Revert D16048264: Add static dispatch mode to reduce mobile code size

Differential Revision: D16048264 Original commit changeset: ad1e50951273 fbshipit-source-id: 69d302929e183e2da26b64dcc24c69c3b7de186b
okly366 · Aug 20, 2019 · bd6cf50 · bd6cf50
1 parent 8ca6220
commit bd6cf50
Show file tree

Hide file tree

Showing 17 changed files with 39 additions and 3,228 deletions.
diff --git a/.circleci/config.yml b/.circleci/config.yml
@@ -100,11 +100,6 @@ pytorch_linux_build_defaults: &pytorch_linux_build_defaults
           NAMED_FLAG="export BUILD_NAMEDTENSOR=1"
         fi
 
-        # dispatch aten ops statically for mobile
-        if [[ ${BUILD_ENVIRONMENT} == *"android"* ]]; then
-          NAMED_FLAG="export USE_STATIC_DISPATCH=1"
-        fi
-
         export COMMAND='((echo "export BUILD_ENVIRONMENT=${BUILD_ENVIRONMENT}" && echo '"$NAMED_FLAG"' && echo "source ./workspace/env" && echo "sudo chown -R jenkins workspace && cd workspace && .jenkins/pytorch/build.sh") | docker exec -u jenkins -i "$id" bash) 2>&1'
         echo ${COMMAND} > ./command.sh && unbuffer bash ./command.sh | ts
 
@@ -300,6 +295,7 @@ caffe2_linux_test_defaults: &caffe2_linux_test_defaults
         export COMMAND='((echo "source ./workspace/env" && echo "sudo chown -R jenkins workspace && cd workspace && ./ci_test_script.sh") | docker exec -u jenkins -i "$id" bash) 2>&1'
         echo ${COMMAND} > ./command.sh && unbuffer bash ./command.sh | ts
 
+
 ##############################################################################
 # Macos build defaults
 ##############################################################################

diff --git a/.circleci/verbatim-sources/linux-build-defaults.yml b/.circleci/verbatim-sources/linux-build-defaults.yml
@@ -37,11 +37,6 @@ pytorch_linux_build_defaults: &pytorch_linux_build_defaults
           NAMED_FLAG="export BUILD_NAMEDTENSOR=1"
         fi
 
-        # dispatch aten ops statically for mobile
-        if [[ ${BUILD_ENVIRONMENT} == *"android"* ]]; then
-          NAMED_FLAG="export USE_STATIC_DISPATCH=1"
-        fi
-
         export COMMAND='((echo "export BUILD_ENVIRONMENT=${BUILD_ENVIRONMENT}" && echo '"$NAMED_FLAG"' && echo "source ./workspace/env" && echo "sudo chown -R jenkins workspace && cd workspace && .jenkins/pytorch/build.sh") | docker exec -u jenkins -i "$id" bash) 2>&1'
         echo ${COMMAND} > ./command.sh && unbuffer bash ./command.sh | ts
 
@@ -236,3 +231,4 @@ caffe2_linux_test_defaults: &caffe2_linux_test_defaults
 
         export COMMAND='((echo "source ./workspace/env" && echo "sudo chown -R jenkins workspace && cd workspace && ./ci_test_script.sh") | docker exec -u jenkins -i "$id" bash) 2>&1'
         echo ${COMMAND} > ./command.sh && unbuffer bash ./command.sh | ts
+
diff --git a/CMakeLists.txt b/CMakeLists.txt
@@ -90,7 +90,6 @@ option(BUILD_CAFFE2_OPS "Build Caffe2 operators" ON)
 option(BUILD_SHARED_LIBS "Build libcaffe2.so" ON)
 option(BUILD_CAFFE2_MOBILE "Build libcaffe2 for mobile (deprecating)" ON)
 option(BUILD_NAMEDTENSOR "Experimental: compile with namedtensor support" OFF)
-option(USE_STATIC_DISPATCH "Use static dispatch for ATen operators" OFF)
 cmake_dependent_option(
     CAFFE2_LINK_LOCAL_PROTOBUF "If set, build protobuf inside libcaffe2.so." ON
     "BUILD_SHARED_LIBS AND BUILD_CUSTOM_PROTOBUF" OFF)
@@ -318,10 +317,6 @@ if(USE_QNNPACK)
   set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DUSE_QNNPACK")
 endif()
 
-if(USE_STATIC_DISPATCH)
-  set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} -DUSE_STATIC_DISPATCH")
-endif()
-
 # ---[ Whitelist file if whitelist is specified
 include(cmake/Whitelist.cmake)
 
