OSSForks/pytorch
1
0
Fork 0
mirror of https://github.com/zebrajr/pytorch.git synced 2025-12-06 12:20:52 +01:00
Code Issues Actions 38 Packages Projects Releases Wiki Activity

Revert D25056091: migrate export_caffe2_op_to_c10.h macros to the new dispatcher registration API

Browse Source 
Test Plan: revert-hammer

Differential Revision:
D25056091 (0ea4982cf3)

Original commit changeset: 0f647ab9bc5e

fbshipit-source-id: e54047b91d82df25460ee00482373c4580f94d50
This commit is contained in:
Brian Hirsh 2020-11-19 19:07:33 -08:00 committed by Facebook GitHub Bot
parent 0ea4982cf3
commit 3c936ecd3c
4 changed files with 70 additions and 54 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

2
c10/macros/Macros.h
View File

@ -77,10 +77,8 @@
* str and ending with a number that varies with the line.
*/
#ifdef __COUNTER__
#define C10_UID __COUNTER__
#define C10_ANONYMOUS_VARIABLE(str) C10_CONCATENATE(str, __COUNTER__)
#else
#define C10_UID __LINE__
#define C10_ANONYMOUS_VARIABLE(str) C10_CONCATENATE(str, __LINE__)
#endif

56
caffe2/core/export_caffe2_op_to_c10.h
View File

@ -8,8 +8,6 @@
#include <ATen/core/grad_mode.h>
#include <ATen/core/op_registration/op_registration.h>
#include <torch/csrc/jit/frontend/function_schema_parser.h>
#include <c10/core/CompileTimeFunctionPointer.h>
#include <torch/library.h>
#include <vector>
namespace caffe2 {
@ -193,52 +191,48 @@ inline FunctionSchema make_function_schema_for_c10(const char* schema_str) {
::caffe2::detail::make_function_schema_for_c10(OperatorSchema); \
return schema; \
} \
TORCH_LIBRARY_FRAGMENT(_caffe2, m) { \
m.def(::caffe2::detail::make_function_schema_for_c10(OperatorSchema)); \
} \
} \
}
#define C10_EXPORT_CAFFE2_OP_TO_C10_CPU_KERNEL_ONLY( \
OperatorName, OperatorClass) \
/* Register call_caffe2_op_from_c10 as a kernel with the c10 dispatcher */ \
TORCH_LIBRARY_IMPL(_caffe2, CPU, m) { \
m.impl("_caffe2::" #OperatorName, \
torch::CppFunction::makeFromBoxedFunction< \
::caffe2::detail::call_caffe2_op_from_c10< \
::caffe2::_c10_ops::schema_##OperatorName, \
OperatorClass>>()); \
}
static auto registry_##OperatorName##_##__COUNTER__ = \
::c10::RegisterOperators().op( \
::caffe2::_c10_ops::schema_##OperatorName(), \
::c10::RegisterOperators::options() \
.kernel<&::caffe2::detail::call_caffe2_op_from_c10< \
::caffe2::_c10_ops::schema_##OperatorName, \
OperatorClass>>(::c10::DispatchKey::CPU));
#define C10_EXPORT_CAFFE2_OP_TO_C10_CPU( \
OperatorName, OperatorSchema, OperatorClass) \
C10_EXPORT_CAFFE2_OP_TO_C10_SCHEMA_ONLY(OperatorName, OperatorSchema) \
#define C10_EXPORT_CAFFE2_OP_TO_C10_CPU( \
OperatorName, OperatorSchema, OperatorClass) \
C10_EXPORT_CAFFE2_OP_TO_C10_SCHEMA_ONLY(OperatorName, OperatorSchema) \
C10_EXPORT_CAFFE2_OP_TO_C10_CPU_KERNEL_ONLY(OperatorName, OperatorClass)
#define C10_EXPORT_CAFFE2_OP_TO_C10_CUDA(OperatorName, OperatorClass) \
/* Register call_caffe2_op_from_c10 as a kernel with the c10 dispatcher */ \
TORCH_LIBRARY_IMPL(_caffe2, CUDA, m) { \
m.impl("_caffe2::" #OperatorName, \
torch::CppFunction::makeFromBoxedFunction< \
::caffe2::detail::call_caffe2_op_from_c10< \
::caffe2::_c10_ops::schema_##OperatorName, \
OperatorClass>>()); \
}
static auto registry_##OperatorName##_##__COUNTER__ = \
::c10::RegisterOperators().op( \
::caffe2::_c10_ops::schema_##OperatorName(), \
::c10::RegisterOperators::options() \
.kernel<&::caffe2::detail::call_caffe2_op_from_c10< \
::caffe2::_c10_ops::schema_##OperatorName, \
OperatorClass>>(::c10::DispatchKey::CUDA));
// You should never manually call the C10_EXPORT_CAFFE2_OP_TO_C10_HIP macro .
// The C10_EXPORT_CAFFE2_OP_TO_C10_CUDA macro from above will be automatically
// rewritten to C10_EXPORT_CAFFE2_OP_TO_C10_HIP by hipify .
#define C10_EXPORT_CAFFE2_OP_TO_C10_HIP(OperatorName, OperatorClass) \
/* Register call_caffe2_op_from_c10 as a kernel with the c10 dispatcher */ \
TORCH_LIBRARY_IMPL(_caffe2, HIP, m) { \
m.impl("_caffe2::" #OperatorName, \
torch::CppFunction::makeFromBoxedFunction< \
::caffe2::detail::call_caffe2_op_from_c10< \
::caffe2::_c10_ops::schema_##OperatorName, \
OperatorClass>>()); \
}
static auto registry_##OperatorName##_##__COUNTER__ = \
::c10::RegisterOperators().op( \
::caffe2::_c10_ops::schema_##OperatorName(), \
::c10::RegisterOperators() \
.options() \
.kernel<&::caffe2::detail::call_caffe2_op_from_c10< \
::caffe2::_c10_ops::schema_##OperatorName, \
OperatorClass>>(::c10::DispatchKey::HIP));
#else
// Don't use c10 dispatcher on mobile because of binary size

