.h>
+
 using c10::RegisterOperators;
 using c10::DispatchKey;
 using c10::Stack;
diff --git a/aten/src/ATen/core/boxing/impl/kernel_stackbased_test.cpp b/aten/src/ATen/core/boxing/impl/kernel_stackbased_test.cpp
@@ -6,6 +6,8 @@
 #include <ATen/core/Tensor.h>
 #include <torch/csrc/jit/frontend/function_schema_parser.h>
 
+#include <ATen/core/LegacyTypeDispatch.h>
+
 using c10::RegisterOperators;
 using c10::DispatchKey;
 using c10::Stack;
diff --git a/aten/src/ATen/core/boxing/impl/make_boxed_from_unboxed_functor_test.cpp b/aten/src/ATen/core/boxing/impl/make_boxed_from_unboxed_functor_test.cpp
@@ -5,6 +5,8 @@
 #include <ATen/core/Tensor.h>
 #include <torch/csrc/jit/frontend/function_schema_parser.h>
 
+#include <ATen/core/LegacyTypeDispatch.h>
+
 using c10::RegisterOperators;
 using c10::OperatorKernel;
 using c10::DispatchKey;
diff --git a/aten/src/ATen/core/op_registration/op_registration_test.cpp b/aten/src/ATen/core/op_registration/op_registration_test.cpp
@@ -16,6 +16,8 @@
 #include <ATen/core/Tensor.h>
 #include <functional>
 
+#include <ATen/core/LegacyTypeDispatch.h>
+
 using c10::RegisterOperators;
 using c10::OperatorKernel;
 using c10::OperatorHandle;
diff --git a/aten/src/ATen/function_wrapper.py b/aten/src/ATen/function_wrapper.py
@@ -129,7 +129,7 @@ def TypedDict(name, attrs, total=True):  # type: ignore
 C10_TENSOR_METHOD_DEFINITION = CodeTemplate("""\
 
 // ${schema_string}
-inline ${return_type} Tensor::${api_name}(${method_formals}) const {
+${return_type} Tensor::${api_name}(${method_formals}) const {
 #ifdef USE_STATIC_DISPATCH
     ${static_dispatch_method_body}
 #else
@@ -165,7 +165,7 @@ def TypedDict(name, attrs, total=True):  # type: ignore
 """)
 
 # In order to rely on the linker to strip unused ops, it requires us to dispatch statically
-# in Functions.h and TensorMethods.h.
+# in Functions.h and TensorMethods.cpp.
 #
 # NB: The default body also needs to apply a variable guard, as in some
 # situations what we think is a default body actually does have an
diff --git a/aten/src/ATen/gen.py b/aten/src/ATen/gen.py
@@ -158,7 +158,7 @@ def check_all_files_written(self):
 BACKEND_SELECT_REGISTER_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/BackendSelectRegister.cpp")
 SCHEMA_REGISTER_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/SchemaRegister.cpp")
 TENSOR_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TensorBody.h")
-TENSOR_METHODS_H = CodeTemplate.from_file(TEMPLATE_PATH + "/TensorMethods.h")
+TENSOR_METHODS_CPP = CodeTemplate.from_file(TEMPLATE_PATH + "/TensorMethods.cpp")
 
 FUNCTIONS_H = CodeTemplate.from_file(TEMPLATE_PATH + "/Functions.h")
 
@@ -402,7 +402,7 @@ def gen_per_op_registration_filename(opname):
 # so that the script runs quickly when we are just querying the
 # outputs
 def declare_outputs():
-    core_files = ['TensorBody.h', 'TensorMethods.h', 'ATenOpList.cpp']
+    core_files = ['TensorBody.h', 'TensorMethods.cpp', 'ATenOpList.cpp']
     for f in core_files:
         core_file_manager.will_write(f)
     files = ['Declarations.yaml', 'TypeDefault.cpp', 'TypeDefault.h',
@@ -513,7 +513,7 @@ def generate_outputs():
 
     core_files = {
         'TensorBody.h': TENSOR_H,
-        'TensorMethods.h': TENSOR_METHODS_H,
+        'TensorMethods.cpp': TENSOR_METHODS_CPP,
         'ATenOpList.cpp': OPS_ALREADY_MOVED_TO_C10_CPP,
     }
 
diff --git a/aten/src/ATen/templates/Functions.h b/aten/src/ATen/templates/Functions.h
@@ -15,6 +15,11 @@
 #include <ATen/TensorUtils.h>
 #include <ATen/Context.h>
 