@@ -467,7 +462,6 @@ include_directories(BEFORE ${PROJECT_SOURCE_DIR})
 include_directories(BEFORE ${PROJECT_BINARY_DIR})
 
 include_directories(BEFORE ${PROJECT_SOURCE_DIR}/aten/src/)
-include_directories(BEFORE ${PROJECT_BINARY_DIR}/aten/src/)
 
 # ---[ Main build
 add_subdirectory(c10)

diff --git a/aten/src/ATen/ScalarOps.h b/aten/src/ATen/ScalarOps.h
@@ -2,7 +2,6 @@
 
 #include <c10/core/Scalar.h>
 #include <ATen/Tensor.h>
-#include <ATen/Functions.h>
 
 // This is in the c10 namespace because we use ADL to find the functions in it.
 namespace c10 {

diff --git a/aten/src/ATen/core/Tensor.h b/aten/src/ATen/core/Tensor.h
@@ -837,24 +837,6 @@ template <typename T, typename... Args>
 Tensor make_tensor(Args&&... args) {
   return Tensor(c10::make_intrusive<T>(std::forward<Args>(args)...));
 }
-
-inline Backend infer_backend(const Tensor & t) {
-  TORCH_CHECK(t.defined(), "undefined Tensor");
-  return tensorTypeIdToBackend(t.type_id());
-}
-inline Backend infer_backend(const TensorList & tl) {
-  TORCH_CHECK(tl.size() > 0, "expected a non-empty list of Tensors");
-  return tensorTypeIdToBackend(tl[0].type_id());
-}
-
-inline bool infer_is_variable(const Tensor & t) {
-  TORCH_CHECK(t.defined(), "undefined Tensor");
-  return t.is_variable();
-}
-inline bool infer_is_variable(const TensorList & tl) {
-  TORCH_CHECK(tl.size() > 0, "expected a non-empty list of Tensors");
-  return tl[0].is_variable();
-}
 } // namespace detail
 
 } // namespace at

diff --git a/aten/src/ATen/core/TensorMethods.h b/aten/src/ATen/core/TensorMethods.h
diff --git a/aten/src/ATen/function_wrapper.py b/aten/src/ATen/function_wrapper.py
@@ -131,12 +131,8 @@ def TypedDict(name, attrs, total=True):  # type: ignore
 # add non-virtual declaration to Tensor.cpp
 TENSOR_METHOD_DEFINITION = CodeTemplate("""\
 inline ${return_type} Tensor::${api_name}(${method_formals}) const {
-#ifdef USE_STATIC_DISPATCH
-    ${mobile_method_body}
-#else
     static auto table = globalATenDispatch().getOpTable("${schema_string}");
     return table->getOp<${return_type} (${formals_types})>(tensorTypeIdToBackend(type_id()), is_variable())(${method_actuals});
-#endif
 }
 """)
 # add a method declaration in Functions.h
@@ -150,33 +146,10 @@ def TypedDict(name, attrs, total=True):  # type: ignore
 # add method definition in Functions.h
 FUNCTION_DEFINITION = CodeTemplate("""\
 static inline ${return_type} ${api_name}(${formals}) {
-#ifdef USE_STATIC_DISPATCH
-    ${mobile_function_body}
-#else
     static auto table = globalATenDispatch().getOpTable("${schema_string}");
     return table->getOp<${return_type} (${formals_types})>(${inferred_backend}, ${inferred_is_variable})(${native_actuals});
-#endif
 }
 """)
-
-# for mobile builds, we rely on the linker to strip unused ops.
-# this requires us to dispatch statically in Functions.h and TensorMethods.h
-MOBILE_FUNCTION_DEFAULT_BODY = CodeTemplate("""\
-${return_call} TypeDefault::${native_type_method_dispatch}(${native_arguments});
-""")
-MOBILE_FUNCTION_SWITCH_BODY = CodeTemplate("""\
-switch(${backend}) {
-    ${mobile_function_switches}
-    default:
-        AT_ERROR("${api_name} not implemented for ", at::toString(${backend}));
-}
-""")
-MOBILE_FUNCTION_SWITCH_STATEMENT = CodeTemplate("""\
-case Backend::${backend}:
-    ${return_call} ${backend}Type::${api_name}(${native_arguments});
-    break;
-""")
-
 # add a native declaration for a native function
 NATIVE_DECLARATION = CodeTemplate("""\
 CAFFE2_API ${return_type} ${native_type_method_dispatch}(${formals_with_defaults});
@@ -185,13 +158,9 @@ def TypedDict(name, attrs, total=True):  # type: ignore
 # special method definition for factory functions in Functions.h that initializes backends
 FACTORY_DEFINITION = CodeTemplate("""\
 static inline ${return_type} ${api_name}(${formals}) {
-#ifdef USE_STATIC_DISPATCH
-    ${mobile_function_body}
-#else
     globalLegacyTypeDispatch().initForBackend(${inferred_backend});
     static auto table = globalATenDispatch().getOpTable("${schema_string}");
     return table->getOp<${return_type} (${formals_types})>(${inferred_backend}, ${inferred_is_variable})(${native_actuals});
-#endif
 }
 """)
 