4
tools/code_analyzer/run_analyzer.sh
View File

@ -19,8 +19,8 @@ echo "Analyze: ${INPUT}"
-op_invoke_pattern="c10::Dispatcher::findSchema" \
-root_symbol_pattern="torch::jit::[^(]" \
-torch_library_init_pattern="^.*TORCH_LIBRARY_init_([^(]+)(\(.*)?$" \
-torch_library_init_pattern="^.*TORCH_LIBRARY_FRAGMENT_init_([^_]+)_[0-9]+(\(.*)?$" \
-torch_library_init_pattern="^.*TORCH_LIBRARY_IMPL_init_([^_]+)_([^_]+)_[0-9]+(\(.*)?$" \
-torch_library_init_pattern="^.*TORCH_LIBRARY_FRAGMENT_init_([^(]+)(\(.*)?$" \
-torch_library_init_pattern="^.*TORCH_LIBRARY_IMPL_init_([^(]+)_([^_]+)(\(.*)?$" \
${EXTRA_ANALYZER_FLAGS} \
"${INPUT}" \
> "${OUTPUT}"

62
torch/library.h
View File

@ -658,16 +658,25 @@ public:
/// within the same namespace cannot be easily put into one macro block
/// (this is mostly the case for custom ops in fbcode that were ported from
/// the old API)
#define TORCH_LIBRARY_FRAGMENT(ns, m) _TORCH_LIBRARY_FRAGMENT(ns, m, C10_UID)
#define TORCH_LIBRARY_FRAGMENT(ns, m) \
static void TORCH_LIBRARY_FRAGMENT_init_ ## ns ## _ ## k (torch::Library&); \
static torch::detail::TorchLibraryInit TORCH_LIBRARY_FRAGMENT_static_init_ ## ns ## _ ## k ( \
torch::Library::FRAGMENT, \
&TORCH_LIBRARY_FRAGMENT_init_ ## ns ## _ ## k, \
#ns, c10::nullopt, __FILE__, __LINE__ \
); \
void TORCH_LIBRARY_FRAGMENT_init_ ## ns ## _ ## k (torch::Library& m)
/// \private
///
/// The above macro requires an extra unique identifier (uid) to prevent variable name collisions
/// This can happen if TORCH_LIBRARY_FRAGMENT is called multiple times with the same namespace
/// in the same translation unit.
/// Note that the TORCH_LIBRARY variant doesn't run into this problem, because it enforces
/// that it can only be called once for a given namespace.
#define _TORCH_LIBRARY_FRAGMENT(ns, m, uid) \
/// This macro should only be used in a few legacy areas.
/// This macro is a version of TORCH_LIBRARY() that:
/// - doesn't enforce that there is only library for the given namespace
/// - takes in a unique identifier that it appends to the names
/// of each static variable that it creates
/// It effectively lets you define multiple TORCH_LIBRARY_FRAGMENT blocks
/// for the same namespace within the same translation unit, avoiding naming collisions.
#define TORCH_LIBRARY_FRAGMENT_UNIQUE(ns, m, uid) \
static void C10_CONCATENATE(TORCH_LIBRARY_FRAGMENT_init_ ## ns ## _, uid) (torch::Library&); \
static torch::detail::TorchLibraryInit C10_CONCATENATE(TORCH_LIBRARY_FRAGMENT_static_init_ ## ns ## _, uid) ( \
torch::Library::FRAGMENT, \
@ -716,25 +725,40 @@ public:
///