+#include <ATen/core/dispatch/Dispatcher.h>
+#include <ATen/TypeDefault.h>
+#include <ATen/CPUType.h>
+#include <ATen/QuantizedCPUType.h>
+
 namespace at {
 
 using native::tensor;
diff --git a/aten/src/ATen/templates/TensorBody.h b/aten/src/ATen/templates/TensorBody.h
@@ -17,6 +17,7 @@
 #include <ATen/core/DeprecatedTypePropertiesRegistry.h>
 #include <ATen/core/DeprecatedTypeProperties.h>
 #include <ATen/core/NamedTensor.h>
+#include <torch/csrc/WindowsTorchApiMacro.h>
 
 namespace caffe2 {
 class Tensor;
@@ -295,7 +296,7 @@ class CAFFE2_API Tensor {
   /// Returns a `Tensor`'s layout. Defined in Type.h
   Layout layout() const noexcept;
 
-  /// Returns a `Tensor`'s dtype (`TypeMeta`). Defined in TensorMethods.h
+  /// Returns a `Tensor`'s dtype (`TypeMeta`). Defined in TensorMethods.cpp
   caffe2::TypeMeta dtype() const noexcept;
 
   /// Returns a `Tensor`'s device.
@@ -669,6 +670,24 @@ class CAFFE2_API Tensor {
   c10::intrusive_ptr<TensorImpl, UndefinedTensorImpl> impl_;
 };
 
+int64_t get_device(Tensor self);
+
+template <typename T>
+auto Tensor::register_hook(T&& hook) const -> Tensor::hook_return_void_t<T> {
+  // Return the grad argument in case of a hook with void return type to have an
+  // std::function with Tensor return type
+  std::function<void(Tensor)> fn(hook);
+  return _register_hook([fn](const Tensor& grad) {
+    fn(grad);
+    return Tensor();
+  });
+}
+
+template <typename T>
+auto Tensor::register_hook(T&& hook) const -> Tensor::hook_return_var_t<T> {
+  return _register_hook(hook);
+}
+
 namespace detail {
 // Helper creator for Tensor class which doesn't requires the users to pass
 // in an intrusive_ptr instead it just converts the argument passed to
diff --git a/aten/src/ATen/templates/TensorMethods.cpp b/aten/src/ATen/templates/TensorMethods.cpp
diff --git a/test/cpp/jit/test_misc.cpp b/test/cpp/jit/test_misc.cpp
diff --git a/torch/csrc/jit/codegen/cuda/kernel.h b/torch/csrc/jit/codegen/cuda/kernel.h
diff --git a/torch/csrc/jit/codegen/cuda/kernel_cache.cpp b/torch/csrc/jit/codegen/cuda/kernel_cache.cpp
diff --git a/torch/csrc/jit/codegen/cuda/kernel_cache.h b/torch/csrc/jit/codegen/cuda/kernel_cache.h

Original file line number	Diff line number	Diff line change
`@@ -9,6 +9,8 @@`
`9`	`9`	`#include <ATen/core/Tensor.h>`
`10`	`10`	`#include <torch/csrc/jit/frontend/function_schema_parser.h>`
`11`	`11`
	`12`	`+#include <ATen/core/LegacyTypeDispatch.h>`
	`13`	`+`
`12`	`14`	`/**`
`13`	`15`	`* This file tests the legacy function-based API for registering kernels.`
`14`	`16`	`*`