@@ -236,8 +205,6 @@ def TypedDict(name, attrs, total=True):  # type: ignore
     ('BFloat16', 'BFloat16', 'BFloat16AccrealNotDefined', True),
 ]
 
-mobile_backends = ['CPU', 'QuantizedCPU', 'SparseCPU']
-
 
 class NYIError(Exception):
     """Indicates we don't support this declaration yet"""
@@ -1092,26 +1059,9 @@ def has_named_tensor_formals(formals):
 
         def gen_tensor_method(option):
             # type: (Any) -> FunctionCode
-            if isinstance(type_method_dispatch, dict):
-                mobile_function_switches = []
-                for backend in mobile_backends:
-                    if backend in type_method_dispatch:
-                        mobile_function_switches.append(MOBILE_FUNCTION_SWITCH_STATEMENT.substitute(
-                            option,
-                            backend=backend,
-                            backend_function=type_method_dispatch[backend],
-                            native_arguments=option['method_actuals']))
-                mobile_method_body = MOBILE_FUNCTION_SWITCH_BODY.substitute(
-                    option,
-                    backend='tensorTypeIdToBackend(type_id())',
-                    mobile_function_switches=mobile_function_switches)
-            else:
-                mobile_method_body = MOBILE_FUNCTION_DEFAULT_BODY.substitute(
-                    option, native_arguments=option['method_actuals'])
-
             return FunctionCode(
-                declaration=TENSOR_METHOD_DECLARATION.substitute(option, mobile_method_body=mobile_method_body),
-                definition=TENSOR_METHOD_DEFINITION.substitute(option, mobile_method_body=mobile_method_body))
+                declaration=TENSOR_METHOD_DECLARATION.substitute(option),
+                definition=TENSOR_METHOD_DEFINITION.substitute(option))
 
         def gen_namespace_function(option, dispatch_tensor, dispatch_options):
             # type: (Any, Optional[str], Any) -> FunctionCode
@@ -1127,28 +1077,10 @@ def gen_namespace_function(option, dispatch_tensor, dispatch_options):
                 option['inferred_is_variable'] = 'false'
             declaration = DEPRECATED_FUNCTION_DECLARATION if option['deprecated'] else FUNCTION_DECLARATION
             fn_declaration = declaration.substitute(option)
-
-            if isinstance(type_method_dispatch, dict):
-                mobile_function_switches = []
-                for backend in mobile_backends:
-                    if backend in type_method_dispatch:
-                        mobile_function_switches.append(MOBILE_FUNCTION_SWITCH_STATEMENT.substitute(
-                            option,
-                            backend=backend,
-                            backend_function=type_method_dispatch[backend],
-                            native_arguments=option['native_actuals']))
-                mobile_function_body = MOBILE_FUNCTION_SWITCH_BODY.substitute(
-                    option,
-                    backend=option['inferred_backend'],
-                    mobile_function_switches=mobile_function_switches)
-            else:
-                mobile_function_body = MOBILE_FUNCTION_DEFAULT_BODY.substitute(
-                    option, native_arguments=option['native_actuals'])
-
             if is_factory_method:
-                fn_definition = FACTORY_DEFINITION.substitute(option, mobile_function_body=mobile_function_body)
+                fn_definition = FACTORY_DEFINITION.substitute(option)
             else:
-                fn_definition = FUNCTION_DEFINITION.substitute(option, mobile_function_body=mobile_function_body)
+                fn_definition = FUNCTION_DEFINITION.substitute(option)
             return FunctionCode(definition=fn_definition, declaration=fn_declaration)
 
         # Emit #ifdef BUILD_NAMEDTENSOR macros for any code generated here

diff --git a/aten/src/ATen/native/native_functions.yaml b/aten/src/ATen/native/native_functions.yaml
@@ -2758,9 +2758,7 @@
 # legacy method
 - func: _dimI(Tensor self) -> int
   variants: method
-  dispatch:
-    SparseCPU: sparse_dim_sparse
-    SparseCUDA: sparse_dim_sparse
+  dispatch: sparse_dim_sparse
   requires_tensor: True
   device_guard: False
 
@@ -2776,9 +2774,7 @@
 # legacy method
 - func: _dimV(Tensor self) -> int
   variants: method
-  dispatch:
-    SparseCPU: dense_dim_sparse
-    SparseCUDA: dense_dim_sparse
+  dispatch: dense_dim_sparse
   requires_tensor: True
   device_guard: False
 

diff --git a/aten/src/ATen/templates/Functions.h b/aten/src/ATen/templates/Functions.h
@@ -68,6 +68,29 @@ inline Tensor from_blob(
   return from_blob(data, sizes, detail::defaultStrides(sizes), [](void*) {}, options);
 }
 