// NB: if the dispatch key is not whitelisted, we simply omit the Library
// call entirely
#define TORCH_LIBRARY_IMPL(ns, k, m) _TORCH_LIBRARY_IMPL(ns, k, m, C10_UID)
/// \private
///
/// The above macro requires an extra unique identifier (uid) to prevent variable name collisions.
/// This can happen if TORCH_LIBRARY_IMPL is called multiple times with the same namespace
/// and dispatch key in the same translation unit.
#define _TORCH_LIBRARY_IMPL(ns, k, m, uid) \
static void C10_CONCATENATE(TORCH_LIBRARY_IMPL_init_ ## ns ## _ ## k ## _, uid) (torch::Library&); \
static torch::detail::TorchLibraryInit C10_CONCATENATE(TORCH_LIBRARY_IMPL_static_init_ ## ns ## _ ## k ## _, uid) ( \
#define TORCH_LIBRARY_IMPL(ns, k, m) \
static void TORCH_LIBRARY_IMPL_init_ ## ns ## _ ## k (torch::Library&); \
static torch::detail::TorchLibraryInit TORCH_LIBRARY_IMPL_static_init_ ## ns ## _ ## k ( \
torch::Library::IMPL, \
c10::guts::if_constexpr<c10::impl::dispatch_key_whitelist_check(c10::DispatchKey::k)>( \
[]() { return & C10_CONCATENATE(TORCH_LIBRARY_IMPL_init_ ## ns ## _ ## k ## _, uid); }, \
[]() { return & TORCH_LIBRARY_IMPL_init_ ## ns ## _ ## k; }, \
[]() { return [](torch::Library&) -> void {}; } \
), \
#ns, c10::make_optional(c10::DispatchKey::k), \
__FILE__, __LINE__ \
); \
void C10_CONCATENATE(TORCH_LIBRARY_IMPL_init_ ## ns ## _ ## k ## _, uid) (torch::Library& m)
void TORCH_LIBRARY_IMPL_init_ ## ns ## _ ## k (torch::Library& m)
/// \private
///
/// This macro should only be used in a few legacy areas.
/// This macro is a version of TORCH_LIBRARY_IMPL() that:
/// - takes in a unique identifier that it appends to the names
/// of each static variable that it creates
/// It effectively lets you define multiple TORCH_LIBRARY_IMPL blocks
/// for the same namespace/dispatch key within the same translation unit,
/// avoiding naming collisions.
#define TORCH_LIBRARY_IMPL_UNIQUE(ns, k, m, uid) \
static void C10_CONCATENATE(TORCH_LIBRARY_IMPL_init_ ## ns ## _ ## k, uid) (torch::Library&); \
static torch::detail::TorchLibraryInit C10_CONCATENATE(TORCH_LIBRARY_IMPL_static_init_ ## ns ## _ ## k, uid) ( \
torch::Library::IMPL, \
c10::guts::if_constexpr<c10::impl::dispatch_key_whitelist_check(c10::DispatchKey::k)>( \
[]() { return & C10_CONCATENATE(TORCH_LIBRARY_IMPL_init_ ## ns ## _ ## k, uid); }, \
[]() { return [](torch::Library&) -> void {}; } \
), \
#ns, c10::make_optional(c10::DispatchKey::k), \
__FILE__, __LINE__ \
); \
void C10_CONCATENATE(TORCH_LIBRARY_IMPL_init_ ## ns ## _ ## k, uid) (torch::Library& m)
// These are variants of the macros above which are to be used for testing (they