+namespace detail {
+
+static inline Backend infer_backend(const Tensor & t) {
+  TORCH_CHECK(t.defined(), "undefined Tensor");
+  return tensorTypeIdToBackend(t.type_id());
+}
+static inline Backend infer_backend(const TensorList & tl) {
+  TORCH_CHECK(tl.size() > 0, "expected a non-empty list of Tensors");
+  return tensorTypeIdToBackend(tl[0].type_id());
+}
+
+static inline bool infer_is_variable(const Tensor & t) {
+  TORCH_CHECK(t.defined(), "undefined Tensor");
+  return t.is_variable();
+}
+static inline bool infer_is_variable(const TensorList & tl) {
+  TORCH_CHECK(tl.size() > 0, "expected a non-empty list of Tensors");
+  return tl[0].is_variable();
+}
+
+
+} // namespace detail
+
 // function definitions are all static inline because
 // they are one-line statically dispatched functions that
 // invoke the actual dynamic dispatch on the correct argument

diff --git a/aten/src/ATen/templates/LegacyTHFunctions.h b/aten/src/ATen/templates/LegacyTHFunctions.h
@@ -2,18 +2,7 @@
 
 // ${generated_comment}
 
-#include <ATen/Context.h>
-#include <c10/core/ScalarType.h>
-#include <c10/core/TensorOptions.h>
-
-namespace c10 {
-class Scalar;
-}
-namespace at {
-struct Generator;
-class Tensor;
-struct Type;
-} // namespace at
+#include <ATen/ATen.h>
 
 namespace at {
 namespace native {

diff --git a/aten/src/ATen/templates/Tensor.h b/aten/src/ATen/templates/Tensor.h
@@ -396,24 +396,6 @@ template <typename T, typename... Args>
 Tensor make_tensor(Args&&... args) {
   return Tensor(c10::make_intrusive<T>(std::forward<Args>(args)...));
 }
-
-inline Backend infer_backend(const Tensor & t) {
-  TORCH_CHECK(t.defined(), "undefined Tensor");
-  return tensorTypeIdToBackend(t.type_id());
-}
-inline Backend infer_backend(const TensorList & tl) {
-  TORCH_CHECK(tl.size() > 0, "expected a non-empty list of Tensors");
-  return tensorTypeIdToBackend(tl[0].type_id());
-}
-
-inline bool infer_is_variable(const Tensor & t) {
-  TORCH_CHECK(t.defined(), "undefined Tensor");
-  return t.is_variable();
-}
-inline bool infer_is_variable(const TensorList & tl) {
-  TORCH_CHECK(tl.size() > 0, "expected a non-empty list of Tensors");
-  return tl[0].is_variable();
-}
 } // namespace detail
 
 } // namespace at

diff --git a/aten/src/ATen/templates/TensorMethods.h b/aten/src/ATen/templates/TensorMethods.h
@@ -11,12 +11,6 @@
 #ifdef BUILD_NAMEDTENSOR
 #include <ATen/NamedTensor.h>
 #endif
-#ifdef USE_STATIC_DISPATCH
-#include <ATen/TypeDefault.h>
-#include <ATen/CPUType.h>
-#include <ATen/QuantizedCPUType.h>
-#include <ATen/SparseCPUType.h>
-#endif
 
 namespace at {
 

diff --git a/aten/src/ATen/templates/TypeDefault.h b/aten/src/ATen/templates/TypeDefault.h
@@ -1,34 +1,12 @@
 #pragma once
 
-// ${generated_comment}
-
-#include <c10/core/TensorOptions.h>
-#include <c10/core/Scalar.h>
-#include <c10/core/QScheme.h>
-#include <c10/core/MemoryFormat.h>
-#include <c10/util/ArrayRef.h>
-#include <c10/util/intrusive_ptr.h>
-#include <torch/csrc/WindowsTorchApiMacro.h>
+#include <ATen/ATen.h>
 
-namespace c10 {
-struct Storage;
-}
+// ${generated_comment}
 
 namespace at {
 
-class Tensor;
-using TensorList = ArrayRef<Tensor>;
-
-class Context;
-struct Generator;
-
-struct Quantizer;
-// This is temporary typedef to enable Quantizer in aten native function API
-// we'll remove them when we are actually exposing Quantizer class
-// to frontend
-using ConstQuantizerPtr = const c10::intrusive_ptr<Quantizer>&;
-
-struct TORCH_API TypeDefault {
+struct CAFFE2_API TypeDefault {
   ${type_method_declarations}
 };